Volume 42, July 2015 ISSN: 0167-6296

JOURNAL OF
HEALTH
ECONOMICS
 JOURNAL OF HEALTH ECONOMICS Publication information: Journal of Health Economics (ISSN 0167-6296). For 2015, volumes 39–44 are scheduled for publication.
Subscription prices are available upon request from the Publisher or from the Regional Sales Office nearest you or from this journal’s website
(http://www.elsevier.com/locate/jhe). Further information is available on this journal and other Elsevier products through Elsevier’s website:
(http://www.elsevier.com). Subscriptions are accepted on a prepaid basis only and are entered on a calendar year basis. Issues are sent by
standard mail (surface within Europe, air delivery outside Europe). Priority rates are available upon request. Claims for missing issues should
Aims and Scope be made within six months of the date of dispatch.
This Journal seeks articles related to the economics of health and medical care. Its scope will include the following
topics: production of health and health services; demand and utilization of health services; financing of health services; Advertising information: If you are interested in advertising or other commercial opportunities please e-mail Commercialsales@elsevier.com
and your enquiry will be passed to the correct person who will respond to you within 48 hours.
measurement of health; behavioral models of demanders, suppliers and other health care agencies; health behaviors
and policy interventions; efficiency and distributional aspects of health policy; and such other topics as the Editors Funding body agreements and policies
may deem appropriate. Applications to problems in both developed and less-developed countries are welcomed. Elsevier has established agreements and developed policies to allow authors whose articles appear in journals published by Elsevier, to
comply with potential manuscript archiving requirements as specified as conditions of their grant awards. To learn more about existing
agreements and policies please visit http://www.elsevier.com/fundingbodies
Editors
J. CAWLEY, Department of Policy Analysis and Department of Economics, Cornell University, Ithaca, NY, USA. Orders, claims, and journal enquiries: Please contact the Elsevier Customer Service Department nearest you:
E-mail: johncawley@cornell.edu St. Louis: Elsevier Customer Service Department, 3251 Riverport Lane, Maryland Heights, MO 63043, USA; phone: (877) 8397126 [toll free
M. CHALKLEY, Centre for Health Economics, University of York, Heslington, York, UK. within the USA]; (+1) (314) 4478878 [outside the USA]; fax: (+1) (314) 4478077; e-mail: JournalCustomerService-usa@elsevier.com
Tokyo: Elsevier Customer Service Department, 4F Higashi-Azabu, 1-Chome Bldg, 1-9-15 Higashi-Azabu, Minato-ku, Tokyo 106-0044, Japan;
E-mail: martin.chalkley@york.ac.uk phone: (+81) (3) 5561 5037; fax: (+81) (3) 5561 5047; e-mail: JournalsCustomerServiceJapan@elsevier.com
M.E. CHERNEW, Department of Health Care Policy, Harvard Medical School, Boston, MA, USA. Singapore: Elsevier Customer Service Department, 3 Killiney Road, #08-01 Winsland House I, Singapore 239519; phone: (+65) 63490222;
E-mail: chernew@hcp.med.harvard.edu fax: (+65) 67331510; e-mail: JournalsCustomerServiceAPAC@elsevier.com
Oxford: Elsevier Customer Service Department, The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, UK; phone: (+44) (1865) 843434;
D. CUTLER, Department of Economics, Harvard University, Cambridge, MA, USA. E-mail: dcutler@harvard.edu fax: (+44) (1865) 843970; e-mail: JournalsCustomerServiceEMEA@elsevier.com
E. MEARA, The Dartmouth Institute of Health Policy & Clinical Practice, Dartmouth College, Lebanon, NH, USA.
E-mail: ellen.r.meara@dartmouth.edu Author enquiries
For enquiries relating to the submission of articles (including electronic submission) please visit this journal’s homepage at
N. RICE, Centre for Health Economics, University of York, Heslington, York, UK. http://www.elsevier.com/locate/jhe. For detailed instructions on the preparation of electronic artwork, please visit http://www.elsevier.com/
E-mail: nigel.rice@york.ac.uk artworkinstructions. Contact details for questions arising after acceptance of an article, especially those relating to proofs, will be provided
L. SICILIANI, Department of Economics and Related Studies, University of York, Heslington, York, UK. by the publisher. You can track accepted articles at http://www.elsevier.com/trackarticle. You can also check our Author FAQs at http://www.
elsevier.com/authorFAQ and/or contact Customer Support via http://support.elsevier.com.
E-mail: luigi.siciliani@york.ac.uk
A.D. STREET, Centre for Health Economics, University of York, Heslington, York, UK. Illustration services
E-mail: andrew.street@york.ac.uk Elsevier’s WebShop (http://webshop.elsevier.com/illustrationservices) offers Illustration Services to authors preparing to submit a manuscript
but concerned about the quality of the images accompanying their article. Elsevier’s expert illustrators can produce scientific, technical and
medical-style images, as well as a full range of charts, tables and graphs. Image ‘polishing’ is also available, where our illustrators take your
Associate Editors image(s) and improve them to a professional standard. Please visit the website to find out more.
J.E. ASKILDSEN, University of Bergen, Bergen, Norway.
K. BAICKER, Harvard School of Public Health, Boston, MA, USA. USA mailing notice: Journal of Health Economics (ISSN 0167-6296) is published bimonthly (January, March, May, July, September and
P.P. BARROS, Universidade Nova de Lisboa, Lisbon, Portugal. November) by Elsevier B.V. (P.O. Box 211, 1000 AE Amsterdam, The Netherlands). Periodicals postage paid at Jamaica, NY 11431 and
additional mailing offices (not valid for journal supplements).
A. BASU, University of Washington, Seattle, WA, USA. USA POSTMASTER: Send change of address to Journal of Health Economics, Elsevier Customer Service Department, 3251 Riverport Lane,
H. BLEICHRODT, Erasmus University, Rotterdam, The Netherlands. Maryland Heights, MO 63043, USA.
R.P. ELLIS, Boston University, Boston, MA, USA.
AIRFREIGHT AND MAILING in USA by Air Business Ltd., c/o Worldnet Shipping Inc., 156-15, 146th Avenue, 2nd Floor, Jamaica, NY 11434,
J. GLAZER, Tel Aviv University, Tel Aviv, Israel. USA.
S. GLIED, Columbia University, New York, NY, USA.
M. GROSSMAN, National Bureau of Economic Research, New York, NY, USA. The paper used in this publication meets the requirements of ANSI/NISO Z39.48-1992
J. GRUBER, Massachusetts Institute of Technology, Cambridge, MA, USA. (Permanence of Paper).
R. KAESTNER, University of Illinois at Chicago, Chicago, IL, USA. Printed by Henry Ling Ltd, Dorchester, UK.
M. KIFMANN, University of Hamburg, Hamburg, Germany.
A. LLERAS-MUNEY, University of California at Los Angeles, Los Angeles, CA, USA.
J. MULLAHY, University of Wisconsin-Madison, Madison, WI, USA.
O.A. O’DONNELL, Erasmus University, Rotterdam, The Netherlands. “For a full and complete Guide for Authors, please go to: http://www.elsevier.com/locate/Jhe”
P. OLIVELLA, Universitat Autònoma de Barcelona, Barcelona, Spain.
C. PROPPER, University of Bristol, Bristol, UK.
A. SCOTT, University of Melbourne, Victoria, Australia.

The Journal of Health Economics has no page charges

doi:10.1016/S0167-6296(15)00058-2
 Journal of Health Economics 42 (2015) A1

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Editorial statement on negative findings

The Editors of the health economics journals named below 2. Authors engaging in “data mining,” “specification searching,”
believe that well-designed, well-executed empirical studies that and other such empirical strategies with the goal of produc-
address interesting and important problems in health economics, uti- ing results that are ostensibly “positive” (e.g. null hypotheses
lize appropriate data in a sound and creative manner, and deploy reported as rejected).
innovative conceptual and methodological approaches compatible
with each journal’s distinctive emphasis and scope have potential Henceforth we will remind our referees of this editorial philos-
scientific and publication merit regardless of whether such stud- ophy at the time they are invited to review papers. As always, the
ies’ empirical findings do or do not reject null hypotheses that may ultimate responsibility for acceptance or rejection of a submission
be specified. As such, the Editors wish to articulate clearly that the rests with each journal’s Editors.
submission to our journals of studies that meet these standards is
encouraged. American Journal of Health Economics
We believe that publication of such studies provides properly European Journal of Health Economics
balanced perspectives on the empirical issues at hand. Moreover, Forum for Health Economics & Policy
we believe that this should reduce the incentives to engage in two Health Economics Policy and Law
forms of behavior that we feel ought to be discouraged in the spirit Health Economics Review
of scientific advancement: Health Economics
International Journal of Health Economics and Management
1. Authors withholding from submission such studies that are oth- Journal of Health Economics
erwise meritorious but whose main empirical findings are highly
likely “negative” (e.g. null hypotheses not rejected).

http://dx.doi.org/10.1016/j.jhealeco.2015.06.002
0167-6296/© 2015 Published by Elsevier B.V.
 Journal of Health Economics 42 (2015) 1–16

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Information disclosure and peer effects in the use of antibiotics

Illoong Kwon ∗ , Daesung Jun
Graduate School of Public Administration, Seoul National University, 1 Gwanak-ro, Gwanak-gu, Seoul 151-742, Republic of Korea

a r t i c l e i n f o a b s t r a c t

Article history: Mandatory information disclosure may allow sellers to observe and respond to other sellers’ attributes
Received 21 June 2013 (seller peer effects) as well as informing consumers of the sellers’ attributes (consumer learning effect).
Received in revised form 3 June 2014 Using the data from mandatory information disclosure of antibiotic prescription rates for the common
Accepted 24 October 2014
cold in Korea, this paper shows that while average prescription rates decreased after the disclosure, more
Available online 24 February 2015
than 30% of the clinics increased their antibiotic prescriptions. Moreover, clinics that were prescribing
relatively fewer antibiotics than other local clinics before the disclosure requirement were more likely to
JEL classification:
increase their prescription rate. The average prescription rates also declined less in markets with stronger
I1
L1
clinic competition. These results are consistent with seller peer effects.
D8 © 2015 Elsevier B.V. All rights reserved.

Keywords:
Information disclosure
Peer effects
Antibiotic overuse

1. Introduction
When sellers have more information than buyers on the
attributes of a product, the sellers can overstate the product’s qual-
ity and overcharge the buyers. Such an information asymmetry
problem can lead to the collapse of markets (Stigler, 1961), dis-
tort investment decisions, and undermine the quality and safety of
products and services including health care, foods, education, and
the environment. Therefore, there is an increasing use of manda-
tory information disclosure as a regulatory mechanism to address
this information asymmetry problem and to improve the quality of
products and services.
However, mandatory information disclosure can reveal the
attributes of products and services not only to consumers but also
to other competing sellers. That is, even though the previous litera-
ture has largely focused on the effects of information disclosure to
consumers, mandatory information disclosure can directly affect
the interaction among sellers. In particular, when a seller learns
that most other sellers were providing lower quality services, the
seller may reduce its quality after the information disclosure.
In this paper, we consider a simple theoretical framework to
distinguish whether (i) consumers learn the attributes of sellers
from mandatory information disclosure and pressure the sellers to
∗ Corresponding author. Tel.: +82 2 880 8551.
E-mail addresses: ilkwon@snu.ac.kr (I. Kwon), dswin27@snu.ac.kr (D. Jun).
improve quality (called consumer learning effects) or (ii) sellers learn
their competitors’ attributes from mandatory information disclo-
sure and influence each other’s quality (called seller peer effects).
Even though consumer learning effects suggest that mandatory
information disclosure should increase the quality of all sellers,
we show that seller peer effects may decrease the quality of some
sellers.
Therefore, when introducing mandatory information disclo-
sure, it is important for policy makers to understand the existence
and the extent of seller peer effects. For example, in markets where
sellers themselves do not know the attributes of other sellers, an
information disclosure policy can introduce seller peer effects as
well as consumer learning effects. Moreover, if the disclosed infor-
mation is difficult for consumers to find or interpret, seller peer
effects can dominate consumer learning effects.
Before we proceed further, it is worth clarifying the definition
of peer effects in this paper. We define peer effects as a situation
where an individual’s behavior or decisions are influenced by oth-
ers’ behavior in a relevant peer group, called “endogenous peer
effects” by Manski (1993). Such peer effects can arise from an intrin-
sic social preference for behaving like others. Such peer effects
can also arise from rational decisions to obtain higher economic
payoffs. For example, following the behavior of others can be cost-
efficient and rational (Bikhchandani et al., 1992). While some may
argue that peer effects arising from social preference are the true
peer effects, we are more interested in whether consumers learn
about sellers’ behavior from mandatory information disclosure or

http://dx.doi.org/10.1016/j.jhealeco.2014.10.008
0167-6296/© 2015 Elsevier B.V. All rights reserved.
2 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16
the competing sellers do. Therefore, in this paper, we do not dis-
tinguish between peer effects based on social preference and peer
effects based on rational or strategic choice.
Empirically, we examine the effects of the 2006 mandatory pub-
lic disclosure of the antibiotic prescription rates for the common
cold of every clinic and hospital in Korea. In 2012, the director gen-
eral of the World Health Organization (WHO) warned that overuse
of antibiotics has led to widespread drug-resistant pathogens that
are more difficult, toxic, and costly to treat.1 However, antibiotics
are still frequently prescribed for the common cold, often because
of patient demands and hospital competition, even though they are
not useful for fighting infections caused by viruses like the com-
mon cold, most sore throats, and bronchitis (Bennett et al., 2011;
Robohm and Ruff, 2012). Thus, on February 9th of 2006, the Min-
istry of Health and Welfare in Korea began disclosing antibiotic
prescription rates for the common cold online through the public
disclosure website of the Health Insurance Review and Assessment
Service (HIRA).
On average, we find that the antibiotic prescription rates for
the common cold have decreased from 60% to 51% after the
information disclosure. Surprisingly, however, we uncover a large
amount of heterogeneity among the clinics. More than 30% of
clinics have increased their antibiotic prescription rates after the
information disclosure. In particular, among clinics whose antibi-
otic prescription rates were in the lowest quartile of local clinics
before disclosure, almost half of them increased their prescription
rates after disclosure. This finding is more consistent with seller (or
clinic) peer effects. That is, when a clinic finds out that other clinics
were prescribing relatively more antibiotics than itself, it is more
likely to increase its antibiotic prescriptions.
Alternatively, consumers may prefer higher antibiotic pre-
scription rates, and may have pressured the lower-than-average
antibiotic prescribing clinics to increase their prescription rates.
However, the evidence shows that for those clinics that were pre-
scribing antibiotics relatively more than other local clinics before
the information disclosure, consumers started visiting those clinics
less after the disclosure. Moreover, in townships where consumers
responded more negatively to the antibiotic prescription rates, the
average antibiotic prescription rates decreased more. These results
suggest not only that consumers learned from the information dis-
closure, but also that informed consumers prefer lower antibiotic
prescription rates for the common cold.
We also find that in townships with relatively more clinics, the
average antibiotic prescription rates after the information disclo-
sure decreased less. This result suggests that stronger competition
led to relatively higher antibiotic prescription rates and that the
clinic peer effects triggered by mandatory information disclosure
have reinforced this competition effect.
Overall, the empirical evidence supports both consumer learn-
ing effects and seller peer effects. The previous literature has
implicitly assumed that sellers can observe their competitors’
attributes even before mandatory information disclosure, and has
focused on consumer learning effects. This paper contributes to
the literature by showing that when sellers cannot observe the
attributes of their competitors’ products and services, mandatory
information disclosure can allow the sellers to learn their com-
petitors’ attributes and potentially trigger perverse peer effects.
Because the seller peer effects can cancel out some of the consumer
learning effects, our results may also explain why some previous
studies have found no significant effect of information disclosure.
1
Available from http://www.cbsnews.com/8301-504763 162-57398949-
10391704/who-antibiotic-overuse-so-prevalent-scraped-knee-could-be-deadly/
2. Background and previous literature
2.1. Information disclosure
Most previous literature has focused on the effect of information
disclosure on consumers, or consumer learning effects. For example,
without mandatory information disclosure, consumers may not be
able to observe the quality of a product. Then, as Akerlof (1970)
shows, firms cannot benefit from high quality and may leave the
market, which can lead to the collapse of the whole market, called
the lemon problem. In this case, quality information disclosure to
consumers would benefit high quality firms, and provide incentives
to improve quality.
Also, quality information disclosure can allow consumers to
identify high quality firms more easily, and make them more sensi-
tive to differences in quality. Then, information disclosure can lead
to more competition among firms and may improve the quality of
products (see, e.g., Stigler, 1961; Butters, 1977; Salop and Stiglitz,
1977; Jin and Leslie, 2003).2
However, the empirical evidence on the effect of information
disclosure on quality (or other performance measures) is generally
mixed. For example, Chipty and Witte (1998) find that information
availability on the quality of child care has no significant effect on
the quality of the care. However, Jin and Leslie (2003) find that
information disclosure on restaurants’ hygiene has significantly
improved their hygiene.
In the Health Care industry, Vladeck et al. (1988) do not find any
significant differences in the occupancy rates between high- and
low-mortality rate hospitals after the release of the HCFA (Health
Care Financing Administration) data on hospital-specific mortal-
ity, while Mennemeyer et al. (1997) do find a small but significant
effect. Longo et al. (1977) examine the impact of an obstetrics con-
sumer report on hospital behavior in Missouri, and find that half of
the hospitals improved the quality of their hospital care. Shekelle
et al. (2008) provide a systematic survey of more recent studies,
but show mixed results as well. In the electricity industry, many
states in the US require electricity providers to disclose price and
fuel mix so that consumers can compare prices and environmen-
tal impacts. However, these disclosure policies have not induced
much consumer switching (Bird, 2009).
Note that the previous literature on mandatory information dis-
closure has mainly focused on the changes in consumers’ behavior
from learning new information on product quality (consumer learn-
ing effect), which can induce the changes in firms’ behavior. Few
studies, however, have considered the direct effect of informa-
tion disclosure on firms’ behavior. Some exceptions include the
studies on the effect of information disclosure on firms’ collu-
sion (see, e.g., Albaek et al., 1997; Njoroge, 2003).3 However, these
studies do not explain why firms often oppose mandatory infor-
mation disclosure.4 Consequently, when the effect of mandatory
2
On the other hand, quality information disclosure may allow consumers to per-
ceive the difference between firms, and increase product differentiation among
firms. Then information disclosure would reduce competition (Nelson, 1974; Jin
and Leslie, 2003).
3
There is also a theoretical literature that shows firms would disclose their quality
voluntarily if they know each others’ quality, called the unraveling effect (Grossman
and Hart, 1980; Milgrom, 1981). Therefore, it is a theoretical puzzle why firms
in reality do not disclose their quality (see Board, 2009). Our empirical evidence
suggests that firms may not know each others’ quality (see also Matthews and
Postlewaite, 1985; Shavell, 1994).
4
For example, in 2000 the National Hospital Association opposed a proposal
to impose mandatory information disclosure on fatal and other serious medical
errors. (CNN News February 22, 2000) In 1998, the National Restaurant Association
strongly opposed the mandatory display of hygiene “grade cards” (Food Council
News, Vol. 5, Issue 1, January 2002). In 2006, the Korean Congress attempted
 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16
information disclosure is insignificant, studies often blame con-
sumers’ mistrust, disinterest, or lack of understanding (see, e.g.,
Hibbard and Jewett, 1997; Marshall et al., 2000).
Jun and Chung (2011) also analyze the effect of information dis-
closure on antibiotic prescription rates in Korea. However, they
focus on the change in average prescription rates, and do not ana-
lyze clinic heterogeneity or consumer response. More importantly,
they do not consider the difference between consumer learning
effects and seller peer effects.
2.2. Peer effects
Information disclosure regulations can provide new informa-
tion to competing firms as well as to consumers. In the presence
of peer effects, learning competitors’ behavior can directly affect
firms’ behavior.
As Fortin et al. (2007) summarize, peer effects can arise for
several reasons. In the context of antibiotic prescription for the
common cold, doctors may feel less guilt and prescribe more antibi-
otics when they find that other doctors are prescribing potentially
ineffective antibiotics as well (the social conformity effect).5 Also,
when doctors find out that their peers are prescribing antibiotics
for the common cold, they can learn that prescribing antibiotics for
the common cold may have benefits without strong side effects,
may not lead to drug-resistant bacteria, and may not induce strong
regulatory resistance (the social learning effect). Finally, when other
doctors prescribe antibiotics to attract more patients instead of
educating the patients on the ineffectiveness and potential harms
of antibiotics, doctors may feel it is unfair and become more likely
to prescribe antibiotics to restore equity (the fairness effect).6
As Schelling (1978) and Akerlof (1980) point out, these peer
effects can generate a “social multiplier effect”, as observing other
doctors’ antibiotic prescription rates would encourage and rein-
force the increase in prescription rates even further (see also
Glaeser et al., 2003; Fischer and Huddart, 2008). Malani et al. (2008)
argue that overuse of antibiotics is due to a social norm estab-
lished among doctors and patients, and is difficult to change in the
short-term.
Empirically, there is increasing evidence of peer effects in
economics.7 However, the observed correlation among clinics’ pre-
scription rates does not necessarily imply peer effects, because
clinics in a given market may be subject to similar unobserved char-
acteristics and shocks (exogenous effects and correlated effects). As
Manski (1993) shows, it is generally difficult to distinguish endoge-
nous peer effects from exogenous or correlated effects.
In this paper, we exploit the introduction of an information
disclosure regulation. Note that the regulation can disclose new
information (on prescription rates in our context) not only to con-
sumers but also to competing clinics. If the regulation does disclose
to pass a bill mandating a notice for antibiotics on the prescription, but failed
mostly due to opposition by the hospitals. (See http://www.yakup.co.kr/news/
index.html?cat=11&cat2=51&cat3=&mode=view&nid=82641&num start=5170
&pmode=, in Korean.)
5
Doctors often unwillingly prescribe antibiotics due to pressure from patients,
knowing that antibiotics are ineffective for the treatment of the common cold. In
other words, prescribing antibiotics for the common cold can have psychic costs to
the doctors. Gordon (1989) and Myles and Naylor (1996) argue that in the case of
tax evasion, individuals can derive a psychic payoff, or feel less guilt, from adhering
to the average pattern of their reference group.
6 reduced susceptibility to penicillin is an alarming 70% in Korea,
Spicer and Becker (1980) show that those who believe that they are treated
unfairly by the tax system are more likely to evade taxes.
7
See, for example, Gaviria and Raphael (2001) for drug use, Wilson (2007) for
cigarette smoking, Sacerdote (2001) for GPA, Carrell et al. (2008) for academic
cheating, Duflo and Saez (2002) for investment decisions, Fortin et al. (2007) for
tax evasion.
3
new information to the competing clinics (that is, the compet-
ing clinics did not know the disclosed information before), in the
presence of peer effects, the disclosed information would trigger
the social multiplier effects. However, the regulation would not
change clinic or market characteristics, and consequently should
not change the exogenous or correlated effects. Thus, the change
in antibiotic prescription rates after the regulation would be due
to the endogenous peer effects, not to the exogenous or correlated
effects. On the other hand, the information disclosure would also
change the consumers’ demand function. Therefore, we still need
to distinguish between the changes due to clinic peer effects and
the changes due to consumer learning effects.
Mas and Moretti (2009) distinguish between two types of
peer effects. When a worker is observed by a high productiv-
ity worker, they find that the productivity of the worker being
observed increases. However, when a worker observes another
highly productive worker, the productivity of the first worker does
not increase. In other words, they find a significant social pres-
sure effect, but no social conformity effect. It is not clear whether
this pattern will hold in other work environments, but their study
shows that it is important to distinguish between the social pres-
sure effect and the social conformity effect in discussing peer
effects.
2.3. Antibiotic overuse and the common cold
The common cold, or Acute Upper Respiratory Tract Infection
(ARTI), is one of the most common illnesses known to humans and
one of the most common reasons patients visit hospitals. Annu-
ally, about $227 million are estimated to be spent on antibiotics
for the treatment of the common cold in the United States.8 How-
ever, antibiotics do not treat upper respiratory infections caused by
viruses like the common cold. Controlled clinical trials have consis-
tently shown that antibiotics therapy does not treat the common
cold. In addition, antibiotics may have caused many complications
and side-effects. In particular, the overuse of antibiotics has led to
a rise in antibiotic-resistant bacteria (Gonzales et al., 2001). Infec-
tions due to penicillin-resistant bacteria are especially difficult to
treat. In the United States, 4–7 billion dollars are spent on the treat-
ment of resistant infections each year (Lautenbach et al., 2001;
Bennett et al., 2011).
The causes for the overuse of antibiotics, especially for the
common cold, are controversial. Patients may demand antibi-
otics to mask symptoms and gain psychological comfort (Butler
et al., 1998). Doctors may overprescribe antibiotics to retain their
patients (Brody, 2005). Also, when it is not clear whether the infec-
tion is caused by a virus (that cannot be treated by antibiotics)
or bacteria (that can be treated by antibiotics), doctors may sim-
ply prescribe antibiotics to avoid time-consuming medical tests to
discern whether the infection is bacterial or viral. Or doctors may
prescribe antibiotics in order to avoid potential lawsuits and con-
flicts for not providing antibiotics when it turns out to be a bacterial
infection later. Currie et al. (2012) show that financial kickbacks
(or rebates) from pharmaceutical companies for prescribing antibi-
otics are also important causes of overprescription in China.
According to OECD Health Data 2011, Korea has the sixth high-
est rate of antibiotic use among OECD countries. In particular, the
average antibiotic prescription rates for the common cold in 2005
were over 60%. Consequently, the prevalence of S pneumoniae with
compared to 25% in the United States (Conly, 1998).
8
Cited from http://www.npcentral.net/ce/colds/cold.shtml.
4 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16
In part to reduce the overuse of antibiotic prescriptions, in 2000
the Korean government prohibited doctors from selling medica-
tions and allowed them to write prescriptions only. Before 2000,
doctors were allowed to dispense/sell medications to their patients.
As a result, under the old system doctors had financial incentives
to prescribe and dispense more medications including antibiotics.
In 2000, the Health Insurance Review and Assessment Service
(HIRA) was established to monitor and encourage proper drug
prescription. Despite these efforts to prevent overuse of antibi-
otics, antibiotic prescription rates have remained high (Ministry
of Health and Welfare, 2006). On January 5th, 2006, the Seoul
Administration Court gave a verdict mandating the disclosure of
the antibiotic prescription rates for the common cold of every clinic
and hospital in Korea.
3. Theoretical framework
3.1. Peer effects
On the supply side, we assume that antibiotic prescription rates
are subject to peer effects among clinics. However, peer effects can
arise only when the behavior (that is, antibiotic prescription rates)
is observable by peers. Ali et al. (2011), for example, show that ado-
lescents’ sleep habits and breakfast consumption (unobservable by
their peers) are not influenced by peers, but that participation in
sports or eating at fast food restaurants (observable by their peers)
are influenced by peers.
Suppose that clinics could not observe other clinics’ antibiotic
prescription rates before the information disclosure. Then manda-
tory information disclosure would allow clinics to observe the
antibiotic prescription rates of other clinics, and trigger peer effects
among the clinics.
According to Mas and Moretti (2009), there can be two types
of peer effects. The first type is the social conformity effect. Before
the information disclosure, when clinics could not observe other
clinics’ antibiotic prescription rates, each clinic would prescribe
antibiotics based on its own norm or expectations of other clin-
ics’ antibiotic prescription rates. After the information disclosure,
if a clinic finds out that many other local clinics were prescribing
relatively more (fewer) antibiotics than itself, the social confor-
mity effect would lead the clinic to increase (decrease) its antibiotic
prescription rates.
Note that the social conformity effect may arise from a social
preference for behaving like others or from rational decisions. For
example, when a clinic finds out that it is prescribing more antibi-
otics than the average, it may reduce its antibiotic prescription rates
because of possible social or regulatory pressure, learning of the
increased danger of drug-resistant bacteria or the potential side
effects of antibiotics.
The second type of peer effect is the social pressure effect that
would pressure clinics to do the socially ‘right’ thing. In the context
of antibiotic prescription for the common cold, the socially right
thing would be not to prescribe antibiotics.
Since the social conformity effects for individual clinics are likely
to cancel each other out, mandatory information disclosure would
lead to less average antibiotic prescription rates among clinics due
to the social pressure effect.
Then, in the case where clinics could not observe other clin-
ics’ antibiotic prescription rates before the information disclosure,
we can summarize the peer effects triggered by the information
disclosure in the following proposition.
Proposition 1. Suppose that mandatory information disclosure
triggers peer effects among clinics. Then,
(i) average antibiotic prescription rates would decrease;
(ii) an individual clinic’s antibiotic prescription rate can decrease if
it is relatively higher than other clinics’ or if the social pressure
effect is sufficiently strong;
(iii) an individual hospital’s antibiotic prescription rate can increase
if it is relatively lower than other clinics’ and if the social pressure
effect is sufficiently weak.
3.2. Consumer learning effects
On the demand side, from mandatory information disclosure,
consumers may learn individual clinics’ actual antibiotic prescrip-
tion rates, and update their belief on each clinic’s quality. We
assume that average consumers regard lower antibiotic prescrip-
tion rates for the common cold as a good signal for the clinic’s
quality. Therefore, after the information disclosure, the market
demand for relatively lower (higher) prescribing clinics would
increase (decrease). Then consumer learning along with clinic com-
petition would lead clinics to reduce their antibiotic prescription
rates. This is called the consumer learning effect.
Note that consumers’ ideal levels of antibiotic prescription rates
for the common cold can differ, and some consumers may want
even higher antibiotic prescription rates for their clinics. However,
market demand reflects the aggregate (or average) of individual
demands. Thus, as long as consumers’ average ideal level of antibi-
otic prescription rates for the common cold is lower than clinics’
actual antibiotic prescription rates, the market would regard lower
antibiotic prescription rates as a good signal for a clinic’s quality.
Yoo et al. (2009) show in a 2009 survey that 80% of Korean con-
sumers think clinics are prescribing too many antibiotics. They also
show that only 10.7% of consumers want antibiotic prescriptions
for the common cold. Therefore, while consumers may want some
level of antibiotic prescription for the common cold, it appears that
most clinics are prescribing more antibiotics than the average con-
sumer wants. For example, as discussed above, doctors may simply
prescribe antibiotics to avoid time-consuming medical tests and
consumer education to determine whether an infection is bacte-
rial or viral. Or doctors may prescribe antibiotics in order to avoid
potential lawsuits and conflicts for not providing antibiotics when
it turns out to be a bacterial infection, or possibly to gain financial
kickbacks from pharmaceutical companies.
In the context of antibiotic prescription rates for the common
cold, mandatory information disclosure can also educate con-
sumers about the difference between viral and bacterial infection
and the ineffectiveness of antibiotics for viral infections such as the
common cold. Thus, mandatory information disclosure can lower
consumers’ ideal level of antibiotic prescription rates for the com-
mon cold as well. Then, low (high) antibiotic prescription rates
for the common cold would become an even better (worse) sig-
nal for clinic quality, and this would put stronger market pressure
on clinics to reduce their antibiotic prescription rates. Also, con-
sumers would become less likely to visit clinics for the common
cold in general, and more likely to get over-the-counter medicine
from pharmacies instead. That is, the total market demand for clinic
visits is likely to decrease.
Moreover, assuming that consumers prefer lower antibiotic pre-
scription rates for the common cold (, which we will test empirically
below), the consumer learning effect should reduce both the aver-
age and individual antibiotic prescription rates of all clinics. In
contrast, from Proposition 1, clinic peer effects predict that some
clinics would increase antibiotic prescription rates, while the aver-
age prescription rates may fall.
Then, in the case where consumers could not observe clin-
ics’ antibiotic prescription rates before the information disclosure,
 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16
(a) Clinic Peer Effect
Prescription
Rate
(before disclosure)
(after disclosure)
Competition
Fig. 1. Effect of information disclosure and interaction with competition.
we can summarize the consumer learning effects triggered by the
information disclosure as follows:
Proposition 2. Suppose that mandatory information disclosure
triggers consumer learning effects. Then,
(i) market demand for relatively lower (higher) prescribing clinics
would increase (decrease);
(ii) total market demand would decrease;
(iii) the average antibiotic prescription rates would decrease;
(iv) individual clinic’s antibiotic prescription rates would decrease.
3.3. Competition effects
Suppose that before the mandatory information disclosure reg-
ulation, neither the clinics nor the consumers could observe other
clinics’ antibiotic prescription rates. As discussed earlier, clinics
may prescribe antibiotics for cost saving or financial kickbacks.
Also, when consumers cannot observe or compare antibiotic pre-
scription rates, prescribing more antibiotics may increase demand
as it can mask the symptoms, and reduce the number of tests. Then,
with more clinics, stronger competition can lead to higher antibi-
otic prescription rates. For example, Fogelberg and Karlsson (2012)
and Bennett et al. (2011) show that stronger competition has a
positive effect on antibiotic prescription rates.
Suppose that after the mandatory information disclosure reg-
ulation, clinics observe other clinics’ antibiotic prescription rates
but consumers cannot, possibly because the disclosed information
is too difficult for consumers to find or interpret. In the presence of
clinic peer effects, stronger competition would lead to even higher
antibiotic prescription rates because if one clinic increases its pre-
scription rates, other clinics would increase their prescription rates
as well, in accordance with the social multiplier effect. That is, while
Proposition 1 predicts that the average antibiotic prescription rates
will fall due to the social pressure effect, if competition among clin-
ics becomes stronger, the prescription rates after the disclosure
would be relatively larger. Therefore, when competition is strong,
information disclosure would decrease antibiotic prescription rates
less in absolute value. (See Fig. 1(a) for an illustration.)
Now suppose that after mandatory information disclosure, con-
sumers can observe and compare the antibiotic prescription rates
of local clinics, and that they prefer lower prescription rates. With
more clinics, consumers have more choices for clinics, and can
switch clinics more easily. Then, after the information disclosure,
stronger competition among clinics would lead to lower antibiotic
prescription rates. That is, if information disclosure triggers con-
sumer learning effects, the average antibiotic prescription rates
5
(b) Consumer Learning Effect
Prescription
Rate
(before disclosure)
(after disclosure)
Competition
would decrease even more when there is stronger clinic compe-
tition in the market. (See Fig. 1(b) for an illustration.)
To summarize,
Proposition 3.
(i) If mandatory information disclosure triggers clinic peer effects,
it would reduce the average antibiotic prescription rates less (in
absolute value) when competition is stronger.
(ii) If mandatory information disclosure triggers consumer learning
effects, it would reduce the average antibiotic prescription rates
even more (in absolute value) when competition is stronger.
As discussed in the beginning, one of the main justifications for
the mandatory information disclosure policy has been based on
the interaction between consumer learning effects and the com-
petition effect. That is, when consumers are informed of the true
quality of products and services, they would choose sellers with
higher quality products and services. Therefore, with mandatory
information disclosure, competition among sellers would force the
sellers to increase the quality of products, and/or drive the low
quality product sellers out of the market.
However, Proposition 3 suggests that if the sellers themselves,
not the consumers, are informed of the true quality of their
competitors’ products and services by mandatory information dis-
closure, stronger competition can reduce the effect of mandatory
information disclosure, which undermines the conventional justi-
fication for a mandatory information disclosure policy.
It is also worth emphasizing that Proposition 3 provides a pos-
sible way to distinguish between consumer learning effects and
clinic peer effects even when individual clinic level prescription
data are not available. Propositions 1 and 2 show that individual
clinics’ antibiotic prescription rates would respond to the informa-
tion disclosure differently depending on whether the information
disclosure triggers clinic peer effects or consumer learning effects.
However, both effects predict that the market average antibiotic
prescription rates will fall. Therefore, unless researchers have data
on individual clinic level antibiotic prescription rates, it can be
difficult to distinguish between clinic peer effects and consumer
learning effects.
Proposition 3, however, shows that the interaction effect
between competition and information disclosure on average
antibiotic prescription rates would differ depending on whether
the clinic peer effect or the consumer learning effect dominates.
Therefore, even with market level data on average antibiotic pre-
scription rates, one can potentially distinguish between the two
effects.
6 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16
4. Data
Empirically, we analyze individual clinics’ antibiotic prescrip-
tion rates for the common cold before and after the mandatory
information disclosure regulation in Korea. In March 2005, People’s
Solidarity for Participatory Democracy (PSPD), one of the largest
civil rights organizations in Korea, petitioned for information dis-
closure on individual clinics’ antibiotic prescription rates, but the
petition was denied by the Ministry of Health and Welfare. In June
2005, PSPD filed a formal administrative litigation against the gov-
ernment. On January 5th, 2006, the Seoul Administration Court
delivered a verdict mandating the disclosure of antibiotic prescrip-
tion rates for acute upper respiratory tract infection, or the common
cold
“in order to protect the rights of citizens in the matter of choos-
ing their own treatment options and also to help consumers
make better-informed choices on their health care providers by
disseminating comparative information on antibiotic prescrip-
tion rates.”
On February 9th, 2006, the Ministry of Health and Welfare began
disclosing the antibiotic prescription rates for the common cold of
every clinic and hospital which had written out more than one-
hundred prescriptions for antibiotics per quarter. This information
was made available online through the public disclosure website
of the Health Insurance Review and Assessment Service (HIRA) at
http://www.hira.or.kr.
Note, however, that visiting the HIRA website online has been
the only practical way for consumers to find out the disclosed
antibiotic prescription rates. Even though Korea has the highest
internet penetration rates among OECD countries, not all con-
sumers are familiar with internet use. Moreover, not all consumers
are aware of the existence of such a website. Therefore, it is not
entirely clear how many consumers have gained new information
through this disclosure. For example, a 2007 survey shows that
only 21.5% of consumers knew about the information disclosure
and only 7% of consumers have actually visited the HIRA website
to check the antibiotic prescription rates. In contrast, 95% of doctors
knew about the information disclosure.9
Fig. 2 shows how the antibiotic prescription rates are being pre-
sented online as of April 2014. Though some of the visual designs
have changed since 2006, the main contents on the webpage have
not changed much.10 From the HIRA main website, consumers can
select a township and display a list of clinics in the selected town-
ship ranked by antibiotic prescription rates, as shown in Fig. 2(a).
Then, consumers can select a clinic and display the antibiotic pre-
scription rates for the common cold of the selected clinic, as shown
in Fig. 2(b). Also, the same webpage shows warnings that antibi-
otics do not treat the common cold and that ‘good’ clinics prescribe
antibiotics only when they are necessary. Note that clinics in the
same township can be considered as a natural peer group. And
consumers may learn not only the antibiotic prescription rates of
their local clinics, but also the fact that antibiotics do not treat the
common cold.11
9
Ministry of Health and Welfare 2007-1-12 press kit “Results from a Survey on
Medical Service and Provision After Information Disclosure of Antibiotics Prescrip-
tion Rates”.
10
As of April 2014, various new performance ratings for several diseases (e.g. dia-
betes), operations (e.g. breast cancer), and prescriptions (e.g. antibiotics for acute
otitis media in children) can also be found and compared on the same website.
However, this information was not disclosed during our sample period (2005–2009).
11
For some patients, such as infants or those with a prior history of lower respira-
tory infections, prescribing antibiotics for the common cold proactively can be the
Our empirical analysis is based on the quarterly individual clinic
level data on the antibiotic prescription rates for the common cold
of every clinic in the city of Seoul, Korea between 2005.Q1 and
2009.Q2. The prescription rate is defined as the number of antibiotic
prescriptions divided by the number of patient visits per quarter.
The denominator and the numerator of the prescription rate are not
reported separately. However, we were able to collect the informa-
tion on the number of patient visits for the common cold between
2005.Q1 and 2006.Q3.
We merged the data with the 2007 clinic characteristics data
that contain the medical specialty, location, number of doctors, and
clinic age. The location shows both the township (“dong”) and the
district (“gu”). There are 25 districts in Seoul, and within these dis-
tricts there are 424 townships in total. We then merged the data
with the 2009 township and district characteristics such as total
population, area size, share of population with age over 65, and
share of college graduates.12 Because clinic and township charac-
teristics do not change much over time during our sample period,
we do not measure them every quarter.
The data contain 19 tertiary general hospitals, 43 general hos-
pitals, 77 hospitals, and 2,802 clinics. Tertiary general hospitals
require referrals, except for emergencies. Also, for the common
cold, most people go to local clinics within walking distance from
their home or work. Therefore, we will restrict our analysis to the
clinics only, resulting in 42,020 observations. However, the results
do not change even when we include all the other hospital types.13
Recall that the prescription rates of the clinics that have written
out less than one-hundred prescriptions for antibiotics per quarter
are not disclosed online. We exclude these observations in most of
our analyses, but will include them as a comparison group in testing
the robustness of our results. Also, clinics that are reported either
before 2006.Q1 only or after 2006.Q1 only are excluded from the
data. These restrictions raise concerns about changes in the compo-
sition of the clinics. Therefore, we will control for clinic fixed effects,
and also replicate our analysis with the fully balanced sample only,
which accounts for 63 percent of the observations.
Table 1 shows the summary statistics of the selected variables.
5. Clinic peer effects
5.1. Change in average antibiotic prescription rates
Fig. 3(a) shows the average prescription rates for the common
cold over time. Recall that the prescription rates were disclosed
to the public online on February 9th, 2006, or the first quarter of
2006. The average prescription rate fell by almost 10% from the
second quarter of 2005 to the second quarter of 2006. To check
whether this decrease in the prescription rate is due to a change in
clinic composition, Fig. 3(b) shows the same graph for the balanced
sample only, but the results do not change much.
In Table 2 we estimate the effect of information disclosure more
formally. “Disclosure” is a dummy variable equal to one if the date is
strictly after 2006.Q1 and zero otherwise. In column (1) of Table 2,
we run a simple OLS model with the disclosure dummy, quarterly
right decision. However, such risk adjustments were not made. On the other hand,
the clinics can prescribe antibiotics to such patients under different diagnostic codes.
12
The township characteristics are available from Seoul Statistics Information.
13
Given that the data report the antibiotic prescription rates for the common cold,
most clinics in the data are specialized in general medicine, internal medicine, pedi-
atrics, and otorhinolaryngology. However, some clinics in the data are specialized in
surgery, tuberculosis, and dermatology. Thus, we will include the clinic or the spe-
cialty fixed effects in our analysis. Also, focusing on general medicine and internal
medicine specialties only (47% of the sample) does not change the results.
 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16 7

Fig. 2. Online disclosure of antibiotic prescription rates for the common cold. Note: The English translations in the figure are made and inserted by the authors.

dummies, and 23 medical specialty dummies.14 The estimated
effect of information disclosure is a 9.67 percentage point reduction
in the average antibiotic prescription rate for the common cold. In
column (2), we control for clinic characteristics and township char-
acteristics. In column (3), we control for township fixed effects. In
column (4), we control for clinic fixed effects. Finally, in column
(5), we use the balanced sample only. These models show that the
estimated effect of information disclosure is in the range of −9.67%
14
Including year dummies or a linear time trend does not change the results.
to −8.98 % . These estimates are close to the estimates from earlier
studies. For example, Jun and Chung (2011) estimate the effect to
be around −9.53% to −6.49 % .
From Propositions 1(i) and 2(iii), the decline in the average
antibiotic prescription rates is consistent with both clinic peer
effects and consumer learning effects. That is, clinics may have
reduced their antibiotic prescription rates either because of social
pressure to do the right thing or because of market pressure to keep
the informed patients.
However, we cannot rule out the possibility that the decline was
driven by other concurrent unobserved shocks such as a sudden
8 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16

Table 1
Summary statistics of selected variable.

N Mean SD Min Max

Antibiotic Antibiotic prescription (%) 42,020 52.97 28.72 0 100

Prescription # of patient visits 17,220 1419 1402 109 29,334

Clinic Medical speciality

Characteristics (2007) # of doctors 41,293 1.18 0.64 0 13
# of doctors and staff 41,296 4.76 4.14 1 79
Hospital age 41,315 11.02 8.74 1 53

Township (dong) # of clinics 42,020 8.74 3.75 1 19

Characteristics (2009) Population 41,878 26,539 8649 1095 51,446
# of clinics per 1000 41,878 3.78 4.59 0.27 73.06
Age over 65 (%) 41,878 9.31 1.96 5.01 17.2
Age under 9 (%) 41,878 8.59 1.64 3.94 14.31

District (gu) College degree (%) 42,020 23.19 7.12 11.74 38.20

decrease in the more severe type of common cold. Jun and Chung
(2011) provide difference-in-difference estimates using those clin-
ics whose antibiotic prescription rates were not disclosed online as
a comparison group. Recall that the clinics that have written out less
than one-hundred prescriptions for antibiotics per quarter were
not disclosed online. In columns (1)–(3) of Table 3, we replicate
the results from Jun and Chung (2011) with the clinic fixed effects.
“Open” is a dummy variable to indicate whether the prescription
rates are disclosed online or not. Note that column (2) of Table 3
shows that the effect of disclosure for the comparison group is
much smaller. Alternatively, in column (3), we use the full sample,
and estimate the effect of the interaction term between disclosure
and the open dummy variable. From these difference-in-difference
estimates, the effect of mandatory information disclosure is about
−5.2 % .
However, even when a clinic’s antibiotic prescription rate is not
disclosed online in one quarter, the clinic’s prescription rates can
be disclosed online in other quarters when the number of its pre-
scriptions becomes more than one hundred. In fact, the prescription
rates of every clinic were disclosed at least once during our sam-
ple period. Then, the difference-in-difference estimates are likely
to underestimate the true effect of the mandatory information dis-
closure. Therefore, unless specified otherwise, we will focus on the
disclosed samples with the clinic fixed effects as in column (4) of
Table 2 for a base specification, and use the difference-in-difference
model in column (3) of Table 3 to check the robustness of the results.
Also, clinics may have known about the mandatory informa-
tion disclosure and changed their prescription rates even before the
actual disclosure on February 9th, 2006. Recall that the litigation
for the information disclosure was filed in June 2005, and that the
verdict for the disclosure was delivered on January 5th, 2006. But
our disclosure dummy variable in Tables 2 and 3 assumes that the
event took place in the beginning of the second quarter in 2006.
Then, our estimates is likely to under-estimate the true effect of
mandatory information disclosure.
However, the court verdict on January 5th, 2006 came as a sur-
prise because the government had argued that prescription rates
were a part of business secrets and were exempted from any

Table 2
Change in antibiotic prescription rates for the common cold (dependent variable = antibiotic prescription rate (%)).

(1) (2) (3) (4) (5)

Disclosure −9.6725*** −9.6682*** −9.5442*** −9.3494*** −8.9836***

(0.3090) (0.2992) (0.2782) (0.1397) (0.1611)
No. of doctors −1.0189*** −1.1080***
(0.3158) (0.3193)
No. of doctors and staff −0.1581*** −0.0802
(0.0518) (0.0517)
Clinic age 0.0019 0.0232
(0.0157) (0.0160)
No. of clinics 0.1124***
(0.0404)
Population (1000) 0.0223
(0.0187)
Age 65 and older (%) 0.4215***
(0.0744)
Quarter = 2 0.7353* 0.6824* 0.6760** 0.6215*** 0.5091***
(0.3760) (0.3634) (0.3375) (0.1657) (0.1929)
Quarter = 3 2.1484*** 1.8128*** 1.8223*** 1.7544*** 1.7320***
(0.3992) (0.3860) (0.3585) (0.1764) (0.2036)
Quarter = 4 −0.0702 −0.1454 −0.1786 −0.1767 −0.2094
(0.3943) (0.3814) (0.3541) (0.1737) (0.2030)
Medical specialities Yes Yes Yes No No
Fixed effect No No Township Clinic Clinic

Observations 42,020 41,160 41,286 42,020 28,035

R-squared 0.0236 0.1053 0.1070 0.1057 0.1089

Notes: Column (3) controls for the township fixed effects. Columns (4) and (5) control for the clinic fixed effects. In column (5), we use the balanced sample only.
*
Significant at 10%.
**
Significant at 5%.
***
Significant at 1%.
 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16 9

.04
(a) full sample
65

.03
prescription rate(%)
60

Density
55

.02
50

.01
2005q1 2006q1 2007q1 2008q1 2009q1

time

(b) balanced sample

0
-100 -50 0 50 100
65

change in prescription rate

prescription rate(%)

Fig. 4. Distribution of the changes in prescription rates.

60

5.2. Clinic heterogeneity

55

Even though the average antibiotic prescription rates have

decreased after the mandatory information disclosure, it turns
50

out that there are large variations. For example, when we plot a
2005q1 2006q1 2007q1 2008q1 2009q1 histogram for the simple difference in the average antibiotic pre-
time
scription rates before and after the information disclosure for each
Fig. 3. Average antibiotic prescription rates.
clinic, Fig. 4 shows that the changes in prescription rates vary
widely across clinics. In particular, 30% of the clinics have increased
their antibiotic prescription rates after the regulation.
information disclosure requirement. Therefore, it is unlikely that Proposition 2(iv) shows that assuming consumers prefer fewer
clinics changed their prescription rates before January 2006. On the antibiotic prescriptions, consumer learning effects should lead to a
other hand, it is possible that some clinics started changing their fall in the prescription rates of every clinic. However, Proposition
prescription rates right after the verdict even before the actual dis- 1(ii) and (iii) show that clinic peer effects can cause individual clin-
closure. In fact, Fig. 3 shows that the average prescription rates ics’ prescription rates to either increase or decrease depending on
started to decline in the first quarter of 2006. Thus, in column (4) of whether their prescription rates are relatively higher or lower than
Table 3, we have excluded the sample from the first quarter of 2006. their peers’. Therefore, Fig. 4 is more consistent with Proposition 1,
Alternatively, in column (5), we redefined the disclosure dummy suggesting the existence of clinic peer effects.
to one if it is strictly after the fourth quarter of 2005. However, the More specifically, Fig. 5 shows each individual clinic’s prescrip-
results do not change much. tion rate over time in one particular township. Note that those
clinics with higher-than-average pre-disclosure prescription rates
Table 3 are more likely to decrease their rates post-disclosure (e.g. ID = E,
Change in antibiotic prescription rates for the common cold: robustness (dependent F). Also, those clinics with lower-than-average prescription rates
variable = antibiotic prescription rate (%)).
pre-disclosure are more likely to increase their rates (e.g. ID = C, D).
(1) (2) (3) (4) (5) These patterns are consistent with the clinic peer effects as dis-
Open = 1 Open = 0 All Open = 1 Open = 1 cussed in Proposition 1(ii) and (iii), even though not all clinics
Disclosure −9.3494*** −3.4096*** −4.0870*** −9.2608*** follow these patterns.
(0.1397) (0.6693) (0.3606) (0.1418) To check whether the patterns in Fig. 5 generalize to other town-
Open 3.4230*** ships, we first create dummy variables, P25, P50, P75, and P100
(0.5727)
where P25 = 1 if a clinic’s antibiotic prescription rate in 2005.Q1
Disclosure × open −5.2779***
(0.3899) is lower than the 25 percentile in a township, P50 = 1 if a clinic’s
Disclosure 1 −9.3696*** antibiotic prescription rate is between the 25 percentile and 50
(0.1492) percentile in a township, and so on. Then, for the townships with
Fixed effect Clinic Clinic Clinic Clinic Clinic
more than 100 observations, we estimate the following model by
Observations 42,020 7,594 49,614 39,400 42,020 each township:
R-squared 0.1057 0.0049 0.0747 0.1068 0.0947

Notes: All regressions include quarterly dummies and clinic fixed effects. Column (1) prescription rateit = ˇ0 + ˇ1 Disclosuret ∗P25i
uses the sample of clinics whose prescription rates are disclosed online (Open = 1).
Column (2) uses the sample of clinics whose prescription rates are not disclosed + ˇ2 Disclosuret ∗P50i + ˇ3 Disclosuret ∗P75i
online because they had less than 100 prescriptions (Open = 0). In column (3), we
use the full sample and estimate the interaction effect between “Disclosure” and + ˇ4 Disclosuret ∗P100i
“Open” dummies. In column (4), we exclude the observations in 2006Q1. In column
(5), we redefine the “Disclosure” dummy to be one if date is strictly after 2005Q4.
+ (Quarter Dummiest ) + ıi + it , (1)
*
Significant at 10%.
**
significant at 5%. where Disclosuret =1 if date is after 2006.Q1, and ıi is a clinic fixed
***
Significant at 1%. effect.
10 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16

A B C D

100
50
0
prescription rate(%)

E F G H
100
50
0

2005q12006q12007q12008q12009q12005q12006q12007q12008q12009q1

I J
100
50
0

2005q12006q12007q12008q12009q12005q12006q12007q12008q12009q1

time
Graphs by clinic ID

Fig. 5. Change in prescription rate: by each clinic in a township.

Fig. 6 shows the histograms for the estimated ˇ1 , ˇ2 , ˇ3 and ˇ4
for each township. Note that ˇ1 measures the change in antibiotic
prescription rates for those clinics whose prescription rates before
the information disclosure were in the lowest 25 percentile in the
township. Likewise, ˇ4 measures the change in antibiotic prescrip-
tion rates for those clinics whose pre-disclosure rates were in the
top 25% in the township.
From Fig. 6(d), a majority (90%) of the township’s top 25% pre-
disclosure clinics decreased their rates post-disclosure. However,
from Fig. 6(a), almost a half (47%) of the lowest 25% pre-disclosure
clinics increased their rates post-disclosure.
Alternatively, in Table 4, we estimate a probit model where the
dependent variable is one if a clinic has increased its antibiotic pre-
scription rates after the information disclosure regulation, and zero

(a) 0<prescription<25th (b) 25th<prescription<50th

.06

.04
.02 .03
.04
Density

Density
.02

.01
0

-100 -50 0 50 100 -100 -50 0 50 100

change in prescription rate change in prescription rate

(c) 50th<prescription<75th (d) 75th<prescription<100

.04
.03

.02 .03
.02
Density

Density
.01

.01
0

-100 -50 0 50 100 -100 -50 0 50 100

change in prescription rate change in prescription rate

Fig. 6. Change in prescription rates: by prescription rate before information disclosure.

 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16 11

Table 4
Increase in antibiotic prescription rates: probit analysis (dependent variable = 1 if a clinic has increased prescription rate after the information disclosure, = 0 otherwise.)

(1) (2) (3) (4)

P50 −0.2753 ***

−0.2709 ***

(0.0843) (0.1047)
P75 −0.4398*** −0.3572***
(0.0913) (0.1125)
P100 −0.6571*** −0.6148***
(0.0881) (0.1111)
Prescription rate −0.8416*** −0.7941***
Ranking (before disclosure) (0.1074) (0.1342)
Medical speciality Yes Yes Yes Yes
Random effect Township Township Township Township

Observations 2272 1538 2272 1538

Notes: All regressions control for number of doctors, clinic age, number of clinics, population, the share of age 60 and older, medical speciality dummies and the township
random effects. In columns (2) and (4), we use the balanced sample only. P50 is equal to one if prescription rate before the information disclosure was between 25 percentile
and 50 percentile, and zero otherwise. P75 and P100 are defined in a similar way.
*
Significant at 10%.
**
Significant at 5%.
***
Significant at 1%.

otherwise. Column (1) of Table 4 shows that clinics are relatively are consistent with our hypothesis that information disclosure
more likely to increase their rates if they were relatively lower regulation can allow the clinics to learn their competitors’ prescrip-
within a township pre-disclosure. Also, when we control for the tion rates and trigger perverse peer effects.
relative ranking of antibiotic prescription rates within a township
before disclosure, column (3) shows similar results. In column (2)
and (4), we restrict the analysis to the balanced sample, but the 5.3. Regression to the mean
results do not change.
Therefore, not all clinics have decreased antibiotic prescription Regression to the mean can be an alternative explanation for
rates after the information disclosure regulation. In particular, clin- the finding that those clinics with lower than average antibiotic
ics with relatively lower pre-disclosure prescription rates are much prescription rates pre-disclosure are more likely to increase pre-
more likely to increase their rates after the regulation. These results scription rates post-disclosure. To evaluate the extent of regression
to the mean bias, we focus on the sample after 2007.Q1. Then we
hypothetically assume that there was an information disclosure in
Table 5
Regression to the mean (dependent variable = antibiotic prescription rates (%)).
2008.Q1. From Fig. 3, the market prescription rates appear to have
reached a new steady state equilibrium by 2007Q1. Thus, a hypo-
(1) (2) (3) thetical disclosure in 2008Q1 should not have any further effect on
(a) 2007Q1–2009Q2 the prescription rates.
Disclosure (hypothetical) 0.1772 3.8382*** 3.2614*** In Table 5(a), we measure the relative ranking (or CDF) of
(0.1324) (0.2923) (0.3125)
each clinic’s prescription rates within its township in 2007.Q1
Ranking (at 2007Q1) 75.7823*** 76.2029***
(1.3518) (1.6701) (before the hypothetical information disclosure), called “Rank-
Disclosure × ranking −6.4400*** −5.0534*** ing”. Then, we estimate the effect of the interaction term between
(0.4431) (0.4717) the hypothetical disclosure dummy and the ranking. Column (1)
in Table 5(a) shows that the hypothetical disclosure has no sig-
Random effect Clinic Clinic Clinic
nificant effect on the antibiotic prescription rates as expected.
Observations 21,854 21,223 15,547 Column (2) in Table 5(a), however, shows that the interaction term
between the hypothetical disclosure dummy and the prescription
(b) 2005Q1–2006Q4
rate ranking has a negative and significant effect. That is, those
Disclosure (actual) −10.3428*** −1.4340*** −1.3909***
(0.1770) (0.3762) (0.4411) clinics with relatively higher prescription rates in 2007.Q1 have
Ranking (at 2005Q1) 71.2326*** 68.5265*** decreased their prescription rates after the hypothetical informa-
(1.3069) (1.6671) tion disclosure in 2008.Q1. Because there was no real disclosure in
Disclosure × ranking −15.0551*** −14.8017***
2008.Q1, this effect is likely to be due to the regression to the mean
(0.5615) (0.6612)
bias.
Random effect Clinic Clinic Clinic For comparison, in Table 5(b), we estimate the same model for
the sample before 2007.Q1. Because the real information disclo-
Observations 19,306 17,363 12,240
sure was implemented in 2006.Q1, the effect of the interaction
Notes: All regressions control for the number of doctors, the number of doctors and
term between the (real) disclosure dummy and the prescription
staff, clinic age, number of clinics, population, share of age 65 and older, quarterly
dummies, and medical speciality dummies. In column (3), only the balanced samples
rate ranking (in 2005.Q1) would include both the regression to the
are included. In (a), “Ranking” is the relative ranking of antibiotic prescription rates mean bias and the real policy effect. From column (2) in Table 5(a),
within the township in 2007Q1 (=1 if the highest, =0 if the lowest). “Disclosure” is the coefficient of the interaction term for the hypothetical dis-
one if time is after 2008Q1. In (b), “Ranking” is the ranking of antibiotic prescription closure is −6.4, while from column (2) of Table 5(b), that for the
rates within the township in 2005Q1 (=1 if the highest, =0 if the lowest). “Disclosure”
real disclosure is −15.05. Therefore, it seems that even after tak-
is one if time is after 2006Q1 and zero otherwise.
*
Significant at 10%. ing into account the regression to the mean bias, those clinics
**
Significant at 5%. with relatively higher antibiotic prescription rates before the (real)
***
Significant at 1%. information disclosure are more likely to reduce their prescription
12 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16

Table 6
Competition and information disclosure (dependent variable = antibiotic prescription rates (%)).

(1) (2) (3) (4)

Open=1 All

Disclosure −9.9817*** −9.6624*** −4.0114*** −19.7177***

(0.1800) (0.2283) (0.3618) (2.7591)
Disclosure × competition 1.6304*** 1.8257***
(0.2998) (0.4443)
Open 3.4005*** −2.3096
(0.5738) (1.5321)
Disclosure × open −6.0444*** 10.0319***
(0.4116) (2.7685)
Disclosure × open × competition 1.7853*** 1.8781***
(0.3385) (0.4461)
Fixed effect Clinic Clinic Clinic Clinic

Observations 41,878 27,949 49,405 28,044

R-squared 0.1067 0.1097 0.0756 0.1105

Notes: All regressions include quarterly dummies and clinic fixed effects. “Competition” is measured by the number of clinics per 1000 population. Columns (1) and (2) are
estimated for those clinics whose antibiotic prescription rates are disclosed (Open=1). Columns (3) and (4) incude those clinces whose antibiotic prescription rates are not
disclosed (Open=0). Columns (2) and (4) use the balanced sample only.
*
Significant at 10%.
**
Significant at 5%.
***
Significant at 1%.

rates after the disclosure.15 Again, these results are consistent with
clinic peer effects.
5.4. Competition effects
As discussed in Proposition 3, competition can have different
effects on the change in antibiotics prescription rates depending
on who is learning from the information disclosure or whether
information disclosure triggers clinic peer effects or consumer
learning effects. If consumers learn from the disclosure, stronger
competition would make the change in the prescription rate more
negative. However, if mainly the clinics learn from the disclosure,
stronger competition would make the change less negative. (See
Fig. 1.)
Thus, in Table 6, we measure competition by the number of clin-
ics per 1000 population, and estimate the effect of competition on
the change in antibiotic prescription rates due to the information
disclosure. Table 6 column (1) shows that the interaction between
the disclosure dummy and competition has a positive and signif-
icant effect on the prescription rates. In column (2), we use the
balanced sample only, but the results are robust. That is, when there
are more competing clinics, the information disclosure regulation
decreases the prescription rates less in absolute value.
In columns (3) and (4), we use the non-disclosed clinics as a
comparison group (open = 0), and estimate the interaction effect
between disclosure and competition. Despite our previous caveat
that this comparison group is likely to underestimate the effect
of information disclosure, the coefficient of the interaction term
among disclosure, open, and competition is positive and signifi-
cant. That is, higher competition reduces the effect of disclosure
in absolute value. Overall, these results are more consistent with
clinic peer effects as discussed in Proposition 3.
Alternatively, the effect of information disclosure may vary sys-
tematically across markets. And our competition measure may be
correlated with some of the other market characteristics. Though
our clinic fixed effects should control for other market character-
istics, they do not control for the interaction effect between the
disclosure dummy and other market characteristics. Therefore, we
will first estimate the heterogeneity in market responses, and ana-
lyze the robustness of the competition effect when controlling for
the interaction effects with other market characteristics.
6. Consumer learning effects
So far we have focused on the evidence for clinic peer effects on
the supply side. In this section, we analyze the extent of consumer
learning effects on the demand side.
6.1. Consumer preference and learning
Recall that from mandatory information disclosure, consumers
can learn not only the antibiotic prescription rates of their local
clinics but also the fact that antibiotics do not treat the common
cold. Then, as discussed in Proposition 2, consumers would visit
clinics less after the information disclosure. Also, if consumers pre-
fer lower antibiotic prescription rates, they would visit those clinics
with relatively high prescription rates even less frequently.
Therefore, in Table 7, we estimate how the number of patient
visits for the common cold has changed after the information dis-
closure, especially depending on the clinics’ relative ranking by
antibiotic prescription rates.16
Column (1) of Table 7 shows that after the information disclo-
sure, the number of patient visits for the common cold decreased
by 15.9%. That is, consumers may have learned that antibiotics do
not treat the common cold, and stopped going to the clinics for the
common cold. For example, as shown in Fig. 2(b), the information
disclosure website displays a clear message that antibiotics do not
treat the common cold.
Moreover, column (2) of Table 77 shows that the coefficient
of the interaction between disclosure dummy and lag ranking of
antibiotic prescription rates is negative and significant.17 That is,
the number of patient visits to those clinics with relatively high
antibiotic prescription rates have decreased even more. This result
suggests that consumers did learn and compare local clinics’ antibi-
otic prescription rates. In particular, this result is consistent with

15
Alternatively, we have used the sample (2005Q1–2005Q4) before the actual
16
disclosure, and repeated the placebo test pretending that there was information As Fig. 2(a) shows, the clinics were disclosed in the reverse order of antibiotic
disclosure in 2005Q2. The results are essentially the same as Table 5(a). For example, prescription rate ranking.
17
the estimated coefficient for the interaction between disclosure and prescription We control for the lag ranking because only the prescription rates in the previous
rate ranking is −6.162. quarter are disclosed online.
 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16 13

Table 7
Patient visits and antibiotic prescription rates (dependent variable = log(number of patient visits for the common cold)).

(1) (2) (3) (4) (5) (6)

All All Before After Before After

Disclosure −0.1591*** −0.0872***

(0.0068) (0.0134)
Ranking (t − 1) 0.2145*** −0.0070 −0.1417***
(0.0248) (0.0320) (0.0458)
Disclsoure × −0.1598***
Ranking (t − 1) (0.0200)
Deviation+ −0.0010 −0.0020**
(0.0007) (0.0008)
Deviation− −0.0002 −0.0029***
(0.0006) (0.0008)
Fixed effect Clinic Clinic Clinic Clinic Clinic Clinic

Observations 17,220 14,473 9550 4923 9550 4923

R-squared 0.2848 0.3769 0.3317 0.6068 0.3319 0.6090

Notes: All regressions include quarterly dummies and clinic fixed effects. Columns (3) and (5) are estimated for data before the information disclosure. Columns (4) and (5)
are estimated for data after the information disclosure. Ranking (t − 1) is the lag relative ranking of antibiotic prescription rates in a township where 0 is the lowest and 1
is the highest. Deviation is the lag difference between the clinic’s antibiotic prescription rate and the township median antibiotic prescription rate. Deviation+ is equal to
Deviation if Deviation > 0 and zero otherwise. Deviation− is equal to Deviation if Deviation < 0 and zero otherwise.
*
Significant at 10%.
**
Significant at 5%.
***
Significant at 1%.

our assumption that average consumers prefer lower antibiotic pre- Table 8
Lag vs. contemporenous ranking of antibiotic prescription rates (dependent vari-
scription rates for the common cold.
able = log(number of patient visits for the common cold)).
In columns (3) and (4), we estimate the model separately for
before and after the information disclosure. From column (3), (1) (2)
Before After
before the information disclosure, the relative ranking of a clinic’s
antibiotic prescription rates has no significant effect on the num- Ranking (t − 1) −0.0067 −0.1486***
ber of patient visits. This result confirms that before the disclosure, (0.0320) (0.0467)
Ranking (t) −0.0168 −0.0471
consumers did not know individual clinics’ antibiotic prescription
(0.0320) (0.0628)
rates. Fixed effect Clinic Clinic
However, from column (4) of Table 7, after the information dis-
Observations 9550 4923
closure, the relative ranking of a clinic’s antibiotic prescription rates R-squared 0.3317 0.6069
has a negative and significant effect. Again, this result is consis-
Notes: All regressions include quarterly dummies and clinic fixed effects. Column
tent with consumer learning effects and consumers’ preference for
(1) is estimated for data before the information disclosure. Column (2) is estimated
lower antibiotic prescription rates. for data after the information disclosure. Ranking (t − 1) is the lag relative ranking of
Alternatively, consumers may avoid clinics that deviate from antibiotic prescription rates in a township where 0 is the lowest and 1 is the highest.
the norm, including those clinics with much lower prescription Ranking (t) is the contemporaneous ranking.
*
rates than average. Thus, we measure the difference between each Significant at 10%.
**
Significant at 5%.
clinic’s prescription rate and the township median, called “Devia- ***
Significant at 1%.
tion”. Then, we define Deviation+ as equal to Deviation if Deviation
is positive and zero otherwise. Likewise, we define Deviation− as
equal to Deviation if Deviation is negative and zero otherwise. evidence suggests that consumers actually prefer lower prescribing
Columns (5) and (6) of Table 7 show that both Deviation+ and clinics.20
Deviation− have negative and significant effects on the number of
patient visits after the information disclosure. And the coefficients 6.2. Consumer learning vs. clinic manipulation
of the two variables are not statistically different.18 That is, con-
sumers seem to prefer lower antibiotic prescription rates for all An important caveat for our analysis is that clinics may have
clinics. manipulated the disclosed antibiotic prescription rates. For exam-
Note that we have interpreted the finding that pre-disclosure ple, because the denominator for antibiotic prescription rate is the
low-prescribing clinics are more likely to increase their rates number of patient visits, clinics can reduce antibiotic prescription
post-disclosure as evidence for clinic peer effects. However, rates by asking patients to visit clinics more frequently. However,
such a finding can also arise if the consumers prefer higher as column (1) of Table 7 shows, the number of patient visits has
antibiotics prescription rates, and insist that low-prescribing decreased, rather than increased, after the information disclosure.
clinics should increase their antibiotics prescription rates. In Alternatively, clinics can change the patients’ diagnostic codes
fact, patient demand for antibiotics has been blamed as the for those patients prescribed with antibiotics. Because only the
main reason for the overuse of antibiotics.19 However, our antibiotic prescription rates for the common cold, or acute upper
respiratory tract infection, are disclosed, clinics can change the
diagnostic codes, for example, to a lower respiratory tract infection
18
The p-value is 0.45. such as pneumonia, and continue to prescribe antibiotics. Then,
19
“A number of factors influenced the tendency to overprescribe [antibiotics],
including mostly patient demand, but also time pressure to end patient visits sooner,
fear of malpractice lawsuits if a prescription is denied,...” Forbes (July 9, 2012)
20
available from http://www.forbes.com/sites/gerganakoleva/2012/07/09/private- Currie et al. (2012) also find that the overuse of antibiotics in China is not demand
physicians-drive-up-antibiotic-resistance-helped-along-by-patients/. driven but is largely a supply-side phenomenon.
14 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16

Table 9
The effect of clinic and consumer characteristics on learning (dependent variable = log(number of patient visits for the common cold)).

(1) (2) (3) (4) (5) (6) (7)

Disclosure −0.0872 ***

−0.0847 ***
−0.0929 ***
−0.0866 ***
−0.0868 ***
−0.0864 ***
−0.0873***
(0.0134) (0.0135) (0.0140) (0.0134) (0.0135) (0.0134) (0.0140)
Ranking (t − 1) 0.2145*** 0.2793*** 0.1550** 0.4100*** −0.1748 0.2077*** 0.0334
(0.0248) (0.0441) (0.0610) (0.0826) (0.1300) (0.0278) (0.1770)
INT (=disclosure × ranking) −0.1598*** −0.2077*** −0.1076*** −0.2932*** −0.0589 −0.1480*** −0.2867***
(0.0200) (0.0242) (0.0290) (0.0348) (0.0488) (0.0204) (0.0655)
INT × clinic age 0.0034*** 0.0032***
(0.0010) (0.0010)
INT × pop density −0.0019* −0.0010
(0.0010) (0.0010)
INT × college grad 0.0057*** 0.0059***
(0.0012) (0.0013)
INT × age 9 under −0.0118** −0.0012
(0.0052) (0.0056)
INT × pediatric −0.0673*** −0.0819***
(0.0220) (0.0234)
Fixed effect Clinic Clinic Clinic Clinic Clinic Clinic Clinic

Observations 14,473 14,251 13,410 14,473 14,425 14,473 13,206

R-squared 0.3769 0.3791 0.3746 0.3783 0.3771 0.3774 0.3795

Note: All regressions include quarterly dummies and clinic fixed effects. Clinic age and pediatric dummy variable are measured by each clinic in 2007. Population density
and share of age 9 under are measured by each township in 2009. The share of college graduates is measured by each district in 2009.
*
Significant at 10%.
**
Significant at 5%.
***
Significant at 1%.

the reported number of patient visits for the common cold would
decline as shown in column (1) of Table 7.
However, if clinics have manipulated the patients’ diagnostic
codes in such a way, both the reported number of patient visits for
the common cold and the antibiotic prescription rates for the com-
mon cold would decline together at the same time at the clinic level.
But both columns (2) and (4) of Table 7 show that when the ranking
of antibiotic prescription rates of a clinic declined, the number of
patient visits for that clinic increased.
Moreover, manipulation of diagnostic codes would affect the
number of patient visits and the antibiotic prescription rates at the
same time. On the other hand, because the antibiotic prescription
rates in the previous quarter are disclosed online, consumer learn-
ing effects suggest that the number of patient visits would depend
on the lag antibiotic prescription rates, not on the contemporane-
ous rates.
Thus, in columns (1) and (2) of Table 8, we control for both
lag ranking of antibiotic prescription rates and contemporane-
ous rates at the same time. Note that the contemporaneous rates
are not significant either before or after the disclosure. However,
the lag prescription rates are significant only after the disclosure.
Therefore, while we cannot rule out the possible manipulation of
disclosed prescription rates, our results appear to be driven by
consumer learning effects rather than manipulation of disclosed
prescription rates.21
6.3. Heterogeneity in consumer learning
We consider consumer learning as a process where consumers
use the disclosed information to (Bayesian) update their belief on
clinics’ quality. Because consumers’ priors and weights on the dis-
closed information are subjective, the extent of learning can be
heterogeneous depending on consumer and clinic characteristics.
21
Alternatively, consumers may have interpreted high antibiotic prescription
rates as a signal for more infectious patients and avoided that clinic. This is unlikely,
however, because the awareness of hospital-acquired infection has unfortunately
been very low in Korea, especially during our sample period.
Since we do not have data on individual consumer charac-
teristics, in Table 9, we estimate how the average consumer
characteristics in each township and clinic characteristics affect
consumer learning. As a benchmark, in column (1) of Table 9,
we estimate the effect of an interaction term between disclosure
dummy and clinics’ antibiotic prescription ranking. The coefficient
of this interaction term measures how much consumers’ response
to the antibiotic prescription ranking has changed after the infor-
mation disclosure, and thus provides a measure for consumer
learning.
In columns (2)–(6), we estimate how this measure for con-
sumer learning depends on clinic and consumer characteristics.
Column (2) shows that the interaction term with the clinic age is
positive and significant. Because consumers respond negatively to
the antibiotics ranking after the information disclosure, this result
implies that the consumer learning is smaller in absolute value if
the clinics are older. One interpretation is that for older clinics, con-
sumers have observed their quality for a long time and have a strong
prior on their quality. Then, the newly disclosed information would
not change consumers’ posterior belief on their quality much. Thus,
when old clinics’ antibiotic prescription ranking increases, con-
sumers may respond less negatively.
Column (3) in Table 9 shows that the interaction term with pop-
ulation density has a negative and significant effect, which implies
that consumer learning is larger in absolute value when population
density within a township is high. Recall that a 2007 government
survey showed that only 7% of consumers have actually visited the
information disclosure website. The significant evidence for con-
sumer learning despite the low website visit rates suggests that
consumers may have learned from some of their neighbors. Thus,
our result is consistent with a hypothesis that in more densely pop-
ulated areas, consumer learning through neighbors is likely to be
more important.
Column (4) shows that the interaction term with the share of
college graduates has a positive and significant effect. That is, the
consumer learning effect is smaller in absolute value in townships
with a more highly educated population. This result may seem
counter-intuitive. However, a 2010 survey shows that more highly
educated respondents are more likely to believe that antibiotics can
 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16 15

Table 10
The effect of clinic and consumer characteristics on learning (dependent variable = antibiotic prescription rates (%)).

(1) (2) (3) (4) (5) (6) (7)

D (=disclosure) −9.3494 ***

−11.9383 ***
−7.4970 ***
−12.3032 ***
−5.6907 ***
−8.2533 ***
−8.9290***
(0.1397) (0.2726) (0.4109) (0.4833) (0.7831) (0.1926) (1.0503)
D × competition 1.5377*** 1.2226*** 1.6463*** 1.2601*** 1.2523*** 0.7942**
(0.3000) (0.3108) (0.2998) (0.3069) (0.2981) (0.3140)
D × clinic age 0.1514*** 0.1505***
(0.0159) (0.0165)
D × pop density −0.1008*** −0.0668***
(0.0153) (0.0157)
D × college grad 0.1000*** 0.0809***
(0.0193) (0.0198)
D × age 9 under −0.4825*** −0.1705*
(0.0857) (0.0908)
D × pediatric −8.4123*** −8.3416***
(0.3495) (0.3666)
Fixed effect Clinic Clinic Clinic Clinic Clinic Clinic Clinic

Observations 42,020 41,189 38,956 41,878 41,878 41,878 38,311

R-squared 0.1057 0.1092 0.1062 0.1073 0.1074 0.1197 0.1223

Note: All regressions include quarterly dummies and clinic fixed effects. Clinic age and pediatric dummy variable are measured by each clinic in 2007. Population density
and share of age 9 under are measured by each township in 2009. The share of college graduates is measured by each district in 2009.
*
Significant at 10%.
**
Significant at 5%.
***
Significant at 1%.

treat the common cold.22 Also, more highly educated consumers
may put more weight on their own prior belief and relatively less
weight on the newly disclosed information by the government.
Thus, the consumer learning effect can be smaller for more highly
educated consumers.
Column (5) shows that the interaction term with the share of
children has a negative and significant effect. That is, consumer
learning is larger in absolute value in townships with more chil-
dren. This result suggests that parents with children are either
putting more weight on the disclosed information or are more sen-
sitive to the perceived quality of clinics. Thus, they respond more
sensitively to the disclosed antibiotics prescription ranking. Simi-
larly, column (6) shows that patient visits to pediatric clinics have
become more sensitive to the antibiotics ranking after the disclo-
sure, compared with the other specialities.
In column (7), we control for all the interactions at the same
time. The qualitative results do not change. But the interaction
with the share of children becomes insignificant largely due to the
control for the interaction with the pediatric dummy variable. The
interaction with population density also becomes insignificant, but
the sign of the coefficient does not change.
These results suggest that there is significant heterogene-
ity in consumer learning, depending on clinic and consumer
characteristics.23
6.4. Heterogeneity in consumer learning and change in antibiotic
prescription rates
The patterns of heterogeneity in consumer learning are impor-
tant as they can help policy makers to predict when the information
disclosure regulation will be more effective. For example, our evi-
dence shows that for older clinics and more educated consumers,
the extent of learning from the information disclosure is smaller.
Thus, we can predict that the antibiotic prescription rates would fall
22
Korea Food & Drug Administration PressKit (2011-04-26).
23
We have also analyzed the effects of the share of elderly in the population and
the size of local tax revenue. But the effects were not significant, and are not reported
due to space constraints.
less after information disclosure regulation for older clinics and for
townships with more educated consumers.
To confirm such predictions, in Table 10, we estimate how the
changes in antibiotic prescription rates vary with clinic and con-
sumer characteristics. More specifically, as in Table 9, we interact
clinic and consumer characteristics with the disclosure dummy
variable. In our base specification, column (1), the average antibi-
otic prescription rates have decreased by 9.34% after the mandatory
information disclosure. Column (2) shows that the interaction term
between the disclosure dummy and clinic age is positive and sig-
nificant. That is, the antibiotic prescription rates decreased less for
older clinics. Note that this pattern is consistent with the pattern
of consumer learning found in column (2) of Table 9.
Likewise, the estimates from columns (3)–(7) of Table 10 are
remarkably consistent with those found in the corresponding
columns (3)–(7) in Table 9. That is, Table 9 shows that the extent of
consumer learning from information disclosure is larger in more
densely populated townships with relatively more children or
pediatric clinics but with fewer college graduates. Then, given
consumer preference for lower antibiotic prescription rates, antibi-
otic prescription rates should decrease more in such townships.
Consistent with this prediction, columns (3)–(7) of Table 10 show
that the antibiotic prescription rates have decreased more in more
densely populated townships with relatively more children or
pediatric clinics but with fewer college graduates. Also note that
the positive effect of the interaction between disclosure and com-
petition is robust in all specifications.
7. Conclusion
Mandatory information disclosure is an increasingly popular
regulatory device to reduce the information asymmetry problem
between sellers and buyers. Consequently, the previous literature
has focused on whether consumers learn from the information
disclosure and pressure sellers to increase their quality, called con-
sumer learning effects. This paper, however, shows that mandatory
information disclosure can also allow sellers to observe their com-
petitors’ attributes, and trigger peer effects among them. More
specifically, in the context of antibiotic prescription rates for the
common cold, this paper shows that some clinics have increased,
instead of decreasing, their antibiotic prescription rates when they
16 I. Kwon, D. Jun / Journal of Health Economics 42 (2015) 1–16
found out that other local clinics were prescribing more antibiotics
than they were. Therefore, even though the average prescription
rates have decreased after the mandatory information disclosure,
the decline was smaller when there were more peer clinics.
In the literature on peer effects, this paper provides an alterna-
tive way of identifying peer effects. While most previous studies
on peer effects have attempted to find an exogenous change in the
behavior of peers, it is difficult to find an exogenous shock that
affects some peers but not others. This paper suggests that the
exogenous change in the observability of peer behavior can be an
alternative way of identifying the peer effects, which can avoid the
“reflection problem” discussed by Manski (1993).
This paper also finds significant evidence for consumer learning
effects. After the information disclosure, consumers were less likely
to visit clinics with higher antibiotic prescription rates for the com-
mon cold. This result suggests that overuse of antibiotics may not
be driven by patient demands, but by hospital competition and an
information asymmetry problem. We also find significant hetero-
geneity in consumer learning. In particular, after the information
disclosure, consumers became more sensitive to the ranking of
antibiotic prescription rates in townships with younger clinics,
higher population density, lower education, and more children or
pediatric clinics. Consequently, we find that antibiotic prescription
rates have also declined more in such townships. These results may
explain why some previous studies have found that the informa-
tion disclosure policy had a significant effect while others have not.
These results also suggest when the information disclosure policy
can be most effective.
One of the limitations of this study is that we do not analyze
what caused clinic peer effects. For example, such peer effects may
arise from rational selfish decisions such as social learning or strate-
gic interactions, or from intrinsic social preference. Such an analysis
could be an interesting topic for future studies.
References
Akerlof, G.A., 1970. The market for ‘lemons’: quality uncertainty and the market
mechanism. Quarterly Journal of Economics 84 (3), 488–500.
Akerlof, G.A., 1980. A theory of social custom, of which unemployment may be one
consequence. Quarterly Journal of Economics 94 (4), 749–775.
Albaek, S., Mollgaard, P., Overgaard, P.B., 1997. Government-assisted oligopoly
coordination? A concrete case. Journal of Industrial Economics 45 (4),
429–443.
Ali, M.M., Heiland, F.W., 2011. Weight-related behavior among adolescents: the role
of peer effects. PLoS ONE 6 (6), e21179.
Bennett, D., Che-Lun Hung, Tsai-Ling Lauderdale, 2011. Health care competition and
antibiotic use in Taiwan. Harris School of Policy 12.
Bikhchandani, S., Hirshleifer, D., Welch, I., 1992. A theory of fads, fashion, custom,
and cultural change as informational cascades. Journal of Political Economy 100
(5), 992–1026.
Bird, L., 2009. Disclosure Issues: Renewable Energy Purchasing. Mimeo.
Board, O., 2009. Competition and disclosure. Journal of Industrial Economics 57 (1),
197–213.
Brody, H., 2005. Patient ethics and evidence-based medicine – the good healthcare
citizen. Cambridge Quarterly of Healthcare Ethics 14, 141–146.
Butler, C.C., Rollnick, S., Maggs-Rappaport, R.F., Slott, N., 1998. Understanding the
culture of prescribing: qualitative study of general practitioners’ and patients’
perception of antibiotics for sore throats. BMJ 317 (7159), 637–642.
Butters, G.R., 1977. Equilibrium distributions of sales and advertising prices. Review
of Economic Studies 44 (3), 465–491.
Carrell, S.E., Malmstrom, F.V., West, J.E., 2008. Peer effects in academic cheating.
Journal of Human Resources 43 (1), 173–207.
Chipty, T., Witte, A.N., 1998. “Effects of Information Provision in a Vertically Differ-
entiated Market,” NBER Working Paper No. 6493.
Conly, J., 1998. Controlling antibiotic resistance by quelling the epidemic of overuse
and misuse of antibiotics. Canadian Family Physician 44, 1769–1784.
Currie, J., Lin, W., Meng, J., 2012. “Antibiotic Abuse in China: Supply or Demand?,”
Working Paper.
Duflo, E., Saez, E., 2002. Participation and investment decisions in a retirement plan:
the influence of colleagues’ choices. Journal of Public Economics 85, 121–148.
Fischer, P., Huddart, S., 2008. Optimal contracting with endogenous social norms.
American Economic Review 98 (4), 1459–1475.
Sara, F., Karlsson, J., 2012. “Competition and Antibiotics Prescription,” IFN Working
Paper No. 939.
Fortin, B., Lacroix, G., Villeval, M.-C., 2007. Tax evasion and social interactions. Jour-
nal of Public Economics 91 (11), 2089–2112.
Gaviria, A., Raphael, S., 2001. School based peer effects and juvenile behavior. Review
of Economics and Statistics 83 (2), 257–268.
Glaeser, E., Shleifer, A., 2003. The rise of the regulatory state. Journal of Economic
Literature 41, 401–425.
Gonzales, R., Malone, D.C., Maselli, J.H., Sande, M.A., 2001. Excessive antibiotic use
for acute respiratory infections in the United States. Clinical Infectious Diseases
33 (6), 757–762.
Gordon, J.P.P., 1989. Individual morality and reputation costs as deterrents to tax
evasion. European Economic Review 33 (4), 7–805.
Grossman, S.J., Hart, O.D., 1980. Disclosure laws and takeover bids. Journal of Finance
35 (2), 323–334.
Hibbard, J.H., Jewett, J.J., 1997. Will quality report cards help consumers? Health
Affairs 16 (3), 218–228.
Jin, G.Z., Leslie, P., 2003. The effect of information on product quality: evidence
from restaurant hygiene grade cards. Quarterly Journal of Economics 118 (2),
409–451.
Jun, D., Chung, G., 2011. Analysis on the effect of information disclosure – antibiotic
prescription rates for the common cold in hospitals and clinics in Seoul. Korean
Association for Policy Studies 20 (2), 109–142.
Lautenbach, E., Patel, J.B., Bilker, W.B., Edelstein, P.H., Fishman, N.O., 2001. Extended-
spectrum ˇ-lactamase-producing Escherichia coli and Klebsiella pneumoniae:
risk factors for infection and impact of resistance on outcomes. Clinical Infec-
tious Diseases 32 (8), 1162–1171.
Longo, D.R., Garland, L.G., Wayne Schramm, Judy Fraas, Barbara Hoskins, Vicky
Howell, 1977. Consumer reports in health care: do they make a difference in
patient care? Journal of the American Medical Association 278 (19), 1579–1584.
Marshall, M.N., Shekelle, P.G., Leatherman, S., Brook, R.H., 2000. The public release of
performance data: what do we expect to gain: a review of the evidence. Journal
of the American Medical Association 283 (14), 1866–1874.
Malani, A., Buchman, T.G., Dushoff, J., Effron, M.B., 2008. Antibiotic overuse: the
influence of social norms. Journal of the American College of Surgeons 265.
Manski, C.F., 1993. Identification of endogenous social effects: the reflection prob-
lem. Review of Economic Studies 60 (3), 531–542.
Mas, A., Moretti, E., 2009. Peers at work. American Economic Review 99 (1), 112–145.
Matthews, S., Postlewaite, A., 1985. Quality testing and disclosure. RAND Journal of
Economics, 328–340.
Mennemeyer, S.T., Morrisey, M.A., Howard, L.Z., 1997. Death and reputation: how
consumers acted upon HCFA mortality information. Inquiry 34 (2), 117–128.
Milgrom, P., 1981. Good news and bad news: representation theorems and applica-
tions. Bell Journal of Economics 12 (2), 380–391.
Myles, G.D., Naylor, R.A., 1996. A model of tax evasion with group conformity and
social customs. European Journal of Political Economy 12 (1), 49–66.
Nelson, P., 1974. Advertising as information. Journal of Political Economy 81 (4),
729–754.
Njoroge, K., 2003. Information pooling and collusion: implications for the livestock
mandatory reporting act. Journal of Agricultural and Food Industrial Organiza-
tion 1.
Robohm, C., Ruff, C., 2012. Diagnosis and treatment of the common cold in pedi-
atric patients. Journal of the American Academy of Physician Assistants 25 (12),
43–47.
Sacerdote, B., 2001. Peer effects with random assignment: results for Dartmouth
roommates. Quarterly Journal of Economics 116 (20), 681–704.
Salop, S., Stiglitz, J., 1977. Bargains and Ripoffs: a model of monopolistically com-
petitive price dispersion. Review of Economic Studies 44 (3), 493–510.
Schelling, T.C., 1978. Micromotives and Macrobehavior. Norton, New York/London.
Shavell, S., 1994. Acquisition and disclosure of information prior to sale. RAND
Journal of Economics, 20–36.
Shekelle, P.G., Lim, Y.-W., Mattke, S., Damberg, C., 2008. Does public release of
performance results improve quality of care? A systematic review. The Health
Foundation, London, UK.
Spicer, M.W., Becker, L.A., 1980. Fiscal inequity and tax evasion: an experimental
approach. National Tax Journal 33 (2), 171–175.
Stigler, G.J., 1961. The economics of information. Journal of Political Economy 69 (3),
213–225.
Vladeck, B.C., Goodwin, E.J., Myers, L.P., Sinisi, M., 1988. Consumers and hospital
use: the HCFA ‘death list’. Health Affair 7 (1), 122–125.
Wilson, J., 2007. Peer effects and cigarette use among college students. Atlantic
Economic Journal 35 (2), 233–247.
Yoo, H.J., Song, E., Lee, K.U., Lee, E.K., Lee, J.A., 2009. Misuse of antibiotics and related
awareness of consumers. Journal of Korean Association for Crisis and Emergency
Management 1, 98–122.
 Journal of Health Economics 42 (2015) 17–28

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Recessions, healthy no more?

Christopher J. Ruhm ∗,1
University of Virginia and National Bureau of Economic Research, Charlottesville, VA 22904-4893, United States

a r t i c l e i n f o a b s t r a c t

Article history: Over the 1976–2010 period, total mortality shifted from strongly procyclical to being weakly or unrelated
Received 24 September 2014 to macroeconomic conditions. The association is likely to be poorly measured when using short (less
Received in revised form 10 March 2015 than 15 year) analysis periods. Deaths from cardiovascular disease and transport accidents continue to
Accepted 11 March 2015
be procyclical; however, countercyclical patterns have emerged for fatalities from cancer mortality and
Available online 20 March 2015
external causes. Among the latter, non-transport accidents, particularly accidental poisonings, play an
important role.
Keywords:
© 2015 Elsevier B.V. All rights reserved.
Mortality
Health
Recessions
Macroeconomic conditions

1. Introduction the state unemployment rate was estimated to decrease total

Health is usually thought to worsen when the economy weak-
ens, but substantial recent research suggests that mortality actually
declines during such periods. Following Ruhm (2000), most recent
studies utilize longitudinal data and panel techniques to control for
many confounding factors, including time-invariant area-specific
determinants and characteristics that vary over time in a uni-
form manner across locations1 . Using data from a variety of
countries and time periods, these investigations provide strong
evidence of procyclical fluctuations in total mortality and sev-
eral specific causes of death2 . In Ruhm’s (2000) analysis of the
U.S., covering 1972–1991, a one percentage point increase in
mortality by 0.5% and motor vehicle and cardiovascular disease
(CVD) deaths by 3.0% and 0.5%, with reductions also observed
for fatalities from influenza/pneumonia, liver disease, non-vehicle
accidents and homicides. By contrast, cancer mortality was unaf-
fected and suicides were estimated to rise by 1.3%3 . Using similar
empirical methods, the procyclicality of total mortality has been
confirmed for Germany (Neumayer, 2004), Spain (Tapia Granados,
2005), France (Buchmueller et al., 2007), Mexico (Gonzalez and
Quast, 2011), Canada (Ariizumi and Schirle, 2012), OECD countries
(Gerdtham and Ruhm, 2006), and Pacific-Asian nations (Lin, 2009)4 .
Motor vehicle and CVD fatalities are procyclical in almost all stud-
ies, with more variation in mortality from other causes5 .

∗ Tel.: +1 434 243 3729.

E-mail address: ruhm@virginia.edu
1
I thank seminar participants at the College of William and Mary, New School for
Social Research, National Bureau of Economic Research, Bureau of Economic Analy-
sis, American Health Economics Conference and the Southeastern Health Economics
Study Group for helpful comments and the University of Virginia Bankard Fund for
financial support.
1
Earlier investigations (e.g. Brenner, 1971, 1979) typically used time series data
for a single geographic location. This research has been criticized on methodological
grounds (e.g. Kasl, 1979; Gravelle et al., 1981) and suffers from the fundamental
problem that any lengthy time-series may contain omitted confounding factors that
are spuriously correlated with health. Ruhm (2012) provides a detailed discussion
of these issues.
2
Mortality rates are the most common proxy for health: they represent the most
severe negative health outcome, are well measured and diagnosis generally does not
depend on access to the medical system. However, changes in non-life-threatening
health conditions are not accounted for. Due to limited data availability, few anal-
yses examine how macroeconomic conditions affect morbidity. Exceptions include
Ruhm (2003) and Charles and DeCicca (2008).
3
Thus, mental health and physical health may move in the opposite direction.
4
Economou et al. (2008) find that total mortality is negatively but insignificantly
related to unemployment rates for 13 EU countries but that the unemployment coef-
ficient reverses sign when controlling health behaviors (smoking, drinking, calorie
consumption) and other potential mechanisms (like pollution rates).
5
Stuckler et al. (2009) obtain evidence from 26 EU countries of positive, negative
and neutral relationships between unemployment rates and suicides, deaths from
transport accidents, and total mortality; however, the statistical methods focus on
rates of changes in mortality and unemployment, making it difficult to compare
the results with other related research. Analyses undertaken as early as the 1920s
uncovered positive relationships between economic activity, total mortality and
several specific causes of death (Ogburn and Thomas, 1922; Thomas, 1927; Eyer,
1977), as have some recent analyses using different methods (e.g. Fishback et al.,
2007; Tapia Granados and Diez Roux, 2009).

http://dx.doi.org/10.1016/j.jhealeco.2015.03.004
0167-6296/© 2015 Elsevier B.V. All rights reserved.
18 C.J. Ruhm / Journal of Health Economics 42 (2015) 17–28
Some investigations suggest that mortality has become less
procyclical or countercyclical in recent years. Using methods and
data similar to Ruhm (2000), Stevens et al. (2011) find that a one
percentage point increase in the state unemployment rate was
associated with a 0.40% reduction in total mortality from 1978 to
1991, but a smaller 0.19% decrease when extending the analysis
through 20066 . McInerney and Mellor (2012) estimate that a one-
point rise in joblessness lowered the mortality rates of persons 65
and over by 0.27% during 1976–1991, but raised them 0.49% from
1994 to 2008. Svensson (2007) uncovers a positive relationship
between Swedish unemployment rates and heart attack deaths
from 1987 to 20037 .
Changes in health behaviors provide a potential mechanism for
the mortality response. Consistent with this, reductions in drink-
ing, obesity, smoking and physical inactivity during bad economic
times have been demonstrated (Ruhm and Black, 2002; Ruhm,
2005; Gruber and Frakes, 2006; Freeman, 1999; Xu, 2013), and
Edwards (2011) shows that individuals spend more time socializ-
ing and caring for relatives during such periods. However, research
using recent data again raises questions about the strength and
direction of these relationships. Charles and DeCicca (2008) indi-
cate that male obesity is countercyclical; Arkes (2009) obtains a
similar result for teenage girls (but not boys); Arkes (2007) shows
that teenage drug use increases in bad times; Dávlos et al. (2012)
uncover a countercyclical pattern for some types of alcohol abuse
and dependence; Colman and Dave (2013) suggest that increased
leisure-time exercise during periods of economic weakness is more
than offset by reductions in work-related physical exertion. Such
findings are provocative although, as shown below, they should be
viewed with skepticism because the analysis periods are too short
(eight years or less) to provide definitive results.
Using U.S. data covering 1976–2010, the present study
examines whether the relationship between macroeconomic con-
ditions and mortality has changed over time. Comparability with
previous investigations is maximized by using empirical meth-
ods that conform closely to that research8 . Three primary results
emerge. First, total mortality has shifted from being strongly pro-
cyclical to being weakly related or unrelated to macroeconomic
conditions. Evidence from prior research that deaths decline when
the economy deteriorates largely reflects the inclusion of early
sample years, when this was the case. Second, the results obtained
using relatively short (less than 15 year) periods show consider-
able instability and should probably be viewed as unreliable. Third,
fatalities due to cardiovascular disease and, to a smaller degree,
transport accidents continue to be procyclical, whereas strong
countercyclical patterns for cancer and some external sources of
death (particularly accidental poisonings) have emerged.
2. Research design
This analysis uses variations of previously employed panel data
methods (e.g. by Ruhm, 2000) to analyze the relationship between
macroeconomic conditions and mortality rates. The estimating
equation is:
 
ln Mkjt = ˛kj + Xjt ˇ + Ujt  + kt + Tjkt + εkjt (1)
6 10
The estimated reduction rises to 0.33% over the 1978–2006 period when using
age-adjusted mortality rates.
7
Using time-series methods for the U.S. from 1961 to 2010, Lam and Piérard
(2014) also argue that total and cardiovascular mortality have become less pro-
cyclical over time, while motor vehicle fatalities remain strongly procyclical.
8
One exception is the use of an uncommonly detailed set of age controls.
where Mkjt is the mortality rate from source k in state j at year
t, U is the state unemployment rate, X a vector of covariates, ˛
a state fixed-effect,  a general time effect, T a state-specific lin-
ear time trend, ε is the error term, and ˆ provides the estimated
macroeconomic effect of key interest9 .
The year effects (kt ) hold constant determinants of death that
vary uniformly across locations over time (e.g. advances in widely
used medical technologies or behavioral norms); the location fixed-
effects (˛kj ) account for those that differ across states but are
time-invariant (such as persistent lifestyle disparities between
residents of Nevada and Utah). Since the supplementary time-
varying state characteristics (Xjt ) do not necessarily control for all
time-varying determinants of death, the models also include state-
specific trends (Tjkt )10 . The 1976–2010 analysis period reflects the
availability of consistent data on state unemployment and mor-
tality rates. The macroeconomic impact is then identified from
within-location variations in mortality rates, relative to changes
in other states and after controlling for demographic characteris-
tics and state-trends. Since the impact of national business cycles is
absorbed by the time effects, discussions of macroeconomic effects
refer to changes within-states rather than at the national level.
One way of investigating whether the impact of macroecono-
mic conditions on mortality has changed is to compare predicted
effects differ across sub-periods. However, since such estimates are
often sensitive to the choice of starting or ending years, two alter-
native strategies are employed. First, models for total mortality are
estimated with differing starting and ending dates, and with vary-
ing lengths of the analysis period. The second, and main, method
specifies analysis periods of fixed duration and then sequentially
estimates models for all alternative sample windows permitted by
the data. Most commonly, 20-year periods are used with results
obtained for 16-windows ranging from 1976–1995 to 1991–2010.
Figures are frequently provided with point estimates (and
sometimes confidence intervals) on the unemployment rate coeffi-
cient presented for each analysis window. Tables are also supplied
showing unemployment coefficients and standard errors for the
first and last of the 20-year periods (1976–1995 and 1991–2010),
denoted by ˆ  and ŝ , respectively, where  equals 1 (2) in the first
(last) period. I test whether the macroeconomic effect has changed
by providing estimates for ˆ = ˆ 2 − ˆ 1 .
Using Eq. (1), (ek ˆ − 1) × 100% provides the predicted per-
centage change in mortality from source k resulting from a one
percentage point increase in the unemployment rate. While these
estimates show the relative size of the macroeconomic effect, they
do not directly indicate changes in the absolute number of pre-
dicted fatalities because, for example, large relative effects may
imply small absolute changes for sources that are responsible for
few deaths. These relative effect sizes are translated into absolute
numbers through estimates of:
 
ek
ˆ
− 1 × k D (2)
where ˆ = ˆ 2k − ˆ 1k , D is the average annual number of deaths
(2222,313) and k is the share of deaths due to source k over the
1976–2010 period.
9
Unemployment rates are used to proxy macroeconomic conditions; however, a
procyclical variation in mortality does not imply that the loss of a job improves
health. To the contrary, Sullivan and von Wachter (2009) show that job loss is
associated with increases in individual mortality rates.
Mortality trends vary considerably across sources of death, with large secular
reductions for total mortality and that from cardiovascular disease and external
sources, a relatively flat trend for cancer, and an increase for other disease deaths.
State-year population weights were also sometimes incorporated but unweighted
estimates are generally preferred (Wooldridge, 1999; Butler, 2000; Solon et al.,
2015) and so are focused upon below.
 C.J. Ruhm / Journal of Health Economics 42 (2015) 17–28
A potential concern is that calculations based on (2) do not
account for the possibility that a portion of the trend in macro-
economic effects on overall mortality rates could reflect secular
changes in the shares of deaths due to specific sources. This was
examined using a variation of the Oaxaca (1973) and Blinder’s
(1973) decomposition method. The method and results, which are
summarized in the online appendix, indicate that almost all of the
macroeconomic effect was due to changes in the coefficients, rather
than in mortality shares, so that predictions obtained from (2) are
useful.
3. Data and descriptive statistics
Annual average state unemployment rates, the main prox-
ies for macroeconomic conditions were obtained from the U.S.
Department of Labor’s Local Area Unemployment Statistics Database
(www.bls.gov/lau/lauov.htm), which provides monthly estimates
of total employment and unemployment rates for census regions
and divisions, states, metropolitan statistical areas, counties, and
some cities11 . Concepts and definitions underlying the LAUS data
come from the Current Population Survey. Mortality data are from
the Center for Disease Control and Prevention’ Compressed Mor-
tality Files (CMF) (www.cdc.gov/nchs/data access/cmf.htm), which
contain information for every death of a U.S. resident including:
state and county of residence, year of death, race and sex, Hispanic
origin (after 1998), age group (16 categories), underlying cause of
death (ICD codes and CDC recodes). Data prior to 1988 are pub-
licly available; those from 1989 to 2010 were obtained by special
agreement with the CDC. Population data (the denominator in the
mortality rate calculations) from 1981 on come from the National
Cancer Surveillance Epidemiology and End Results (SEER) program
(http://www.seer.cancer.gov/data)12 . These were supplemented
by census estimates, included in the CMF files, for 1976–1980.
In addition to total annual mortality rates, sex-specific death
rates were constructed, as were fatality rates for five age groups
(<25, 25–44, 45–64, 65–74, and ≥75 year olds) and deaths from
major diseases and external causes13 . The SEER data were addi-
tionally used to construct independent variables for the share of
the state population who were female, nonwhite, Hispanic, and
aged <1, 1–19, 45–54, 55–64, 65–74, 75–84 and ≥85 years old14 .
The analysis of cause-specific mortality introduces complica-
tions. From 1976 to 1978, cause of death was categorized using the
8th revision of the International Classification of Diseases (ICD-8
codes). ICD-9 codes were used between 1979 and 1998, and ICD-10
categories since 1999. Crosswalks have been established between
11
Some recent studies of macroeconomic patterns of health behaviors have ana-
lyzed county-level or MSA data (e.g. Charles and DeCicca, 2008; An and Liu, 2012).
This has potential advantages (e.g. examining smaller regional economies) and
disadvantages (e.g. greater measurement error). For this investigation, the major
disadvantage is that a consistent data series of county unemployment rates only
begins in 1990 and the Department of Labor cautions against using county level
data prior to that time. Also, Lindo (2015) provides evidence that the health effects
of macroeconomic conditions are understated when using more disaggregated (e.g.
county rather than state) data. Preliminary analysis revealed similar results using
state and county data starting in 1990.
12
The SEER data are designed to supply more accurate population estimates for
intercensal years than standard census projections, and to adjust for population
shifts in 2005, resulting from Hurricanes Katrina and Rita. Differences between the
SEER and CMF population estimates are miniscule prior to 2000 but are sometimes
reasonably large (up to 3%) after 2003.
13
I examined other age-specific death rates, including infant mortality rates, in
preliminary analysis, but focus on these age groupings since the large majority of
deaths occur to those who are relatively old.
14
Hispanic population shares are not provided prior to 1981. Therefore, shares for
1976–1980 were extrapolated as a linear trend for changes occurring between 1981
and 1986.
19
ICD-8 and ICD-9 and between ICD-9 and ICD-10 coding systems;
however, the correspondence is imperfect. These issues are typi-
cally minor when looking at broad causes of death (e.g. those from
cardiovascular disease) but are important for many specific sources
of mortality. To provide information on this, the National Cen-
ter for Health Statistics has calculated “estimated comparability
ratios” indicating the relative number of deaths in 1996 attributed
to a specific cause using ICD-9 and ICD-10 classifications (Anderson
et al., 2001) and, similarly, for 1976 using ICD-8 versus ICD-9 codes
(Klebba and Scott, 1980).
When the estimated comparability ratios are close to one (i.e.
a similar number of deaths are reported using either ICD system),
issues of data comparability are likely to be minor and well cap-
tured by the inclusion of regression year fixed-effects. For example,
the estimated comparability ratios are 1.013 and 1.003 for CVD
and cancer fatalities, when using ICD-8 and ICD-9 codes, and 0.998
and 1.007 for ICD-9 and ICD-10 categories. However, the poten-
tial problems are greater for some numerically important causes
of death, and for others that have been analyzed in previous
research15 . Due to these concerns, the analysis of disease mortality
is restricted to the major categories of CVD and malignant neo-
plasms, as well as a generic grouping of all other disease types16 . A
fuller investigation is provided for subcategories of external deaths,
including those from transport accidents, non-transport (other)
accidents, intentional self-harm (suicide), and homicide/legal
intervention. Because non-transport accidents will be shown to be
particularly important, separate analysis is conducted for the sub-
components: falls, drowning/submersion, smoke/fire/flames, and
poisoning/exposure to noxious substances17 .
Appendix Table A1 details the ICD codes used to classify causes
of death. Means and sample standard errors for mortality rates (per
100,000 population)and state characteristics are detailed in the
online appendix. Appendix Table A2 illustrates how the sources
of death changed over the analysis period, showing numbers
and shares of fatalities during 1976–1995 and 1991–2010. As
expected, given increased life expectancy, the proportion of mor-
tality accounted for by the elderly has grown substantially. Declines
in the share of cardiovascular deaths has been offset by increases in
mortality from other diseases. The fraction from external sources
changed little, with reductions from fatal transport accidents and
homicides being compensated for by increases in non-transport
accidents, particularly poisoning deaths.
4. The declining procyclicality of total mortality
Fig. 1 supplies three ways of examining whether the procycli-
cality of total mortality has diminished over time, by estimating Eq.
(1) for different time periods. Solid lines show point estimates and
dotted lines the 95% confidence intervals.
Fig. 1A displays unemployment rate coefficients where the anal-
ysis period begins in 1976 and ends in years ranging between 1995
and 2010. The magnitude of the estimated macroeconomic effect
declines monotonically but modestly as the sample is extended to
more recent periods, ranging from −0.0043 when the last year is
1995 to −0.0034 when it is 2010. All of the coefficients are sig-
nificantly different from zero and these results, which are largely
15
For instance, the ICD-10 to ICD-9 comparability ratios are 0.698, 1.232 and 1.554
for influenza/pneumonia, kidney disease (nephritis, nephrotic syndrome, nephro-
sis) and Alzheimer’s disease.
16
Ruhm (2013) provides a preliminary analysis examining three sub-components
of CVD, five categories of malignant neoplasms and five types of other diseases.
17
These account for 65% of deaths due to non-transport accidents. The most
important remaining category, “other and unspecified transport accidents and their
sequelae”, is not comparable over time.
20 C.J. Ruhm / Journal of Health Economics 42 (2015) 17–28
Fig. 1. Unemployment coefficients for total mortality using different analysis samples. (A) Sample begins in 1976 and continues through specified year. (B) Sample begins
in specified Year and Continues through 2010. (C) Analysis sample covers 20-year period.
similar to those of previous research, do not alter the conclusion
that mortality is procyclical.
The sensitivity of findings to the choice of sample periods can
be seen more explicitly in Fig. 1B, where the sample always ends
in 2010 but the starting year varies between 1976 and 1991. The
unemployment coefficient attenuates from −0.0034 for the entire
sample period to between −0.0029 and −0.0009 when the starting
year is 1978 or later18 . Perhaps more importantly, the data fail to
reject the null hypothesis of no macroeconomic effect for periods
beginning after 1988.
Fig. 1C displays results using 20-year sample windows begin-
ning in the specified year. The left-most entry shows that the
unemployment coefficient for 1976–1995 is −0.0043, while the
farthest right result shows that it is −0.0010 for 1991–2010. Total
mortality is significantly procyclical (negative unemployment rate
coefficients) for all 20-year windows starting between 1976 and
1987, but the predicted effect diminishes steadily for windows
beginning after 1982 and is small and insignificant for those
starting in 1988 through 1991. This pattern is not caused by
18
The full sample estimate is in line with previous results. Ruhm (2000) obtains a
slightly larger 0.5% reduction in total mortality but the current estimate is close to
the 0.3% decrease obtained in Stevens et al.’s (2011) preferred specification.
especially negative health consequences of the great recession of
2007–2009.
The choice of 20-year sample windows is arbitrary and may
conceal an increased procyclical variation of mortality toward the
end of the data period. This possibility is investigated in Fig. 2,
which replicates Fig. 1C, but for periods of between 5-years and
20-years. Two findings deserve mention. First, at shorter durations,
the estimates become more volatile and less precise. For instance,
when using 5-year windows, the unemployment coefficients fluc-
tuate wildly for even small changes in timing (e.g. from 0.0120
for 1996–2000 to −0.0077 for 1999–2003) but almost always fail
to reject the null hypothesis of no macroeconomic effect. Sec-
ond, the standard errors have typically increased for more recent
samples. As a result, the estimates obtained using 10-year or 15-
year analysis windows, while less volatile than those using 5-year
periods, still lack sufficient precision to determine whether the
possible partial reversion of the macroeconomic effects in recent
years (toward more procylical mortality) is real or reflects sta-
tistical noise19 . An important implication is that the findings of
19
When using 10-year periods, the average standard error is 46% larger for analysis
windows beginning between 1989 and 2001 than for those starting between 1976
and 1988 (0.0018 versus 0.0013).
 C.J. Ruhm / Journal of Health Economics 42 (2015) 17–28 21

Fig. 2. Unemployment coefficients for total mortality using different sample windows.

many recent investigations of macroeconomic variations in health

outcomes and behaviors should be viewed with extreme cau-
tion because the analysis periods are too short to provide reliable
estimates20 .
Fig. 3 shows that the estimated change in the relationship
between macroeconomic conditions and total mortality is fairly
insensitive to controlling for state-specific linear time trends or
weighting the data. (This figure, and those that follow, show point
estimates for 20-year analysis windows that begin in the year spec-
ified on the X-axis.) In all cases, the estimated procyclicality of
mortality declined over time, with most of the change dating from
the early to mid-1980s and with statistically insignificant unem-
ployment rate coefficients for some or all recent analysis windows.
Removing state-specific trends makes this effect more pronounced
than in the “preferred” specifications, which use unweighted data
and control for state trends; weighting the data and including
trends makes it somewhat less so21 . In all cases, the change in the
Fig. 3. Unemployment coefficients for total mortality using alternative estimation
methods.

20
unemployment coefficient between 1976–1995 and 1991–2010 is
Charles and DeCicca’s (2008) analysis of male obesity used data from 1997 to
2001; Arkes’ (2007, 2009) investigation of teenage body weight utilized informa-
positive and statistically significant.
tion from 1997 to 2004, Dávlos et al.’s (2012) study of alcohol abuse and dependence Table 1 shows the unemployment coefficients, from estimat-
compared 2001–2002 and 2004–2005, Colman and Dave’s (2013) research on work ing Eq. (1), for total, sex-specific and age-specific mortality. Results
and leisure-time physical activity covered 2003–2010, Cotti and Tefft’s (2011) anal-
ysis of alcohol-related vehicle fatalities used data from 2003 to 2009, and Tekin et al.
(2013) investigated a variety of health outcomes and behaviors from 2005 to 2011.
21
For example, the unemployment rate coefficient (standard error) for the and −0.0019 (0.0009) for unweighted data without trends, weighted data without
1991–2010 analysis window is 0.0022 (0.0017), 0.0039 (0.0020), −0.0010 (0.0010) trends, unweighted data with trends and weighted data with trends.
22 C.J. Ruhm / Journal of Health Economics 42 (2015) 17–28
Table 1
Estimated macroeconomic effects on specific sources of mortality.
Type of mortality 1976–1995
All −0.0043 (0.0009) ***
Sex-specific
Males −0.0044 (0.0010)***
Females −0.0041 (0.0010)***
Age-specific (Years)
<25 −0.0165 (0.0025)***
25–44 −0.0078 (0.0028)***
45–64 −0.0020 (0.0011)*
65–74 −0.0037 (0.0009)***
≥75 −0.0041 (0.0010)***
Note: Dependent variable is the natural log of the specified state mortality rate, obtained from the Compressed Mortality Files, for 1976 to 2010 (n = 1785). The first two
columns show the coefficient on the state unemployment rate for 20-year subsamples (n = 1020) covering 1976–1995 and 1991–2010. The regressions also include vectors
of state and year dummy variables, state-specific linear time trends, and controls for the share of the state population who are: female, nonwhite, Hispanic, and aged <1,
1–19, 45–54, 55–64, 65–74, 75–84 and ≥85 years old. The third column shows the difference between the unemployment coefficients for the 1991–2010 and 1976–1995
subsamples. Robust standard errors, clustered at the state level, are shown in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1.
in the first column refer to 1976–1995, those in the second to
1991–2010, with the final column showing the difference between
the two. A one point rise in unemployment predicts a statistically
significant 0.43% reduction in total mortality during 1976–1995
compared to a small and insignificant 0.10% decrease in 1991–2010
(see the first row). The 0.33% difference between these two periods
is statistically significant and indicates that the procyclicality of
mortality has largely disappeared in recent years22 .
To address the possibility that the observed secular trends
reflect a change in the relationship between unemployment rates
and macroeconomic conditions, rather than in the health effects of
economic conditions, I estimated specifications that controlled for
nonemployment (the percentage of the 16 and over civilians unem-
ployed or out of the labor force) rather than unemployment rates23 .
In these models a one percentage point increase in the nonem-
ployment rate predicted a statistically significant 0.39% reduction
in total mortality during 1976–1995 and an insignificant 0.02%
decrease in 1991–201024 . The highly significant 0.37% difference
is slightly larger than that obtained using unemployment rates.
The remainder of Table 1 and Fig. 4 summarize subgroup analy-
ses, stratified by gender and age. In Fig. 4, and subsequently, thicker
lines indicate sources with relatively high mortality shares25 . The
evidence suggests larger secular changes in macroeconomic effects
for men than women. In 1976–1995, a one point rise in unem-
ployment predicted a 0.44% reduction in male mortality and a
0.41% decrease for females. This effect completely disappeared
by 1991–2010 for men but fell only half as much for women.
The declining procyclicality of mortality has been particularly pro-
nounced for males since 1982, while showing a steadier reduction
for females (see Fig. 4A).
22
Coefficients on the other time-varying state level covariates are provided in the
online appendix.
23
For instance, declines in labor force participation rates were particularly pro-
nounced during the “great recession” that began in 2007, when compared to other
economic downturns (Shierholz, 2012).
24
A one percentage point increase in the unemployment rate predicts a 0.73
percentage point rise in nonemployment over the full period, in models control-
ling for state demographic characteristics, time trends, and state and year dummy
variables. Changes in interstate migration are also unlikely to explain the results.
Migrants tend to be healthy and to move from areas of higher to lower unemploy-
ment rates (Halliday, 2007), introducing a countercyclical mortality effect. Mortality
might therefore have become less procyclical if migration rates were increasing over
time. However, migration rates instead peaked around 1980 and have fallen sharply
since then (Malloy et al., 2011).
25
For example, in Fig. 4B, the line for ≥75 year olds is thick because they account for
51% of mortality, from 1976 to 2010, whereas that for <25 year olds is thin because
they are responsible for less than 4% of deaths.
1991–2010 Difference
−0.0010 (0.0010) 0.0033 (0.0012)***
−0.0001 (0.0012) 0.0043 (0.0015)***
−0.0018 (0.0011) 0.0022 (0.0014)
0.0024 (0.0035) 0.0189 (0.0037)***
0.0016 (0.0035) 0.0094 (0.0042)**
0.0018 (0.0013) 0.0038 (0.0017)**
−0.0019 (0.0012) 0.0018 (0.0014)
−0.0019 (0.0013) 0.0022 (0.0012)*
The trends also appear to be relatively pronounced for the young
and middle-aged. We are unable to reject the null hypothesis of a
zero unemployment rate effect in 2010 for all age groups. Specif-
ically, a one percentage point increase in joblessness reduced the
predicted death rates of <25, 25–45, 45–64, 65–74, and ≥75 years
old by 1.6%, 0.8%, 0.2%, 0.4%, and 0.4% in 1976–1995 but increased
them by 0.2%, 0.2%, 0.2%, −0.2%, and −0.2% in 1991–2010. We are
unable to reject the null hypothesis of a zero unemployment rate
effect in 1991–2010 for all age groups. The patterns for all possi-
ble 20-year windows are qualitatively similar (see Fig. 4B). The 95%
confidence intervals (not shown) exclude positive unemployment
rate coefficients for all two-decade periods beginning prior to 1986
for ≥75 year olds, before 1989 for <25 year olds, and earlier than
1983 for those aged 65–74. Conversely, a zero coefficient is rarely
rejected for 45–64 year olds.
5. Heterogenous effects across sources of death
Table 2 and Figs. 5 and 6 stratify disease versus external sources
of death, and then separately examine three disease and four exter-
nal causes. The three disease categories: cardiovascular, cancer
and other diseases, accounted for 42%, 23% and 29% of deaths
over the 1976–2010 period. The four external sources: transport
accidents, other (non-transport) accidents, suicides and homicides
were responsible for 2.2%, 2.4%, 1.4% and 0.9%. Finally, four specific
types of non-transport accidents are considered – falls, drown-
ing/submersion, smoke/fires/flames and poisoning/exposure to
noxious substances – which constituted 0.7%, 0.2%, 0.2% and 0.5%
of fatalities.
Levels and trends of the macroeconomic effects differ markedly
for mortality from disease versus external causes. A one point
rise in joblessness lowered predicted disease mortality by 0.33%
in 1976–1995 versus 0.14% in 1991–2010, a stastically insignifi-
cant change. By contrast, a much larger 1.5% reduction in external
deaths was estimated for the earlier years versus a statistically sig-
nificant 0.8% increase in the later ones, and the difference is highly
significant. Fig. 5 suggests an almost monotonic but modest atten-
uation over time in the unemployment coefficient for deaths from
disease, with statistically significant negative estimates obtained
for all two decade periods starting before 1986. Conversely, the
predicted effect for external causes was negative and fairly sta-
ble for 20-year periods beginning between 1976 and 1982, but
attenuated steadily for analysis windows starting from 1982 to
1990, with a statistically insignificant effect for those with first
years between 1984 and 1989, and significantly positive unem-
ployment rate coefficients obtained for those initiating in 1990 or
1991. These results help to explain the sharp reversal in the effects
 C.J. Ruhm / Journal of Health Economics 42 (2015) 17–28 23

Fig. 4. Unemployment coefficients for sex-specific and age-specific mortality. (A) Sex-specific mortality. (B) Age-specific mortality.

Table 2
Estimated macroeconomic effects on cause-specific mortality.

Cause of death 1976–1995 1991–2010 Difference

Diseases −0.0033 (0.0010)*** −0.0014 (0.0010) 0.0019 (0.0013)

Cardiovascular disease −0.0036 (0.0013)*** −0.0041 (0.0017)** −0.0005 (0.0019)
Cancer 0.0002 (0.0013) 0.0027 (0.0011)** 0.0024 (0.0011)**
Other diseases −0.0060 (0.0020)*** −0.0011 (0.0036) 0.0048 (0.0027)*

External causes −0.0148 (0.0020)*** 0.0078 (0.0028)*** 0.0226 (0.0023)***

Transport accidents −0.0265 (0.0035)*** −0.0086 (0.0048)* 0.0180 (0.0030)***
Other accidents −0.0173 (0.0033)*** 0.0086 (0.0049)* 0.0259 (0.0048)***
Suicides 0.0041 (0.0031) 0.0171 (0.0046)*** 0.0130 (0.0054)**
Homicides −0.0063 (0.0065) 0.0160 (0.0115) 0.0223 (0.0100)**

Other accidents −0.0173 (0.0033)*** 0.0086 (0.0049)* 0.0259 (0.0048)***

Falls −0.0119 (0.0053)** 0.0016 (0.0084) 0.0135 (0.0095)
Drowning/submersion −0.0015 (0.0065) 0.0143 (0.0155) 0.0159 (0.0147)
Smoke/fire/flames −0.0308 (0.0099)*** −0.0019 (0.0179) 0.0289 (0.0169)*
Poisoning/noxious −0.0148 (0.0161) 0.0422 (0.0251)* 0.0570 (0.0290)**

Note: See note on Table 1. *** p < 0.01, ** p < 0.05, * p < 0.1.

Fig. 5. Unemployment coefficients for disease versus external of deaths.

24 C.J. Ruhm / Journal of Health Economics 42 (2015) 17–28
Fig. 6. Unemployment coefficients for deaths from specific diseases and external causes. (A) Specific diseases. (B) External causes. (C) Other accidents.
of macroeconomic conditions on deaths of younger persons, who
disproportionately die from external causes26 .
5.1. Diseases
There are striking disparities across types of diseases. Cancer
mortality was unrelated to the economy in 1976–1995 but strongly
countercyclical by 1991–2010, whereas CVD mortality remained
strongly procyclical throughout (see the top panel of Table 2 and
Fig. 6A). The procyclicality of other disease mortality declined over
time but the change between 1976–1995 and 1991–2010 was not
quite significant at the 0.05 level27 .
26
For example, <45 year olds accounted for 56% of external deaths (from 1976 to
2010) but less than 10% of total mortality.
27
This result is sensitive to weighting. Using weighted data, the unemployment
rate coefficient was −0.0029 in 1976–1995 and −0.0020 in 1991–2009. The differ-
ence was 0.0009 with a standard error of 0.0031.
Research for earlier time periods (e.g. Ruhm, 2000; Neumayer,
2004; Miller et al., 2009) documents a strong procyclicality of
cardiovascular deaths but with little macroeconomic variation in
cancer fatalities, and attributes this to the likelihood that short-
term behavior changes (e.g. smoking, diet and exercise) more
strongly influence the risk of CVD than cancer deaths. A conceivable
explanation for the findings just described is that the relation-
ship between macroeconomic conditions and health behaviors has
remained relatively stable, while cancer mortality has become
more sensitive to the availability of financial resources and access
to (procyclical) health care due to improvements in expensive med-
ical treatments and technologies28 .
28
The cost per cancer case rose from $47,000 in 1983 to $70,000 in 1999 (Philipson
et al., 2012), with many expensive new medical treatments and chemotherapy
agents coming into use in the 1990s and early 2000s (Cutler, 2008). The contin-
ued procyclicality of CVD mortality could also occur for other reasons, such as a
(stable) deleterious health effect of air pollution (Heutel and Ruhm, 2013).
 C.J. Ruhm / Journal of Health Economics 42 (2015) 17–28 25

Table 3
Change in effect of macroeconomic conditions on predicted number of deaths.

Type of mortality Share of deaths Predicted in no. of deaths

Point estimate 95% Conf. Interval

All deaths 1.00000 7253*** 1882–12,637

Sex-specific
Males 0.5121 4900*** 1625–8185
Females 0.4879 2402 −487–5298

Age-specific (years)
<25 0.0389 1655*** 1009–2305
25–44 0.0572 1202** 138–2276
45–64 0.1865 1586** 177–3000
65–74 0.2031 799 −454–2054
≥75 0.5140 2565* −206–5343

Diseases 0.9293 3988 −1221–9210

CVD 0.4155 −466 −3846–2927
Cancer 0.2263 1277** 128–2329
Other diseases 0.2875 3104* −326–6552

External causes 0.0707 3582*** 2845–4322

Transport accidents 0.0218 880*** 593–1168
Other accidents 0.0243 1418** 889–1942
Suicide 0.0138 402** 75–733
Homicide 0.0094 469** 56–891

Other accidents 0.0243 1418*** 899–1942

Falls 0.0069 207 −78–497
Drowning 0.0019 67 −54–191
Fires 0.0019 124* −18–271
Poisoning 0.0054 709** 1–1459

Note: Predicted changes are for a one-percentage point increase in unemployment. Share of deaths is for 1976–2010. “Predicted  in # of deaths” is calculated as (ek ˆ
−
1) × k D, where  —  2 −  1 , for   the predicted unemployment coefficient in period , with  = 1 in 1976–1995 and  = 2 in 1991–2010; D is the average annual number
of deaths during 1976–2010 (2222,313); k is the share of deaths from source k. 95% confidence intervals are estimated as (e(k±1.96×s
ˆ k ) − 1)k D, for sk the standard error

on ˆ k .
***
p < 0.01,
**
p < 0.05,
*
p < 0.1.

5.2. External causes
There is considerable heterogeneity in the effects for specific
sources of external deaths (see the second panel of Table 2 and
Fig. 6B). One of the most consistent previous research findings is
that transport fatalities are procyclical29 . This effect persists but has
weakened recently, with a one percentage point rise in the unem-
ployment rate predicting a 2.6% decrease in 1976–1995 versus a
0.9% reduction in 1991–2010. Suicides increase with joblessness,
consistent with most prior studies, and this effect has strengthened
over time: a one point growth in unemployment was associated
with an insignificant 0.4% rise is suicides during 1976–1995 versus
a highly significant 1.7% increase in 1991–201030 .
The most noteworthy finding is that fatal non-transport acci-
dents have switched from being strongly procyclical to sharply
countercylical: a one point rise in unemployment reduced pre-
dicted mortality rates by 1.7% in 1976–1995 but increased them
0.9% in 1991–2010, with nearly monotonic growth over time. The
parameter estimates were negative and significant for all 20-year
periods starting prior to 1985 but insignificantly positive for those
beginning after 1987.
The bottom panel of Table 2 and Fig. 5C provide additional speci-
ficity on non-transport accident deaths, showing that the secular
trends are dominated by changes for accidental poisonings, where
the unemployment coefficient rose from −0.0148 in 1976–1995
to 0.0422 in 1991–2010, with a strong countercyclical pattern
emerging for 20-year windows beginning after the early 1980s.
There are more modest changes for deaths from falls or drownings
and the procyclicality of fatalities from fires largely disappears in
recent years but, as shown later, this is always a relatively minor
source of mortality. Given these results, accidental poisonings
receive special attention below.
6. Predicted changes in number of deaths
I next demonstrate that external sources of deaths, especially
those for non-transport accidents and among these accidental poi-
sonings, play a key role in explaining the declining macroeconomic
responsiveness of total mortality. All numerical calculations are
based on Eq. (2) and refer to a one percentage point increase
in unemployment. The discussion focuses on predicted secular
changes in macroeconomic effects, rather than levels for a single
analysis period.
The first row of Table 3 shows that a one point increase in
unemployment predicts 7253 more fatalities in 1991–2010 than in
1976–199531 . As noted, the procyclicality of mortality weakened
over time more for men than women, so that males account for
two-thirds of the overall change in the macroeconomic effect (see

29
Previous analyses have often examined motor vehicle deaths, which consti-
tuted 94% of transport accident fatalities from 1976 to 2010. Transport deaths are
considered here because they are coded more consistently across time.
30
The unemployment coefficient was positive in all 20-year windows and statis-
tically significant for those beginning after 1987.
31
The unemployment coefficient was −0.00428101 for 1976–1995 and
−0.00102266 in 1991–2010. The resulting difference of .00325834 implies around
0.33% more fatalities, or 7253 additional deaths per year, based on 2222,313 fatali-
ties annually: (exp[−0.00102266–−0.00428101] − 1) × 2222,313 = 7252.86.
26 C.J. Ruhm / Journal of Health Economics 42 (2015) 17–28
Fig. 7. Trends in accidental poisoning mortality, by age.
the second and third rows of Table 3)32 . When decomposing by
age, the most striking finding is the extent to which the declining
procyclicality is concentrated among the relatively young: persons
under <25 are responsible for less than 4% of deaths but 23% of
the secular trend in the macroeconomic impact; those under 45
(65) comprise less than 10% (30%) of fatalities but almost two-fifths
(over three-fifths) of the predicted change over time.
The remainder of Table 3 examines specific causes of mortal-
ity. A one percentage point rise in unemployment predicts almost
3600 more annual deaths from external sources in 1991–2010 than
1976–1995, or 49% of the change in total mortality. This occurs even
through only 7% of all deaths are due to such causes, and helps to
explain the large changes for males and <45 year olds (for whom
external deaths account for over 40% of mortality). Cancer and other
diseases also play a role, although the estimates are imprecise for
the former and sensitive to the use of sampling weights for the
latter33 . By contrast, cardiovascular disease – the number one killer
– explains none of the secular change, as the unemployment rate
coefficient actually becomes more negative in later years.
The third panel of Table 3 provide a more detailed decompo-
sition of external deaths. Non-transport accidents are of special
interest because the unemployment coefficient switches from large
and negative to positive, accounting for over 1400 additional deaths
annually, or 40% of the predicted rise in external cause mortality.
This increase is 11% larger than for all cancers, even though non-
transport accidents constitute only 2% of all fatalities, versus 22%
for malignant neoplasms. Transport accidents are also important,
explaining 880 extra deaths per year. The effects on suicides and
homicides are in the same direction but of considerably smaller
magnitude.
The bottom panel of Table 3 presents separate results for four
sources of non-transport accidents, which together explain around
two-thirds of deaths from this cause. The role of accidental poison-
ings is remarkable. Although accounting for just 0.5% of deaths, a
one point rise in unemployment is predicted to result in 709 more
annual poisoning fatalities in 1991–2009 than in 1976–1995, or
half the change in non-transport accident fatalities, 20% of that for
external deaths and almost 10% of the overall mortality effect.
32
Since separate (unconstrained) models are estimated for different sources of
death, the total contribution of changes in predicted group-specific mortality can
sum to more or less than the effect predicted for total mortality.
33
With weighted data, a one point unemployment increase predicts 580 more
deaths from other diseases in 1991–2010 than in 1975–1995
7. Discussion
The strong procyclical pattern of mortality present in the 1970s
and 1980s has been largely eliminated in recent years. The pat-
tern varies across sources of deaths, with much larger secular
changes observed for external than disease causes. All types of
external deaths became less procyclical or more countercyclical
during the analysis period, with particularly large changes for
non-transport accidents, and within this category, for accidental
poisonings. Among diseases, cardiovascular mortality continues to
fall sharply when the economy deteriorates, whereas cancer deaths
have became substantially countercyclical. These findings are rel-
evant not only for understanding of the production of health but
also for measuring the size and effects of business cycle fluctua-
tions. Egan et al. (2013) argue that procyclical mortality implies
that business cycle fluctuations are milder than when calculated
using standard GDP measures, but this may have become less true
in recent years.
Some estimates are sensitive to changing the starting and end-
ing dates of analysis. Such parameter instability is particularly
problematic when the sample window is short – probably anything
less than 15 years – raising concerns about the findings of many
recent related investigations that have used brief (often less than
10 year) timespans. One contribution of this study is to provide par-
simonious methods of illustrating the sensitivity of the results the
length of the analysis window, and to the first and last years exam-
ined. Another caveat is that specific sources of death are implicitly
treated here as being independent of each other, although some
prior research (Yeung et al., 2014) identifies potential correlations
between them. This is not an issue for the analysis of total mortality
but may be important when considering specific causes of death as
competing risks.
Mechanisms for the previously observed procyclical variation
in mortality remain poorly understood, so it is speculative as to
why the relationship has changed in recent years. Two possibilities
are intriguing. First, the change dates to (20-year) analysis periods
beginning in the early 1980s, which precisely coincides with the
reduction in macroeconomic volatility that has been referred to
as the “Great Moderation” (Stock and Watson, 2003; Bernanke,
2004)34 . This raises the possibility that the mortality patterns here
are part of a broader change in the effects of short-term changes
34
Productivity also shifted from procyclical to acyclical or slightly countercyclical
at about the same time (Galí and van Rens, 2010).
 C.J. Ruhm / Journal of Health Economics 42 (2015) 17–28 27

in macroeconomic performance, or in the role of unemployment economic weakness has long been associated with diminished
rates as a proxy for macroeconomic conditions. With regards to the mental health (Ruhm, 2000, 2003; Charles and DeCicca, 2008;
latter, it is noteable that the residual variation in state-year unem- Bradford and Lastrapes (2014) and, to the extent these drugs are
ployment rates after including controls (one minus the R-squared now being taken to address this, the increased procyclicality of
from regressing unemployment rates on state and year dummy poisoning deaths may be a physical manifestation of what was
variables, state-specific time trends, and the time-varying state previously a mental health problem36 .
demographic characteristics) fell from 0.177 in 1976–1995 to 0.094
in 1999–2010. There is also suggestive evidence that the procycli-
cality of mortality might have increased slightly in the most recent
Appendix A. Appendix
analysis periods, that include the severe 2007–2009 recession.
Second, the emerging importance of accidental poisoning fatal-
Tables A1 and A2.
ities occurred at the same time that deaths from this source
increased dramatically for young and middle-aged adults (see
Fig. 7). Over 90% of poisoning fatalities are now due to drug
overdoses, with particularly important roles for prescription opi- Appendix B. Supplementary data
oids (such as hydrocodone and oxycodone) and benzodiazepines
(Warner et al., 2011; Ruhm, 2015). The higher death rates reflect Supplementary data associated with this article can be found,
greater availability of these drugs raising the ease of self-injury in the online version, at http://dx.doi.org/10.1016/j.jhealeco.
and accidental death during bad economic times35 . Moreover, 2015.03.004.

Table A1
Definitions of specific causes of mortality.

Variable Description ICD-8 (1976–1978) ICD-9 (1979–1998) ICD-10 (1999–2010)

Cancer Malignant neoplasms 140–209 140–208 C00–C97

CVD Major cardiovascular diseases 390–448 390–448 I00–I78
Heart Diseases of the heart 390–398, 402, 404, 410–429 390–398, 402, 404–429 I00–I09, I11, I13, I20–I51
Transport Transport accidents 800–848, 940–941 800–848, 929.0, 929.1 V02–V99, Y85
Other Ac Other (non-transport) accidents 850–939, 942–949 850–928, 929.2–949 W00–X59, Y86
Falls Accidents: falls 880–887 880–888 W00–W19
Drowning Accidents: drowning/submersion 910 910 W65–W74
Fires Accidents: smoke/fire/flames 890–899 890–899 X00–X09
Poison Accidents: poisoning/noxious substances 850–879, 924 850–869, 924.1 X40–X49
Suicide Suicide (intentional self-harm) 950–959 950–959 X60–X84, Y87.0
Homicide Homicide and legal intervention 960–978 960–978 X85–Y09, Y87.1, Y35, Y89.0

Table A2
Sources of death by time period.

Source of death All Years 1976–1995 1991–2010

# % # % # %

All deaths 2222,313 100.0 2081,936 100.0 2367,352 100.0

Males 1138,011 51.2 1097,329 52.7 1181,266 49.9
Females 1084,302 48.8 984,607 47.3 1186,086 50.1

Age of death (years)

<25 86,549 3.9 97,528 4.7 74,871 3.2
25–44 127,161 5.7 126,536 6.1 134,595 5.7
45–64 414,460 18.6 399,677 19.2 418,713 17.7
65–74 451,301 20.3 472,501 22.7 437,723 18.5
≥75 1142,362 51.4 985,091 47.3 1301,093 55.0

Cause of death
Cardiovascular 923,419 41.6 957,085 46.0 892,420 37.7
Cancer 502,882 22.6 463,278 22.3 548,661 23.2
Other diseases 638,973 28.8 510,290 24.5 765,354 32.3
External causes 157,039 7.1 151,283 7.3 160,917 6.8
Transport accidents 48,545 2.2 50,581 2.4 45,750 1.9
Other accidents 54,044 2.4 45,419 2.2 60,350 2.5
Falls 15,225 0.7 12,721 0.6 17,217 0.7
Drowning/submersion 4181 0.2 4676 0.2 3565 0.2
Smoke/fires/flame 4229 0.2 4984 0.2 3402 0.1
Poison/noxious substance 12,090 0.5 5938 0.3 17,225 0.7
Suicide 30,755 1.4 29,391 1.4 32,171 1.4
Homicide 22,560 1.1 22,560 1.1 20,117 0.8

Note: Table shows average deaths per year for the specified age group or cause.

35
For example, per capita opioid sales more than tripled between 1999 and 2010
36
(Paulozzi, 2012). See Ruhm (2013) for a more extensive discussion of these issues.
28 C.J. Ruhm / Journal of Health Economics 42 (2015) 17–28
References
An, Ruopeng, Liu, Junyi, 2012. Local labor market fluctuations and physical activity
among adults in the United States 1990–2009. In: ISRN Public Health, 2012,
Article 318610., pp. 1–7.
Anderson, Robert, N., Arialdi, M., Miniño, Donna, L., Hoyert, Harry, M., 2001. Rosen-
berg comparability of cause of death between ICD-9 and ICD-10: preliminary
estimates. National Vital Statistics Reports 49 (2), 1–32.
Ariizumi, Hideki, Schirle, Tammy, 2012. Are recessions really good for your health?
Evidence from Canada. Social Science and Medicine 74, 1224–1231.
Arkes, Jeremy, 2007. Does the economy affect teenage substance use? Health Eco-
nomics 16 (1), 19–36.
Arkes, Jeremy, 2009. How the economy affects teenage weight. Social Science and
Medicine 68 (11), 1943–1947.
Bernanke, Ben, 2004. The Great Moderation, www.federalreserve.gov/boarddocs/
speeches/2004/20040220/ (retrieved 6 March 2015).
Blinder, Alan S., 1973. Wage discrimination: reduced form and structural estimates.
Journal of Human Resources 8 (4), 436–455.
Bradford, W. David, Lastrapes, William D., 2014. A prescription for unemployment?
Recessions and the demand for mental health drugs. Health Economics 23 (11),
1301–1325.
Brenner, M. Harvey, 1979. Mortality and the national economy. The Lancet 314
(8142), 568–573.
Brenner, M. Harvey, 1971. Economic changes and heart disease mortality. American
Journal of Public Health 61 (3), 606–611.
Buchmueller, Tom, Grignon, Michel, Jusot, Florence, 2007. Unemployment and mor-
tality in France, 1982-2002. In: Center for Health Economics and Policy Analysis
Working Paper 07-04. McMaster University.
Butler, J.S., 2000. Efficiency results of MLE and GMM estimation with sampling
weights. Journal of Econometrics 96 (1), 25–37.
Charles, Kerwin Kofi, DeCicca, Philip, 2008. Local labor market fluctuations and
health: is there a connection and for whom? Journal of Health Economics 27
(6), 1532–1550.
Colman, Gregory J., Dave, Dhaval M., 2013. Exercise physical activity, and exertion
over the business cycle. Social Science and Medicine 93 (September), 11–20.
Cotti, Chad, Tefft, Nathan, 2011. Decomposing the relationship between macroeco-
nomic conditions and fatal car crashes during the great recession: alcohol- and
non-alcohol-related accidents. B. E. Journal of Economic Analysis and Policy 11
(1), 1–48.
Cutler, David M., 2008. Are we finally winning the war on cancer? Journal of Eco-
nomic Perspectives 22 (4), 3–26.
Dávlos, María E., Fang, Hai, French, Michael T., 2012. Easing the pain of an economic
downturn: macroeconomic conditions and excessive alcohol consumption.
Health Economics 21 (11), 1318–1335.
Economou, Athina, Nikolau, Agelike, Theodossiou, Ioannis, 2008. Are recessions
harmful to health after all? Evidence from the European Union. Journal of Eco-
nomic Studies 35 (5), 368–384.
Edwards, Ryan, 2011. American Time Use Over the Business Cycle. University of New
York, Mimeo City.
Egan, Mark L., Mulligan, Casey B., Philipson, Tomas J., 2013. Adjusting measures of
economic output for health: is the business cycle countercyclical? In: National
Bureau of Economic Research Working Paper No. 19058.
Eyer, Joseph, 1977. Prosperity as a cause of death. International Journal of Health
Services 7 (1), 125–150.
Fishback, Price V., Haines, Michael R., Kantor, Shawn, 2007. Births deaths and new
deal relief during the great depression. Review of Economics and Statistics 89
(1), 1–14.
Freeman, Donald G., 1999. A note on economic conditions and alcohol problems.
Journal of Health Economics 18 (5), 661–670.
Galí, Jordi, van Rens, Thijs, 2010. The vanishing procyclicality of labor productivity,.
In: Kiel Institute Working Paper No. 1641 (August).
Gerdtham, Ulf-G, Ruhm, Christopher J., 2006. Deaths rise in good economic times:
evidence from the OECD. Economics and Human Biology 43 (3), 298–316.
Gonzalez, Fidel, Quast, Troy, 2011. Macroeconomic changes and mortality in Mexico.
Empirical Economics 40 (2), 305–319.
Gravelle, H.S.E., Hutchinson, G., Stern, J., 1981. Mortality and unemployment: a
critique of Brenner’s time-series analysis. The Lancet 318 (8248), 675–679.
Gruber, Jonathan, Frakes, Michael, 2006. Does falling smoking lead to rising obesity?
Journal of Health Economics 25 (2), 183–197.
Halliday, Timothy J., 2007. Business cycles migration and health. Social Science and
Medicine 64 (7), 1420–1424.
Heutel, Garth, Ruhm, Christopher J., 2013. Air pollution and procyclical mortality.
In: National Bureau of Economic Research Working Paper No. 18959.
Kasl, Stanislav V., 1979. Mortality and the business cycle: some questions about
research strategies when utilizing macro-social and ecological data. American
Journal of Public Health 69 (8), 784–788.
Klebba, A. Joan, Scott, Joyce H., 1980. Estimates of selected comparability ratios based
on dual coding of 1976 death certificates by the eighth and ninth revisions of
the international classifications of diseases. Monthly Vital Statistics Report 28
(11), 1–19.
Lam, Jean-Paul, Piérard, Emmanuelle, 2014. The Time-Varying Relationship Between
Mortality and Business Cycles in the U.S. University of Waterloo, Mimeo (June).
Lin, Shin-Jong, 2009. Economic fluctuations and health outcome: a panel analysis of
Asia-Pacific countries. Applied Economics 41 (4), 519–530.
Lindo, Jason M., 2015. Aggregation and the relationship between unemployment
and health. Journal of Health Economics 40 (2), 83–96.
Malloy, Raven, Smith, Christopher L., Wozniak, Abigail, 2011. Internal migration in
the United States. Journal of Economic Perspectives 25 (3), 173–196, Summer.
McInerney, Melissa, Mellor, Jennifer M., 2012. Recessions and seniors’ health, health
behaviors, and healthcare use: analysis of the medicare beneficiary survey. Jour-
nal of Health Economics 31 (5), 744–751.
Miller, Douglas L., Page, Marianne E., Huff Stevens, Ann, Filipski, Mateusz, 2009.
Why are recessions good for your health? American Economic Review 99 (2),
122–127.
Neumayer, Eric, 2004. Recessions lower (some) mortality rates. Social Science &
Medicine 58 (6), 1037–1047.
Ogburn, William F., Thomas, Dorothy S, 1922. The influence of the business cycle
on certain social conditions. Journal of the American Statistical Association 18
(139), 324–340.
Oaxaca, Ronald, 1973. Male–female wage differentials in urban labor markets. Inter-
national Economic Review 14 (3), 693–709.
Paulozzi, Leonard J., 2012. Prescription drug overdoses: a review. Journal of Safety
Research 43 (4), 283–289.
Philipson, Thomas, Ebner, Michael, Lakdawalla, Darius N., Corral, Mitra, Conti, Rena,
Goldman, Dana P., 2012. An analysis of whether higher health care spending in
the United States versus Europe is ‘worth it’ in the case of cancer. Health Affairs
31 (4), 667–675.
Ruhm, Christopher J., 2000. Are recessions good for your health? Quarterly Journal
of Economics 115 (2), 617–650.
Ruhm, Christopher J., 2003. Good times make uou sick. Journal of Health Economics
22 (4), 637–658.
Ruhm, Christopher J., 2005. Healthy living in hard times. Journal of Health Economics
24 (2), 341–363.
Ruhm, Christopher J., 2012. Understanding the relationship between macroecono-
mic conditions and health. In: Andrew, M. Jones (Ed.), Elgar Companion to Health
Economics. , second ed. Edward Elgar, Cheltenham, UK, pp. 5–14.
Ruhm, Christopher J., 2013. Recessions, healthy no more? In: NBER Working Paper
No. 19287 (August).
Ruhm, Christopher J., 2015. Drug Poisoning Deaths in the United States 1999–2012.
University of Virginia, Mimeo.
Ruhm, Christopher J., Black, William E, 2002. Does drinking really decrease in bad
times? Journal of Health Economics 21 (4), 659–678.
Shierholz, Heidi, 2012. Labor force participation: cyclical versus structural changes
since the start of the great recession. In: Economic Policy Institute Issue Brief
No. 333 (May 24).
Solon, Gary, Haider, Steven J., Wooldridge, Jeffrey, 2015. What are we weighting for?
Journal of Human Resources (forthcoming).
Stevens, Ann Huff, Miller, Douglas L., Page, Marianne, Filipski, Mateusz, 2011. The
best of times the worst of times: understanding procyclical mortality. In: NBER
Working Paper No. 17657.
Stock, James H., Watson, Mark W., 2003. Has the business cycle changed and why?
In: Gertler, M., Rogoff, K. (Eds.), NBER Macroeconomics Annual 2002. MIT Press,
Cambridge, pp. 159–218.
Sullivan, Daniel, von Wachter, Till, 2009. Job displacement and mortality: an analysis
using administrative data. Quarterly Journal of Economics 124 (3), 1265–1306.
Stuckler, David, Basu, Sanjay, Suhrcke, Marc, Coutts, Adam, McKee, Martin, 2009.
The public health effect of economic crisis and alternative policy responses in
Europe: an empirical analysis. The Lancet 374 (9686), 315–323.
Svensson, Mikael, 2007. Do not go breaking your heart: do economic upturns really
increase heart attack mortality? Social Science and Medicine 65 (4), 833–841.
Tapia Granados, José A., 2005. Recessions and mortality in Spain, 1980–1997. Euro-
pean Journal of Population 21 (4), 393–422.
Tapia Granados, José A., Diez Roux, Ana V, 2009. Life and death during the
great depression. Proceedings of the National Academy of Sciences 106 (41),
17290–17295.
Tekin, Erdal, McClellan, Chandler, Jean Minyard, Karen, 2013. Health and health
behaviors during the worst of times: evidence from the great recession. In: NBER
Working Paper No. 19234.
Thomas, Dorothy Swaine, 1927. Social Aspects of the Business Cycle. Alfred A. Knopf,
New York, NY.
Warner, Margaret, Chen, Hui Li, Makuc, Diane M., Anderson, Robert N., Miniño,
Arialdi M., 2011. Drug poisoning deaths in the United States, 1980–2008. In:
NCHS Data Brief No. 81. National Center for Health Statistics, Hyattsville, MD.
Wooldridge, Jeffrey, 1999. Asymptotic properties of weighted M-estimators for vari-
able probability samples. Econometrica 67 (6), 1385–1406.
Xu, Xin, 2013. The business cycle and health behaviors. Social Science and Medicine
77 (January), 126–136.
Yeung, Gary Y.C., van den Berg, Gerard J., Lindeboom, Marrten, Portrait., France R.M,
2014. The impact of early-life economic conditions on cause-specific mortality
during adulthood. Journal of Population Economics 27 (3), 895–919.
 Journal of Health Economics 42 (2015) 29–43

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Education and health: The role of cognitive ability夽

Govert E. Bijwaard a,b , Hans van Kippersluis c,d,e,∗ , Justus Veenman c,f
a
Netherlands Interdisciplinary Demographic Institute (NIDI-KNAW/University of Groningen), PO Box 11650, 2502 AR The Hague, The Netherlands
b
IZA, Bonn, Germany
c
Erasmus School of Economics, Erasmus University Rotterdam, PO Box 1738, 3000 DR Rotterdam, The Netherlands
d
Tinbergen Institute, Rotterdam, The Netherlands
e
Netspar, Tilburg, The Netherlands
f
ERCOMER, Utrecht/Rotterdam, The Netherlands

a r t i c l e i n f o a b s t r a c t

Article history: We aim to disentangle the relative impact of (i) cognitive ability and (ii) education on health and mortality
Received 17 September 2013 using a structural equation model suggested by Conti et al. (2010). We extend their model by allowing
Received in revised form 21 July 2014 for a duration dependent variable (mortality), and an ordinal educational variable. Data come from a
Accepted 7 March 2015
Dutch cohort born between 1937 and 1941, including detailed measures of cognitive ability and family
Available online 17 March 2015
background in the final grade of primary school. The data are linked to the mortality register 1995–2011,
such that we observe mortality between ages 55 and 75. The results suggest that at least half of the
JEL classification:
unconditional survival differences between educational groups are due to a ‘selection effect’, primarily
C41
I14
on the basis of cognitive ability. Conditional survival differences across those having finished just primary
I24 school and those entering secondary education are still substantial, and amount to a 4 years gain in life
expectancy, on average.
Keywords: © 2015 Elsevier B.V. All rights reserved.
Education
Cognitive ability
Mortality
Structural equation model
Duration model

1. Introduction
Disparities in health and life expectancy across educational
groups are striking and pervasive, and are considered one of
夽 Van Kippersluis gratefully acknowledges funding from the National Institute on
Aging (NIA) under grant R01AG037398, from NETSPAR under the project “Income
and health, work and care across the life cycle II”, and from the Netherlands Organi-
zation of Scientific Research (NWO Veni grant 016.145.082). The authors acknowl-
edge access to linked data resources (DO 1995–2011) by Statistics Netherlands (CBS).
We thank Mars Cramer and Mirjam van Praag for help in accessing the original data
(see https://easy.dans.knaw.nl/ui/datasets/id/easy-dataset:39042), and are grateful
to Mars Cramer, and attendants from the Empirical Health Economics conference in
Munich 2013, the International Health Economics Association conference in Sydney
2013, the Multistate Event History Analysis in Hangzhou 2012, and seminar partici-
pants at the Chinese University of Hong Kong, Cornell University, Erasmus University
Rotterdam, and the University of Southern California for helpful comments.
∗ Corresponding author at: Erasmus School of Economics, Erasmus University
Rotterdam, PO Box 1738, 3000 DR Rotterdam, The Netherlands. Tel.: +31 10 4088837.
E-mail addresses: bijwaard@nidi.nl (G.E. Bijwaard), hvankippersluis@ese.eur.nl
(H. van Kippersluis), veenman@ese.eur.nl (J. Veenman).
the most compelling and well established facts in social science
research (Mazumder, 2012). Even in an egalitarian country such
as the Netherlands, with a very accessible health care system, the
difference in life expectancy between the university educated and
those who finished only primary school is 6–7 years (CBS, 2008). It is
commonly assumed that a large part of this association derives from
the causal effect of education on health outcomes. An abundant list
of possible mechanisms was proposed, among which occupational
demands, health behavior, and the ability to process information
are the most commonly mentioned (Ross and Wu, 1995; Cutler
and Lleras-Muney, 2008).
Yet, the association between education and health could also
stem from (i) ‘reverse causality’, in which childhood ill-health con-
strains educational attainment (Behrman and Rosenzweig, 2004;
Case et al., 2005) and (ii) confounding ‘third factors’ such as ability,
parental background and time preference that influence both edu-
cation and health outcomes (Fuchs, 1982; Auld and Sidhu, 2005;
Deary, 2008).
Studies based on natural experiments in education, such as
changes in compulsory schooling laws, overcome the difficulty of

http://dx.doi.org/10.1016/j.jhealeco.2015.03.003
0167-6296/© 2015 Elsevier B.V. All rights reserved.
30 G.E. Bijwaard et al. / Journal of Health Economics 42 (2015) 29–43
separating the direct causal effect of education from third factor
effects. The estimates based on these studies point towards a small
effect (Lleras-Muney, 2005; Oreopoulos, 2006; Van Kippersluis
et al., 2011; Meghir et al., 2013), or even insignificant effect of edu-
cation on health and mortality (Arendt, 2005; Albouy and Lequien,
2008; Mazumder, 2008; Braakmann, 2011; Clark and Royer, 2013).
This suggests that confounding factors may well play an impor-
tant role in shaping the strong association between education and
health.
Surprisingly little research in economics has investigated the
contribution of early childhood abilities and childhood social back-
ground in shaping the association between education and health.1
Some recent economic studies report associations between child-
hood cognitive and non-cognitive abilities, and health outcomes
at ages 30–40 using the British Cohort Study (Murasko, 2007), the
U.K. National Child Development Study (Carneiro et al., 2007), the
U.S. National Longitudinal Study of Youth 1979 (Auld and Sidhu,
2005; Kaestner and Callison, 2011), or the Dutch ‘Brabant data’
(Cramer, 2012). It is established that cognitive ability and some
non-cognitive factors such as self-esteem and conscientiousness
are associated with health outcomes. Nonetheless, hardly anything
is known about (i) the relative impact of education and childhood
abilities on health outcomes, and in turn (ii) how much of the
association between education and health is explained by these
cognitive and non-cognitive abilities.
A notable contribution to the literature is a recent series of
papers by Conti and Heckman (2010), Conti et al. (2010, 2011),
and Heckman et al. (2014) who, using the British Cohort Study
and the National Longitudinal Study of Youth (NLSY79), estimate a
structural equation model in which the interdependence between
education, health, and two latent factors capturing cognitive and
non-cognitive abilities is explicitly modeled. The authors show that
for most health outcomes around half of the association between
education and health is driven by cognitive and non-cognitive
abilities and early childhood social background. The other half is
interpreted as the causal effect of education on health.
While the series of papers by Conti, Heckman and co-authors
provided a significant contribution to the literature, there are two
notable limitations. First, the health outcomes are measured at age
30, an age at which health differences by education may not have
fully materialized. In fact, disparities in health and mortality seem
to peak around middle-age (Cutler and Lleras-Muney, 2008). Sec-
ondly, the health measures are all self-reported, which may bias
the estimates since education is related to subjective health per-
ceptions (Bago d’Uva et al., 2008).
In this paper, we aim to disentangle the effects of education and
cognitive ability on health outcomes. We will use the so-called ‘Bra-
bant data’ – a representative cohort of primary school sixth graders
in the Dutch province of Noord-Brabant – that has detailed infor-
mation on cognitive ability and social background measured back
in 1952. Three follow-up surveys in 1957, 1983 and 1993 contain
information on education, employment, and self-reported health.
We have linked these data to the mortality register 1995–2011,
such that the impact on mortality can be analyzed.
The contribution of this paper is threefold. First, we study the
relative impact of cognitive ability and education on mortality, as
an objective health indicator. The second contribution is that, in
contrast to existing studies that measure health outcomes at ages
30–40, we observe mortality during ages 55–75. Finally, we extend
the structural equation model by Conti et al. (2010) by allowing for
a duration dependent variable (mortality).2
1
See Gottfredson (2004) for an overview of the epidemiological literature.
2
Savelyev (2012) developed a similar structural equation model for mortality as
ours, yet using a discrete-time hazard model and not taking into account dynamic
The results show that for most ages, cognitive ability and family
socioeconomic status explain around half of the raw differences in
mortality across educational groups. Stated otherwise, education
remains important in determining mortality even after controlling
for cognitive ability, family socioeconomic status, and a range of
other background variables. The conditional survival differences
across educational groups are even remarkable, and amount to a
4-year gain in life expectancy for those entering at least secondary
school compared to those that dropped out after primary school.
This paper is structured as follows. Section 2 presents the Bra-
bant data including the available register data from Statistics
Netherlands, Section 3 presents the structural equation model that
we will use to disentangle the relative contributions of cognitive
ability and education on health outcomes. Section 4 presents the
results and Section 5 discusses them.
2. Data and descriptive statistics
The data are from a Dutch cohort born between 1937 and 1941.
Very detailed information about individual intelligence, social
background and school achievement is available for 5823 individ-
uals. The survey was held in the spring and summer of 1952 among
pupils of the sixth (last) grade of primary schools in the Dutch
province of Noord-Brabant, and hence is referred to as the ‘Bra-
bant data’. One-fourth of the province population was sampled;
mainly by including every fourth child from the schools’ list of
pupils.3 Hartog (1989) investigated the data and found no reason to
doubt representativeness. A selective dropout of pupils before par-
ticipating in the data collection does not exist, as primary school
was compulsory and enforcement of school attendance was strict
(Dronkers, 2002).
Follow-up surveys took place in 1957, 1983 and 1993.4 In 1957
only a sub-sample – those who scored above-average on six tests
– of the original cohort was interviewed about the school careers
between 1952 and 1957 to particularly investigate school career
choices of the most intelligent half of the cohort. In 1983 and 1993
attempts were made to trace all initial respondents of the Brabant-
cohort to investigate labour market behavior, with overall response
rates of around 45 percent. The sample is reduced to 2998 individ-
uals who have measurements in 1952 and in either 1983 or 1993,
or both.5
The Brabant data are subsequently linked to administrative
records from Statistics Netherlands. The basis for this linkage is
identifying information on ZIP code, date of birth, and sex, provided
in 1993 by Dutch municipalities, which includes information on all
individuals living in the Netherlands. The administrative records
are available since 1995. Because of the two-year discrepancy only
86 percent of the 2998 individuals could be traced in the munici-
pality register in 1995, leaving us with a working sample of 2579
individuals. Administrative records include the mortality register
and the municipality register for the years 1995–2011 inclusive.
The mortality register is used to identify drop out due to death in
selection of ability due to differential survivorship. His study is based on the Terman
data, a cohort of individuals with IQ beyond 140. Hence, apart from differences in
the model specification, his focus is on an extraordinary sample corresponding to
the 99.6th percentile of the intelligence distribution, with very limited variation
in cognitive ability. Not surprisingly, he examines the effect of higher education
whereas we focus on secondary education.
3
Some schools had school years beginning in April rather than in September. For
these schools, half the pupils of half the schools were included in the sample, which
yielded 369 observations on a total of 5823 (Hartog, 1989).
4
Mathijssen and Sonnemans (1958), Hartog and Pfann (1985), Van Praag (1992),
and Hartog et al. (2002). The complete questionnaire is included in Van Praag (1992)
‘Brabantse zesdeklassers, 1952–2010’.
5
In Section 4.2 it is verified that selective attrition does not affect our results.
 G.E. Bijwaard et al. / Journal of Health Economics 42 (2015) 29–43
the follow-up period. Demographics are obtained from the munic-
ipality register.
2.1. Dependent variables
Our outcome variable is Mortality, which is identified from the
mortality register in the period 1995–2011. Given that most pupils
are born around 1940, this implies that we follow mortality from
age 55 until 75.6 In our sample, 409 individuals, or 16 percent, died
during the period 1995–2011. Close to 50 percent died from cancer,
25 percent from cardiovascular diseases, and 8 percent from respi-
ratory diseases such as COPD and pneumonia. External causes such
as accidents comprise only two percent, as do mental disorders (e.g.
dementia), diseases of the digestive system (e.g. liver cirrhosis) and
diseases of the nervous system (e.g. Parkinson).
2.2. Independent variables
Our main independent variable of interest is Education, here
defined as the highest level of education attended, in three cat-
egories: (1) Lower Education, including those who attended at
most (extended)7 primary school, (2) Lower Vocational Education,
including those who attended at most lower vocational educa-
tion such as the lower agricultural school or lower polytechnic
schools, and (3) At least General Secondary School, including those
who attended lower general secondary school, higher general
secondary school, and higher vocational education or university.
Education is retrieved mainly from the 1983 and 1993-survey
variables on the highest level of education attended. The maxi-
mum of the two defines Education, and where missing we update
our educational variable with information from the 1957 sur-
vey.
Table 1 presents descriptive statistics and shows that 14 per-
cent did not continue school after primary school forming the Lower
Education category, 35 percent only attended Lower Vocational Edu-
cation, and the other 51 percent attended At least General Secondary
School. Fig. 1 shows the Kaplan–Meier survival curves for a binary
indicator of education with threshold at Lower Education, and sep-
arately for the three education categories. It is clear that the largest
survival differences are between those with only primary school
and those above primary school, and that the difference grows with
age to around ten percentage points near age 75.
Our second independent variable is Cognitive Ability. In the Bra-
bant data there are two measurements for cognitive ability, both
measured in the final grade of primary school (i.e. around age 12):
(i) the Raven Progressive Matrices Test, and (ii) a Vocabulary test
(picking synonyms).8 The timing of the intelligence tests implies
that the plausible feedback effects from education to cognitive abil-
ity (Deary and Johnson, 2010; Brinch and Galloway, 2012; Meghir
et al., 2013) will be seen as an education effect, while the impact
6
Of the Dutch population 1940 cohort, only 6.8 percent died between the
ages of 12 and 55 – Human Mortality Database, University of California, Berkeley
(USA), and Max Planck Institute for Demographic Research (Germany). Available
at www.mortality.org or www.humanmortality.de (data downloaded on July 30,
2012).
7
At the time, pupils had to stay in school for at least 8 years, or until they reached
the age of 14. Since regular primary school only consisted of 6 grades, some schools
offered an additional 2-year extended primary school (“vglo”).
8
The data also contain the so-called LO-IV test, which consists of six sub-tests:
regularities in series of numbers, analogies in figures, analogies in words, and sim-
ilarities between concepts (equal, not-equal, cause). Since the quality of this test
has been questioned (Hartog et al., 2002, p. 5) we will not use it in our analyses.
There is also information on grades for specific courses (Dutch language, mathe-
matics (arithmetics), history, physics, geography, health sciences, and traffic), but
since these are not clean measures of cognitive ability and are relative to others in
one’s classroom, we choose not to use these grades.
31
of the cognitive ability endowment in the final grade of primary
school on educational choice and later-life mortality is seen as a
selection effect.9
The IQ p.m. (‘progressive matrices’) test focuses on mathemat-
ical ability and is a replication of the British Progressive Matrices
test, designed by Raven (1958). It is considered to be a ‘pure’ mea-
surement of problem solving abilities, as it does not require any
linguistic or general knowledge (Dronkers, 2002). Hence, the Raven
test is supposed to measure fluid or analytic intelligence (Carpenter
et al., 1990). In this sense, the test can be compared to Spearman’s
g test (1927). The term g refers to the determinants of the common
variance within intelligence tests, being the core issue of intelli-
gence measurement (Carpenter et al., 1990).
Table 1 shows that the ability test designed by Raven has an
average of 102, with standard deviation of 13 while the vocabulary
test is 101, on average, with standard deviation 13. The correlation
between the Raven test and the vocabulary test is 0.38. This sug-
gests that while there seems to be some overlap between the two
measurements, the tests additionally gauge some idiosyncratic part
of cognitive ability. Therefore, we will use both measurements to
build a comprehensive latent factor of cognitive ability. In a robust-
ness check we solely use the Raven test to see whether the results
differ.
2.3. Control variables
Apart from a fairly standard set of demographic control variables
such as Age, whether Male, and Birth Rank, we also have informa-
tion about the social and school environment of the individuals.
Most of these variables are reported by the School principal. Family
Socioeconomic Status is measured in three categories from lowest to
highest depending on father’s occupation.10 We additionally know
whether the child had to work in the parent’s farm or company,
defining the binary indicator Child Works, which potentially sig-
nals part of the childhood health status. In this (historical) case,
however, the variable is mainly dependent on the parents having a
firm.
Available information regarding the school includes School Type
and the Number of Teachers. Repeat defines the number of classes
that children had to repeat. Further, we know the Teacher’s Advice
regarding further education of the child, and the Preference of the
Parents concerning the education of the pupil, categories of which
are defined in Table 1, which also includes descriptive statistics.
We have no information about childhood health status, which
prevents us from investigating the possibility of reverse causality
from health to education in our sample. The sample is comprised
of pupils who made it to the final grade of primary school. Hence,
pupils with severe health problems impairing going to school in
the first place will not be present in our sample. Moreover, in the
1983 wave of the survey male respondents were asked whether
they served in the military. The main reason for disqualification of
9
It should be emphasized however that there could be unobserved factors cor-
related to both cognitive ability in the final grade of primary school and later-life
mortality, which our measure of cognitive ability would be picking up.
10
We classify lower administrative, agricultural, industrial, and other lower work-
ers, and the disabled into the Lowest Socioeconomic Status. If the School Principal
considered the family antisocial, the family is also classified into the Lowest Socioe-
conomic Status. Intermediary personnel, self-employed farmers, self-employed
craftsmen, and the retired are categorized into the Intermediate Socioeconomic Sta-
tus (following Cramer, 2012). Teachers, executives and academics are classified into
the Highest Socioeconomic Status. In case father’s occupation is missing, we use
father’s education for individuals in the 1957 survey. Father’s education is classified
into 3 levels, which we directly translate into the three socioeconomic statuses. We
use mother’s education in case the father died or was not present in the household.
32 G.E. Bijwaard et al. / Journal of Health Economics 42 (2015) 29–43

Table 1
Descriptive statistics of the Brabant data sample.

Variable Average Standard deviation Number of observations

Dependent variables
Mortality 0.16 0.35 2579
Independent variables
Education
Lower education 0.14 0.34 2537
Lower vocational education 0.34 0.48 2537
At least general secondary school 0.51 0.35 2537
Raven p.m. test 102.04 13.28 2579
Vocabulary test 101.42 12.87 2579
Control variables
Male 0.58 0.49 2579
Birth rank 2.50 2.55 2412
Family socioeconomic status
Lowest 0.53 0.50 2409
Middle 0.44 0.50 2409
Highest 0.03 0.16 2409
Child works 0.28 0.45 2256
School religion
Roman-Catholic 0.76 0.43 2518
Protestant 0.19 0.40 2518
Special 0.03 0.17 2518
Public 0.02 0.13 2518
Number of teachers 6.92 2.47 2452
Repeat
No repetition of grade 0.64 0.48 2462
Repeated once 0.27 0.45 2462
Repeated twice or more 0.09 0.28 2462
Teacher’s advice
Continue primary school 0.24 0.43 2429
Lower vocational education 0.38 0.48 2429
Lower secondary education 0.24 0.43 2429
Higher secondary education 0.14 0.20 2429
Preference of the parents
Work in family company 0.13 0.33 2200
Paid work without vocational education 0.20 0.28 2200
Paid work with vocational education 0.27 0.44 2200
General secondary education 0.41 0.49 2200

Notes: Author’s calculations on the basis of the Brabant data linked to the municipality register and the mortality register.

compulsory military duty is health problems.11 Since the fraction of
individuals having served in the military is almost identical across
educational levels, this provides some indirect evidence that health
differences across educational levels were minimal during teenage
years. We furthermore refer to Conti et al. (2010) who showed that
in their sample childhood health, as measured by childhood height,
was not an important determinant of educational choice. The lack
of information on childhood health should therefore not be a major
source of concern.
3. Methodology
Our empirical approach is an extension of the structural equa-
tion framework developed by Conti et al. (2010). We briefly
describe the Conti et al. model, after which we will present our two
extensions: allowing for an objective duration dependent variable
(mortality), and introducing an ordinal educational choice. Finally,
we explain how we disentangle the effects of cognitive ability and
education on the health outcomes.
3.1. Basic structural equation model
The basic Conti et al. model allows a way of modeling the inter-
relationships between abilities, education and health outcomes,
where individuals potentially make their educational decisions
11
Other reasons were exemption owing to one’s brother’s service, grounds of
conscience, or personal indispensability (e.g. Van Schellen and Nieuwbeerta, 2007).
depending on the perceived health gains. Hence, the educa-
tional choice is endogenous, and in practice it is assumed that
selection into schooling can be fully accounted for by using
observed characteristics and unobserved ability. The model con-
sists of three parts: (i) a binary educational choice depending
on latent abilities and other covariates, (ii) potential outcomes
depending on the choice of education, latent abilities, and other
covariates, and (iii) a measurement system for the latent abili-
ties.
The binary indicator for education Di is defined as 1 if individual
i took any education beyond the compulsory schooling age, and 0
if not:

1 if Di∗ ≥ 0
Di = (1)
0 otherwise
where we assume Di∗ is an underlying latent utility which is con-
tinuous and linear, and depends on latent abilities , and observed
characteristics XD :
Di∗ = XiD + ˛D i + iD (2)
with D being an error term independent of XD and . We assume
that D is normally distributed, which implies that we have a probit
model for the educational choice. We fix the variance at 1 since the
variance is not identified in a probit model.
The second part is the potential outcomes part, in which
there are two potential outcomes Yi1 and Yi0 , where the former
is the outcome in case the individual chose to pursue education
beyond what is compulsory, and the latter is the outcome in case
the individual dropped out of school right after the compulsory
 G.E. Bijwaard et al. / Journal of Health Economics 42 (2015) 29–43
Fig. 1. Kaplan–Meier survival function by education level in two categories (top) and three categories (bottom).
schooling age. Both Yi1 and Yi0 depend on latent ability , and on
observed characteristics XY :
Yi1 = ˇ1 XiY + ˛1 i + i1
Yi0 = ˇ0 XiY + ˛0 i + i0
with (0 , 1 ) independent of XY and , independent of iD con-
ditional on XY and , and both follow a normal distribution with
variance 12 and 02 , respectively.
The final part of the model is the measurement equation, where
one or two measurements, Mik (k = 1, 2), implicitly define the latent
ability :
Mik = ık XiM + ˛Mk i + iM k
with Mk independent of XM and . We assume that Mk is normally
2 .
distributed with variance M
k
3.2. Allowing for a duration outcome as dependent variable
While the basic model is useful in disentangling the relative
contributions of education and abilities on continuous and binary
health outcomes, it does not allow for a duration outcome like
survival till death.
In our extended model, the first part is the same, defining a
binary educational choice as in (1) and (2), placing the cut-off at
Lower Education (primary school). Hence, in our model individ-
uals face the choice of quitting after primary education (D = 0),
or enrolling into secondary education (D = 1). The measurement
33
equation for latent ability is defined by (5), where we have two
measurements for latent cognitive ability.
It is common practice to define the potential outcomes of a
(3) duration variable like mortality in terms of the hazard that the out-
come of interest occurs.12 We define (1) (t) as the hazard rate for
(4)
an individual with education level beyond primary school (Di = 1),
and (0) (t) as the hazard rate for an individual with an education
level equal to primary school (Di = 0). We assume a Gompertz pro-
portional hazard model for the two potential hazards, which has
been shown to be an accurate representation of mortality between
the ages of 30 and 80 (e.g. Gavrilov and Gavrilova, 1991; Cramer,
2012). Both potential hazards depend on the latent ability ,13 and
observed characteristics XY :
(5)
(0) (t|X Y , ) = exp (a0 t + ˇ0 XiY + ˛0 i ) (6)
(1) (t|X Y , ) = exp (a1 t + ˇ1 XiY + ˛1 i ) (7)
12
We can use a duration model with potential outcomes because the endogenous
education choice is determined before mortality plays a major role: mortality can be
largely ignored for young ages. If the education choice would still play a role during
higher mortality rates the model for educational choice should take selective sur-
vival effects into account. Then a ‘timing-of-events’ model could be a better model,
see Abbring and van den Berg (2003).
13
The latent ability in the hazard is similar to including unobserved heterogeneity
in the hazard, and for identification the unobserved heterogeneity needs to have a
finite mean. The mean of the unobserved heterogeneity term in our model, e˛ , only
depends on ˛ and is finite when ˛ is finite.
34
The effect of latent ability on the hazard is captured by ˛0 and
˛1 . The corresponding potential survival rates are
  t 
(0) Y (0) Y tion. However, Gaussian quadrature can approximate this one
S (t|X , ) = exp −  (s|X , )ds
0 dimensional integral very well. Hence, we estimate the parame-
  t 
S (1) (t|X Y , ) = exp − (1) (s|X Y , )ds
0
Without additional restrictions on the distribution of the latent
factors the model is not identified. However, because we have
an intrinsically non-linear duration outcome instead of a linear
outcome, the Ledermann bound on the number of measurements
compared to the number of latent factors does not apply. Identi-
fication of our model is closely related to the identification in a
mixed proportional hazard (MPH) model, where we assume that
the unobserved heterogeneity has a log-normal distribution. A MPH
model is identified when the unobserved heterogeneity term has
finite mean and is independent of the other observed factors (Elbers
and Ridder, 1982). When we assume a normal distribution for the
k−1 < Di∗ ≤
k (13)
latent ability, ∼N(0, 2 ), the implied unobserved heterogeneity in
 
0 =− ∞ and
K =∞. Then, assuming normally distributed D ,
the hazard (Eqs. (6) and (7)) has mean exp
k . Each education level now has a corresponding
identification ˛j or 2 needs to be fixed. We choose to fix 2 = 1.
Thus the latent ability follows a standard normal distribution.14
An important feature of duration data is that for some individu-
als we only know that he or she survived up to a certain time (often
the end of the observation window). In this case an individual is
(right) censored, i = 0, and we use the survival function instead of
the hazard in the likelihood function. Another feature of duration
data is that only individuals are observed having survived up to a
certain age. In our case, mortality follow-up is only available from
age 55 onwards. In this case the individuals are left-truncated, and
we need to condition on survival up to the age of first observation,
t0 .
The likelihood contribution of individual i in our duration
model is
Li = (j) (t)i S (j) (t)/S (j) (t0 ),
(j)
j = 0, 1
With left-truncated data the distribution of latent ability among
the survivors (up to the left-truncation time) changes. When
only individuals are observed that have survived until age t0 the
likelihood contribution is
 the two levels of educational attainment, where conditioning is

i
Li = ˚ XiD + ˛D  · (1) (t|X Y , ) S (1) (t|X Y , )/S (1) (t0 |X Y , )

1−Di
× ˚ −XiD − ˛D  · (0) (t|X Y , )
i
S (0) (t|X Y , )/S (0) (t0 |X Y , )
G.E. Bijwaard et al. / Journal of Health Economics 42 (2015) 29–43
with h() is a normal distribution with variance 2 = 1. The
maximum likelihood estimation of the parameters involves the
calculation of an integral that does not have an analytical solu-
(8)
ters using maximum likelihood on the basis of Gaussian quadrature
(9) approximation.15
3.3. Allowing for an ordered discrete educational choice
Usually education is available in more than two categories with a
natural ordering of the alternative education levels. As a robustness
check (see Section 4.2), we extend the standard model to account
for this type of ordinal independent variable, where the starting
point is, again, an index model with a single latent variable given as
in (2). Assume there are K education levels and define Di as the indi-
cator of education that takes value k if the individual has reached
education level k:
Di = k if
where
1 2 2
˛ 
2 j 
, for j = 0, 1. For we have an ordered probit model with (K − 1) additional thresh-
old parameters,
potential Gompertz hazard (k) , that depends on exogenous char-
acteristics XY and on the unobserved latent ability, , i.e.,
(k) (t|X Y , ) = exp (ak t + ˇk XiY + ˛k i ) (14)
3.4. Disentangling the effects of ability and education
At the individual level, the main estimate of interest is the sur-
vival difference across the two educational levels, S(1) (t) − S(0) (t),
where S(1) (t) denotes the survival time up to age t for individuals
with at least secondary education (D = 1), and S(0) (t) is the survival
time up to age t for those with primary school only (D = 0). We
are interested in the expected value of this identity for a given
(sub)population. In the sample, the difference in the Kaplan–Meier
survival curves is the unconditional survival difference between the
(10) two levels of educational attainment, E[S(1) (t) − S(0) (t)]. This uncon-
ditional difference can be interpreted as the association between
education and mortality.
Here we are interested to what extent this association is driven
by cognitive ability and other control variables. Using the estimated
parameters, we define the conditional survival difference between
Di
based on cognitive ability and the other control variables, as fol-
lows:
 
 
E S (1) (t) − S (0) (t)|X = x,  = c dF X, (x, c) (15)

2 M − ık XiM − ˛Mk 

1 ik
× dH(|T > t0 ) (11) where X are the covariates, and  is the value of latent cognitive
Mk Mk
k=1
ability. We integrate over the joint distribution of the covariates

with the distribution of the latent abilities conditional on survival

up to t0
 D


˚ Xi + ˛D  S (1) (t0 |X Y , ) + ˚ −XiD − ˛D  S (0) (t0 |X Y , ) h()
dH(|T > t0 ) = 

 (12)
D D
˚ Xi + ˛D  S (1) (t0 |X Y , ) + ˚ −Xi − ˛D  S (0) (t0 |X Y , ) h() d

14
In principle restricting the distribution of  to a normal distribution is not nec-
essary. In line with the literature on MPH models a discrete distribution with finite
points of support would be an alternative choice (Heckman and Singer, 1984). How-
ever, using a normal distribution assumes a continuum of ability values rather than
a finite number, and the distribution of intelligence is generally found to be close to
normal (Gottfredson, 1997).
15
Gaussian quadrature is a numerical integration method based on Hermite poly-
nomials (Press et al., 1993). It provides an efficient approximation for evaluating
indefinite integrals based on normal distributions (Butler and Moffitt, 1982). A sim-
ilar method has been applied before in survival analysis (Lillard, 1993).
 G.E. Bijwaard et al. / Journal of Health Economics 42 (2015) 29–43
and latent ability, FX, (x, c).16 Note that these conditional survival
differences are conditional on surviving to the initial age, which is
55 in our case.
Unfortunately, the integrals cannot be solved analytically, as the
dimension of the covariates X is too large. Another issue is that the
comparison of the survival functions involves the counterfactual of
surviving with another education level. Hence in order to illustrate
the conditional survival differences we resort to simulation.17 For
each education level we simulate the survival of 10,000 individu-
als. To each individual we assign observed characteristics based on
the empirical distribution in the sample. The simulation procedure
consists of four steps:
1. Draw a vector of parameter estimates assuming that the esti-
mator is normally distributed around the point estimates with a
variance-covariance matrix equal to the estimated one.
2. Compute the conditional hazard rates based on these parameter
values and individual characteristics using (6) and (7), condi-
tional on the value of the latent ability.
3. Determine the unconditional survival function for every individ-
ual and for the whole age-range from 55 to 100 on the basis of
equations (8) and (9), and by integrating out the latent ability
through Gaussian quadrature methods.
4. Calculate the average (over the 10,000 individuals) survival at
each age (with steps of a month).
We repeat these steps 100 times to obtain 100 independent obser-
vations of the survival function for each education level.
With this information, we can compute the fraction of individu-
als who are still alive at a certain age for the two educational groups
(both the average and the variance). This defines the conditional
survival difference between the two educational groups, since we
condition on cognitive ability and the other covariates. The simu-
lations also allow us to compute life expectancy (and its standard
error) separately for the two educational groups, by multiplying
the survival function with the age steps.
In order to illustrate the relative importance of education and
cognitive ability, we decompose the unconditional survival differ-
ences from the Kaplan–Meier curves in Fig. 1 into the conditional
survival difference and a residual, which is a selection effect on the
basis of cognitive ability and the other observable factors. Mathe-
matically,
  cation, with children from families with a higher socioeconomic
E S (1) (t) − S (0) (t)
 
 
dren who went to protestant or other schools, as compared to
= E S (1) (t) − S (0) (t)|X = x,  = c dF X, (x, c) + ε X, 
(16)
where the LHS represents the unconditional survival differences
represented by the Kaplan–Meier survival curve, the first part of
the RHS is the conditional survival difference defined in (15), and
ε(X, ) represents the selection effect on the basis of observable
characteristics X and cognitive ability . Note that this selection
effect is the combination of actual selection bias and selection based
on perceived gains of secondary education.
16
Since the conditional survival differences may well be very different for indi-
viduals in different parts of the education distribution, we additionally define the
conditional survival difference for those with D = j, j = 0, 1 as follows:
 
  percent cut-off. The point estimates suggest that the effect of a
E S (1) (t) − S (0) (t)|X = x,  = c, D = j dF X,|D=j (x, c)
17
Cockx and Picchio (2012) use a similar simulation procedure to obtain predic-
tions and the standard errors for non-linear combinations of the parameters. Elbers
et al. (2003) used a related method to account for counterfactual uncertainty.
35
For the ordinal education measure the procedure is very similar.
We have three potential hazards and three possible survival func-
tions, one corresponding to each educational level. Although there
are more possibilities now to compare the educational groups, we
choose to focus on two binary comparisons of the particular edu-
cational level to the educational level directly preceding it. Hence,
we estimate two different conditional survival differences: (i) lower
vocational education compared to primary education only and (ii)
at least general secondary education compared to lower vocational
education.
4. Results
Our baseline specification is the survival model with a binary
education variable and two measurements for cognitive ability.
We estimate the model by maximizing the likelihood in (11), and
present the results in Section 4.1. Exogenous factors influencing
the outcome, XY in (6) and (7), include male, whether the child is
working, family socioeconomic status, and birth rank. Factors addi-
tionally influencing the measurements of cognitive ability, XM in (5),
include school type and the number of teachers at school. Finally,
on top of the exogenous variables affecting the outcome and intel-
ligence, additional factors influencing the educational choice, XD in
(2), include the teacher’s advice, whether a grade was repeated, and
the preference of the parents.
4.1. Main results
Table 2 contains the parameter estimates of the model. The
first column shows that our latent factor of cognitive abil-
ity strongly influences the educational choice, as expected. The
probability of entering secondary school can be derived from
the impact of the latent factor and is already beyond 0.6 for
those with the lowest cognitive abilities, and gradually increases
towards 1.0 for those with the highest cognitive abilities (see
Fig. 2).
Conditional on other observed characteristics such as parental
preference, teachers advice, and family socioeconomic status,
males were less likely to enter secondary school, as are chil-
dren who had to work in the family business during primary
school. Family socioeconomic status is a strong predictor of edu-
status significantly more likely to enter secondary school. Chil-
those who went to catholic schools, were more likely to enter
secondary school. Strong predictors of educational choice are the
teacher’s advice and the preference of the parents. Children who
repeated one or more grades were less likely to enter secondary
school.
Columns 2 and 3 show that on both measurements of cognitive
ability girls did slightly better, and children from families with a
higher socioeconomic status had higher scores. School characteris-
tics such as the school type and the number of teachers also relate
to the test scores.
The final two columns of the table present the determinants
of mortality across the two educational groups. While the point
estimates of the effect of cognitive ability on mortality are neg-
ative as expected, the effects do not reach statistical significance
at the 10 percent level, although the p-values are close to the 10
one-standard deviation increase in cognitive ability is to reduce
the mortality hazard by 18% and 28%, for those with primary
school only and those with at least secondary education, respec-
tively. These results are comparable to both the results presented
in the review by Batty et al. (2007) for the Scottish Mental survey
36 G.E. Bijwaard et al. / Journal of Health Economics 42 (2015) 29–43

Table 2
Duration model – binary education variable, two measurements for ability.

Outcome Education Raven Test Vocabulary Hazard Hazard

Test
D M1 M2 (0) (1)
Cognitive ability
˛ 0.36*** 9.63*** 10.37*** −0.33 −0.20
(0.09) (1.49) (1.58) (0.27) (0.13)
Constant term
c 2.13*** 3.59*** 4.69*** −11.68*** −10.71***
(0.21) (0.74) (0.69) (1.50) (0.80)
a 0.11*** (0.02) 0.09*** (0.01)
Control variables
Male −0.25*** −0.93* −0.87* 0.33 0.66***
(0.08) (0.53) (0.49) (0.25) (0.12)
Child is working – base is “No”
Yes −0.29*** −3.84*** −7.14*** 0.34 0.15
(0.09) (0.63) (0.58) (0.26) (0.14)
Missing −0.32*** −1.07 −2.52*** −0.87 0.12
(0.14) (0.90) (0.83) (0.57) (0.19)
Family socioeconomic status – base is “Low”
Middle 0.42*** 2.55*** 2.24*** −0.34 0.00
(0.10) (0.54) (0.50) (0.30) (0.12)
High 0.42 4.16*** 4.68*** −0.34 0.44
(0.46) (1.64) (1.52) (0.30) (0.29)
Missing −0.54*** −4.39*** −7.63*** −0.67 0.17
(0.18) (1.30) (1.20) (0.59) (0.30)
Birthrank – base is “First”
Second −0.15 0.53 −0.02 −0.17 −0.00
(0.12) (0.79) (0.73) (0.38) (0.16)
Third or Fourth −0.09 −0.22 −2.70*** 0.09 −0.19
(0.11) (0.73) (0.68) (0.35) (0.16)
Fifth or higher −0.09 −3.02*** −4.52*** 0.09 −0.26*
(0.11) (0.73) (0.68) (0.35) (0.16)
Missing 0.11 −0.63 0.47 1.13* −0.65*
(0.31) (1.47) (1.36) (0.62) (0.34)
School religion – base is “Catholic”
Protestant 0.31*** 0.62 2.59***
(0.11) (0.68) (0.63)
Other 0.42** 5.19*** 7.32***
(0.20) (1.13) (1.04)
Number of teachers – base is “5–8 teachers”
≤4 −0.16 −3.81*** −3.16***
(0.10) (0.73) (0.67)
9–12 0.05 0.37 0.42
(0.10) (0.63) (0.59)
Missing 0.33 0.81 0.66
(0.22) (1.30) (1.21)
Teacher’s advice – base is “Lower vocational school”
Continued primary school −0.22**
(0.09)
Lower general secondary school 0.42**
(0.17)
Higher general secondary school 0.39
(0.26)
Missing −0.56**
(0.25)
Repeat grade – base is “None”
Once −0.30***
(0.09)
Twice or more −0.74***
(0.12)
Missing 0.74*
(0.42)
Preference of the parents – base is “Only vocational education”
Work in own company −0.78***
(0.19)
Work without education −1.29***
(0.19)
Work with education −0.88***
(0.20)
General secondary school −0.28
(0.18)
Missing −0.85***
(0.18)
Notes: Author’s calculations on the basis of the Brabant data linked to the municipality register and the mortality register.
*
p-value < 0.1.
**
p-value < 0.05.
***
p-value < 0.01.
 G.E. Bijwaard et al. / Journal of Health Economics 42 (2015) 29–43
Fig. 2. Relationship between cognitive ability and the binary measure for education.
for individuals born in 1921 whose intelligence was measured at
age 11, with mortality follow-up for 65 years (reduction in haz-
ard rate 21%), and the results presented by Batty et al. (2009) for
a cohort of one million Swedish whose IQ was measured at age
18 and who were followed for 20 years (reduction in hazard rate
25%). Males have a higher hazard of dying compared to females,
although the effect is only statistically significant among the higher
educated.
The coefficients in Table 2 allow computing the conditional
survival difference across educational groups, as described in Sec-
tion 3.4. Fig. 3 shows these conditional survival differences for all
age groups from 55 to 75 years of age. The survival difference
between the two educational groups, conditional on family back-
ground and cognitive ability, is positive and increases with age.
Note that the confidence intervals are fairly wide, such that the
conditional survival differences only reach statistical significance
at higher ages. The sizes can be interpreted as percentage point dif-
ferences in the survival probability at a certain age. Hence, around
age 70, entering secondary school is associated with a 2 percentage
point increase in the survival probability (90% confidence interval
0.12
Conditional survival difference
90% upper and lower bound
0.10
0.08
difference in survival
0.06
0.04
0.02
0.00
−0.02
55 60 65 70
age
Fig. 3. Conditional survival differences (see Eq. (15)) by age and binary education
category variable.
37
−1 to 7). Around age 74, entering secondary school is associated
with a 5 percentage point increase in the survival probability (90%
confidence interval 1–10).
If we extrapolate the estimated survival functions outside of
our observed age window, the simulations allow computing the
estimated differences in life expectancy for the average individ-
ual in the sample. This provides an alternative summary measure.
The life expectancy of those only finishing primary school is 82.86
(standard error 1.37), compared with 87.15 (1.11) for those having
finished at least secondary school,18 a statistically significant dif-
ference of 4.29 years (1.58). This implies that entering secondary
school is associated with an increase of more than 4 years in life
expectancy, which is within the bandwidth of the raw survival dif-
ference of 5 years across individuals with primary and secondary
education. It has to be acknowledged, however, that this estimate is
based upon extrapolation and hence on relatively strong functional
form assumptions.
We decompose the unconditional differences in the
Kaplan–Meier survival curves from Fig. 1 into a conditional
difference and a selection effect based on cognitive ability and
other observable control variables.19 Fig. 4 shows that at early
ages mortality differentials are mainly due to selection effects,
while after age 60 the importance of education increases. For
most ages, the selection effect is responsible for around half of the
unconditional differences in survival across educational groups. It
has to be acknowledged though that these conclusions are based
on the point estimates, uncertainty around which is relatively
large in this case.
To gauge the importance of cognitive ability in the selection
effect, we additionally ran all models without the latent factor
for cognitive ability. The results show that the conditional sur-
vival differences are larger in a model without cognitive ability.20
This is an indication that cognitive ability plays an important role
in the selection effect. It is tempting to decompose the selection
effect into a selection due to cognitive ability and a selection on
other observable characteristics. The selection on other observable
characteristics can be computed as the difference between (i) the
unconditional difference from the observed Kaplan–Meier survival
rate of the two education levels and (ii) the conditional survival dif-
ference from the model without cognitive abilities. The selection
on cognitive ability can then be easily computed as the difference
between the total selection effect and the part of the selection effect
attributed to other observable characteristics. This is illustrated in
Fig. 5, which shows that cognitive ability explains the largest part
of the selection effect. In fact, selection on other observable factors
is even negative between ages 60 and 70. We have to emphasize,
however, that this interpretation should be taken with care as cog-
nitive ability could be correlated to other control variables in the
model, and as such the selection effect may not be additive.
4.2. Robustness checks
In a model with an ordinal educational choice, corresponding
to the three levels in the definition of Education in Section 2, the
18
These outcomes are reasonably close to the gender-education specific estimated
life expectancies conditional on surviving to age 55 that Statistics Netherlands
presents on the basis of the 2012 mortality risks (see http://www.cbs.nl/en-GB/
menu/themas/gezondheid-welzijn/cijfers/extra/resterende-gezonde-
levensverwachting.htm?Languageswitch=on). Men and women having finished
primary education can expect to live up to 78.1 and 82.2, respectively. Men and
75 women having finished higher education can expect to live up to 83.3 and 87.1,
respectively.
19
The corresponding graphs using the distribution of X those with D = 1 and D = 0
are very similar.
20
Results are available upon request.
38 G.E. Bijwaard et al. / Journal of Health Economics 42 (2015) 29–43

0.10
Conditional survival difference
Selection Effect

0.08
0.06
0.04
0.02
0.00

55 57 59 61 63 65 67 69 71 73
age

Conditional survival difference

Selection Effect
0.10
0.05
0.00
−0.05

55 57 59 61 63 65 67 69 71 73 75

age

Fig. 4. Decomposition of unconditional difference in the Kaplan–Meier survival function into conditional differences and a selection effect based on observed characteristics
and cognitive ability, with binary education variable and two measurements for cognitive ability.

coefficient estimates of the exogenous variables are very simi-
lar to the ones presented for the binary educational variable.21
Fig. 6 presents the conditional survival differences for the three
different educational levels. It is clear that there is a large, but
insignificant, conditional survival difference between lower voca-
tional school (level 2) and primary school (level 1). At age 75,
those who only attended primary school are around four per-
centage points more likely to die than those who attended lower
vocational school. The conditional survival difference between gen-
eral secondary school and lower vocational school is practically
zero.
21
All results not presented and the details of the models used in this section are
available upon request.
If we decompose the unconditional survival differences
between the three educational groups into a conditional survival
difference and a selection effect, we obtain Fig. 7. This graph
shows that the conditional survival difference between primary
and vocational education is positive and becomes larger than the
selection effect from age 70 onwards, in line with the findings of
the dichotomous indicator for education. The conditional survival
difference between vocational and higher education is negligible.
Taken together, Figs. 6 and 7 clearly indicate that the largest dif-
ference is between those having finished primary school and those
beyond primary school, such that the dichotomization in the pre-
vious subsection seems justified.
While mortality is an objective, and in some sense ‘the ulti-
mate’, health outcome, the influence of education and cognitive
ability may differ depending on the health outcome used. In the
1993 wave of our Brabant survey, hence around age 53 for our
 G.E. Bijwaard et al. / Journal of Health Economics 42 (2015) 29–43 39

0.05
Conditional survival difference
Selection effect (other)
Selection effect (cognitive skills)

0.04
0.03
0.02
0.01
0.00
−0.01
−0.02

55 57 59 61 63 65 67 69 71 73
age

Conditional survival difference

Selection Effect (other)
0.10

Selection Effect (cognitive skills)

0.05
0.00
−0.05

55 57 59 61 63 65 67 69 71 73 75
age

Fig. 5. Decomposition of observed difference in the Kaplan–Meier survival function into conditional differences and a selection effect due to observed characteristics and
cognitive ability, and other selection effects based on observed characteristics only, with binary education variable and two measurements for cognitive ability (with 90%
confidence intervals, below).

sample, a subjective assessment of one’s health was asked to the

Fig. 6. Conditional survival differences by age, ordinal education variable, two mea-
surements for cognitive ability.
respondents in five categories, i.e. ‘poor’, ‘sometimes good, some-
times bad’, ‘fair’, ‘good’, and ‘very good’. We estimated the model
described in Section 3, now allowing for an ordinal dependent vari-
able, to check robustness to our main outcome measure, and to
compare our results to the literature.
We estimated the conditional difference in the probability to
report any of the five categories between the two educational
groups. The conditional difference in the probabilities for the
categories ‘poor’ and ‘sometimes good, sometimes bad’ are signif-
icantly lower with −0.03 (standard error 0.007) and −0.08 (0.011)
respectively. The conditional difference in the probabilities for
the categories ‘fair’ and ‘very good’ are very close to zero, while
the conditional difference in the probability of reporting to be in
‘good’ health between those having finished only primary school
and those entering secondary education is large (and significant)
and amounts to a 15 (standard error 0.011) percentage points
increase.
When comparing our results to the literature, we confirm the
findings of Hartog and Oosterbeek (1998) that both education and
cognitive ability affect self-reported health. Conti and Heckman
(2010) used a binary indicator for ‘poor health’ and found that half
of the raw differences in poor health is due to a treatment effect of
education and the other half was selection. We find that selection
is not as important as in our mortality analyses, while education
does play a large role in explaining the raw differences in health
levels in a model with five health levels and binary education, see
40 G.E. Bijwaard et al. / Journal of Health Economics 42 (2015) 29–43

Conditional survival difference primary to vocational education

Selection Effect primary to vocational education
Conditional survival difference vocational to higher education

0.06
Selection Effect vocational to higher education

0.04
0.02
0.00
−0.02

55 57 59 61 63 65 67 69 71 73
age

Fig. 7. Decomposition of observed difference in the Kaplan–Meier survival function into conditional differences and a selection effect based on observed characteristics and
cognitive ability, with ordinal education variable and two measurements for cognitive ability (with 90% confidence intervals: lower left primary to vocational education;
lower right vocational to higher education).

Conditional difference
0.15

Fig. 8. This suggests that the relative contributions of education and Selection effect
selection effects may well differ across objective health measures
and the subjective health measures that are commonly used in the
literature.
0.10

Since the sample size is somewhat small we chose not to

present all results separately by gender. Yet, since both educa-
tional choices and survival are obviously dependent on gender, we
0.05

ran all models separately for males and females. Strong disparities
in survival across educational groups exist for both males and
females. This can be distracted from Fig. 9 where the height of
0.00

the bar indicates the unconditional survival differences across the

two educational groups. While unconditional survival differences
across educational groups are larger for females, the relative
−0.05

importance of education, derived from the decomposition of the

raw survival differences is higher for males.
One could argue that the Raven progressive matrices test is a
−0.10

purer measurement of cognitive ability and should be used inde-

pendently from the vocabulary test. We ran all analyses for both
the binary and the ordinal educational classification, using only the
poor sometimes bad fair good very good
Raven test as a measure of cognitive ability. The results were very
similar. Fig. 10 depicts the conditional survival difference for the health status
model without latent ability, a model with one measurement and
Fig. 8. Decomposition of observed difference in the self-reported health into con-
a model with two measurements.
ditional differences and a selection effect based on observed characteristics and
Even though the initial sample in 1952 was found to be repre- cognitive ability, with binary education variable and two measurements for cogni-
sentative for the Dutch population at that time, more than half of tive ability (with 90% confidence bars).
the sample is lost between 1952 and our observation period that
 G.E. Bijwaard et al. / Journal of Health Economics 42 (2015) 29–43 41

Conditional survival difference Conditional survival difference

0.14

0.14
Selection effect Selection effect

0.12

0.12
0.10

0.10
0.08

0.08
0.06

0.06
0.04

0.04
0.02

0.02
0.00

0.00
55 57 59 61 63 65 67 69 71 73 75 55 57 59 61 63 65 67 69 71 73 75

age age
0.15

Conditional survival difference Conditional survival difference

0.20
Selection Effect Selection Effect
0.10

0.15
0.05

0.10
0.05
0.00

0.00
−0.05

−0.05
−0.10

55 57 59 61 63 65 67 69 71 73 75 55 57 59 61 63 65 67 69 71 73 75

age age

Fig. 9. Decomposition of observed difference in the Kaplan–Meier survival function into conditional differences and a selection effect based on observed characteristics and
cognitive ability, for males (left) and females (right).

starts in 1995. This could lead to an attrition bias, if attrition is non- again did not find any deviation. All these results are summarized
random. Unfortunately, we do not have access to the original data in Fig. 11 and the detailed estimation results are available upon
files such that we cannot investigate attrition directly. However, request.
Hartog (1989) investigated the non-response for the 1983 survey Finally, we varied the observed characteristics in the model.
and found no attrition bias in a wage analysis.22 Since the sample First by including additional variables among the exogenous vari-
in 1983 has been shown to be representative, we reran all analy- ables such as family size, number of children, additional school
ses on just the respondents that were observed in 1983 and found characteristics (e.g. whether restricted to girls, restricted to boys,
no substantial changes in the results. This suggests that selective or mixed), and whether both parents were still alive. These vari-
attrition does not affect our results. ables were not statistically significant in any of the models, and
The data contains information about children from different did not alter the results. Second, we also checked robustness to
years of birth. Most of them born in earlier years had to repeat excluding individuals with item non-response on some of the
a class and their average cognitive skills are lower. This could observed characteristics, in which case too the results remain
be a potential source of selection, if staying back was due to low similar.
cognitive ability, or, worse, to health reasons. We ran a robustness
check in which we excluded individuals born in 1937, 1938 and
1941. The results show that the conditional survival difference 5. Discussion
hardly deviates from the base model, if anything the difference
even becomes larger. Finally, we included an indicator for the This paper estimates to what extent survival differences across
year of measurement of the education level (1983 or 1993), and educational groups are due to a ‘selection effect’ based on cognitive
ability and other background variables. We extend the structural
equation model of Conti et al. (2010) to allow for a duration depen-
dent variable and an ordinal educational choice, and estimate the
22
Following Hartog (1989) we investigated whether the attrition between 1993 model on the basis of a Dutch cohort born around 1940 for which
and 1995 was related to observed characteristics. Literally all explanatory variables
including education, family background, and intelligence were not related to attri-
we observe mortality between ages 55 and 75. Most important
tion. The only exception was self-reported health; a worse health status increased conclusion is that the selection effect based on cognitive ability is
the probability of attrition between 1993 and 1995. responsible for around half of the raw differences in survival. Yet,
42 G.E. Bijwaard et al. / Journal of Health Economics 42 (2015) 29–43
Fig. 10. Conditional survival differences by age, binary education variable. Base: two
measurements for cognitive ability; Model without latent skills and model with 1
measurement for cognitive ability (Raven test).
Base
0.8
Fig. 11. Conditional survival differences by age, binary education variable using
alternative models. Base: two measurements for cognitive ability; 1983: only
respondents observed in 1983; non-cog: non-cognitive skills included; 1939–40:
only respondents born in 1939–40; edudum: dummy for measurement of education
included.
even conditional on cognitive ability and a wide range of individ-
ual characteristics, survival differences between individuals having
finished only primary school and those who entered at least sec-
ondary education are still substantial, and correspond to a 4 year
difference in life expectancy.
Even though we analyze mortality between ages 55 and 75
rather than self-reported health at age 30, our findings are in line
with the results presented by Conti et al. (2010). Due to this striking
similarity in findings, irrespective of the health measures and
samples used, two tentative conclusions regarding the education-
health gradient are emerging. First, at least half of the raw
association between education and health is due to confounding
‘third factors’, of which cognitive ability proved very important in
our analysis, while Conti et al. (2010) and Savelyev (2012) stress the
importance of non-cognitive factors, in particular conscientious-
ness. Second, even after controlling for cognitive ability, family
socioeconomic status, and a range of other background variables,
education seems to remain important in determining mortality.
This suggests that at least part of the educational differences in
health outcomes is due to a genuine, causal effect of education on
health.
However, a limitation of our data is the absence of direct mea-
surements of non-cognitive ability. Hence, we cannot rule out that
specific non-cognitive factors influence both education and health,
such that our ‘conditional survival difference’ across educational
groups cannot be interpreted as – and is likely to be an upper
bound to – the causal effect of education on mortality. Moreover,
we may overestimate the influence of cognitive ability if correlated
non-cognitive abilities are omitted from the model.23
As a starting point for sorting out this important issue, we
can use the teacher’s advice regarding secondary education of the
child. It is presumably a function of both the cognitive and non-
cognitive abilities of the pupil. Hence, one may argue that while
controlling for cognitive ability, the teacher’s advice is a proxy
for non-cognitive abilities. When allowing the teacher’s advice
to influence mortality directly, on top of being a determinant of
educational choice, the conditional survival differences become
smaller, as illustrated in Fig. 11. This evidence corroborates our
main conclusion that the selection effect explains at least half
of the association between education and mortality. It also sug-
gests that the remaining causal effect of education on mortality
is likely to be smaller when taking non-cognitive abilities into
account.
Another limitation is the relatively small sample size, which nat-
urally compromised the precision of our results. Care should be
taken in interpreting the exact fraction of survival differences into a
selection effect and a part attributable to education, given the large
uncertainty around these estimates. While the conditional survival
differences across educational groups from age 73 onwards as well
as the extrapolated differences in life expectancy across educa-
tional groups are statistically significant, we cannot rule out the
absence of conditional differences in survival across educational
groups at younger ages.
A fruitful avenue for future research would be to investigate
the effect of both education and cognitive abilities on health out-
comes using a more elaborate set of non-cognitive abilities. In such
research the literature could benefit from our structural equation
model that allows for a duration dependent variable like mortality,
and an ordinal independent variable such as educational attain-
ment.
Appendix A. Supplementary data
Supplementary data associated with this article can be
found, in the online version, at http://dx.doi.org/10.1016/j.
jhealeco.2015.03.003.
23
Although the literature suggests that most non-cognitive abilities are uncorre-
lated with IQ (Borghans et al., 2011; Savelyev, 2012).
 G.E. Bijwaard et al. / Journal of Health Economics 42 (2015) 29–43 43

References Elbers, C., Ridder, G., 1982. True and spurious duration dependence: the iden-
tifiability of proportional hazards models. Review of Economic Studies 49,
403–410.
Abbring, J., van den Berg, G., 2003. The nonparametric identification of treatment
Fuchs, V.R., 1982. Time preference and health: an exploratory study. In: Fuchs, V.
effects in duration models. Econometrica 71 (5), 1491–1517.
(Ed.), Economic Aspects of Health. The University of Chicago Press, Chicago.
Albouy, V., Lequien, L., 2008. Does compulsory education lower mortality? Journal
Gavrilov, L.A., Gavrilova, N.S., 1991. The Biology of Life Span: A Quantitative
of Health Economics 28 (1), 155–168.
Approach. Harwood Academic Publisher, New York, ISBN 3-7186-4983-7.
Arendt, J.N., 2005. Does education cause better health? A panel data analysis
Gottfredson, L., 1997. Mainstream science on intelligence: an editorial with 52 sig-
using school reforms for identification. Economics of Education Review 24 (2),
natories, history, and bibliography. Intelligence 24, 13–23.
149–160.
Gottfredson, L., 2004. Intelligence: is the epidemiologists’ elusive fundamental cause
Auld, M.C., Sidhu, N., 2005. Schooling, cognitive ability and health. Health Economics
of social class inequalities in health? Journal of Personality and Social Psychology
14 (10), 1019–1034.
86 (1), 174–199.
Bago d’Uva, T., O’Donnell, O., van Doorslaer, E., 2008. Differential health reporting by
Hartog, J., 1989. Survey non-response in relation to ability and family background:
education level and its impact on the measurement of health inequalities among
structure and effects on estimated earnings functions. Applied Economics 21,
older Europeans. International Journal of Epidemiology 37 (6), 1375–1383.
387–395.
Batty, G.D., Deary, I.J., Gottfredson, L.S., 2007. Premorbid (early life) IQ and later
Hartog, J., Pfann, G., 1985. Vervolgonderzoek Noord-Brabantse zesdeklassers 1983,
mortality risk: systematic review. Annals of Epidemiology 17 (4), 278–288.
Verantwoording van hernieuwde gegevensverzameling onder Noordbrabantse
Batty, G.D., Wennerstad, K.M., Smith, G.D., Gunnell, D., Deary, I., Tynelius, P., Ras-
zesdeklassers van 1952. University of Amsterdam, Amsterdam.
mussen, F., 2009. IQ in early adulthood and mortality by middle age: cohort
Hartog, J., Oosterbeek, H., 1998. Health, wealth and happiness: why pur-
study of 1 million Swedish men. Epidemiology 20 (1), 100–109.
sue a higher education? Economics of Education Review 17 (3),
Behrman, J.R., Rosenzweig, M.R., 2004. Returns to birthweight. Review of Economics
245–256.
and Statistics 86 (2), 586–601.
Hartog, J., Jonker, N., Pfann, G., 2002. Documentatie Brabant data. Netherlands Insti-
Borghans, L., Golsteyn, B.H.H., Heckman, J.J., Humphries, J.E., 2011. IQ, Achievement,
tute for Scientific Information Services, Amsterdam.
and Personality. University of Maastricht (Unpublished manuscript).
Heckman, J.J., Humphries, J.E., Veramendi, G., Urzua, S., 2014. Education Health and
Braakmann, N., 2011. The causal relationship between education, health and health
Wages. NBER Working Paper No. 19971.
related behaviour: evidence from a natural experiment in England. Journal of
Heckman, J.J., Singer, B., 1984. A method for minimizing the impact of distribu-
Health Economics 30 (4), 753–763.
tional assumptions in econometric models for duration data. Econometrica 52,
Brinch, C.N., Galloway, T.A., 2012. Schooling in adolescence raises IQ scores. Pro-
271–320.
ceedings of the National Academy of Sciences of the United States of America
Kaestner, R., Callison, K., 2011. Adolescent cognitive and non-cognitive correlates of
109 (2), 425–430.
health. Journal of Human Capital 5 (1), 29–69.
Butler, J.S., Moffitt, R., 1982. A computationally efficient quadrature procedure for
Lillard, L.A., 1993. Simultaneous equations for hazards: marriage duration and fer-
the one-factor multinomial probit model. Econometrica 50 (3), 761–764.
tility timing. Journal of Econometrics 56, 189–217.
Carneiro, P., Crawford, C., Goodman, A., 2007. The impact of early cognitive and
Lleras-Muney, A., 2005. The relationship between education and adult mortality in
non-cognitive skills on later outcomes. In: CEE DP 92.
the United States. Review of Economic Studies 72, 189–221.
Carpenter, P.A., Just, M.A., Shell, P., 1990. What one intelligence test measures: a
Mathijssen, M.A.J.M., Sonnemans, G.J.M., 1958. Schoolkeuze en schoolsucces bij
theoretical account of processing in the Raven progressive matrices test. Psy-
VHMO en ULO in Noord-Brabant. Zwijssen, Tilburg.
chological Review 97 (3), 404–431.
Mazumder, B., 2008. Does education improve health: a reexamination of the evi-
Case, A., Fertig, A., Paxson, C., 2005. The lasting impact of childhood health and
dence from compulsory schooling laws. Economic Perspectives 33 (2).
circumstance. Journal of Health Economics 24 (2), 365–389.
Mazumder, B., 2012. The effects of education on health and mortality. Nordic Eco-
Centraal Bureau voor de Statistiek, CBS, 2008. Hoogopgeleiden leven lang en gezond.
nomic Policy Review 2012, 261–301.
In: Gezondheid en zorg in cijfers. CBS.
Meghir, C., Palme, M., Simeonova, E., 2013. Education, Cognition and Health: Evi-
Clark, D., Royer, H., 2013. The Effect of Education on Adult Mortality and Health:
dence from a Social Experiment. NBER Working Paper 19002.
Evidence from Britain. American Economic Review 103 (6), 2087–2120.
Murasko, J.E., 2007. A lifecourse study on education and health: the relationship
Cockx, B., Picchio, M., 2012. Are short-lived jobs stepping stones to long-lasting jobs?
between childhood psychosocial resources and outcomes in adolescence and
Oxford Bulletin of Economics and Statistics 74 (5), 646–675.
young adulthood. Social Science Research 36 (4), 1348–1370.
Conti, G., Heckman, J.J., 2010. Understanding the early origins of the education-
Oreopoulos, P., 2006. Estimating average and local average treatment effects of edu-
health gradient: a framework that can also be applied to analyze
cation when compulsory school laws really matter. American Economic Review
gene-environment interactions. Perspectives on Psychological Science 5 (5),
96 (1), 152–175.
585–605.
Press, W.H., Flannery, B.P., Teukolsky, S.A., Vetterling, W.T., 1993. Numerical Recipes
Conti, G., Heckman, J.J., Urzua, S., 2010. The education-health gradient. American
in C: The Art of Scientific Computing, 2nd ed. Cambridge UP, Cambridge.
Economic Review Papers and Proceedings 100, 234–238.
Raven, J.C., 1958. Mill Hill Vocabulav Scale, 2nd ed. H.K. Lewis, London.
Conti, G., Heckman, J.J., Urzua, S., 2011. Early Endowments, Education, and Health.
Ross, C.E., Wu, C.-L., 1995. The links between education and health. American Soci-
University of Chicago, Department of Economics (Unpublished manuscript).
ological Review 60 (5), 719–745.
Cramer, J.S., 2012. Childhood Intelligence and Adult Mortality, and the Role of Socio-
Savelyev, P.A., 2012. Conscientiousness, Education, and Longevity of High-Ability
Economic Status. Tinbergen Institute Discussion Paper 2012-070/4.
Individuals. Vanderbilt University, Department of Economics (Unpublished
Cutler, D., Lleras-Muney, A., 2008. Education and health: evaluating theories and
manuscript).
evidence. In: House, J.S., Schoeni, R.F., Kaplan, G.A., Harold, P. (Eds.), Making
Spearman, C., 1927. The Abilities of Man: Their Nature and Measurement. Macmillan,
Americans Healthier: Social and Economic Policy as Health Policy. Russell Sage
New York.
Foundation, New York.
Van Kippersluis, H., O’Donnell, O., van Doorslaer, E., 2011. Long run returns to edu-
Deary, I., 2008. Why do intelligent people live longer? Nature 456, 175–176.
cation: does schooling lead to an extended old age? Journal of Human Resources
Deary, I., Johnson, W., 2010. Intelligence and education: causal perceptions drive
46 (4), 695–721.
analytic processes and therefore conclusions. International Journal of Epidemi-
Van Praag, M., 1992. Zomaar een dataset: ‘Noordbrabantse zesde klassers’, Een pre-
ology 39, 1362–1369.
sentatie van 15 jaar onderzoek. University of Amsterdam, Amsterdam.
Dronkers, J., 2002. Bestaat er een samenhang tussen echtscheiding en intelligentie?
Van Schellen, M., Nieuwbeerta, P., 2007. De invloed van de militaire dienst-
Mens & Maatschappij 77 (1), 25–42.
plicht op de ontwikkeling van crimineel gedrag. Mens & Maatschappij 82 (1),
Elbers, C., Lanjouw, J.O., Lanjouw, P., 2003. Micro-level estimation of poverty and
5–27.
inequality. Econometrica 71 (1), 355–364.
 Journal of Health Economics 42 (2015) 44–63

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

The intensive margin of technology adoption – Experimental

evidence on improved cooking stoves in rural Senegal夽
Gunther Bensch a , Jörg Peters a,b,∗
a
Rheinisch-Westfälisches Institut für Wirtschaftsforschung (RWI), Essen, Germany
b
AMERU, University of the Witwatersrand, Johannesburg, South Africa

a r t i c l e i n f o a b s t r a c t

Article history: Today, almost 3 billion people in developing countries rely on biomass as primary cooking fuel, with
Received 30 July 2014 profound negative implications for their well-being. Improved biomass cooking stoves are alleged to
Received in revised form 12 March 2015 counteract these adverse effects. This paper evaluates take-up and impacts of low-cost improved stoves
Accepted 12 March 2015
through a randomized controlled trial. The randomized stove is primarily designed to curb firewood
Available online 20 March 2015
consumption, but not smoke emissions. Nonetheless, we find considerable effects not only on firewood
consumption, but also on smoke exposure and, consequently, smoke-related disease symptoms. The
JEL classification:
reduced smoke exposure results from behavioural changes in terms of increased outside cooking and a
C93
I12
reduction in cooking time. We conclude that in order to assess the effectiveness of a technology-oriented
O12 intervention, it is critical to not only account for the incidence of technology adoption – the extensive
O13 margin – but also for the way the new technology is used – the intensive margin.
Q53 © 2015 Elsevier B.V. All rights reserved.

Keywords:
Household air pollution
Energy access
Technology adoption
Development economics
Biomass fuel

1. Introduction mostly firewood. The collection of and cooking with firewood is

In developing countries, almost 3 billion people rely on tradi-
tional biomass-based fuels for their daily cooking purposes. In rural
sub-Saharan Africa, virtually all households cook with biomass,
夽 We thank Mark Andor, Manuel Frondel, Rachel Griffith, Michael Grimm,
Subhrendu Pattanayak, Fiona Ross, Christoph M. Schmidt, Maximiliane Sievert,
and in particular Colin Vance for helpful comments. Participants of the Nordic
Conference in Development Economics, Gothenburg/Sweden in June 2012, the Cen-
tre for the Studies of African Economies conference in Oxford/United Kingdom in
March 2013 and research seminars at University of Göttingen/Germany and Witwa-
tersrand University Johannesburg/South Africa provided valuable input. Financial
support from the German Federal Ministry for Economic Cooperation and Develop-
ment (BMZ) through the Independent Evaluation Unit of Deutsche Gesellschaft für
Internationale Zusammenarbeit (GIZ) is gratefully acknowledged. Peters gratefully
acknowledges the support of a special grant (Sondertatbestand) from the German
Federal Ministry for Economic Affairs and Energy and the Ministry of Innovation,
Science, and Research of the State of North Rhine-Westphalia.
∗ Corresponding author at: RWI, Hohenzollernstrasse 1–3, 45128 Essen, Germany.
Tel.: +49 0201 8149 247; fax: +49 0201 8149 200.
E-mail address: peters@rwi-essen.de (J. Peters).
associated with various negative effects on the living conditions of
the poor. According to the World Health Organization (WHO), the
emitted smoke is the leading environmental cause of death and
is responsible for 4.3 million premature deaths every year – more
deaths than are caused by malaria or tuberculosis (WHO, 2014;
Martin et al., 2011). Medical research throughout the last decades
found links between air pollution induced by open fires and vari-
ous illnesses including pneumonia, chronic obstructive pulmonary
disease (COPD), and eye infections, but also stunted growth of chil-
dren, tuberculosis, and cardiovascular diseases (see Armstrong and
Campbell, 1991; Campbell et al., 1989; Dherani et al., 2008; Kan
et al., 2011; McCracken et al., 2011; Pandey, 1984a,b; Pandey et al.,
1989). Furthermore, biomass usage for cooking is a major source of
climate-relevant emissions (Shindell et al., 2012).
Improved biomass cooking stoves (ICSs) are often believed to
be a game changer for cooking in developing countries. It is in this
context that the United Nations set out the Sustainable Energy for
All initiative with the ambitious goal of globally universal adoption
of clean cooking stoves and electricity by 2030. There is, however, a
wide range of ICSs with different levels of sophistication that have

http://dx.doi.org/10.1016/j.jhealeco.2015.03.006
0167-6296/© 2015 Elsevier B.V. All rights reserved.
 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63
strong implications for smoke emissions and thus cleanliness. It
is hence still a matter of ongoing debate under which conditions
ICSs can be considered as clean, also compared to modern fuels like
electricity and gas.1
This paper presents findings from a Randomized Controlled Trial
(RCT) among 253 households in twelve villages in Senegal to ana-
lyze behavioural responses and impacts following the introduction
of an ICS. The ICS, which was assigned free of charge, is a low-
cost and maintenance-free portable clay-metal stove. It is produced
in a fairly standardized way by local manufacturers (potters and
whitesmiths) in their workshops and is marketed at a retail price
of around 10 US$. The stove has an expected life span of one to
three years before it deteriorates and has to be replaced. It has
already been widely used in large governmental dissemination pro-
grammes in urban and rural Africa. As such, this is the first study
to assess a type of ICS whose design is geared towards fuel savings,
ease of use, affordability and, hence, large-scale applicability, but
one that lacks specific health-conducive technical features such as
a cleaner burning process or a chimney. Without further changes in
cooking behaviour, the reduction in particulate matter emissions
that the randomized ICS can technically achieve would probably be
insufficient to affect the health of users. This is due to the non-linear
particulate exposure–response relation found in medical research
suggests that large reductions in smoke exposure are required in
order to ensure positive health effects (see, for example, Ezzati and
Kammen, 2001; Pope et al., 2011; Burnett et al., 2014).
The main impact indicators of this study are firewood consump-
tion, time use, respiratory disease symptoms and eye infections.
They are supplemented by various indicators along the results
chain of the intervention with regard to cooking behaviour. Effects
on these indicators were assessed 12 months after randomiza-
tion following a baseline study in November 2009. The behavioural
changes we look at – firewood usage patterns and smoke exposure
– can be expected to materialize already in the first few months
after ICS adoption. The changes in these indicators we observe after
one year of ICS ownership therefore reflect impacts to be expected
in the long run – as long as people continue to use the ICS and
replace it by a new one once it is not functional anymore. The
third wave of interviews in March 2013 is used to track the longer-
term usage behaviour and the stove’s durability at the end of what
technically is the life span of the ICS.
A couple of factors contribute to a high external validity of this
RCT for the African context: the study was implemented in an unob-
trusive way in order to ensure that we observe real-world cooking
behaviour. It was designed and conducted in cooperation with the
ICS dissemination programme of the Government of Senegal, so
that an upscaling of the intervention under real-world conditions
would be possible. Furthermore, the dominating cooking fuel in
our study area is firewood, which is also the case in most other
African countries (Bonjour et al., 2013). Firewood scarcity in our
study region and, consequently, the incentive to use more effi-
cient stoves is pronounced and comparable to other dry areas in
non-equatorial Africa.2
We find that the ICSs are taken up by virtually all households
and intensively used, even after three and a half years. For the most
part, people only give up using the stove when it is not functional
anymore and not because they lose interest in using it. We further-
more observe substantial effects on firewood consumption, which
1
See World Bank (2011) for a more detailed discussion of different types of
improved cooking stoves and Martin et al. (2011) for a recent overview on the
improved stoves and air pollution policy debate.
2
External validity and potential challenges to it are discussed further in Section
3.5 and Appendix D.
45
confirm savings rates determined in lab tests. In addition, we find a
decrease in early indicators for respiratory diseases and eye infec-
tions. These effects on people’s health status cannot be explained
only by the take-up of the new ICS and the firewood savings, but
rather by an additional reduction in smoke exposure due to more
outside cooking and a reduced cooking time that is enabled by the
new stove.
Our findings add to the existing body of evidence on ICS impacts,
which so far is mainly represented by two RCTs: the RESPIRE study
in Guatemala (see, for example, Smith-Sivertsen et al., 2004, 2009;
Díaz et al., 2007; Smith et al., 2011) and a study conducted by J-Pal
in India (Hanna et al., 2012).3 Both studies used stationary chim-
ney ICSs that are installed in the user’s kitchen, with the difference
that the RESPIRE stoves are of higher quality, thus more expen-
sive (100–150 US$), and require less maintenance than those used
in the Hanna et al. (2012) study. A more detailed comparison of
technical features of the ICSs used in the different studies is pro-
vided in Appendix A. While the RESPIRE study detects a substantial
reduction in household air pollution and a reduction in the risk
of respiratory disease symptoms and eye problems, Hanna et al.
observe reductions in smoke inhalation only in the first year but
not over a four year time horizon. This is mainly driven by mainte-
nance being more and more neglected over time, which leads to a
weak performance and low usage rates after some years.
Against this background, our paper is the first to add evidence on
how people use an adapted and simple ICS in an unsupervised setup
that is deemed to represent a more realistic study environment
than the highly controlled medical trials conducted for RESPIRE.
Our study contributes to the literature by providing compelling
evidence that such a simpler and cheaper ICS can actually also
trigger substantial impacts – if cooking behaviour also changes.
Conceptually, these results confirm the findings of Hanna et al.:
Looking at the technical features of an ICS is not enough, since the
real-world behaviour of users strongly co-determines the results.
Unlike Hanna et al., though, we find that behavioural adaptations
to a simple ICS may trigger sizable positive health effects.
These differences in findings of the two studies show the poten-
tials of disseminating ICS that are adapted to the target population
and that facilitate cleaner cooking. The stove used in the Hanna et al.
study requires regular maintenance, for which people in turn need
to be trained (which not all of them were), while the stove random-
ized for our study is maintenance-free. Furthermore, our portable
stove is well adapted to the local cooking habits, whereas the stove
distributed in Hanna et al. interferes more with local cooking habits
by requiring people to cook inside, which they are not accustomed
to. In this sense, the stove in our study increases the number of
choice variables for the users, while the one used in Hanna et al.
decreases it.
In this broader behavioural context, our study adds to a
nascent strand in the health economics literature studying adop-
tion behaviour of households for health relevant technologies and
goods such as bednets (Cohen and Dupas, 2010; Tarozzi et al.,
2014), point-of-use drinking water disinfectants (Luby et al., 2008;
Kremer et al., 2009), deworming drugs (Kremer and Miguel, 2007),
condoms (e.g. Kamali et al., 2003), or a range of such technolo-
gies (Wendland et al., 2015). More specifically, it demonstrates
3
In addition to these two studies, further evidence with mixed results exists for
China (Mueller et al., 2013; Yu, 2011), Mexico (Masera et al., 2007) and urban Senegal
(Bensch and Peters, 2013). Burwen and Levine (2012) conducted an RCT in Ghana
using a very simple mud stove. As a major difference to the present study as well as
the RESPIRE and the J-Pal study, tests in a controlled field lab setting already find that
the stove does not perform better than the traditional ones. The poor performance
is also reflected in low usage rates after a few months.
46 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63
that the analysis of technology adoption and related promotion
programmes should encompass both a technical and an economic
perspective, not only an assessment of the mechanical perfor-
mance. This is in line with the concept of intensive and extensive
margins of behaviour that has recently been brought into the
debate on public health interventions (see Dupas, 2011): It is not
only the mere technology adoption that counts (extensive margin).
Rather, the full effect can only be determined if the way the new
technology is used is accounted for as well, the intensive margin.
The remainder of the paper is organized as follows: Section
2 reviews the country and intervention background and outlines
the research design including the identification strategy. Section 3
presents the study results for all our impact indicators, and Section
4 concludes.
2. Programme background and methodological approach
2.1. Improved stove dissemination and cooking fuels in Senegal
Despite its seeming superiority to traditional biomass cooking,
the ICS technology has not made significant inroads into African
households. There may be various reasons for this, which are
comprehensively discussed in Rehfuess et al. (2014) and Lewis
and Pattanayak (2012). One explanation relevant for the rural
setting is that firewood can typically be collected for free so that
most of the benefits of ICS usage are not monetary ones. This
makes it more difficult for households to finance the investment
given liquidity and credit constraints. On the supply side, the
stove design may fail to meet user needs in preparing local dishes
with available fuels and cooking utensils. Earlier programmes in
various African countries relied on subsidies for ICS production
or distributed them for free. Most of these programmes did not
succeed, however, in triggering sustainable ICS usage. Based on
such experience, development practitioners frequently argue that
people do not appreciate and use ICS that they receive as a gift
and, consequently, reject the option of distributing ICSs for free
(Barnes et al., 1994; Martin et al., 2011).
This is also the spirit underlying the ICS dissemination pro-
gramme Foyer Amélioré au Sénégal (FASEN), which is implemented
by the Senegalese Ministry of Energy in cooperation with Deutsche
Gesellschaft für Internationale Zusammenarbeit (GIZ).4 In contrast
to earlier ICS interventions, FASEN focuses on establishing a sus-
tainable and autonomous market for ICSs by testing performance,
training producers and distributors, and supporting communica-
tion and promotional campaigns. Similar to other countries, FASEN
so far concentrated its ICS dissemination on charcoal ICSs in urban
areas.
The main ICS type disseminated by FASEN since 2006, the Jam-
baar, is also used in the present RCT. It is a portable single-pot stove
with a fired clay combustion centre enclosed by a metal casing.
Owing to basic design improvements of the Jambaar compared
to traditional stoves, the woodfuel burns more efficiently and the
heat is better conserved and focused towards the cooking pot.
Both charcoal and firewood models exist. We chose the firewood
Jambaar for our experiment as firewood is the dominant fuel
in rural Senegal with 89% of rural households using it as their
primary cooking fuel (ANSD, 2006). In rural areas ICSs have not
been available so far. Stove types used here are either three-stone
stoves available at zero cost or traditional metal stoves and open
fire grills that can be bought for between 500 and 2500 CFA Francs,
4
GIZ provides technical assistance on behalf of the German Federal Ministry for
Development and Economic Cooperation (BMZ) and is one of the largest bilateral
development agencies in the world.
which is equivalent to 1–5 US$ (see Appendix B for pictures of
the ICS and other stove types used in the study region). The GIZ
programme intends to expand its activities to rural areas and
expects the price of the Jambaar for the rural market to be around
4000–5000 CFA Francs (8–11 US$), which is well below the prices
of the more sophisticated ICS technologies widely disseminated in
Latin America or Asia.
Cooking fuels are an issue of major importance in the daily
life of Senegalese households. Households have the custom to
cook inside, which leads to a higher exposure to smoke emissions
than outside cooking. WHO (2009) holds household air pollution
induced by solid fuel usage for cooking accountable for 6300 pre-
mature deaths every year in Senegal alone. Apart from agricultural
land clearance, wood usage for cooking purposes is moreover the
most important driving force of ongoing deforestation in the mostly
arid and Sahelian country (see WEC/FAO, 1999; Tappan et al., 2004;
FAO, 2005a,b). A constant population growth of 2.6% per year puts
further pressure on fuelwood resources. As a consequence, house-
holds face an increasing scarcity of fuelwoods: firewood collection
is becoming increasingly time-consuming, while fuelwood prices
are rising. This circumstance applies particularly to the Bassin
Arachidier, the study area of this evaluation, situated some 200 km
southeast of Dakar.
2.2. Impact indicators
The first impact indicator of our study is the household consump-
tion of firewood. This indicator aggregates each dish cooked in a
typical week, with a dish being one component of a meal that is pre-
pared on a separate stove, for example rice and sauce. We thereby
account for the fact that several stoves may be used simultaneously
for the preparation of a single meal. The rationale for this indicator
is that a reduction in firewood consumption not only has immedi-
ate implications for wood scarcity and deforestation pressures, but
is also a strong intermediate indicator for other ultimately relevant
impacts such as health and time use.
Impacts on health and time use are examined directly. We
investigate the indicator time spent by household members on fire-
wood collection and cooking and the prevalence of diseases that are
potentially related to firewood usage. For this purpose, we look
at symptoms that are likely to be affected in the short-term after
smoke emissions are reduced; these are captured by the indicators
household member with symptoms of respiratory diseases and house-
hold member with eye problems. We examine this indicator both on
the household level and the household member level. For respira-
tory diseases, these symptoms are cough, asthma, or difficulty in
breathing. They indicate acute respiratory infections and chronic
obstructive pulmonary diseases, which are the leading causes of
mortality and diseases induced by exposure to air pollution from
solid fuels (Ezzati and Kammen, 2002). Exposure to particles could
be detected as a causal agent of these and other serious respiratory
diseases such as lung cancer or pneumonia (see Duflo et al., 2008b;
Pattanayak and Pfaff, 2009).
Respiratory diseases and eye problems are elicited on a self-
reporting basis: respondents are asked to give information on those
household members who exhibited the symptoms of interest in
the six months preceding the interviews. While such self-reported
health indicators are sometimes viewed with concern because of
potential measurement errors, the literature supports their appli-
cation by highlighting the correlation with actual illnesses (see
Idler and Benyamini, 1997; Miilunpalo et al., 1997; Peabody et al.,
2006; Butrick et al., 2010). In particular, if specific symptoms are
asked about precisely as was done in this study, respondents can
be expected to report accurately. A deterioration in recall accuracy
of reported morbidity as found in Das et al. (2012) and Kjellsson
 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63
et al. (2014) is a concern in this study but would only reduce the
precision of our health estimates and not induce any bias.
To record firewood consumption and cooking time, the person
responsible for cooking is asked to specify the number of people
cooked for and the types of stoves used for every meal throughout
a typical day. For each stove application, we then record the
cooking duration and the cooking fuel type. In case of firewood,
the cooking person is additionally asked to pile up the amount
of firewood used for the respective stove application, which is
then weighed with scales. In combination with information on the
frequency with which the respective stoves are used throughout a
typical week, this data serves to determine the weekly household
consumption of firewood. Enumerators crosschecked stove usage as
part of the interviews by verifying which stove was currently in use
or had been used recently. The indicator time spent by household
members on cooking aggregates the self-reported cooking duration
for all meals of a typical day, whereas the time spent by household
members on firewood collection aggregates the spells in which
household members are occupied with gathering firewood in the
course of a week.
Technically achievable savings rates for the Jambaar (referred
to as ICS in the following) have already been determined in con-
trolled cooking tests (CCTs), where a cooking person prepares the
same meal on both a traditional stove and an improved stove in
order to compare the woodfuel consumption of both stove types.
However, the effective savings in real-life households might devi-
ate from such laboratory field tests for various reasons summarized
by Bensch and Peters (2013).5 The deficiencies of CCTs can be over-
come by evaluating the woodfuel consumption based on a survey
among a larger sample of households in which the diversity and
dynamics of real-life cooking practises are captured. This is what is
done in the present paper.
2.3. Identification strategy
We employ two approaches to estimate the impact of ICS usage
in this experimental setup. The intention-to-treat effect (ITT) is
obtained by simply comparing mean values of impact indicators
for the treatment and control group, without accounting for non-
compliance from households that were assigned to the treatment
group but for some reason do not use the ICS. In our case, the ITT
serves to estimate the effect of providing the ICS for free to house-
holds who do not yet own one. The average treatment effect on
the treated (ATT), by contrast, accounts for non-usage in the treat-
ment group and potential take-up in the control group and thereby
serves to estimate the impact of effective ICS usage. For this pur-
pose, instrumental variable (IV) estimations are applied with the
random assignment into the treatment group as an instrument for
the effective usage of the ICS. In our case, ITT and ATT are very simi-
lar given the high compliance rate in the treatment group and given
that only one household in the control group acquired an ICS from
another source. Although RCTs allow for a simple comparison of the
impact indicators at the time of the follow-up, the precision of the
estimates can be increased by controlling for other household char-
acteristics that have been collected in a baseline survey. We there-
fore implement both the ITT and ATT approach with and without
5 6
For example, the tests frequently concentrate on the main meal only and they
cannot account for the fact that households might prepare more hot meals because
cooking becomes cheaper due to the higher efficiency of the ICS (or less exhausting in
terms of firewood collection) – a phenomenon, which is referred to as the rebound
effect in the energy economics literature (see Frondel et al., 2008; Herring et al.,
2009).
47
controlling for baseline household characteristics such as educa-
tion and income using Ordinary Least Squares (OLS) regression.
In order to shed more light on how reductions in firewood con-
sumption are induced by ICS usage, we also do an OLS regression on
the individual dish level, additionally controlling for a set of poten-
tial dish- and meal-specific confounders such as the number of
people cooked for. This dish-level regression has to be interpreted
with some care, since – in spite of the random ICS assignment – the
households that received a new stove can still choose whether to
use the ICS or a traditional stove for the respective meal. This choice
might then be driven by unobservable factors, which would distort
the savings estimates if the unobservables are also correlated with
firewood consumption.
Finally, we employ probit regressions on the health status of
households and of individual household members. In principle,
these estimations might as well suffer from some endogeneity
induced by intra-household bargaining processes: healthier and
more powerful women might bargain themselves out of cooking
with the dirtier stove and into cooking with the cleaner ICS (see
Pitt et al., 2006). This potentially leads to a spurious correlation
between ICS ownership and improvements in the health status. In
our context, though, this is very unlikely, since the assignment to
the cooking duty does not seem to be a result of short-term nego-
tiations, but it is rather determined by cultural norms with one
or two women per household being continuously responsible for
cooking. Even if post-randomization selection processes occurred,
they would be uncovered by the health indicators we use, because
we observe both the people responsible for cooking and those who
are not.
2.4. RCT design and implementation
The study design followed the guidelines on the implementa-
tion of RCTs provided in Duflo et al. (2008a). The first decision that
had to be taken was the level on which to randomize the treatment
– the village or the individual household. In the present case, it is
sensible to randomize on the household level, since the decision
about whether to adopt an ICS is taken in the household and not
on a regional level. Furthermore, our impact indicators are mea-
sured on household level (or below). One reason to randomize on
the village level instead of the household level would be to account
for spillover effects. These are expected to be negligible, since the
ICSs are only used by the households themselves and the penetra-
tion rate per village envisaged in this RCT is too low to affect, for
example, local firewood supply.
The next decision regards the sample size, both in terms of
households and villages. We determined the sample size based on
a power calculation focusing on the indicator firewood savings. We
approximated the relevant parameters ex-ante using the data col-
lected for the quasi-experimental study presented in Bensch and
Peters (2013). Taking into account these parameters and the prob-
ability of being assigned to the treatment group, we obtained a
required sample size of 250 households spread across 12 villages
(see Appendix C). We selected villages that are far away from GIZ-
supported ICS producers in order to avoid treatment contamination
that might occur if households randomly assigned to the control
group obtain an ICS independently.6 Furthermore, we selected the
Two further channels exist through which the treatment may be contaminated.
First, treatment households may share their stoves with control households. This
did not occur. Second, the two household groups may exchange about determinants
of respiratory health, for example. Yet, the treatment did not involve any awareness
raising and cooking is also a rather private issue, as stated in open interviews, that
seems less of a talking point in women’s conversations. As a consequence, only
48 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63
Baseline survey Stove allocat
a ion
and lottery and user instructions
i visiits
Nove
ember 2009 few days
after th
he lottery
Fig. 1. Steps in RCT implementation.
12 villages from the target region of a planned GIZ rural elec-
trification intervention so that we could introduce the study as
preparatory field work related to the electrification project and,
thereby, reduce attention paid to the randomization.
In November 2009, we conducted the baseline survey among
253 randomly sampled households (see Fig. 1 for the timeline of
the RCT). Information was gathered using a structured question-
naire covering the socio-economic dimensions that characterize
the relevant living conditions of the households. Since the study
also served as a baseline in the context of the envisaged electrifi-
cation intervention – a solar home system dissemination project
– a particular focus of the questionnaire was on energy sources
(including electricity) and energy services (including cooking).
Consequently, the cooking-related parts of the interviews did
not draw particular attention. This is important to avoid auspices
biases and Hawthorne effects (see Appendix D). We complemen-
tarily gathered qualitative information in focus group discussions
and semi-structured interviews with key informants such as
women’s groups, stove and charcoal producers, teachers, regional
administrators, and village chiefs.
The random assignment was put into practice through a lot-
tery directly following the baseline interviews. We presented the
prizes of this lottery, an ICS or a 5 kg bag of rice, as recompense
for participation in the baseline study. Participants were therefore
not aware of being part of an experiment. The connotation of the
ICS receivers as the treatment group and the bag of rice receivers
as a control group was not communicated to the participants.7 In
order to increase trust in the fairness of the lottery, we conducted it
in each of the villages directly after completing the interviews and
informed the households immediately about which recompense
they would get. Hence, we applied simple stratified randomization
with the villages as the stratification criterion. Of the 253 house-
holds interviewed for the baseline, 98 received an ICS and 155 a
bag of rice. The rice and ICSs were distributed within three days of
the baseline interview. The households that were drawn to get an
ICS received a brief 15-min introduction on how to use the stove.8
The ICS and rice bag distribution as well as the instruction were
done by field workers who were involved in the preparation of the
electrification project and who were visiting the village anyhow.
No specific village gathering was organized. The ICS was presented
minor contamination effects are conceivable that, furthermore, would rather lead
to an underestimation of effects.
7
The average rice consumption per capita in Senegal is 84 kg per year (GAIN,
2011). Hence, the bag of rice received by the control group corresponds to 0.5% of
annual rice consumption for the average household size in our sample and will most
likely not affect any of our impact indicators that were measured one year after the
distribution.
8
Because many other ICS types require more extensive maintenance and more
usage instructions, one might think of these instructions as a treatment in its own
right, which might be introduced as a random second treatment arm. For our ICS,
this is however not the case. Given the simplicity in the use of the ICS and given that
it is virtually maintenance free, additionally randomizing the instruction within the
treatment group in our case would not make a difference.
Intermeediary Follow-up ICS us
sage
Survey tracking sur
s vey
1, 2 and 7 months November March 2013
2
after allocation
o 2010
as a fuel-saving device, which requires a few precautions. House-
holds were, for example, informed that, in contrast to open fires
for which people typically use large branches or even trunks, the
firewood has to be chopped first in order to fit the relatively small
fuel feed entrance of the ICS. In line with what real-world users are
told about this type of ICS, households were briefly informed about
the convenience co-benefits of fuel savings, which are a quicker
cooking process, less smoke and a cleaner kitchen (if cooking is
done indoors). No information about potential repercussions on
the health status was provided. The complete instructions on the
functioning and proper usage of the ICS and related information
provided are presented in Appendix E.
Between the baseline and the follow-up phase, local community
workers conducted three preparatory visits in the survey villages
for the planned electrification project. It is worth highlighting that
the electrification intervention was not implemented in any of
the sampled villages before the end of this study. Furthermore,
electricity is virtually never used for cooking in rural Africa, in par-
ticular not in the case of solar home systems whose capacity is
not sufficient for cooking purposes. Once in the field, the commu-
nity workers additionally checked if ICS households were using the
ICS and whether they had encountered technical problems (which
were in any case very rare). Again, no further treatment in terms of
awareness raising or usage encouragement was undertaken. While
a few of the households were not yet making frequent use of their
new stove one month after ICS allocation, by the time of the second
visit virtually all ICS households cooked regularly on the ICS.9 For
the follow-up phase at the end of 2010, the same structured ques-
tionnaire was used as in the baseline phase. Attrition was very low:
only four households either could not be located or had moved out
of the village, three in the control and one in the treatment group.
None of the households refused to participate in the follow-up
survey.
We excluded two groups of households from the analysis: four
households with affiliated Quran schools, where usually between
50 and 150 students live and eat and which are therefore not com-
parable to family households, as well as households that prior to
the study had already received improved stoves other than the ICS
used in the RCT from urban relatives. These six treatment and ten
control households cannot be expected to have bought another ICS
in a non-RCT world and therefore do not represent the population
of interest. They were originally included in the randomization
only because they were a priori not clearly discernible and since we
conducted the randomization on-site and directly after the survey.
No further restrictions were made on who to include in the sample
9
It is not likely that the delayed take-up was triggered by the visits or in any way
related to them. Instead, the visits revealed that, first, a few housewives travelled
outside the village and therefore had not used the ICS so far. Second, some women
needed to adapt to the quicker cooking with the ICS, which at the beginning created a
feeling of insecurity. Third, some households were reluctant at the beginning as they
wanted to preserve their ICS and used it only sparsely. Fourth, a few polygamous
households needed some time to decide on who would use the ICS and when.
 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63 49

Table 1
Baseline characteristics of randomly assigned ICS owners and non-owners.

Treatment Control p-value for test on difference in

Mean (sd) Mean (sd) means (2) − (1)
(1) (2) (3)

Socio-economic characteristics
Household size 12.88 (5.55) 12.94 (5.82) 0.94
Family structure (%) 0.96
Extended family 77.8 74.8
Nuclear family 15.6 18.0
Couple or monoparental family 6.6 7.2
Household head is of Wolof ethnicity (%) 52.2 50.4 0.92
Father with more than one wife (1 = yes) 33.7 30.2 0.58
Father’s education level (%) 0.88
None 12.5 9.8
Alphabetization 77.3 77.4
Primary 5.7 7.5
At least secondary 4.5 5.3
Main wife’s education level (%) 0.84
None 41.6 39.6
Alphabetization 51.7 53.9
Primary 6.7 5.8
At least secondary 0.0 0.7
Telecommunication expenditures (CFAF) 4250 (3830) 5090 (8640) 0.43
Ownership of bank account (1 = yes) 0.08 0.06 0.55
Household receives remittances (1 = yes) 0.42 0.45 0.65
Thatched roof (1 = yes) 0.67 0.67 0.97
Wall material of house is stone or brick (1 = yes) 0.49 0.51 0.75
Flooring material is soil (1 = yes) 0.36 0.27 0.16
Land is completely owned by household (1 = yes) 0.94 0.93 0.62
Ownership of sheep (1 = yes) 0.62 0.63 0.87
Number of mobile phones owned 1.86 (1.31) 2.04 (2.15) 0.48
Main wife is member of an association (1 = yes) 0.71 0.73 0.71

Father’s primary activity (%) 0.57

Subsistence farming 79.3 83.4
Services and manufacturing 16.1 14.3
Retirement 4.6 2.3
Cooking-related characteristics
Most utilized stove type (%) 0.33
Open fire (three-stones or Os† ) 72.2 72.7
Traditional metal wood stove 24.4 26.6
LPG stove 3.3 0.7
Stove usage in times per week 21.01 (4.10) 21.43 (4.98) 0.50
Firewood consumption per dish (kg)
Three-stones 4.85 (2.14) 5.14 (2.57) 0.37
Os† 4.84 (2.32) 4.84 (2.84) 1.00
Per capita and dish firewood consumption of three-stones and Os, 0.50 (0.32) 0.47 (0.26) 0.52
main dishes lunch and dinner (kg)
Firewood provision (%) 0.95
Only collected 76.1 76.3
Only bought 1.2 0.7
Both collected and bought 22.7 23.0

Number of observations 90 139

Note: sd – standard deviation; p-values are determined by means of t- and chi-square tests
†
The Os is a stove in which an open fire burns between three metal feet.

or the analysis. Altogether, the sample used for the subsequent
impact analysis in Sections 3.2–3.4 comprised 229 households.
As a robustness check shows, not discarding these two groups
of households and, hence, performing the analysis with all 249
households for which baseline and follow-up data is available does
not change any of our findings, neither when applying ITT nor ATT.
In March 2013, approximately three and a half years after the
randomization, an ICS usage tracking survey among the households
that had received an ICS was conducted by enumerators famil-
iar with the ICS. All but one of the 90 ICS households included
in the impact analysis could be retrieved for this interview wave.
In addition to asking the households simple usage questions, the
enumerators recorded their own assessment on the condition of
the ICS. The results of this usage tracking survey are presented in
Section 3.5.
3. Results
3.1. Socio-economic conditions and cooking behaviour
The primary purpose of this section is to scrutinize the balancing
of the two randomized groups, since we abstained from explic-
itly balancing them through re-randomization before assigning the
ICSs. The second purpose is to illustrate the socio-economic envi-
ronment in which the RCT was implemented. Table 1 documents
the baseline socio-economic and cooking-related characteristics of
the 229 households before stove distribution. On average, house-
holds consist of 13 members; household size varies in a range
between 2 and 42 persons per household. Larger households are
more common: 78% are extended families and 16% nuclear families
(two parents plus children). Four in five households are subsistence
50 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63
Fig. 2. Distribution of non-farm income at baseline.
Fig. 3. Distribution of farm income at baseline.
farmers, the majority of them living in houses with thatched roofs.
As can be seen from the p-values in the right-hand column, two-
sided tests of equality of the values for the two compared groups
do not reveal statistically significant differences. The groups are
balanced in the relevant observable characteristics. In addition,
Figs. 2 and 3 show the distribution in non-agricultural and agricul-
tural income: the treatment and the control group strongly overlap.
Accordingly, a two-sample Kolmogorov–Smirnov test cannot reject
the null of identical distributions at the 10 percent level.10
Regarding the baseline stove usage patterns reflected in the
table, two stove types dominate rural kitchens in Senegal: open
fires (three-stone stoves or Os in which the open fire burns between
metal feet) and traditional metal stoves, the Malagasy and Cire. LPG
stoves are rarely used in rural Senegal; in our sample only three
households mainly use LPG for cooking. 90% of dishes are prepared
with firewood. Around 15% of all meals are prepared with more
than one stove, primarily to prepare rice on one stove and a sauce
on a second one. On average, each household prepares 21 hot dishes
10
We additionally ran a probit regression to check the correlation between ICS
allocation and the joint set of cooking-related as well as socio-economic character-
istics and village dummies. As part of this estimation, we performed a Likelihood
Ratio chi-square test with the null hypothesis that all of the regression coefficients
are simultaneously equal to zero. The p-value of 0.98 validates the findings from the
univariate comparisons of no correlation. All tests have as well been carried out with
the original sample of 253 baseline households, for which statistically significant
differences cannot be observed either.
Table 2
Utilization rates of different stove types at follow-up.
Treatment Control
Open fire 19.5% 70.5%
Traditional metal wood stove 7.7% 24.1%
ICS 69.1% 0.7%†
LPG stove 3.7% 4.7%
Average total number of stove applications per 25.3 22.6
household and week
Note: The shares represent the ratio between the number of times the respective
stove type is used and the total number of stove applications per household and
week.
†
ICS usage among the control group is due to the fact that one household which
was not randomly assigned to receive an ICS acquired one individually after the
randomization.
per week using one of its stoves. As sometimes more than one
stove is used for one meal, the range of weekly stove applications
is between 14 and 49.
The follow-up data on stove usage shows no changes in the
control group: the most often used stove types are three-stone
stoves (53%), traditional metal stoves (25%) or Os (20%). Accord-
ingly, the savings potentials of ICS usage are relatively high with
73% of households mainly using open fire stoves in the absence of
an ICS. For the treatment group, the follow-up data shows that the
ICSs have achieved broad acceptance among users. There are only
two non-compliers: one ICS was completely broken in an accident
and one household did not use the new stove. Otherwise, as many
as 95% of the distributed ICSs are used at least seven times per week;
for 85% of treatment households the ICS became the predominantly
used stove. The proportion of individual dishes prepared with the
different stove types also mirrors this usage pattern (see Table 2).
As such, our set-up mimics the most likely scenario where treat-
ment households have one ICS at their disposal and continue to use
less efficient traditional stoves, because one stove is not sufficient
to prepare the required amount of food or because the ICS is too
small for the pot sizes used in a few large households. The table also
shows that treatment households increased the number of dishes
prepared. This is probably not due to rebound effects (see footnote
**), since the total number of hot meals cooked does not increase
in the treatment group and households reported that the quantity
and type of food prepared has not changed since receiving the ICS.
Instead, the increase simply reflects the fact that ICS households
have an additional stove at their disposal such that the different
components of one meal that were formerly prepared on a single
stove are now prepared on two stoves.
3.2. Firewood consumption
ITT and ATT estimates for the household consumption of fire-
wood indicator are calculated both with and without the baseline
household-level control variables taken from Table 1: in addition
to income, telecommunication expenditures are used as a proxy for
living standards. Bank account ownership is used as a proxy for the
household’s access to credits and ability to pay. Housing condi-
tions as a wealth indicator are captured by whether the flooring
material in the household is soil and whether the wall material is
stone or brick. As another wealth metric, we include a dummy indi-
cating sheep ownership. The results do not change if other wealth
and socio-economic indicators shown in the table are included. As
suggested in Bruhn and McKenzie (2009), we additionally include
village dummies in order to account for the stratified randomiza-
tion. According to our findings presented in columns (1) and (2) of
Table 3, firewood savings are substantial, with around 27 kg being
saved per week in every household after introduction of the ICS.
These are ITT results. ATT estimates differ only marginally being
 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63 51

Table 3
Effect of ICS usage on firewood consumption per week and per dish.

Estimator: Coefficient (Standard Error in parentheses)

Ordinary Least Squares, ITT

Dependent variable: Firewood consumption per week in kg Firewood weight per dish in kg

Variable (1) (2) (3) (4)

Dish variables†
Dish is cooked on open fire Ref. Ref.
Dish is cooked on ICS −1.99*** (0.24) −2.04*** (0.24)
Dish is cooked on traditional metal stove −0.03 (0.53) Ref.
Main dish 1.00*** (0.33)
Short cooking (<30 min) −0.94*** (0.19)

Meal variables†
Number of people the meal is cooked for (in terms of the logarithm 1.96*** (0.57)
of adult equivalents)
Lunch Ref.
Breakfast −1.66*** (0.18)
Dinner −0.32*** (0.10)
Multiple stoves −0.14 (0.33)

Household variables
Household with ICS −26.78*** (6.33) −26.96*** (6.10)
Average number of people cooked for (in terms of the logarithm of 42.79*** (14.36)
adult equivalents)
Father has formal education 2.63 (8.27) 0.01 (0.31)
Mother has formal education 5.95 (5.39) 0.23 (0.20)
Household income (in logarithmic terms) 1.25 (2.48) 0.03 (0.07)
Telecommunication expenditures (in logarithmic terms) 0.56 (0.97) 0.03 (0.03)
Bank account ownership 1.66 (16.85) 0.46 (1.18)
Flooring material is soil −17.63** (7.09) −0.60** (0.25)
Wall material of house is stone or brick 2.55 (9.83) 0.42 (0.36)
Ownership of sheep −7.31 (7.80) −0.32 (0.29)
Association membership of the mother −7.02 (7.93) −0.71** (0.29)

Village dummies Included Included Included Included

Constant −13.60 (42.09) 82.56*** (8.83) −0.77 (1.64) 3.90*** (0.36)

Mean of treatment group 60.80 (3.92) 60.69 (3.24) 2.28 (0.16) 2.25 (0.12)
Mean of control group 87.58 (4.68) 87.65 (5.03) 4.27 (0.17) 4.29 (0.21)
Savings rate (%) 30.6 30.8 46.7 47.5

Number of observations 228 228 627 633

Adjusted R-squared 0.25 0.13 0.43 0.18
F-test 4.83*** 3.71*** 16.18*** 9.44***

Note: Computations on household level (columns 1 and 2) are performed with heteroskedasticity corrected standard errors accounting for heterogeneity in treatment
responses; standard errors for the dish-level estimations (columns 3 and 4) are clustered by household.
†
For an explanation of the dish- and meal-level control variables, see Bensch and Peters (2013).
*
Significance level of 10%.
**
Significance level of 5%.
***
Significance level of 1%.

slightly higher. As these observations hold in the same way for the
other impact indicators, we will only present the more conserva-
tive ITT estimates in the following (ATT estimates can be taken
from Table F1 in Appendix F). Inserting in the regression the values
1 and 0 for the binary treatment variable and average values for the
covariates gives us the absolute ICS consumption values shown at
the bottom of the table. This implies that 30% of the households’
total firewood consumption is saved.
This is clearly less than the 40–50% found in CCTs. As noted
above, rebound effects as one potential driver for the difference
to CCT results do not seem to play a role. Another likely reason is
the fact that treatment households do not switch completely to ICS
usage and still prepare parts of their meals on traditional stoves.
In order to assess the savings potentials in case they would fully
switch to ICS usage, we additionally compare the firewood con-
sumption for dishes prepared on an ICS in the treatment group to
dishes prepared on traditional stove types in the control group.
Even though the analysis of firewood savings on the dish level may
be endogenous, it provides an upper bound estimate of savings
potentials where households had access to several ICSs to poten-
tially abandon traditional stoves completely. These estimations
furthermore provide insights into how the savings materialize,
since they make it possible to examine the influence of dish- and
meal-specific factors. Table 3 shows in columns (3) and (4) the
results for the OLS regression that controls for household charac-
teristics and characteristics specific to the stove application. The
results reveal the differential effects of various dish- and meal-
specific variables whose coefficient signs are as expected and
reflect consistent firewood consumption figures across dish types.
The R2 of 0.43 for the estimation including control variables in col-
umn 3 indicates that a good part of the variation in the dependent
variable can be explained by observable factors. The statistically
highly significant ICS coefficient would imply an average ICS sav-
ings rate of 47%. It is, thus, in the range of the CCT results.
An unbiased alternative to come up with a firewood savings
estimate for the case of adopting ICS for the entire range of
stove applications is to perform a slightly adapted version of
the IV estimation in the calculation of the ATT for total firewood
consumption. We now instrument a new treatment variable, ICS
usage intensity, by the random assignment. Usage intensity is
coded as a continuous variable obtained by dividing the number of
dishes prepared on an ICS by the total number of dishes prepared
52 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63
Table 4
Effect of ICS usage on time expenditures.
Treatment
Mean (se)
(1)
Duration of firewood collection per week (min) 719 ((75.1 )
Number of observations† 86
Cooking duration per day (min) 251
Number of observations 90
Note: All values derived from ITT estimations with heteroskedasticity corrected standard errors (in parentheses) and including village dummies; se – standard error.
†
For the firewood collection indicator, the nine missing observations (5 control and 4 treatment) are due to households that were not able to specify the firewood collection
time spells.
***
Significance level of 1%.
in the respective household. It thus ranges from 0% to 100%. The
resulting Wald estimator yields an average rate of 43.8–45.0%
(with and without controls). This unbiased IV estimate still suffers
from the fact that treatment households increased the number of
stove applications over which they spread the food preparation.
Nevertheless, we can conclude that if all meals in a household
were cooked on an ICS, the savings rate obtained in columns (1)
and (2) of Table 3 could well be around 40%.
The results on firewood consumption turn out to be robust
to outliers just like the results for the time and health indicators
assessed in the following two sections. The two robustness checks
we applied were, first, to estimate the median of the dependent
variables by using quantile regression techniques and, second, to
exclude outliers defined as values more than two standard devia-
tions away from the mean. The results can be taken from Table F2
in Appendix F.
3.3. Time use
As many as 96% of all households collect at least part of the fire-
wood they use for cooking. A reduction in firewood consumption is
likely to lead to households spending less time on firewood collec-
tion. In fact, the reduction in the aggregate time spent by household
members on firewood collection is approximately two and a half
hours per week, which corresponds to 16–17% (Table 4). The reduc-
tion, though, is statistically only borderline significant (p-values of
between 0.15 and 0.19 for ITT with and without controls), a finding
that does not seem to be fully consistent with the reduction in total
firewood consumption of around 30% found in the previous section.
Still, it is not surprising that time savings are less pronounced than
savings in firewood. One reason for the lower savings is that ICS-
using households might just collect less wood during one excursion
instead of reducing the number of excursions. The lack of statisti-
cal significance of the difference might be due to inaccuracies in
the time usage variable, which increases the standard error and,
thus, reduces power. The inaccuracies are induced, for example, by
the fact that 31% of households collect the wood on their own land
while farming, which makes it difficult to disentangle time spent on
the task of collection from time spent on ordinary field work. Also,
some households use a variety of wood supply strategies depending
on different factors, most notably the season: some, for example,
do not collect the firewood every week but instead hold a stock that
is typically replenished before the rainy season.
ICS households might moreover save time because cook-
ing is facilitated and quicker. In qualitative interviews women
repeatedly pointed out that the ICS allows them to regulate the
temperature more easily, which, in turn, makes it easier to do
other things while cooking. The cooking duration of all three meals
throughout a typical day decreases significantly by more than
75 min (Table 4), where preparation of an individual meal on an
Control Difference in means (2) − (1) Regression-adjusted difference in means
Mean (se) Mean (se) p-Value (H0 : Diff = 0) Mean (se) p-Value (H0 : Diff = 0)
(2) (3) (4) (5) (6)
867 ((69.4 ) 148 (103.3 ) 0.15 136 (102.0 ) 0.19
134
333 81 (21.8 ) 0.00*** 75 (20.7 ) 0.00***
139
ICS takes around one and a half hours. These savings far exceed the
time that households additionally invest in cutting the firewood
into smaller pieces, which takes not more than 15 min/day. Due
to a lack of local job and business opportunities, a shift of time
towards income-generating activities cannot be observed. The
cooking women do not seem to sleep more either, since their time
awake differs by mere 5 min between the two compared groups.
Qualitative discussions rather suggest that the facilitation of the
cooking task helps them to execute household duties in a less
hurried way and to take more rest during the day.
3.4. Health
The negative effect of firewood usage on people’s health may
be alleviated by ICS usage via two channels. First, the reductions
in firewood consumption found in Section 3.2 can be expected to
reduce harmful smoke emissions, although it is – as discussed in
the introduction – unclear whether simple ICSs like those used
in this RCT reduce smoke emission sufficiently to induce positive
health effects. Second, exposure to the emitted smoke might be
reduced, either via reductions in the cooking duration (as found
in Section 3.3) or if cooking behaviour changes because of the new
stove. In general, smoke exposure is very high in rural Senegal,
with around two-thirds of the household members responsible
for cooking staying next to the stove most of the time they are
cooking. Furthermore, the vast majority of households cook inside,
predominantly in a separate kitchen. While in the control group the
proportion of outside cookers stays stable, in the treatment group
it doubles from 11% to 23% between baseline and follow-up. The
main reason for this can be traced to the fact that the ICS better
shields the fuel from wind than three-stone stoves; also, from the
households’ perspective, wind and dust are indeed the main draw-
back. In addition, the ICS requires less supervision, allowing the
cook to dedicate more of her attention to other tasks away from
the smoke source.
Virtually all persons responsible for cooking are women, on
average two per household with no difference between treatment
and control. We examine whether chronic symptoms of respiratory
diseases and eye infections prevail among the women respon-
sible for cooking and, as placebo outcomes, among the women
not responsible for cooking and male household members. We
first look at two dummy variables: at least one household mem-
ber with symptoms of respiratory diseases and at least one household
member with eye problems take the value one if at least one house-
hold member of the respective group reports having suffered from
these symptoms at some point in the last six months before the
interview. The results are displayed in Table 5 and indicate the
share of households for which these variables take the value one.
The gender-differentiated data provides for striking indications of
health effects: for women responsible for cooking, 9.0% of treated
 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63 53

Table 5
Effect of ICS usage on health status.

Treatment Control Difference in means (2)-(1) Regression-adjusted difference in means

Mean Mean Mean p-Value (H0 : Diff = 0) Mean p-Value (H0 : Diff = 0)
(1) (2) (3) (4) (5) (6)

Household level analysis

†
Respiratory system disease (%)
Any woman responsible for cooking 9.0 17.7 8.7 0.07*
Any male 7.9 6.5 −1.4 0.69
Any woman not responsible for cooking 3.5 6.2 2.7 0.38
†
Eye problems (%)
Any woman responsible for cooking 4.5 14.0 9.5 0.02**
Any male 4.5 7.3 2.8 0.40
Any woman not responsible for cooking 8.1 7.1 −1.0 0.78
Number of observations‡ 86–90 127–139

Individual level analysis

Cook in household shows symptoms of . . . (%)
A respiratory system disease 4.7 11.8 7.1 0.01** 6.9 0.01***
Eye problems 2.9 9.8 6.9 0.01** 5.7 0.00***
Non-cooking person in house hold shows symptoms of. . . (%)
A respiratory system disease 1.7 2.0 0.3 0.63 0.5 0.46
Eye problems 1.9 2.0 0.1 0.87 0.2 0.73
Number of observations 778 1199

Note: Standard errors for the household level estimations are heteroskedasticity corrected, those for individual household member level estimations are clustered by
household, all estimations include village dummies in order to account for the stratified randomization.
†
ITT with inclusion of controls is not shown in this table, since for some control variables (bank account ownership, flooring material, village dummies) failure is perfectly
predicted in the estimated probit regressions.
‡
Differences in the number of observations are due to a few missing values and some households without any woman not responsible for cooking.
§
The values in this analysis are marginal means and marginal effects derived from estimations that can be found in regression form in Appendix F, Table F3. They are
conventionally calculated at the mean of the other independent variables taking into account the particularities of calculating margins for interaction terms in non-linear
models and conditioning on household members who are cooks.
*
Significance level of 10%.
**
Significance level of 5%.
***
Significance level of 1%.

households report at least one of them suffering from respiratory
disease symptoms. The corresponding value for the control group
of 17.7% is almost twice as large – with this difference being statis-
tically significant.
If we look at the same proportion for male household members,
who usually do not spend time around the cooking spot, treatment
and control group households do not differ significantly from each
other, nor do we find a difference for women not responsible for
cooking. The same pattern is observable for eye infections: 14.0%
of households report that at least one woman responsible for cook-
ing suffers from eye problems in the control group compared with
4.5% in the treatment group. The difference is significantly different
from zero. No such statistically significant difference is observed for
men and women not responsible for cooking. With respect to the
potential bargaining into or out of cooking selection processes out-
lined in Section 2.3, one would expect changes in prevalence rates
in the group of women who are not responsible for cooking if that
bargaining process was strong. However, this is not the case.
The bottom of Table 5 refers to results derived from ITT pro-
bit regressions for the same disease symptoms on the level of
individual household members.11 We now look at the dummy
variables household member with symptoms of respiratory diseases
and household member with eye problems, which take the value
one if the respective household member reports having suffered
from these symptoms at some point in the last six months before
the interview. We find prevalence rates of between 3% and 12%
on individual level.12 The results confirm the findings of the
household level estimations. In the group of household members
responsible for cooking the prevalence rates for both respiratory
disease symptoms and eye infections go down by almost seven
percentage points. Significance levels are even more pronounced
with p-values of 0.01 for both estimations with and without control
variables respectively reflecting the more accurate definition of
the indicator and the larger sample size. The estimations as well
corroborate that the treatment has no effect at all on the group of
household members not responsible for cooking.
Altogether, while the reduction in smoke due to fuel savings
might be too modest to trigger perceivable health effects by itself, it
is likely that the combination with the change in cooking behaviour
enabled by the ICS explains the observed improvements in health
indicators: the ICS facilitates outdoor cooking, the cooking duration
is reduced, and the cooking and combustion process requires less
supervision.
3.5. Impact sustainability and upscaling the intervention
Hitherto we have found quite strong and robust evidence for
high take-up and impacts of ICS usage after one year that are, given
the experimental set-up, internally valid. Internal validity, though,
is only a necessary condition for high policy relevance. The decisive
questions in a next step are, first, whether these usage rates and
impacts persist over time, second, whether the intervention yields

11
We abstain from showing ATT estimations here, since the specification requires
interacting the treatment status with the dummy variable that indicates the cooking
responsibility. We would thus need to instrument the ICS uptake and the interaction
term, respectively. Using the random assignment as instrument for both ICS uptake
and also in the interaction term (which is a controversial procedure) does not deliver
any result in our case, since the estimations do not converge.
12
Comparable data on respiratory system diseases and eye problems for Sub-
Saharan-African countries is very sparse (van Gemert et al., 2011). Studies with
indicator definitions that come closest to ours show comparable levels in these
health problems (ANSD and ICF International, 2012; Adeloye et al., 2013).
54 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63
Fig. 4. Decline in the percentage of ICS users among randomized households.
benefits that outweigh the costs and if so, third, whether it can be
upscaled.
In order to assess the sustainability of the observed impacts we
conducted an ICS usage tracking survey three and a half years after
the random assignment. This enables us to examine the durability
of the randomized ICS under day-to-day rural cooking conditions
and the usage behaviour over the full life-span of the ICS. In this sec-
ond follow-up round, we did not collect information on impacts,
because a majority of the stoves would have already exceeded
their useable lifetimes. For statistical power reasons, this reduced
sample size would have made an examination of impacts difficult.
Considering an expected life span of one to three years, the propor-
tion of 49% of treatment households still using the randomized ICS
can, nevertheless, be considered surprisingly high. In the enumer-
ators’ appraisal, half of these ICS were still in good condition. The
proportion of dishes prepared with an ICS among ICS users declined
only slightly from 70% in 2010 to 62%. As can be seen in Fig. 4, those
treatment households who do not use the ICS anymore (51%) only
slowly ceased to use their ICS. All of them have done so because
the stove has deteriorated and 90% of them still used their ICS two
years after randomization.13
Against this background of persisting usage behaviour we con-
duct a simple cost–benefit analysis. The costs of the ICS are
represented by the market price of around 10 US$. For a conser-
vative estimate of the benefits, to begin with, we only account for
reductions in firewood consumption. We take the average price
of 0.02 US$/kg of firewood paid by firewood-purchasing house-
holds at the time of the follow-up survey as an upper bound of
the shadow price for collected firewood. Valuing the firewood that
ICS users save compared to traditional stove users shows that the
savings amount to 2.03 US$ per month. Even with a lower shadow
price for collected firewood, it is obvious already at this stage that
the benefits of ICSs outweigh the costs by far over its life span. If
health benefits and the reduction in cooking duration were taken
into account, the benefits would be even greater. Similarly, bene-
fits would turn out to be larger when social costs were additionally
included, i.e. forest degradation, village air pollution, and carbon
emissions. As a consequence, upscaling the intervention seems to
be economically sensible.
However, some challenges for external validity of the RCT need
to be considered when transferring the results to an upscaled
intervention or to other regions. In Appendix D, we discuss the
aspects raised by Duflo et al. (2008a): general equilibrium effects,
13
Within the complete investigation period of three and a half years, the ICS was
destroyed in two cases, once because of heavy rainfall and once because the kitchen
wall collapsed. In four cases, the ICS was stolen.
Hawthorne and John Henry effects as well as possible limitations to
generalizations beyond our specific intervention and sample. Over-
all, the external validity of this RCT is quite high. In particular, the
fact that our field experiment was implemented in an unobtrusive
way enables us to transfer the findings to a non-experimental set-
up. In terms of transferability of the high take-up rates, the severe
firewood scarcity in our study area may have increased the incen-
tives for households to effectively use the ICS. Take-up in more
biomass-abundant regions could hence be lower. Another driver of
high usage rates is the fact that we are working with a type of ICS
that is adapted to the rural conditions not only in our study area
but also beyond.
In sum, if permanent access to ICS is ensured and provided that
the ICS is slightly modified in response to potentially different cook-
ing habits elsewhere (e.g. pot sizes or cooking fuel), our findings are
transferable to different populations in (Western) Africa.
4. Discussion and conclusion
In this paper we evaluated take-up behaviour and impacts of
improved cooking stoves (ICSs) in rural Senegal by means of a ran-
domized controlled trial (RCT). ICSs are widely seen as an option for
developing countries to combat the devastating effects of wood-
fuel usage for cooking purposes on people’s health, work load as
well as the environment. The first finding is that ICS take-up was
close to 100% among the randomly assigned households and that
people only cease to use the ICS if it deteriorates. This sustain-
ably high take-up rate comes as a surprise, since it is often argued
among development practitioners that people would not use ICSs
for which they have not paid. It also constitutes a major difference
to the findings in Hanna et al. (2012). Major reasons for this are
probably differences in how convenient and advantageous the ICS
technology is from the household perspective and to which degree
the ICS has a better performance than the existing stove portfo-
lio. First, the ICS used in our study is maybe closer to the regular
cooking habits of the target population. It is easier to use, does
not require any particular maintenance and due to its portability
households can decide themselves where to cook. Second, wood
scarcity is probably higher in our study area thereby increasing the
relevance of an ICS. Third, more than a fourth of the households
in the study in India already also used cleaner fuels like electricity
and gas before the randomization so that the randomized ICS did
not necessarily represent an improvement for them.
The firewood savings were found to be statistically significant
and substantial. They amount to around 30% per week in the most
likely scenario where households have one ICS and continue to
use traditional stoves complementarily. If these complementar-
ily used traditional stoves were also replaced by ICSs, the savings
could increase further up to around 40%. Such a reduction in fire-
wood consumption is an important impact in an arid country like
Senegal, where forests are permanently under pressure and fire-
wood provision is a daily hardship for rural women. Moreover,
the CO2 that is sequestered in both dead wood and green wood
is set free with obvious implications for climate change processes.
Deforestation and forest degradation are in fact a relevant source
of global CO2 emissions. IPCC (2013) estimates that net land-use
change, mainly deforestation, is responsible for about 10% of the
total anthropogenic CO2 emissions. To the extent woodfuel usage
contributes to these processes, dissemination of ICS as used in this
study can help to reduce such losses of carbon sinks.
We also observe a reduction in firewood collection time, but this
is only borderline significant. Furthermore, we find that cooking
duration is decreased significantly by over 20%. In addition, the
cooking process is facilitated so that the time the cook needs to be
 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63
in direct proximity to the cooking spot is reduced. Together with
an increase in outdoor cooking, this leads to an evident reduction
in exposure to harmful smoke. Consequently, we also find a clear
indication of a decrease in respiratory disease symptoms and eye
problems, with a drop of around 9 percentage points each for the
women responsible for cooking.
Our self-reported health outcomes might of course feed criti-
cism that objective indicators such as individual particulate matter
exposure as measured in the RESPIRE study deliver more accurate
information. Apart from the high costs of executing such a sur-
vey, there is also a trade-off between the increased accuracy and a
Hawthorne effect. Study participants can be expected to behave
differently if they are asked to wear exposure monitoring tools
for 24 h, for example. Hence, self-reported and objective measure-
ments can rather be seen as complements. In addition, one might
suspect an auspices or courtesy bias in our data where respon-
dents express their gratitude for having received the ICS or expect
additional benefits from a satisfied implementing agency. In their
stove study in Ghana, Burwen and Levine (2012) suspect that this
effect biases their results, since the positive effects on self-reported
health they observe are not plausible given that smoke exposure
is not reduced. However, this bias is not likely in the present case,
since participating households were not aware of the study’s focus
on ICSs. Even if some households noticed the role the ICS played in
this study, they were unlikely to relate its usage to health outcomes.
The fact that we did not observe any health effect among house-
hold members not responsible for cooking strongly underpins this
view. Hence, different from the Burwen and Levine (2012) study,
placebo outcome indicators corroborated our findings. Finally, the
magnitude of observed savings is in the range of what is expected
based on laboratory tests and, thus, does not feed the suspicion of
biased responses.
Altogether, the substantial and statistically significant impacts
on different levels of indicators including positive external effects
such as reduced deforestation and household air pollution substan-
tiate the efforts that the international community dedicates to the
dissemination of ICSs. The findings on the health level fit into the
concept of intensive and extensive margins of behaviour that has a
longer tradition in agricultural economics (Feder et al., 1985) and
has recently been brought into the debate on public health-relevant
behaviour in developing countries (see Dupas, 2011). The present
analysis suggests that not only the extensive margin of cooking
should be addressed by disseminating cleaner stoves, but also the
intensive margin by, for instance, raising awareness of the need to
reduce smoke exposure. This behavioural dimension should also be
taken into account by the Global Alliance for Clean Cookstoves and
the United Nations in outlining future policies to increase access
to improved or clean cooking stoves. Even ICSs that still emit con-
siderable amounts of smoke might trigger positive health effects if
they also induce exposure-relevant behavioural changes.
The almost universal take-up among randomly assigned ICS
owners suggests that if they have an easy opportunity to obtain an
ICS that is adapted to local cooking habits people also use it. A sim-
ple back-of-the-envelope cost–benefit calculation further made it
clear that investing in an ICS would be a profitable investment
55
from the point of view of the individual households. The inter-
play of cash and credit constraints, the lack of information, and
the fact that in many cases the women responsible for cooking do
not manage the household budget, all this however raises doubts
about whether households would be able and willing to pay the
market price for ICSs, even if the stoves were readily available on
the market. The experience from long-standing pilot dissemination
activities in neighbouring rural areas in Senegal seems to support
the presumption that the majority of rural households would prob-
ably stick to the cheaper traditional three-stone or metal stoves.14
As the strategy of promoting the creation of sustainable ICS mar-
kets has already proven to be difficult in urban areas, where fuels
are purchased and ICS benefits are clearly monetary ones, it can be
expected to require even more efforts and resources in rural areas.
In combination, the high take-up and the positive external
effects of ICS usage observed in this study would suggest that more
direct options of ICS promotion should be reconsidered. This could
mean, for example, directly subsidizing the production of ICSs in
rural areas so that end-user prices can compete with traditional
stoves. If the findings can be confirmed in other rural areas, it might
even be an option to distribute ICSs directly to the households,
either for free or at a very low, symbolic price. While this would
be in contrast to the strategies pursued by most ICS dissemination
programmes, and many practitioners are opposed to a free distribu-
tion policy, the empirical literature provides evidence from other
field experiments that supports the idea. Paying a positive price
does not necessarily lead to higher usage rates of health-relevant
goods (Cohen and Dupas, 2010; Tarozzi et al., 2014), charging cost-
sharing prices substantially reduce take-up (Kremer and Miguel,
2007) and there is only weak evidence yet that price serves to allo-
cate the health-relevant goods to those with the most need (Okeke
et al., 2013).
Any ICS promotion policy has to be designed in close coop-
eration with local stakeholders, putting particular effort into the
choice of technically and culturally appropriate ICS models. Insti-
tutions have to be created to sustain the distribution of direct
subsidies for the ICSs, thereby avoiding the flash-in-the-pan effect
that has been observed in unsuccessful earlier ICS subsidization
programmes.
As these recommendations can only be an interim conclusion,
further research on the take-up behaviour and on the impacts of
ICS usage has to follow up in other regions and potentially other
seasons as well. The indication of positive health effects of the
simpler ICS used in this RCT calls for taking into account cooking
behaviour in these studies. As evidenced by the lower take-up of
ICSs in the Hanna et al. (2012) study in India, the results may vary
in different environments and if other ICS types are used. In addi-
tion, further experimental studies should examine the mechanisms
behind take-up behaviour, such as the households’ willingness-to-
pay for ICSs, but also the role of credit constraints, information,
and woodfuel scarcity. Such research efforts can substantiate –
or contradict – the findings in this study and will thereby help
to decide under which circumstances and to which degree sub-
sidies might in fact be required to encourage rural people to obtain
ICSs.
14
See also Miller and Mobarak (2013) for evidence on low purchase rates of ICS in
Bangladesh.
56 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63

Appendix A. Technical features of improved cookstove used in different studies

Study reference Stove type/model name Combustion Fuel type Feed type Chimney Portability Approx. Further stove
chamber type cost (US$) references
This study Jambaar wood Ceramic Wood Continuous No Yes 10 GIZ (2011a)
Bensch and Peters (2013) Jambaar charcoal Ceramic Charcoal Batch fed No Yes 9–19 GIZ (2011b)
Burwen and Levine (2012) Council of Scientific and Industrial Mud Wood Continuous Yes No <10 –
Research (CSIR) improved stove
Hanna et al. (2012) Appropriate Rural Technology Institute Mud Wood Continuous Yes No 12.5 –
(ARTI) improved stove
Masera et al. (2007) Patsari stove Mud/brick Wood Continuous Yes No 35 Kshirsagar and
Kalamkar (2014)
Miller and Mobarak (2013) Bangladesh Council of Scientific and Clay Wood Continuous E: no E: Yes E: $5.8 Mobarak et al.
Industrial Research (BCSIR) “efficiency” C: Yes C: No C: $10.9 (2012)
(E) and “chimney” (C) stove
RESPIRE Plancha mejorada Brick Wood Continuous Yes No 100–150 Díaz (2008)
Notes: All listed stoves are direct combustion stoves with natural draft. Further main (and more advanced) combustion types are gasifier and rocket type direct combustion;
forced draft is an alternative to natural draft. In addition, the combustion chamber may be metallic.

Appendix B. Stove types used in the survey area

Stove type/model name Combustion chamber type Fuel type Feed type Chimney Portability Approx. cost (US$)

Three-stone stoves None Biomass Continuous No Yes –

Os None Biomass Continuous No Yes 1–2
Cire khatach Metal Crop residues Batch fed No Yes 3–5
Cire wood Metal Wood Continuous No Yes 3–5
Malagasy stove Metal Charcoal (wood) Continuous No Yes 3–5
Jambaar Wood Ceramic Wood Continuous No Yes 10
 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63 57

Appendix C. Power calculation Appendix D. External validity

Since information on our primary impact variable, firewood
consumption, was not available in existing data sets for the
target region of our study, we took data collected in the quasi-
experimental study presented in Bensch and Peters (2013) from
urban Senegal to approximate the relevant parameters (prospec-
tive power analysis). After the follow-up survey, we verified these
parameters by rerunning the analysis with the actual baseline data
for those households included in the analysis (retrospective power
analysis).The sample size n is given by the following formula:
1 2
r (sd1 + sd22 )
n = D[(Z˛ + Zˇ )2 ] widespread ICS usage leads to a sizable reduction in firewood
(X2 − X1 )2
Table C1 provides the description, the values and the sources of
the different parameters. The decisive parameter to be defined by
the researcher is the minimum detectable effect size (ES), which
reflects the smallest relative reduction in woodfuel consumption
that we are able to detect at the given significance level (see Bloom,
1995). While the CCT suggest an effect size of 40%, we chose a
minimum detectable effect size of 30% in order to account for the
possibility of an overestimated effect size in the CCT. We defined
the probability of being assigned to the control group to be 60% and
that for the ICS treatment group to be 40%.
Taking these parameters into account, we obtain a required
sample size of around 200 households, as is indicated in the last row
of the column for the prospective analysis in Table C1. In order to
account for the sensitivity of the different parameters in the power
calculation and potential attrition or non-compliance, we built in a
cushion and increased the number of households to be interviewed
to 250.
With respect to health and time savings impacts, the sample size
required to measure significant effects tends to be substantially
higher. The reason is that the effect on respiratory diseases, for
example, can be expected to be less pronounced. The implication
of this is that the power of our study is not necessarily sufficient to
detect all relevant health and time savings effects.
External validity prevails if a study’s findings can be transferred
from the study population to the policy population. In other words,
external validity is concerned with whether findings obtained from
a small sample group represent the wider population in real world
situations. In the following, we discuss how our RCT design took
into account the three dimensions of external validity as defined
by Duflo et al. (2008a): general equilibrium effects, Hawthorne and
John Henry effects as well as possible limitations to generalizations
beyond our specific intervention and beyond our sample.
General equilibrium effects may occur in the present case if
demand and, in turn, to a reduction in the costs of firewood pro-
vision, either because prices decrease or because firewood is less
scarce and easier to collect. This might induce households to con-
sume more of the now cheaper fuel. Although this would bring
welfare benefits such as more hot meals, from a public health and
resource saving perspective this might be considered an adverse
second-round effect. Since most households in rural Senegal col-
lect firewood and do not buy it, this effect can be expected to be
less pronounced than for market-based energy sources.
Another major risk to the external validity of RCT results is if par-
ticipants change their behaviour because they know that they are
participating in an experiment or are somehow under observation.
While so-called Hawthorne effects (if treatment group members
change their behaviour) or John Henry effects (if control group mem-
bers change their behaviour) can never be ruled out completely,
we reduced the risk considerably through various precautionary
measures: first, we embedded the interviews in a baseline survey
for an electrification intervention under preparation in the studied
areas (the intervention was not implemented in any of the sampled
villages before the end of this study). The applied questionnaire
covered a comprehensive set of socio-economic and energy-related
dimensions such as electricity so that attention was not focused
primarily on cooking-related parts of the interviews. Second, the
lottery was framed as a reward for all households to recompense

Table C1
Table C1 Parameters for power calculation.

Description Value Source

Prospective Retrospective

D = 1 + (m + 1) Design effect, accounting for the loss of variation in the data if clustered 1.59 2.25 Household data†
with instead of simple random sampling is used
 Intra-cluster correlation, i.e. the proportion of the overall variance with 0.031 0.069 Household data
respect to firewood consumption explained by within-village (cluster)
variance in the data
m Mean number of interviewed households per cluster (village) 20 229/12 = 19.1 Defined
Z˛ Critical value (Z-score) for a given level of confidence ˛ reflecting the 1.96 (˛ = 5%) 1.96 (˛ = 5%) Defined (conventional)
probability that the null hypothesis is rejected given that it is in fact true
Zˇ Z-score for a given level of confidence ˇ reflecting the probability that the null 0.84 (ˇ = 80%) 0.84 (ˇ = 80%) Defined (conventional)
hypothesis is rejected given that it is in fact false
R Ratio of treatment and control observations (ICS owners to non-owners) 0.66 90/139 = 0.65 Lottery outcome defined in
sampling design
sd1 Standard deviation of firewood consumption of ICS non-owners 0.266 0.259 Household data
sd2 Standard deviation of firewood consumption of ICS owners 0.186 0.181 Implicitly defined through
minimum detectable effect size
(see below)
X1 Per capita firewood consumption of ICS non-owners (in kg) 0.384 0.411 Household data
X2 Expected per capita firewood consumption of ICS owners (in kg) 0.269 0.288 Implicitly defined through
minimum detectable effect size
(see below)
ES = |X2 –X1 |/X1 Minimum detectable effect size 30% 30% Defined based on experiences
with laboratory tests
n = n (ICS owners) + Result of power calculation: required minimum sample sizes for treatment 192 = 76 + 116 229 = 90 + 139
n (non-owners) and control group
†
Household data refers to the data from the urban quasi-experimental study (“prospective”) and to the baseline data from the present study (“retrospective”) to corroborate
the calculations of the prospective analysis.
58 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63

them for participation in the electrification baseline survey, a sim-

ilar procedure as applied by De Mel et al. (2008) in an RCT on
business grants among micro-enterprises in Sri Lanka. Third, all
survey activities were conducted in an unobtrusive way by local
interviewers and community workers.15
According to Duflo et al. (2008a), three problems may hamper
a valid generalization beyond the specific programme and sample.
First, it may be that the particular care with which the random-
ized treatment was implemented makes it difficult to upscale the
intervention. As outlined in Section 2.4 and the instructions given
to participants (see Appendix E) we keep to what real-world users
are told about the randomized ICS. Furthermore, we conducted
the study together with the Government of Senegal and GIZ and
thereby mimicked a typical ICS dissemination intervention. Sec-
ond, the question arises as to whether we can transfer the results
to a slightly modified intervention. Here, the fact that we dis-
tributed the ICS for free deserves some attention as usage behaviour
might change if households need to pay for the ICS. If a change
can be suspected when households with sufficient willingness-to-
pay self-select into the treatment, then most practitioners would
expect an intensification of usage and, thus, also impacts. Yet, usage Fig. E1. Location of survey sites.
intensity is already high so that no substantial increase can be
expected. The third point is the particularity of the study popu- stoves (see Fig. E1 and Table E1). Within the villages, all households
lation. The most important characteristics here are the fuels used were eligible. They were randomly sampled and none of the sam-
for cooking and their availability. Firewood is the dominant cooking pled households refused to participate in the RCT (see also Fig. E2).
fuel in our sample as it is in major parts of rural Sub-Saharan Africa.
98% of sample households use firewood as their primary cooking E.3. Instructions given to participants
fuel. The national average for rural Senegal is slightly lower at 89%
(ANSD, 2006). Across Sub-Saharan Africa this value amounts to 87% On the day of the ICS distribution, households were reminded
(UNDP/WHO, 2009).16 The reason for these slightly lower numbers via phone in order to make sure the person responsible for cooking
of firewood usage is that the national rural averages include peri- in the household was present. A local staff member with several
urban areas, where charcoal is also used. Firewood usage patterns years of experience in ICS usage training had a meeting with those
in rural Africa excluding peri-urban areas will be very much the women who had received an ICS. In the local language Wolof, he
same as in our sample for the vast majority of countries. presented the ICS as a fuel-saving device and briefly informed about
Firewood availability in our study area is typical for large parts convenience co-benefits: a quicker cooking process, less smoke,
of interior Western Africa and dry savannah regions in general. All and a cleaner kitchen. He verbally informed the women about the
the households in our target area use firewood, which is the case functioning of the stove and the proper utilization. He explained
in virtually all rural areas in Africa. Take-up rates and consequently to them that the clay inlay of the ICS serves the purpose of storing
impacts might change, though, in regions in which firewood is more the heat and that it could easily break if the embers were doused
abundantly available (e.g. the southern region of Senegal) or in with water; instead they were told to put the fire out with sand
which cleaner fuels are already available such as in urban Africa on the ground. Moreover, unlike with open fires for which people
or in parts of rural Asia (see Hanna et al., 2012 for an example). typically use entire branches or even trunks, the firewood has to be
chopped in order to fit the fuel feed entrance of the ICS. He advised
them not to use pot sizes that are too big for the stove and not
Appendix E. Experimental design
to move the pot when it is placed on the stove. Households were
also given a leaflet summarizing these instructions (Fig. E3). This
E.1. Design of the field experiment
is all regular information that is also provided by ICS traders in a
non-RCT setup. In addition, in order to avoid ICS misuse the women
The original design of the experiment was drafted in an incep-
were also asked not to share the ICS with other households or lend
tion report for the Independent Evaluation Unit of Deutsche
it to other women. From a methodological point of view, this was
Gesellschaft für Internationale Zusammenarbeit (GIZ) and finalized
also intended to avoid treatment contamination.
on August 20, 2009. It was concretized during an in-country prepa-
ration mission between October 12 and 22, 2009 and is outlined in Table E1
Section 2.4 of this paper (‘RCT design and implementation’). Table E1 List of survey sites.

Village Rural community

E.2. Selection and eligibility of participants Pethie Djilor
Keur Mandao Djilor
We selected twelve villages from the target region of a planned Ndoffane Ndarry Djilor
GIZ rural electrification intervention in Foundiougne District that Keur Omar Djilor
Goudeme Sidy Djilor
are far away from GIZ-supported producers of improved cooking
Darou Keur Mor Khoredi Diossong
Thiamene Ndiagnene Diossong
Simong Bambara Nioro A. Tall
15 Ndiayene Kad Nioro A. Tall
See Zwane et al. (2011) for an examination of how being surveyed might affect
Keur Bacar Santhie Keur S. Diané
response behaviour. The authors generally call for an unobtrusive method of data
Keur Maniane Keur S. Diané
collection.
16 Nema Bah Toubacouta
See Bonjour et al. (2013) for individual country estimates.
G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63 59

Fig. E2. Participant flow.

60 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63

Fig. E3. Leaflet provided to households that received an ICS.

 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63 61

Appendix F. Additional estimation results

Table F1
Table F1 ATT results for household level indicators on firewood consumption, time expenditures, and health.
Difference in means Regression-adjusted difference in means

(se) p-Value (H0 : Diff = 0) Mean (se) p-Value (H0 : Diff = 0)

(1) (2) (3) (4)
Firewood consumption per week (kg) 27.74 (6.53) 0.00*** 27.64 (5.94) 0.00***
Duration of firewood collection per week (min) 153 ((102.9 ) 0.14 140 ((96.2 ) 0.15
Cooking duration per day (min) 84 ((22.1 ) 0.00*** 77 ((20.3 ) 0.00***
Respiratory system disease (%)
Any woman responsible for cooking 9.1 0.05* 9.2 0.05**
Any male −1.0 0.77 −1.3 0.71
Any woman not responsible for cooking 2.7 0.39 3.1 0.32
Eye problems (%)
Any woman responsible for cooking 9.9 0.02** 10.0 0.01**
Any male 2.5 0.45 2.5 0.44
Any woman not responsible for cooking −1.5 0.70 −0.8 0.82
Note: All computations are performed with heteroskedasticity corrected standard errors accounting for heterogeneity in treatment responses and include village dummies;
se – standard error.
*
Significance level of 10%.
**
Sgnificance level of 5%.
***
Significance level of 1%.

Table F2
Table F2 Outlier analysis for household and dish level indicators.
Outlier analysis using median regressions Outlier analysis using outlier exclusion

Difference in means Regression-adjusted difference Difference in means Regression-adjusted difference

in means in means

(se) p-Value (H0 : Mean (se) p-Value (H0 : (se) p-Value (H0 : Mean (se) p-Value (H0 :
Diff = 0) Diff = 0) Diff = 0) Diff = 0)
(1) (2) (3) (4)
Firewood consumption per week (kg) 26.50 (4.63) 0.00*** 26.51 (3.51) 0.00*** 19.15 (4.41) 0.00*** 18.53 (4.45) 0.00****
Firewood weight per dish (kg) 1.54 (0.19) 0.00*** 1.77 (0.13) 0.00*** 1.50 (0.12) 0.00*** 1.61 (0.13) 0.00***
Duration of firewood collection per week (min) 60 (58 ) 0.30 76 (66 ) 0.25 123 (66 ) 0.06* 117 (62 ) 0.06*
Cooking duration per day (min) 68 (21.6 ) 0.00*** 77 ((16.0 ) 0.00*** 40 ((17.1 ) 0.02** 40 ((17.2 ) 0.02**
Note: Median regressions are quantile regressions that determine the median of the dependent variable conditional on the values of the independent variables. For outlier
exclusion, outliers are defined as values more than two standard deviations away from the mean. All values are computed using robust standard errors; se – standard error.
*
Significance level of 10%.
**
Significance level of 5%.
***
Significance level of 1%.

Table F3
Table F3 Probit regression on health status of household members.
Estimator: Coefficient (Standard Error in parentheses)
Probit, ITT

Dependent variable: Household member with respiratory system Household member with eye problem
disease

Variable (1) (2) (3) (4)

ICS dummy −0.11 (0.16) −0.08 (0.16) −0.06 (0.17) −0.02 (0.15)
Household member is responsible for cooking 0.85*** (0.18) 0.87*** (0.16) 0.68*** (0.16) 0.77*** (0.15)
Household member is responsible for cooking × ICS dummy −0.41 (0.28) −0.43 (0.28) −0.62** (0.29) −0.59** (0.28)

Further household member variables

Household member’s sex 0.03 (0.17) 0.30* (0.16)
Household member’s age 0.01** (0.00) 0.02*** (0.00)

Household variables
Average number of people cooked for (in terms of the logarithm of adult equivalents) −0.34** (0.16) −0.17 (0.17)
Father has formal education −0.13 (0.17) −0.26 (0.25)
Mother has formal education 0.19 (0.12) 0.20 (0.14)
Household income (in logarithmic terms) 0.01 (0.04) 0.01 (0.06)
Telecommunication expenditures (in logarithmic terms) 0.04* (0.02) 0.00 (0.02)
Bank account ownership −0.50 (0.44) −0.52 (0.38)
Wall material of house is stone or brick −0.09 (0.14) −0.30** (0.15)
Ownership of sheep 0.04 (0.13) −0.08 (0.14)
Association membership of the mother 0.08 (0.12) −0.09 (0.14)

Village dummies Included Included Included Included

Constant −2.02*** (0.58) −2.07*** (0.20) −2.41*** (0.77) −2.14*** (0.22)

Number of observations 1977 1977 1977 1977

p-Value of interaction term 0.143 0.119 0.031 0.037
Pseudo R-squared 0.131 0.103 0.176 0.090
Note: The household control variable flooring material is not included, since it predicts failure perfectly in the estimated probit regressions; standard errors are clustered by
household.
*
Significance level of 10%.
**
Significance level of 5%.
***
Significance level of 1%.
62 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63
References
Adeloye, D., Chan, K.Y., Rudan, I., Campbell, H., 2013. An estimate of asthma preva-
lence in Africa: a systematic analysis. Croatian Medical Journal 54 (6), 519–531.
ANSD (Agence Nationale de la Statistique et de la Démographie), 2006. Résul-
tats du troisième recensement général de la population et de l’habitat (2002):
Rapport National de présentation, http://www.ansd.sn/publications/rapports
enquetes etudes/enquetes/RGPH3 RAP NAT.pdf (last accessed 03.03.10).
ANSD (Agence Nationale de la Statistique et de la Démographie), ICF International,
2012. Enquête Démographique et de Santé à Indicateurs Multiples au Sénégal
(EDS-MICS) 2010–2011. ANSD and ICF International, Calverton, MD, USA.
Armstrong, J.R., Campbell, H., 1991. Indoor air pollution exposure and lower
respiratory infections in young Gambian children. International Journal of Epi-
demiology 20 (2), 424–429.
Barnes, D.F., Openshaw, K., Smith, K., Van der Plas, R., 1994. What Makes People
Cook with Improved Biomass Stoves? A Comparative International Review of
Stove Programs, World Bank Technical Paper No. 242. World Bank.
Bensch, G., Peters, J., 2013. Alleviating deforestation pressures? Impacts of improved
stove dissemination on charcoal consumption in urban Senegal. Land Economics
89 (4), 676–698.
Bloom, H., 1995. Minimum detectable effect size – a simple way to report the sta-
tistical power of experimental designs. Evaluation Review 19 (5), 547–556.
Bonjour, S., Adair-Rohani, H., Wolf, J., Bruce, N.G., Mehta, S., Prüss-Ustün, A., Lahiff,
M., Rehfuess, E.A., Mishra, V., Smith, K.R., 2013. Solid fuel use for household
cooking: country and regional estimates for 1980–2010. Environmental Health
Perspectives 121 (7), 784–790.
Bruhn, M., McKenzie, D., 2009. In pursuit of balance: randomization in practice
in development field experiments. American Economic Journal: Applied Eco-
nomics 1 (4), 200–232.
Burnett, R.T., Pope III, C.A., Ezzati, M., Olives, C., Lim, S.S., Mehta, S., Shin, H.H., Singh,
G., Hubbell, B., Brauer, M., Anderson, H.R., Smith, K.R., Balmes, J.R., Bruce, N.G.,
Kan, H., Laden, F., Prüss-Ustün, A., Turner, M.C., Gapstur, S.M., Diver, W.R., Cohen,
A., 2014. An integrated risk function for estimating the global burden of disease
attributable to ambient fine particulate matter exposure. Environmental Health
Perspectives 122 (4), 397–403.
Burwen, J., Levine, D.E., 2012. A rapid assessment randomized-controlled trial of
improved cookstoves in rural Ghana. Energy for Sustainable Development 16
(3), 328–338.
Butrick, E., Peabody, J., Solon, O., DeSalvo, K., Quimbo, S., 2010. A compari-
son of objective biomarkers with a subjective health status measure among
children in the Philippines. Asia-Pacific Journal of Public Health 24 (4),
565–576.
Campbell, H., Armstrong, J.R., Byass, P., 1989. Indoor air pollution in developing
countries and acute respiratory infection in children. Lancet 1, 1012.
Cohen, J., Dupas, P., 2010. Free distribution or cost-sharing? Evidence from a ran-
domized Malaria prevention experiment. Quarterly Journal of Economics 125
(1), 1–45.
Das, J., Hammer, J., Sánchez-Paramo, C., 2012. The impact of recall periods on
reported morbidity and health seeking behaviour. Journal of Development Eco-
nomics 98 (1), 76–88.
De Mel, S., McKenzie, D., Woodruff, C., 2008. Returns to capital in micro-enterprises:
evidence from a field-experiment. Quarterly Journal of Economics 123 (4),
1329–1372.
Dherani, M., Pope, D., Mascarenhas, M., Smith, K.R., Weber, M., 2008. Indoor air
pollution from unprocessed solid fuel use and pneumonia risk in children aged
under 5 years: a systematic review and meta-analysis. Bulletin of the World
Health Organization 86 (5), 390–398.
Díaz, E., Smith-Sivertsen, T., Pope, D., Lie, R., Diaz, A., McCracken, J., Arana, B., Smith,
K., Bruce, N., 2007. Eye discomfort, headache and back pain among Mayan
Guatemalan women taking part in a randomized stove intervention trial. Journal
of Epidemiology & Community Health 61 (1), 74–79.
Díaz, E., (Ph.D. dissertation) 2008. Impact of Reducing Indoor Air Pollution on
Women’s Health. RESPIRE Guatemala-Randomised Exposure Study of Pollu-
tion Indoors and Respiratory Effects. Department of Public Health and Primary
Health Care, The University of Bergen.
Duflo, E., Glennerster, R., Kremer, M., 2008a. Using randomization in development
economics research: a toolkit. In: Schultz, P., Strauss, J. (Eds.), Handbook of
Development Economics. North Holland, Amsterdam, pp. 3895–3962.
Duflo, E., Greenstone, M., Hanna, R., 2008b. Indoor air pollution, health and economic
well-being. Sapiens Journal 1 (1), 1–9.
Dupas, P., 2011. Do teenagers respond to HIV risk information? Evidence from a
field experiment in Kenya. American Economic Journal: Applied Economics 3
(1), 1–34.
Ezzati, M., Kammen, D.M., 2001. Indoor air pollution from biomass combustion and
acute respiratory infections in Kenya: an exposure–response study. The Lancet
358, 619–624.
Ezzati, M., Kammen, D.M., 2002. Household energy, indoor air pollution, and health
in developing countries: knowledge base for effective interventions. Annual
Review of Environment and Resources 27, 233–270.
Feder, G., Just, R.E., Zilberman, D., 1985. Adoption of agricultural innovations in
developing countries: a survey. Economic Development and Cultural Change
33 (2), 255–298.
FAO (Food and Agriculture Organization of the United Nations), 2005a. State of the
World’s Forests. Food and Agriculture Organization of the United Nations, Rome.
FAO (Food and Agriculture Organization of the United Nations, 2005b. Global Forest
Resources Assessment. Food and Agriculture Organization of the United Nations,
Rome.
Frondel, M., Peters, J., Vance, C., 2008. Identifying the rebound: evidence from a
German household panel. Energy Journal 29 (4), 154–163.
GAIN (Global Agricultural Information Network), 2011. Senegal, Grain and Feed
Annual. West Africa Rice Annual, http://gain.fas.usda.gov/Recent%20GAIN%
20Publications/Grain%20and%20Feed%20Annual Dakar Senegal 5-6-2011.pdf
(last accessed 08.02.12).
GIZ (Gesellschaft für Internationale Zusammenarbeit), 2011a. Firewood Jam-
bar Stove, Senegal, https://energypedia.info/wiki/File:GIZ HERA 2011 Jambar
Bois Senegal.pdf (last accessed 25.11.14).
GIZ (Gesellschaft für Internationale Zusammenarbeit), 2011b. Charcoal Jam-
bar Stove, Benin, Kenya, Senegal. https://energypedia.info/images/b/b1/
GIZ HERA 2011 Jambar Charbon Senegal.pdf (last accessed 25.11.14).
Hanna, R., Duflo, E., Greenstone, M., 2012. Up in Smoke: The Influence of Household
Behavior on the Long-run Impact of Improved Cooking Stoves, CEEPR WP 2012-
008. MIT Center for Energy and Environmental Policy Research.
Herring, H., Sorrell, S., Elliott, D., 2009. Energy Efficiency and Sustainable Consump-
tion – The Rebound Effect. Palgrave Macmillan, New York.
Idler, E., Benyamini, Y., 1997. Self-assessed health and mortality: a review of twenty-
seven community studies. Journal of Health and Social Behavior 38 (1), 21–37.
IPCC (Intergovernmental Panel on Climate Change), 2013. Climate Change 2013:
The Physical Science Basis. Contribution of Working Group I to the Fifth Assess-
ment Report of the Intergovernmental Panel on Climate Change. Cambridge
University Press, Cambridge/New York.
Kamali, A., Quigley, M., Nakiyingi, J., Kinsman, J., Kengeya Kayondo, J., Gopal, R.,
Ojwiya, A., Hughes, P., Carpenter, L.M., Whitworth, J., 2003. Syndromic manage-
ment of sexually-transmitted infections and behaviour change interventions
on transmission of HIV-1 in rural Uganda: a community randomised trial. The
Lancet 361 (9358), 645–652.
Kan, X., Chiang, C.Y., Enarson, D.A., Chen, W., Yang, J., Chen, G., 2011. Indoor solid
fuel use and tuberculosis in China: a matched case–control study. BMC Public
Health 11 (1), 1–7.
Kjellsson, G., Clarke, P., Gerdtham, U.-G., 2014. Forgetting to remember or remem-
bering to forget: A study of the recall period length in health care survey
questions. Journal of Health Economics 35, 34–46.
Kremer, M., Miguel, E., 2007. The illusion of sustainability. Quarterly Journal of
Economics 122 (3), 1007–1065.
Kremer, M., Miguel, E., Mullainathan, S., Null, C., Zwane, A.P., 2009. Making Water
Safe: Price, Persuasion, Peers, Promoters, or Product Design? Mimeo.
Kshirsagar, M.P., Kalamkar, V.R., 2014. A comprehensive review on biomass cook-
stoves and a systematic approach for modern cookstove design. Renewable &
Sustainable Energy Reviews 30, 580–603.
Lewis, J.J., Pattanayak, S.K., 2012. Who adopts improved fuels and cookstoves? A
systematic review. Environmental Health Perspectives 120 (5), 637–645.
Luby, S., Mendoza, C., Keswick, B., Chiller, T.M., Hoekstra, R., 2008. Difficulties in
bringing point-of-use water treatment to scale in rural Guatemala. The Ameri-
can Journal of Tropical Medicine and Hygiene 78 (3), 382–387.
Martin II, W.J., Glass, R.I., Balbus, J.M., Collins, F.S., 2011. A major environmental
cause of death. Science 334 (6053), 180–181.
Masera, O., Edwards, R., Armendariz, C., Berrueta, V., Johnson, M., Rojas Bracho, L.,
Riojas-Rodríguez, H., Smith, K.R., 2007. Impact of Patsari improved cookstoves
on indoor air quality in Michoacán, Mexico. Energy for Sustainable Development
11 (2), 45–56.
McCracken, J., Smith, K., Stone, P., Díaz, A., Arana, B., Schwartz, J., 2011. Intervention
to lower household wood smoke exposure in Guatemala reduces ST-segment
depression on electrocardiograms. Environmental Health Perspectives 119 (11),
1562–1568.
Miilunpalo, S., Vuori, I., Oja, P., Pasanen, M., Urponen, H., 1997. Self-rated health
status as a health measure: the predictive value of self-reported health status
on the use of physician services and on mortality in the working-age population.
Journal of Clinical Epidemiology 50 (5), 517–528.
Miller, G., Mobarak, M., 2013. Gender Differences in Preferences, Intra-household
Externalities, and Low Demand for Improved Cookstoves, NBER Working Paper
18964. National Bureau of Economic Research.
Mobarak, A.M., Dwivedi, P., Bailis, R., Hildemann, L., Miller, G., 2012. Low
demand for nontraditional cookstove technologies. Proceedings of the National
Academy of Sciences of the United States of America 109 (27), 10815–10820,
http://dx.doi.org/10.1073/pnas.1115571109.
Mueller, V., Pfaff, A., Peabody, J., Liu, Y., Smith, K.R., 2013. Improving stove evalu-
ation using survey data: who received which intervention matters. Ecological
Economics 93, 301–312.
Okeke, E.N., Adepiti, C.A., Ajenifuja, K.O., 2013. What is the price of prevention? New
evidence from a field experiment. Journal of Health Economics 32 (1), 207–218.
Pandey, M.R., 1984a. Prevalence of chronic bronchitis in a rural community of the
hill region of Nepal. Thorax 39, 331–336.
Pandey, M.R., 1984b. Domestic smoke pollution and chronic bronchitis in a rural
community of hill region of Nepal. Thorax 39, 337–339.
Pandey, M.R., Smith, K.R., Boleij, J.S.M., Wafula, E.M., 1989. Indoor air pollution in
developing countries and acute respiratory infection in children. The Lancet 1,
427–429.
Pattanayak, S., Pfaff, A., 2009. Behavior, environment, and health in developing
countries: evaluation and valuation. Annual Review of Resource Economics 1,
183–217.
 G. Bensch, J. Peters / Journal of Health Economics 42 (2015) 44–63
Peabody, J.W., Nordyke, R.J., Tozija, F., Luck, J., Munoz, J.A., Sunderland, A., DeSalvo,
K., Ponce, N., McCulloch, C., 2006. Quality of care and its impact on population
health: a cross-sectional study from Macedonia. Social Science & Medicine 62
(9), 2216–2224.
Pitt, M.M., Rosenzweig, M.R., Hassan, M.N., 2006. Sharing the Burden of Disease:
Gender, the Household Division of Labor and the Health Effects of Indoor Air
Pollution. Mimeo.
Pope, C.A., Burnett, R.T., Turner, M.C., Cohen, A., Krewski, D., Krewski, D., Jerrett,
M., Gapstur, S.M., Thun, M.J., 2011. Lung cancer and cardiovascular disease
mortality associated with ambient air pollution and cigarette smoke: shape of
the exposure–response relationships. Environmental Health Perspectives 119,
1616–1621.
Rehfuess, E.A., Puzzolo, E., Stanistreet, D., Pope, D., Bruce, N., 2014. Enablers and
barriers to large-scale uptake of improved solid fuel stoves: a systematic review.
Environmental Health Perspectives 122 (2), 120–130.
Shindell, D., Kuylenstierna, J.C.I., Vignati, E., van Dingenen, R., Amann, M., Klimont,
Z., Anenberg, S.C., Muller, N., Janssens-Maenhout, G., Raes, F., Schwartz, J., Falu-
vegi, G., Pozzoli, L., Kupiainen, K., Höglund-Isaksson, L., Emberson, L., Streets,
D., Ramanathan, V., Hicks, K., Kim Oanh, N.T., Milly, G., Williams, M., Demkine,
V., Fowler, D., 2012. Simultaneously mitigating near-term climate change and
improving human health and food security. Science 335, 183–189.
Smith, K.R., McCracken, J.P., Weber, M.W., Hubbard, A., Jenny, A., Thompson, L.M.,
Balmes, J., Díaz, A., Arana, B., Bruce, N., 2011. Effect of reduction in household
air pollution on childhood pneumonia in Guatemala (RESPIRE): a randomised
controlled trial. The Lancet 378 (9804), 1717–1726.
Smith-Sivertsen, T., Díaz, E., Pope, D., Lie, R.T., Díaz, A., McCracken, J.P., Bakke, P.,
Arana, B., Smith, K.R., Bruce, N., 2009. Effect of reducing indoor air pollution
on women’s respiratory symptoms and lung function: the RESPIRE randomized
trial, Guatemala. American Journal of Epidemiology 170 (2), 211–220.
Smith-Sivertsen, T., Díaz, E., Bruce, N., Díaz, A., Khalakdina, A., Schei, M.A.,
McCracken, J.P., Arana, B., Klein, R., Thompson, L.M., Smith, K.R., 2004. Reducing
indoor air pollution with a randomized intervention design – a presentation of
the stove intervention study in the Guatemalan highlands. Norsk Epidemiologi
14 (2), 137–143.
63
Tappan, G., Sall, M., Wood, E.C., Cushing, M., 2004. Ecoregions and land cover trends
in Senegal. Journal of Arid Environments 59 (3), 427–462.
Tarozzi, A., Mahajan, A., Blackburn, B., Kopf, D., Krisham, L., Yoong, J., 2014. Micro-
loans, insecticide-treated bednets, and malaria: evidence from a randomized
controlled trial in Orissa, India. American Economic Review 104 (7), 1909–1941.
UNDP/WHO (United Nations Development Programme and World Health Organi-
zation), 2009. The Energy Access Situation in Developing Countries – A Review
Focused on the Least Developed Countries and Sub-Saharan Africa. United
Nations Development Programme, New York.
van Gemert, F., van der Molen, T., Jones, R., Chavannes, N., 2011. The impact of
asthma and COPD in sub-Saharan Africa. Primary Care Respiratory Journal 20
(3), 240–248.
WEC/FAO (World Energy Council and Food and Agriculture Organization of the
United Nations), 1999. The Challenge of Rural Energy Poverty in Developing
Countries. World Energy Council, London.
Wendland, K.J., Pattanayak, S.K., Sills, E.O., 2015. National-level differences in the
adoption of environmental health technologies: a cross-border comparison
from Benin and Togo. Health Policy Planning 30 (2), 145–154.
WHO (World Health Organisation), 2014. Burden of Disease from House-
hold Air Pollution for 2012, http://www.who.int/phe/health topics/outdoorair/
databases/FINAL HAP AAP BoD 24March2014.pdf (last accessed 26.04.14).
WHO (World Health Organization), 2009. Country Profile of Environmental Burden
of Disease – Senegal. World Health Organisation, Geneva.
World Bank, 2011. Household Cookstoves, Environment, Health, and Climate
Change. A New Look on an Old Problem. World Bank, Washington.
Yu, F., 2011. Indoor air pollution and children’s health: net benefits from stove and
behavioral interventions in rural China. Environmental and Resource Economics
50 (4), 495–514.
Zwane, A.P., Zinman, J., Van Dusen, E., Pariente, W., Null, C., Miguel, E., Kremer, M.,
Karlan, D.S., Hornbeck, R., Giné, Y., Duflo, E., Devoto, F., Crepon, B., Banerjee,
A., 2011. Being surveyed can change later behavior and related parameter esti-
mates. Proceedings of the National Academy of Sciences of the United States of
America 108 (5), 1821–1826.
 Journal of Health Economics 42 (2015) 64–80

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

The effect of medical marijuana laws on adolescent and adult use of

marijuana, alcohol, and other substances夽,夽夽
Hefei Wen a,∗ , Jason M. Hockenberry a,b , Janet R. Cummings a
a
Emory University, Department of Health Policy and Management, 1518 Clifton Road, Atlanta, GA 30322, United States
b
National Bureau of Economic Research (NBER), 1050 Massachusetts Avenue, Cambridge, MA 02138, United States

a r t i c l e i n f o a b s t r a c t

Article history: We estimate the effect of medical marijuana laws (MMLs) in ten states between 2004 and 2012 on ado-
Received 22 May 2014 lescent and adult use of marijuana, alcohol, and other psychoactive substances. We find increases in the
Received in revised form 23 February 2015 probability of current marijuana use, regular marijuana use and marijuana abuse/dependence among
Accepted 13 March 2015
those aged 21 or above. We also find an increase in marijuana use initiation among those aged 12–20. For
Available online 23 March 2015
those aged 21 or above, MMLs further increase the frequency of binge drinking. MMLs have no discernible
impact on drinking behavior for those aged 12–20, or the use of other psychoactive substances in either
JEL classification:
age group.
I18
K32 © 2015 Elsevier B.V. All rights reserved.

Keywords:
Medical marijuana law
Marijuana use
Alcohol use
Natural experiment

As of February 2015, 23 states and the District of Columbia
have implemented medical marijuana laws (MMLs), which permit
marijuana use for medical purposes. Three states (i.e., Maryland,
Minnesota, and New York) adopted MMLs during 2014, and an
additional 11 states1 passed pro-medical marijuana legislation.
Medical marijuana bills have also been considered in many of the
remaining states and are likely to land on the legislative agenda
in more states in the near future. Understanding the behavioral
and public health implications of this evolving regulatory envi-
ronment is critical for the ongoing implementation of MMLs and
future iterations of marijuana policy reform. Despite the growing
夽 The authors gratefully acknowledge the helpful comments on earlier drafts of
this study from Sara J. Markowitz and David H. Howard. All errors are our own.
夽夽 The authors declare that they have no relevant or material financial interests
that relate to the research described in this study. The study was approved by the
Emory University Institutional Review Board (IRB) through an expedited review
procedure.
∗ Corresponding author. Tel.: +1 4047911709.
E-mail addresses: hwen2@emory.edu (H. Wen), jason.hockenberry@emory.edu
(J.M. Hockenberry), jrcummi@emory.edu (J.R. Cummings).
1
11 states with pro-medical marijuana legislation include Alabama, Florida, Iowa,
Kentucky, Mississippi, Missouri, North Carolina, South Carolina, Tennessee, Utah,
and Wisconsin.
consensus about the relief medical marijuana can bring for a range
of serious illnesses, concerns have been voiced that MMLs may
give rise to increased marijuana use in the general population and
increased use of other substances. Legislative and public attention
have focused on these issues, but the empirical evidence is limited.
We contribute to the literature on the effects of marijuana
liberalization policies by examining the effect of the implemen-
tation of MMLs in ten states between 2004 and 2012 on a variety
of substance use outcomes including marijuana use, alcohol use,
pain medication misuse, and hard drug use in both adolescent
and adult populations. To tease out the potential causal effect of
MML implementation, we exploited the geographic identifiers in a
restricted-access version of the National Survey on Drug Use and
Health (NSDUH) micro-level data and estimated two-way fixed
effects models with state-specific linear time trends and a rich set
of individual- and state-level covariates.
We find that implementation of an MML leads to a relative 14
percent increase in the probability of past-month marijuana use
and a 15 percent increase in the probability of almost daily/daily
marijuana use among adults aged 21 or above. For this age group,
MML implementation also results in a 10 percent increase in the
probability of marijuana abuse/dependence. Among adolescents
and young adults aged 12–20, we find a 5 percent increase in

http://dx.doi.org/10.1016/j.jhealeco.2015.03.007
0167-6296/© 2015 Elsevier B.V. All rights reserved.
 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80
the probability of past-year marijuana use initiation attributable
to MML implementation.
In addition to the increases in marijuana use, implementation
of an MML also increases the frequency of binge drinking among
those aged 21 or above, partially through increasing simultaneous
use of the two substances. In contrast, MML implementation does
not affect underage drinking among those aged 12–20. In both age
groups, non-medical use of prescription pain medication, heroin
use, and cocaine use are unaffected.
Overall, our findings indicate that state implementation of an
MML increases marijuana use, but has limited impacts on other
types of substance use (i.e., underage drinking, pain medication
misuse, and hard drug use), except for binge drinking among adults
of legal drinking age.
The article proceeds as follows. Section 1 provides background
information on medical marijuana and MMLs, outlines the theoret-
ical framework, and summarizes the existing literature. Section 2
describes the data sources, variable measurement, and identifica-
tion strategy. Section 3 presents the estimated policy effects, and
the robustness checks. Concluding remarks are given in the last
section of the article.
1. Background
1.1. Medical marijuana law and potential risks and medical value
of marijuana
In the last two decades, growing evidence has lent support to
the efficacy and safety of marijuana as medical therapy to alle-
viate symptoms and treat diseases (see, for instance, Ben Amar,
2006; Campbell and Gowran, 2007; Krishnan et al., 2009; Pertwee,
2012; Gloss and Vickrey, 2012). This growing body of clinical evi-
dence on marijuana’s medicinal value has propelled many states
toward a more tolerant legal approach to medical marijuana. In
1996, California signed the Compassionate Use Act into law (Propo-
sition 215) and became the first state in the U.S. to permit the
medical use of marijuana. And since then a total of 23 states and the
District of Columbia have passed MMLs. These laws are intended to
protect patients from state prosecution for their medical marijuana
use (Hoffmann and Weber, 2010).2
Typically under an MML, a patient with an eligible condition
should first obtain recommendation from a qualified doctor for
the use of marijuana in medical treatment. With the doctor’s rec-
ommendation for medical marijuana use, the patient can then be
issued a medical marijuana patient identification card by the state.
The patient ID cardholder and his/her caregivers are allowed to
possess a certain amount of marijuana through cultivation at home
and/or purchase from a nonprofit retail dispensary licensed by
the state (in some states called “compassionate center”).3 As such,
MMLs in principle should only provide restricted legal protection
2
In contrast to the state MMLs, federal law continues to prohibit marijuana use
for any purpose since the enactment of the Controlled Substances Act (CSA) of
1970. A 2005 Supreme Court decision (Gonzales v. Raich) reaffirmed that federal
law enforcement has the authority to prosecute patients for medical marijuana use
in accordance with state laws (Gostin, 2005). It is only recently that the Obama
administration and the Department of Justice clarified the position that federal
law enforcement resources should not be dedicated to prosecuting persons whose
actions comply with their states’ permission of medical marijuana (Hoffmann and
Weber, 2010). This change in the prosecutorial stance would strengthen the legiti-
macy of existing MMLs and pave the way for the passage of new MMLs.
3
Several more recent MMLs have taken innovative twists that are intended to
tighten the regulation on access to medical marijuana. For instance, New York’s
2014 MML is the first in the U.S. to allow doctors in qualified hospitals to prescribe
medical marijuana instead of recommending it. By allowing for medical marijuana
prescription, the law in effect imposes more responsibility on the participating
65
and access to marijuana for a select group of patients. In practice
however, the laws may have a spillover effect on marijuana use in
the non-patient population.
The spillover effect may arise from four dimensions of the
existing MMLs that create a de facto legalized environment for
marijuana use in the general population (Pacula et al., 2013). First,
although all MMLs specify a list of conditions that are eligible for
medical marijuana,4 most MMLs include in the list a generic term
“chronic pain”, rather than specific diseases causing the pain (e.g.,
neuropathy, fibromyalgia, rheumatoid arthritis, etc.) (Pacula et al.,
2013). The interpretation of “chronic pain” can go far beyond the
original legislative intent, analogous to the practice of off-label
prescribing of other medications. Because pain can often be non-
descript and difficult to verify clinically, a recreational user may
pretend to be a pain patient in order to obtain a prescription for
medical marijuana.
Second, some MMLs do not require establishment of a reg-
istry/renewal system to assess and monitor patient eligibility for
medical marijuana. This, coupled with the loosely-defined eligibil-
ity criteria, further blurs the boundary between the patient and the
non-patient population (Cohen, 2010).
Third, MMLs provide medical marijuana patients with access
to the drug by allowing licensed retail dispensaries and/or home
cultivation. These supply channels exist in a legal grey area and
may proliferate as a result of the reduced threat of prosecution
under the MMLs (Pacula et al., 2010). In particular, Anderson
et al. (2013) provided empirical evidence that MMLs have led to a
substantial increase in the supply of high-grade marijuana. As mar-
ijuana supply rises, it may become prohibitively expensive for law
enforcement to ensure that the entire supply of marijuana intended
for medical purpose ends up in the hands of legitimate patients,
akin to how prescription opioids eventually find their way into the
street drug market. This spillover to the non-patient population is
likely to occur in places where marijuana possession is decriminal-
ized, prosecution of a marijuana offense is local law enforcement’s
“lowest priority”, and federal interference in marijuana regulation
is limited (Sekhon, 2009).
In addition to those specific components of the law, an MML
as a whole symbolizes liberalization of marijuana policy, which
in turn, may give rise to the underestimation of the risks associ-
ated with marijuana use and the normalization of marijuana use
for recreational purposes (Hathaway et al., 2011).
1.2. Literature on the effect of MML on marijuana use in the
general population
Empirical evidence is inconclusive with respect to the effect of
an MML on marijuana use in the general population. A review of
this line of literature is beyond the scope of our paper. We direct
readers to Chu (2014) for a comprehensive review. Briefly, how-
ever, we note that the mixed findings from the previous studies can
be explained by the heterogeneity between different age groups
examined and the variation in specific state laws covered by the
studies.
Studies on youths generally find no significant effect of an MML
on youth marijuana use (e.g., Harper et al., 2012; Lynne-Landsman
et al., 2013; Anderson et al., 2011, 2012). The most comprehensive
evidence comes from Anderson et al. (2011, 2012), which brings
doctors for certifying patients’ medical need, as a doctor can be charged with a
felony for prescribing marijuana to an ineligible patient.
4
California is the only exception that allows medical marijuana for any condition
“for which marijuana provides relief” and leaves the interpretation almost entirely
to the discretion of doctors.
66 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80
together several commonly used data sets and covers an 18-year
period from 1993 to 2011. The study findings suggest that imple-
menting an MML does not lead to a significant increase in marijuana
use among youths. Compared to the literature on youth marijuana
use, the existing literature on the adult population is relatively thin
and limited in scope and rigor (e.g., Harper et al., 2012; Anderson
et al., 2011).
In addition to the potential heterogeneity in the response to an
MML across age groups, MMLs may not be treated as a homoge-
nous set of laws between states and across time. The variation
in specific states laws implemented during different periods may
help reconcile the mixed findings from the previous studies. To
explore this potential heterogeneity, a recent study by Pacula
et al. (2013) uses the same data sets as Anderson et al. (2011,
2012) but replaces a single dichotomous MML indicator with a
set of indicators that represent key provisions of MMLs. Although
none of the estimates using a dichotomous MML indicator are
significant, the MMLs that include a provision requiring patient
registry/renewal are found to lower the marijuana use rates and
marijuana-related treatment admissions. This protective effect of
the patient registry/renewal requirement, however, is offset by
another provision of MMLs that allows licensed retailors to dis-
pense marijuana to medical marijuana patients. The third MML
provision this study examines, the home cultivation provision, has
inconsistent and sometimes counterintuitive effects on marijuana
use. These study findings are informative as to the importance of
distinguishing between MML provisions and recognizing the vari-
ation in state MMLs. A caveat, however, is that although Pacula
et al. (2013) take a more nuanced approach to the classification of
MMLs, they lump youths and adults together in their full-sample
analysis. As a result, the aforementioned age heterogeneity may be
obscured.
1.3. Spillover from marijuana use to the use of alcohol and other
substances
On top of the spillover of marijuana use from medical marijuana
patients to the non-patient population, the potential interdepen-
dence of substance use may lead to a further spillover from
marijuana use to the use of other psychoactive substances.5 Assum-
ing marijuana has a downward sloping demand curve, the effect
of an MML on marijuana use should be unequivocally positive.
The effect on other substance use, however, can be positive or
negative, depending on the relative magnitude of the income
and substitution effects (Chaloupka and Laixuthai, 1997; Pacula,
1998). Specifically, contemporaneous substitution of marijuana for
another substance in response to the implementation of an MML
is most likely to occur for substances that have pharmacologi-
cal effects most similar to that of marijuana. A complementary
relationship, on the other hand, is most likely to occur between
marijuana and another substance if their combined use produces a
synergistic interaction (Moore, 2010). In addition to the contempo-
raneous relationship between marijuana use and other substance
use, there may also be a progression from the demand for mari-
juana to the craving and thus future demand for a more powerful
substance with more intense and longer-lasting effects (Kandel,
1975, 2002).
5 7
However, if the increased marijuana use arising from an MML is not for recre-
ational purpose (i.e., “intoxication”) but for medical purpose only, the use of other
substances is unlikely to be affected.
1.3.1. Relationship between marijuana use and alcohol use
Marijuana and alcohol target many common neural pathways in
human brains (Maldonado et al., 2006). On the one hand, marijuana
use produces rewarding and sedative effects that are comparable
to the effect of alcohol use (Boys et al., 2001; Heishman et al., 1997),
especially low-dose alcohol consumption6 (King et al., 2011). In this
case, when MML lowers the cost of marijuana use, an individual
may substitute marijuana for alcohol to achieve a similar expe-
rience such as a general sense of well-being, with perhaps fewer
immediate negative physical symptoms (e.g. hangovers).
On the other hand, the overall intoxication experience may
be enhanced by the simultaneous use of marijuana and alco-
hol together. Evidence suggests that ethanol, especially when
consumed in high doses, can facilitate the absorption of delta 9-
tetrahydrocannabinol (THC) (Boys et al., 2001). In a randomized
control trial (RCT) conducted by Lukas and Orozco (2001), partici-
pants reported significantly more episodes and longer durations of
euphoria when consuming marijuana together with high doses of
alcohol. The enhanced euphoria following simultaneous consump-
tion of alcohol and marijuana may subsequently lead to a greater
urge to drink even more. Such a scenario points toward a competing
hypothesis that marijuana and alcohol, especially high-dose alco-
hol consumption, are complements rather than substitutes. In this
case, an MML may result in the increased use of both substances.
The takeaway of these pharmacologic findings is that whether
marijuana and alcohol are substitutes or complements may depend
on individual motives for substance use. For instance, those who
only expect a mild feeling of happiness and relaxation from sub-
stance use may consume one of the substances in place of the other.
In contrast, those seeking intense euphoria would consume the two
substances together, perhaps in higher doses.
1.3.2. Relationship between marijuana use and other substance
use
Marijuana is also widely portrayed as a “gateway” drug, essen-
tially inducing the use of drugs with more serious health, legal and
social consequences (Kandel, 1975, 2002). One hypothesized path-
way is through pharmacological mechanisms: once users tolerate
the psychoactive effects of marijuana use, they may crave and seek
out more powerful drugs with more intense and longer-lasting
effects. This pharmacological mechanism would thus predict an
increase in subsequent use of hard drugs such as heroin and cocaine
attributable to the implementation of an MML.
An alternative to this pharmacological mechanism is that the
observed sequence from marijuana use to hard drug use may sim-
ply reflect common predisposing factors rooted in genes or in the
environment coupled with an exposure opportunity mechanism
through which marijuana users may be introduced to a shared
market or subculture of hard drugs (Morral et al., 2002; Wagner
and Anthony, 2002a). If predisposing factors and exposure oppor-
tunities are the primary mechanisms that lead users to transition
from marijuana use to hard drug use, an MML should not result in
an increase in hard drug use because the predisposing factors and
exposure opportunities7 for hard drug use remain unaffected.
In contrast to the concern about MML’s “gateway” effect, there
has been evidence that increased access to medical marijuana
resulting from an MML may benefit certain individuals by reducing
their opioid use. For instance, marijuana may provide analgesia for
6
High-dose alcohol consumption, in contrast, tends to lower sedation and
heighten stimulation (King et al., 2011).
The existing MMLs help marijuana users gain access to the drug through medi-
cal marijuana dispensaries and home cultivation, which are unlikely to expose the
marijuana users to the market or subculture of hard drugs.
 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80
patients with chronic pain (Lynch and Campbell, 2011). Thus, those
who have already received opioid pain medication may experience
improved pain relief and lower their opioid dose after they com-
mence medical marijuana treatment. In addition, those who would
have otherwise initiated opioid analgesics may choose medical
marijuana instead (Abrams et al., 2011). Furthermore, marijuana
may also benefit those with opioid misuse (i.e., non-medical use)
by easing withdrawal symptoms and facilitating recovery (Scavone
et al., 2013). Therefore, one would expect states with MMLs to
see a reduction in prevalence of opioid use, or other downstream
benefits such as reduced overdose mortality (Bohnert et al., 2011;
Bachhuber et al., 2014).
1.4. Literature on the relationship between marijuana use and
the use of alcohol and other substances
Through increased marijuana use, a further consequence of an
MML could also be the spillover to alcohol use and the use of other
psychoactive substances. Identification of the spillover effect in
an observational study hinges on the isolation of the exogenous
variation in substance use arising from policy/price shocks from
the endogenous variation due to “common factors” or “exposure
opportunities.”
Previous studies have exploited changes in state excise taxes
on beer (Pacula, 1998), the minimum legal drinking age (MLDA)
(DiNardo and Lemieux, 2001; Yörük and Yörük, 2011, 2013; Crost
and Guerrero, 2012) composite market prices of alcohol (Saffer
and Chaloupka, 1999) and market prices of cocaine (Saffer and
Chaloupka, 1999; DeSimone and Farrelly, 2003) to tease out the
exogenous changes in the use of alcohol or cocaine as well as the
downstream use of marijuana. Although they generally find a direct
policy/price effect on the use of the target substance itself (e.g., alco-
hol and cocaine) that follows a downward sloping demand curve,
the downstream effect on marijuana use is mixed. Chaloupka and
Laixuthai (1997), DiNardo and Lemieux (2001), Crost and Guerrero
(2012), and Crost and Rees (2013) find evidence for a substitution
between marijuana and alcohol. However, Pacula (1998), Saffer
and Chaloupka (1999), and Yörük and Yörük (2011) find evidence
supporting the complementarity hypothesis between marijuana
and alcohol. Moreover, evidence from Saffer and Chaloupka (1999)
and DeSimone and Farrelly (2003) suggests a complementarity
between marijuana and cocaine.
Not only is there a lack of consistent evidence, it is also dif-
ficult to extrapolate the effect of an MML on the use of other
substances from the estimated reduced-form effect of policy/price
related to the other substances on the use of marijuana. This diffi-
culty arises out of the nature of the underlying Marshallian demand
function, which does not require symmetric relationships between
substances (i.e., from substance A to B vs. from substance B to A), nor
does it require symmetric responses to policy/price changes (i.e.,
permissive policy/lower price vs. restrictive policy/higher price).
Thus it is possible for marijuana to be a substitute for alcohol when
alcohol regulations become more restrictive but for alcohol be a
complement to marijuana when marijuana policies become more
permissive.8
8
This asymmetric relationship between marijuana use and alcohol use may come
into play in the context of the minimum legal drinking age (MLDA) vs. an MML: a
teenager under the MLDA cannot legally acquire either alcohol or marijuana and
may resort to illegal supply channels, whereas an experienced marijuana user living
in a MML state can get both marijuana and alcohol with little effort. In essence, when
identifying the relationship between marijuana and alcohol, using different policies
may capture the decisions made by different groups from different choices set. Thus,
the results from one policy may not applicable to another policy setting.
67
Within the context of MMLs, Anderson et al. (2013) provide evi-
dence that states with MMLs see a reduction in alcohol-related
traffic fatalities, alcohol consumption and beer sales. However, the
authors do not have data on changes in marijuana use, thus their
findings do not necessarily imply that marijuana is a substitute for
(or a complement to) alcohol. In fact, when taking into account
the key provisions of MMLs, the replication study by Pacula et al.
(2013) suggests that the findings from traffic fatalities and alcohol
consumption are more consistent with a complementarity hypoth-
esis. Nonetheless, the authors are only able to assess two outcomes
related to alcohol consumption, which limits the scope of their
study.9
Another piece of evidence in the context of MMLs comes from
Bachhuber et al. (2014), which assesses the mortality rate related
to opioid overdose. The authors find a 25 percent reduction in the
annual rate of opioid overdose mortality between 1999 and 2010 in
states with MMLs compared to those without such laws. However,
the unaccounted state heterogeneity in the underlying prevalence
of opioid use or trajectory of overdose deaths may also contribute
to the reduced mortality rate. Therefore, the reduction in opioid
overdose mortality rate may not necessarily imply a substitution
between marijuana and opioids.
In sum, the majority of the literature on the relationship
between marijuana use and the use of alcohol and other substances
relies on policy/price shocks other than MMLs for identification.
Evidence from this line of literature is inconsistent and may not
extrapolate to the effect of an MML. Existing literature in the con-
text of MML, however, is relatively thin and limited in scope and
rigor.
1.5. Significance of our study
To inform the current debate on MMLs and marijuana lib-
eralization policies in general, we examine the effect of state
implementation of MMLs between 2004 and 2012 on marijuana
use, alcohol use, pain medication misuse, and hard drug use in both
adolescent and adult populations. Our study advances the existing
literature by: (i) providing one of the first estimates of the effect of
MML implementation on adult marijuana use based on micro-level
nationally-representative data, as well as the updated estimates for
adolescent marijuana use based on the most recent data; (ii) esti-
mating the effect of MML implementation on a variety of substance
use outcomes with differential elasticities and expected harms; (iii)
estimating the contemporaneous relationship between marijuana
and alcohol and other substances within the context of MMLs; (iv)
estimating explicitly the heterogeneous policy effects of key MML
provisions between different age groups.
2. Methods
2.1. Data sources
We pooled nine years of cross-sectional data from a restricted-
access version of the National Survey on Drug Use and Health
(NSDUH) 2004–2012 (CBHSQ, 2013). NSDUH is a nationally and
9
The first outcome in Pacula et al. (2013), any current alcohol use, may not carry
as much weight as binge or heavy drinking in terms of health consequences and
policy implications, especially for adults of legal drinking age. The other outcome,
specialty alcohol abuse treatment admissions, may not show a clear picture of the
alcohol abuse/dependence prevalence, since more than 90 percent of Americans
who suffer from alcohol abuse/dependence do not receive any treatment for their
conditions. Furthermore, a large proportion of those receiving the treatment only
receive it in a self-help group (e.g., Alcoholics Anonymous) or in a primary care
setting as opposed to a specialty alcohol abuse treatment setting (SAMHSA, 2013).
68 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80
state-representative10 survey sponsored by the Substance Abuse
and Mental Health Services Administration (SAMHSA), and the
primary source of information on substance use behavior by
the U.S. civilian, noninstitutionalized11 population aged 12 or
above. The majority of the NSDUH interview is conducted by self-
administrated audio computer-assisted self-interviewing (ACASI),
a highly private and confidential mode that encourages honest
reporting of substance use and other sensitive behaviors (Johnson
et al., 2010). The response rates range from 73 percent to 76 percent
between 2004 and 2012.
2.2. Variable measurement
2.2.1. Marijuana use outcomes
We created five outcomes related to marijuana use: (i) a
dichotomous indicator assessing whether a respondent used mar-
ijuana during the past month prior to the interview; (ii) another
dichotomous indicator assessing whether a respondent used mar-
ijuana “almost daily or daily”, defined as more than 20 days of
marijuana use during the past month; (iii) the number of mari-
juana use days among past-month marijuana users, which is an
conditional frequency ranging from 1 to 3012 ; (iv) a dichotomous
indicator for using marijuana for the first time during the past
year13 ; and (v) a dichotomous indicator for being classified as abus-
ing or being dependent on marijuana during the past year according
to DSM-IV diagnostic criteria. The DSM-IV defines past-year sub-
stance abuse/dependence as a maladaptive pattern of substance
use leading to clinically significant impairment and distress during
the past year. The impairment and distress related to substance use
can be manifested by symptoms such as tolerance, withdrawal, use
of a substance in a larger amount or over a longer period of time
than intended, continued substance use in dangerous situations,
interference with major obligations, etc. (APA, 2000) (Appendix 1).
2.2.2. Alcohol use outcomes
Empirical evidence suggests that marijuana can be a substi-
tute for and a complement of alcohol, depending on individual
motives of substance use and doses of consumption. Lower-dose
alcohol consumption for mild happiness and relaxation is hypoth-
esized to be replaced by marijuana use (King et al., 2011), whereas
higher-dose alcohol consumption for intense euphoria is hypoth-
esized to be accompanied by marijuana use (Lukas and Orozco,
2001). In this regard, we studied any alcohol use as well as binge
drinking.14 Binge drinking, in the NSDUH, is defined as having
five or more drinks on the same occasion on at least one day
10
The NSDUH sampling frame is state-based, with an independent, multistage
area probability sample within each state and the District of Columbia. The eight
states with the largest population (i.e., California, Florida, Illinois, Michigan, New
York, Ohio, Pennsylvania, and Texas) have an annual sample size of about 3600
each. For the remaining 42 states and the District of Columbia, each has a sample
size of about 900 annually.
11
Institutionalized individuals (e.g. in jails/prisons or hospitals), homeless or tran-
sient persons not in shelters, and military personnel on active duty were excluded
from the NSDUH sample.
12
The majority of past-month marijuana users either use marijuana on a few occa-
sions or use it regularly, with a very small proportion of marijuana users between
these two extremes. Therefore, we assessed both the average change in the fre-
quency of marijuana use days and the change in the right tail of the frequency
distribution (i.e., almost daily/daily marijuana use).
13
Marijuana use initiation is examined in an “at-risk” sample, which excludes
those who first tried marijuana more than a year prior to the interview thus no
longer at risk of initiating marijuana use during the preceding year.
14
Carpenter and Dobkin (2009), for instance, find evidence for the differential elas-
ticity of alcohol demand along the distribution of drinking intensity and frequency.
during the past month.15 We created the following measures for
alcohol use: (i) the total amount of drinks consumed during the
past month,16 (ii) the unconditional frequency of binge drinking
days, and (iii) the probability of being classified as having alcohol
abuse/dependence during the past year according to the DSM-IV
criteria. We also created two dichotomous indicators to assess: (iv)
whether a respondent engaged both in marijuana use and in binge
drinking during the past month, and (v) whether a respondent used
marijuana while drinking alcohol (i.e., on the same occasion) dur-
ing the past month.17 These two measure of simultaneous use of
marijuana and alcohol can provide further insight into the contem-
poraneous complementarity between the two substances.
2.2.3. Other substance use outcomes
In light of the previous evidence suggesting a substitution
between marijuana and opioids (Bachhuber et al., 2014) and
a complementarity between marijuana and cocaine (Saffer and
Chaloupka, 1999; DeSimone and Farrelly, 2003), we focused our
analysis on non-medically used prescription pain medication,18
heroin, and cocaine. NSDUH defines “non-medical use” as the
intentional use of a medication without a prescription, in a way
other than as prescribed, or simply for the experience or feeling that
it causes. NSDUH does not include questions about legitimate pain
medication used according to the prescription. We created three
dichotomous indicators for the probability of: (i) past-year non-
medical use of prescription pain medication, (ii) past-year heroin
use, and (iii) past-year cocaine use.
2.2.4. MML-implementation indicator
The recent launch of the Data Portal system by the CBHSQ pro-
vides us with access to state identifiers and interview dates in
a restricted-access version of the NSDUH micro-level data, thus
enabling us to create a dichotomous indicator for the imple-
mentation of a MML in a given state during a given period. As
summarized in Table 1, between 2004 and 2012, MMLs came
into effect in ten states at various time points. We assigned the
MML-implementation indicator a value of 1 for each full month
subsequent to the effective date of the laws, and a value of 0 for
the remaining periods and for the control states.19 Control states
include eight states that had an MML in place prior to 2004 (i.e.,
“always MML states”) and those that did not have any MML by the
end of 2012 (i.e., “no MML states”).20
15
A commonly used alternative defines “binge drinking” as five or more drinks
for men and four or more drinks for women consumed on one occasion (Wechsler
et al., 1995). Our estimates are robust to this gender-specific definition (not shown).
16
One drink refers to a can or a bottle of beer, a glass of wine or a wine cooler, a
shot of liquor, or a mixed drink with liquor in it.
17
The question about simultaneous use of marijuana and alcohol is not included
in the NSDUH 2004 and 2005 surveys, while the MMLs in Vermont and Montana
both came into effective in 2004. Thus we cannot estimate the effect of these two
states’ implementation of the MMLs on this outcome.
18
NSDUH attempts to capture all types of pain medication by including in its
questionnaire a list of commonly prescribed and misused pain medications in
their generic names (e.g., Codeine, Oxycodone, Hydrocodone, Morphine, Hydromor-
phone, Fentanyl, Tramadol, etc.), brand names (e.g., OxyContin, Vicodin, MSContin,
Dilaudid, Duragestic, Ultram, etc.) and street names, along with an open-ended
question about other pain medications.
19
Note that most previous studies based on annual surveys were only able to
estimate year-on-year policy effect. In our study, we linked the NSDUH interview
dates with the MML effective dates and matched the month-to-month implemen-
tation window of the MMLs with the behavior window of the NSDUH respondents.
This approach minimizes the potential measurement error from misclassification
of pre-MML and post-MML behaviors.
20
We also estimated two alternative model specifications: the first specification
classifies the “always MML states” as the control states, whereas the second specifi-
cation classifies the “no MML states” as the control states. As shown in Appendix 2,
 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80 69

Table 1
Implementation and key provisions of state medical marijuana laws (MMLs).

Approved date Effective date Key statutory provisions

Non-specific pain Patient registry Retail dispensarya Home cultivation

2004–2012 (10 states)

Vermont 2004/05 2004/07 2007/07 2004/07 n/a 2004/07
Montana 2004/11 2004/11 2004/11 n/a n/a 2004/11
Rhode Island 2005/06 2006/01 2006/01 2006/01 2009/07 2006/01
New Mexico 2007/03 2007/07 n/a 2007/07 2007/07 2007/07
Michigan 2008/11 2008/12 2008/12 n/a n/a 2008/12
New Jersey 2010/01 2010/10b 2010/10 2010/10 2010/10 n/a
District of Columbia 2010/05 2010/07 n/a 2010/07 2010/07 n/a
Arizona 2010/11 2011/04 2011/04 2011/04 2011/04 2011/04
Delaware 2011/05 2011/07 2011/07 2011/07 2011/07 n/a
Connecticut 2012/05 2012/05c n/a 2012/05 n/a n/a

1996–2003 (8 states)
California 1996/11 1996/11 1996/11 n/a 1996/11 1996/11
Washington 1998/11 1998/11 1998/11 n/a n/a n/a
Oregon 1998/11 1998/12 1998/12 2007/01 n/a 1998/12
Alaska 1998/11 1999/03 1999/03 1999/03 n/a 1999/03
Maine 1999/11 1999/12 n/a 2009/12 2009/12 1999/12
Hawaii 2000/06 2000/12 2000/12 2000/12 n/a 2000/12
Colorado 2000/11 2001/06 2001/06 2001/06 2001/06 2001/06
Nevada 2000/11 2001/10 2001/10 2001/10 n/a 2001/10

Note:
Maryland passed two laws in 2003 and in 2011 favorable to medical marijuana, albeit not legalizing it.
a
Despite the allowance for retail medical marijuana dispensary under the laws, only four states actually opened their first dispensaries between 2004 and 2012, including
Colorado (2005/07), New Mexico (2009/06), Maine (2011/04), and New Jersey (2012/12).
b
The effective date of New Jersey MML is 2010/07 as specified in the statute, while the state governor Chris Christie delays its implementation.
c
Most sections of Connecticut MML came into effect from its passage (2012/05), while a few sections on 2012/10.

In addition to examining the effect of implementation of an
MML as a whole, Pacula et al. (2013) recognize the importance of
scrutinizing the potential heterogeneous effects between individ-
ual components of an MML. As highlighted in their study, four key
components that may be included in an MML and lead to hetero-
geneity in the policy effect are: (i) “non-specific pain” provision,
which lists a generic “chronic pain” in the eligible conditions for
medical marijuana, rather than specifying diseases causing the
pain; (ii) “patient registry” provision, which requires a patient
registry/renewal system; (iii) “retail dispensary” provision, which
allows licensed marijuana retailors to dispense marijuana legally
to medical marijuana patients; and (iv) “home cultivation” pro-
vision, which allows qualified patients and caregivers to grow a
certain amount of marijuana plants indoors for the patients’ own
medical use. Accordingly, we created four indicators each rep-
resenting the inclusion of a key MML provision. Note that for
an MML state, the inclusion date of a MML provision may differ
from the effective date of the MML, as the state may include the
provision in the original statute, add it in a subsequent amend-
ment, or not include it in the law until the end of the study
period.
2.2.5. Covariates
We controlled for individual-level and state-level factors that
are correlated with both the individual choice to use substances
and with state decisions about MMLs. Individual-level covariates
for adolescents and adults include a rich set of sociodemographic
characteristics. State-level covariates include three time-varying
measures reflecting the fluctuation in state economic conditions:
(i) unemployment rate, (ii) average personal income, and (iii)
median household income of the state, as well as two additional
measures reflecting relevant changes in state policy environment.
the estimated policy effects on the main outcomes are very similar across the mod-
els.
One major policy change during the study period concerns state
implementation of beer taxes.21 The other policy change is mari-
juana decriminalization/depenalization: Massachusetts, California,
and several cities and counties in other states relaxed penalties for
recreational marijuana use or placed it “the lowest law enforce-
ment priority.” We therefore created a dichotomous indicator for
the implementation of a decriminalization/depenalization policy
in a given state during a given month.22 Table 2 provides descrip-
tive summary for the individual-level and state-level covariates
discussed above.
2.3. Identification strategy
To identify the effect of MML implementation on individual
marijuana use, alcohol use, pain medication misuse, and hard drug
use, we estimated the following two-way fixed effects models:
Y ist = ˇ0 + ˇ1 MML st + ˇ2 X 1ist + ˇ3 X 2st + s + t + s t + εist (1)
where i denotes an individual, s denotes the state, and t denotes the
year. Yist represents the substance use outcomes. MMLst is the pol-
icy indicator for the implementation of an MML in a state s during
a year t. X1ist is the full vector of individual-level covariates. X2st is
the full vector of state-level covariates. The two-way fixed effects
are captured in our models by s and  t to account for the time-
invariant state heterogeneity as well as the national secular trend
21
We did not control for the market price of heroin or cocaine. The most com-
monly used source is the U.S. Drug Enforcement Administration’s System to Retrieve
Information from Drug Evidence (STRIDE) data set. Empirical studies often find that
STRIDE prices are not predictive or only weakly predictive of drug use (Horowitz,
2001). As French and Popovici (2011) pointed out, “part of difficulty here is that con-
ventional prices for illicit drug are not readily available and alternative measures
are not yet found.” Nonetheless, fluctuations in heroin prices and cocaine prices
are unlikely to be correlated with the MML implementation, thus omitting these
variables is unlikely to bias our results.
22
For lack of policy variations during the study period, the effect of a decriminal-
ization/depenalization policy itself cannot be precisely estimated.
70 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80

Table 2
Descriptive summary of individual- and state-level covariates, sampling-weight adjusted.

Age 12–20 Age 21+

MML states No and always MML states MML states No and always MML states

Mean S.D. Mean S.D. Mean S.D. Mean S.D.

Panel A: individual-level covariates

# Age 16.0 (2.53) 16.0 (2.56) 48.0 (16.8) 47.5 (16.8)
% Male 51.4 (50.0) 51.2 (50.0) 48.1 (50.0) 48.0 (50.0)
Race/Ethnicity: Non-Hispanic White (ref.)
% Hispanic/Latino 18.8 (39.1) 19.1 (39.3) 13.1 (33.8) 13.3 (33.9)
% Non-Hispanic African Black 12.8 (33.4) 15.0 (35.7) 10.0 (29.8) 11.5 (31.9)
% Non-Hispanic Asian 3.46 (18.3) 4.37 (20.5) 3.63 (18.7) 4.56 (20.8)
% Other Origins 4.15 (19.9) 2.71 (16.2) 2.43 (15.4) 1.87 (13.6)
Self-Reported Health: Excellent (ref.)
% Very Good 42.6 (49.5) 41.5 (49.3) 36.5 (48.1) 35.9 (48.0)
% Good 21.0 (40.7) 21.4 (41.0) 27.1 (44.4) 27.7 (44.8)
% Fair/Poor 4.26 (20.2) 4.15 (19.9) 12.8 (33.5) 13.9 (34.6)
Cigarette Smoking: Non-Smoker (ref.)
% Non-Daily Smoker 11.3 (31.7) 11.2 (31.5) 9.01 (28.6) 8.75 (28.3)
% Daily Smoker 6.75 (25.1) 6.36 (24.4) 15.8 (36.4) 16.1 (36.7)
Health Insurance: Uninsured (ref.)
% Private Health Insurance 65.9 (47.4) 60.9 (48.8) 71.7 (45.0) 68.3 (46.5)
% Medicaid 22.9 (42.0) 24.5 (43.0) 8.59 (28.1) 8.14 (27.3)
% Other Health Insurance 3.15 (17.5) 4.02 (19.6) 7.82 (26.8) 8.84 (28.4)
Family Income: >200% FPL (ref.)
% Living 100–200% FPL 19.8 (39.9) 22.9 (42.0) 17.0 (37.6) 19.3 (39.5)
% Living <100% FPL 19.4 (39.6) 22.1 (41.5) 10.3 (30.4) 12.0 (32.5)
Urbanicity: Non-CBSA (ref.)
% Living in a Micropolitan 8.52 (27.9) 10.0 (30.0) 8.29 (27.6) 10.1 (30.0)
% Living in a Metropolitan 87.6 (33.0) 83.8 (36.8) 87.6 (33.0) 83.3 (37.3)
Marital Status: Married (ref.)
% Never Married 22.5 (41.8) 21.3 (41.0)
% Separated/Divorces 6.91 (25.4) 6.42 (24.5)
% Widowed 13.7 (34.3) 14.4 (35.1)
Education Attainment: College Graduate (ref.)
% Some College 24.9 (43.2) 25.5 (43.6)
% High School Graduate 29.3 (45.5) 29.9 (45.8)
% Less than High School 13.5 (34.2) 15.3 (36.0)
College Enrollment: Not Enrolled (ref.)
% Part-Time Enrolled 3.81 (19.1) 3.31 (17.9)
% Full-Time Enrolled 3.86 (19.3) 4.33 (20.4)
Employment: Full-Time Employed (ref.)
% Part-Time Employed 12.6 (33.2) 12.3 (32.8)
% Unemployed 4.49 (20.7) 4.15 (20.0)
% Not in Labor Force 29.0 (45.4) 29.2 (45.5)

Panel B: state-level covariates

% Unemployment Rate 7.36 (2.60) 6.84 (2.36) 7.36 (2.60) 6.86 (2.36)
$ Average Personal Income (10 K) 4.05 (0.90) 3.86 (0.54) 4.07 (0.90) 3.87 (0.54)
$ Median Household Income (10 K) 5.64 (0.89) 5.25 (0.67) 5.65 (0.90) 5.24 (0.68)
Beer Tax Rates
$ Specific Excise Tax (per gallon) 0.18 (0.07) 0.28 (0.25) 0.18 (0.07) 0.28 (0.25)
% Ad Valorem Tax (on-premises) 0.35 (1.84) 0.76 (3.05) 0.36 (1.88) 0.79 (3.10)
% Ad Valorem Tax (off-premises) 0.14 (1.06) 0.73 (2.97) 0.15 (1.15) 0.75 (3.01)

Number of observations ≈46,700 ≈222,800 ≈56,400 ≈267,500

and common shocks related to substance use. We also included
state-specific linear time trends s t to account for the unobserved
state-level factors that evolve over time at a constant rate (e.g.,
social norms and public sentiments related to substance use).
Standard errors were clustered at the state level to correct
for the serial correlation. The clustered standard errors allow for
arbitrary within-state correlation in error terms but assume inde-
pendence across the states (Bertrand et al., 2004).23
We stratified the sample into two age groups, adolescents and
young adults aged 12–20 (N ≈ 269,500) and adults aged 21 or above
(N ≈ 323,900). We chose age 21 as the cut-off point in light of the
previous evidence of an age 21 discontinuity in both alcohol use
and marijuana use (Crost and Guerrero, 2012; Yörük and Yörük,
2011, 2013). We tested four cut-off points in our analyses, age 18,
age 21, age 25 and age 30. Only the age 21 stratification, which

of the NSDUH sampling design would suppress the state-clustering adjustment.

23
It is worth noting that NSDUH employs a multistage (stratified cluster) design
for the sample selection. The sampling design elements include survey weights, vari-
ance estimation cluster replicates and variance estimation stratum. The descriptive
statistics were adjusted for these survey design elements to make the analytic sam-
ple representative of the U.S. population. However in regression analysis, using the
STATA “svy” procedure to adjust for the weighting, clustering, and stratification
When considering the choice between the two, Solon et al. (2013) noted that theo-
retically “neither strictly dominates the other (in identifying the population average
effect)” (Solon et al., 2013, p. 21). Furthermore, in our study, the results from both the
unweighted, state-clustering adjusted models and the weighted, sampling-design
adjusted were similar (Appendix 3). Therefore, we report the unweighted, state-
clustering adjusted estimates.
 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80
also coincides with the legal drinking age, produces significant and
meaningful differences in the estimated policy effect between age
groups.
We estimated Probit regressions for the dichotomous depend-
ent variables in our study. The other three discrete dependent
variables we study (i.e., the conditional frequency of marijuana
use days, the number of alcohol drinks, and the unconditional fre-
quency of binge drinking days) possess positive skewness and/or
“excess zeroes” compared to a standard normal distribution, which
requires a more flexible estimation approach than an ordinary least
squares (OLS) estimation. A generalized linear model (GLM) with
a gamma distribution and log link24 was estimated for the total
amount of drinks during the past month among those aged 21 or
above. For the total amount of drinks among those aged 12–20,
on the other hand, we estimated a two-part model using Pro-
bit in the first part and GLM (gamma distribution and log link)
in the second part. Because there is an explicit decision process
regarding legality of alcohol consumption among those under 21,
we use the TPM to model the decision to engage in underage
drinking and the quantity consumed conditional upon deciding
to engage in underage drinking as separate processes. We fol-
lowed the same logic when estimating the frequency variables.
Considering the underlying decision processes and the propor-
tions of zero values, we estimated a zero-truncated negative
binomial regression25 for the conditional frequency of marijuana
use days and a zero-inflated negative binomial regression26 for
the unconditional frequency of binge drinking days in both age
groups.
For ease of interpretation, we converted the coefficient of MMLst
in each of the estimations to the average marginal effect calculated
at MMLst = 0 and the observed values of other covariates.
3. Results
3.1. Estimated effect of MML implementation on marijuana use
Fig. 1 shows an upward trend in past-month marijuana use rates
among adults aged 21 or above in parallel with the implementation
of MMLs. A relative increase in adult marijuana use in MML states
emerges immediately after the laws take effect, and persists at
24
The selection of distribution family under the GLM was made based on the
modified Park test results.
25
The likelihood ratio test for overdispersion rejects a Poisson distribution in favor
of a binomial distribution.
26
The likelihood ratio tests for overdispersion reject a Poisson distribution in favor
of a binomial distribution. Furthermore, the Vuong tests for zero-inflation confirm
our choice of a zero-inflated model instead of an ordinary negative binomial model.
The zero-inflated Poisson/negative binomial model assumes that the sample con-
sists of two distinct groups of people: one group whose counts are generated by the
standard Poisson/negative binomial model, and the other group, so-called “abso-
lute zero” group, who have zero probability of a count greater than zero; observed
zeroes can come from either group (Greene, 2011; Wang, 2003). The absolute zero
group, in our case, may be those who abstain from alcohol for religious, cultural,
familial or other reasons. Thus, this group of people, as distinct from the majority
of people who drink alcohol at least occasionally, have “absolute zero” risk of binge
drinking.
An alternative to a zero-inflated regression is a hurdle model (i.e., a TPM for counts)
with first-part Probit and second-part zero-truncated negative binomial. A practical
challenge, however, is that cluster-adjusted standard errors are difficult to compute
when combining the first- and second-part estimates from a hurdle model (Belotti
et al., 2014). Nonetheless, the point estimates for the combined effects we obtained
from the hurdle models (not shown) were very similar to the zero-inflated negative
binomial estimates from our main analyses. In another set of sensitivity analyses, we
also treated the count variables as continuous and estimated the combined marginal
effects and their cluster-adjusted standard errors using the STATA command “TPM”
(Belotti et al., 2014). The TPM estimates (not shown) were slightly larger and more
significant than the zero-inflated negative binomial estimates.
71
least three years afterwards. Among adolescents and young adults
aged 12–20, however, the corresponding trend in past-month mar-
ijuana use rates is not consistent. Bear in mind that the relative
trends shown in Fig. 1 are equivalent to unadjusted DD estimates
that only partial out the two-way fixed effects (i.e., time-invariant
state heterogeneity and national secular trend in past-month mar-
ijuana use), but do not adjust for the individual- and state-level
covariates or state-specific linear trends. Nonetheless, this observa-
tional trend-comparison suggests a potential association between
MML implementation and increased current marijuana use among
adults aged 21 or above, but not among adolescents and younger
adults.
Table 3 presents the marginal effects of MML implementation
on the four marijuana use outcomes, adjusted for the two-way fixed
effects, the full vector of individual- and state-level covariates, and
the state-specific linear trends. Among adults aged 21 or above,
the implementation of an MML increases the probability of using
marijuana during the past month by 1.32 percentage points (Panel
B, Column 1, Row 1). This percentage point change can be trans-
lated into a 14 percent relative increase from a baseline predicted
marijuana use probability of 9.33 percentage points.
The NSDUH data do not allow us to distinguish between medical
marijuana patients and the non-patient population. Nonetheless,
according to the registry data (Anderson et al., 2013), the number
of registered medical marijuana patients accounts for an average of
0.8 percent of the population across the five MML states on which
the registry information is available. Therefore, the 1.3 percentage
point increase in the probability of marijuana use we find among
adults aged 21 or above is not likely to come exclusively from
an increase in use among registered patients. Though we cannot
test this directly, it suggests that there may also be a considerable
spillover effect of MML implementation on recreational marijuana
use or self-medication by the non-patient population.
Among adults aged 21 or above, we also find a 0.58 percent point
or a 15 percent increase in the probability of almost daily/daily
marijuana use (Panel B, Column 2, Row 1) attributable to MML
implementation. Among adolescents and young adults aged 12–20,
in contrast, no change in the probability or frequency of past-month
marijuana use can be attributed to MML implementation (Panel A,
Columns 1–3).
With regard to marijuana use initiation during the preceding
year, MML implementation leads to 0.32 percentage point or a
5 percent increase in the probability of first-time marijuana use
among adolescents and young adults aged 12–20 (Panel A, Col-
umn 4, Row 1). Yet, the lack of a policy effect on the probability
and frequency of past-month marijuana use among this age group
suggests that many of these young people may be engaging in
experimental use with relatively low health, behavioral, and social
consequences. In other words, these findings are consistent with a
scenario in which adolescents and young adults aged 12–20 who
experiment with marijuana use in response to an MML are not
transitioning to regular use, at least in the short term.
In contrast to the findings among adolescents and younger
adults, we find no change in marijuana use initiation among those
aged 21 or above (Panel B, Column 4) as a result of MML imple-
mentation, despite the aforementioned significant increases in any
past-month marijuana use and almost daily/daily use (Panel B,
Columns 1 and 2). These findings suggest that the adults who
respond to an MML by increasing current and regular use come
largely from those who first tried marijuana long before its medi-
cal use was permitted. After the introduction of an MML that helped
reduce costs of marijuana use (i.e., market prices as well as non-
market health, legal and social consequences), those with prior
marijuana use experience would likely reinitiate or increase their
marijuana use.
72 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80

Fig. 1. Pre- and post-trend in past-month marijuana use rates in medical marijuana law (MML) states relative to the control states. Note: The differences in past-month
marijuana use rate are equivalent to unadjusted difference-in-differences (DD) estimates that partial out the two-way fixed effects, but not adjust for individual- and state-
level covariates or state-specific linear trends. The time 0 is centered at the period when each medical marijuana law (MML) state started to implement its law, so the time
1 represents the first full month subsequent to the effective date of an MML. We calculate the differences between each of the MML state and the control states during each
month, and average them across all 10 MML states and over a 3-month period (to smooth the fluctuations in the monthly rate). Whiskers indicate 95% confidence intervals.

3.2. Estimated effect of MML implementation on alcohol use
To the extent that alcohol is a complement or substitute to mar-
ijuana, the effect of MML implementation on marijuana use may
spread to alcohol use (Table 4). Our estimates indicate that, among
adults aged 21 or above, MML implementation is not associated
with the total number of drinks (Panel B, Column 1), but positively
associated with the frequency of binge drinking. Our estimates
indicate an effect size of 0.16 more binge drinking days or a relative
increase of 10 percent (Panel B, Column 2, Row 1). The spillover
increase in binge drinking implies a complementary relationship
between marijuana use and high-dose alcohol consumption among
adults aged 21 or above. Not only is this contemporaneous com-
plementarity reflected in the independent measures of marijuana
use and binge drinking, it is further confirmed by the measure of
simultaneous use of the two substances. Among adults aged 21 or
above, we find a 1.44 percentage point or a 22 percent increase in
the probability of both marijuana use and binge drinking during
the past month (Panel B, Column 3, Row 1) and a 0.82 percent-
age point or a 18 percent increase in the probability of marijuana
use while drinking (i.e., in the same occasion) as a result of MML
implementation (Panel B, Column 4, Row 1).
Among adolescents and young adults aged 12–20, we find no
significant change in any measure of alcohol use (Panel A), which
suggests that the increased marijuana use initiation we reported
previously is unlikely to spread to underage drinking.
3.3. Immediate and delayed effect of MML implementation on
other downstream outcomes
In addition to marijuana use and binge drinking, MML
implementation may have a spillover effect on marijuana
abuse/dependence, alcohol abuse/dependence, non-medical use of
prescription pain medication, and the use of hard drugs such as
heroin and cocaine. The progression from marijuana use and binge
drinking to these downstream outcomes may be a gradual transi-
tion (Wagner and Anthony, 2002b). As such, we estimated not only
the contemporary policy effect but also the one-year and two-year
lagged policy effect (Table 5).
The effect arguably most salient to the public health impli-
cations of MMLs is the effect on marijuana abuse/dependence
among adults aged 21 or above. We found a delayed policy effect
on increasing the probability of marijuana abuse/dependence by
a relative 10 percent (Panel B, Column 1, Rows 2 and 3). The
 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80 73

Table 3
Estimated marginal effect of implementation and provisions of medical marijuana laws (MMLs) on marijuana use.

Marijuana use outcomes (1) (2) (3) (4)

% past-month % marijuana Cond. # marijuana % marijuana use
marijuana use daily/almost daily use use days initiation

Panel A: age 12–20

MML implementation −0.43 (0.48) −0.25 (0.17) −0.28 (0.45) 0.32** (0.16)
MML provisions
∼ Non-specific pain −0.05 (0.41) −0.46 (0.28) −0.74 (0.44) 0.43 (0.26)
∼ Patient registry −0.74 (0.63) −0.14 (0.27) 0.28 (0.36) 0.04 (0.23)
∼ Retail dispensary 0.89** (0.34) −0.20 (0.36) −0.46 (0.59) 0.45 (0.33)
∼ Home cultivation 0.12 (0.61) 0.43 (0.27) 0.93 (0.56) 0.18 (0.24)

Baseline predicted mean [10.68] [3.52] [12.29] [6.47]

Number of observations ≈269,500 ≈269,500 ≈28,600 ≈213,900

Panel B: age 21+

MML Implementation 1.32** (0.58) 0.58** (0.26) 0.17 (0.64) 0.15 (0.23)
MML provisions
∼ Non-specific pain 1.56** (0.73) 0.86** (0.42) 0.28 (0.88) 0.31 (0.78)
∼ Patient registry −0.45 (0.73) −0.35 (0.52) 0.55 (0.76) −0.05 (0.44)
∼ Retail dispensary −0.12 (0.79) −0.09 (0.64) −0.67 (0.85) 0.07 (0.53)
∼ Home cultivation 0.55 (0.76) −0.10 (0.41) −0.48 (0.76) 0.02 (0.77)

Baseline predicted mean [9.33] [3.78] [14.15] [0.92]

Number of observations ≈323,900 ≈323,900 ≈30,500 ≈161,400

Note:
Standard errors in parentheses are clustered at the state level.
Baseline predicted mean is calculated as the average of predicted probabilities/counts when setting MMLst to 0 and leaving the other covariates as the observed values.
*Significant at the 10 percent level.
**
Significant at the 5 percent level.
***Significant at the 1 percent level.

Table 4
Estimated marginal effect of implementation and provisions of medical marijuana laws (MMLs) on alcohol use.

Alcohol use outcomes (1) (2) (3) (4)

# past-month total # binge drinking % marijuana use % marijuana use
alcohol drinks days and binge drinking while drinking

Panel A: age 12–20

MML implementation −0.03 (1.74) 0.04 (0.18) −0.63 (0.39) −0.38 (0.49)
MML provisions
∼ Non-specific pain −0.54 (2.86) 0.03 (0.03) −0.59 (0.40) 0.07 (0.67)
∼ Patient registry −0.53 (2.51) −0.05 (0.04) −0.39 (0.80) −0.42 (0.63)
∼ Retail dispensary 0.65 (2.04) 0.06 (0.06) 0.54 (0.47) 0.87 (0.74)
∼ Home cultivation 0.52 (3.26) −0.01 (0.04) −0.03 (0.56) −0.65 (0.64)

Baseline predicted mean [7.76] [0.66] [6.41] [4.10]

Number of observations ≈269,500 ≈269,500 ≈269,500 ≈207,900

Panel B: age 21+

MML implementation 0.95 (1.18) 0.16** (0.08) 1.44*** (0.35) 0.82* (0.45)
MML provisions
∼ Non-specific pain 0.65 (1.56) 0.19* (0.10) 1.03** (0.49) 1.23** (0.59)
∼ Patient registry 0.27 (1.22) −0.17 (0.11) −0.10 (0.57) −0.33 (0.55)
∼ Retail dispensary 0.47 (1.18) 0.07 (0.06) 0.21 (0.58) −0.03 (0.66)
∼ Home cultivation 0.18 (1.52) 0.13 (0.09) 0.20 (0.62) −0.57 (0.74)

Baseline predicted mean [18.69] [1.52] [6.44] [4.45]

Number of observations ≈323,900 ≈323,900 ≈323,900 ≈250,500

Note:
Standard errors in parentheses are clustered at the state level.
Baseline predicted mean is calculated as the average of predicted probabilities/counts when setting MMLst to 0 and leaving the other covariates as the observed values.
*
Significant at the 10 percent level.
**
Significant at the 5 percent level.
***
Significant at the 1 percent level.

increase in marijuana abuse/dependence of such magnitude is of
concern. It suggests that those who used marijuana in response
to MML implementation are at high risk of progressing to abuse/
dependence.
For both age groups, we found neither an immediate nor a
delayed effect of MML implementation on other downstream out-
comes including alcohol abuse/dependence, non-medical use of
prescription pain medication, heroin use and cocaine use.
3.4. Policy heterogeneity between key MML provisions
Our main estimates, in essence, capture the average policy effect
across all ten MMLs implemented between 2004 and 2012. How-
ever, the policy effect of each of these laws may not necessarily
have the same magnitude or even the same direction. As noted by
Pacula et al. (2013), four key MML provisions, namely the ambiguity
in “non-specific pain”, the requirement for patient registry/renewal
74 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80

Table 5
Estimated immediate and delayed marginal effect of implementation of medical marijuana laws (MMLs) on marijuana abuse/dependence, alcohol abuse/dependence,
prescription pain medication misuse, cocaine use, and heroin use.

Downstream outcomes (1) (2) (3) (4) (5)

% marijuana % alcohol % prescription % cocaine use % heroin use
abuse/dependence abuse/dependence painkiller misuse

Panel A: age 12–20

MML contemporary −0.07 (0.34) −0.12 (0.52) −0.05 (0.22) 0.03 (0.14) 0.008 (0.06)
MML lags
∼ 1-Year lag −0.10 (0.24) −0.22 (0.54) 0.03 (0.35) 0.01 (0.23) −0.01 (0.07)
∼ 2-Year lag 0.03 (0.27) −0.28 (0.34) −0.08 (0.42) −0.01 (0.18) −0.05 (0.11)

Baseline predicted mean [4.59] [7.77] [8.26] [2.41] [0.26]

Number of observations ≈269,500 ≈269,500 ≈269,500 ≈269,500 ≈269,500

Panel B: age 21+

MML contemporary 0.19 (0.13) 0.65 (0.44) −0.02 (0.39) 0.06 (0.18) 0.01 (0.11)
MML lags
∼ 1-Year lag 0.25** (0.11) 0.37 (0.35) −0.05 (0.21) −0.11 (0.20) 0.007 (0.08)
∼ 2-Year lag 0.23* (0.12) 0.25 (0.55) −0.09 (0.14) −0.09 (0.21) 0.005 (0.09)

Baseline predicted mean [2.30] [10.87] [6.65] [3.28] [0.32]

Number of observations ≈323,900 ≈323,900 ≈323,900 ≈323,900 ≈323,900

Note:
Standard errors in parentheses are clustered at the state level.
Baseline predicted means in square brackets are calculated as the average of predicted probabilities/counts when setting MMLst to 0 and leaving the other covariates as the
observed values.
*
Significant at the 10 percent level.
**
Significant at the 5 percent level.
***Significant at the 1 percent level.

system, the allowance for retail dispensaries, and the permission
for home cultivation, may have different implications for peo-
ple’s marijuana use behavior. Specifically, the “patient registry”
provision may in effect reduce marijuana use in the general pop-
ulation. This protective effect of the “patient registry” provision,
however, can be offset by the effect of “retail dispensary” provi-
sion which increases marijuana use significantly. In contrast to
Pacula et al. (2013), our study finds no consistent protective or
offsetting effect in either provision (Tables 3 and 4, Panels A and
B, Rows 2–4). A plausible explanation is the discrepancy between
the time when a “patient registry” provision or a “retail dispen-
sary” provision was included a state’s MML and the time when the
state’s registry/renewal system or its legal dispensaries actually
began to operate (Anderson and Rees, 2014). Due to the contro-
versy and complexity surrounding its implementation, the time lag
between the effective date of a “retail dispensary” provision and
the actual opening of the first medical marijuana store may be par-
ticularly long.27 Although we find no consistent effect of “patient
registry” or “retail dispensary”, we observe a consistent and sig-
nificant effect of the “non-specific pain” provision on increasing
marijuana use, binge drinking and simultaneous use of marijuana
and alcohol among adults aged 21 or above. The observed effect
of “non-specific pain” provision suggests that including a generic
term “chronic pain” in the eligible conditions for medical marijuana
may extend the patient base to adults with less severe conditions
or possibly those who pretend to be pain patients. Nonetheless,
considering the limited policy variations across the four MML pro-
visions during our study period, the estimated individual effects of
these provisions should be interpreted with caution.28
3.5. Policy endogeneity of MML adoption
There is a geographic concentration of MMLs states that have
adopted MMLs are all in the West and Northeast. This geographic
similarity raises concern that there may be some past disturbances
in marijuana use in these regions leading to their adoption of
MMLs and not accounted for by the state fixed effects and the
state-specific linear trends. In other words, MML adoption may
be endogenous to marijuana use. To check for this potential pol-
icy endogeneity, specifications with a series of lagged and leading
indicators for adopting an MML were estimated for the probabil-
ity of past-month marijuana use (Table 6). We find that only the
contemporary and 6-month lagged policy indicators had signifi-
cant effects, and the indicators for approved but not implemented
MMLs and the 12-month policy lag had moderate albeit imprecisely
estimated effects. All the leads had small and statistically insignif-
icant effects (Panel B, Column 2). These estimates suggest that it
is in fact the policy shock from adopting an MML that drives the
changes in marijuana use, rather than some past disturbances in
marijuana use that drive the adoption of an MML.
3.6. State-aggregate effect of MML implementation
To further check the robustness of our individual-level esti-
mates with regard to serial correlation, we aggregated the data to
the state level and estimated the effect of MML implementation on
state-level prevalence rates of our main individual-level findings.29

27
Anderson and Rees (2014) pointed out that, for instance, Colorado included
a “retail dispensary” provision in its original MML effective in 2001, but medical
marijuana dispensaries did not become commonplace until 2009. Moreover, Maine
and Rhode Island added “retail dispensary” provisions to their MMLs in 2009, but
the first legal dispensary in Maine did not open until 2011 and the first Rhode Island
dispensary did not open until 2013.
28
From a statistical standpoint, a substantial policy effect from one or two states
could potentially account for the overall findings. We tested for the heterogeneous
policy effect between states by replacing the single indicator for MML implementa-
tion with ten separate indicators for MML implementation in each of the MML states.
We find, in most cases, across-the-board significant policy effects in the same direc-
tion, albeit with varied effect sizes (Appendix 4). We cannot come to a conclusion,
therefore, as to whether the heterogeneous policy effect comes from states’ unique
experiences with implementing the MMLs or their inclusion/exclusion of certain
provisions.
29
In Columns 1 and 3 of Table 7, we clustered the standard errors at the state
level; while in Columns 2 and 4, we removed the time-series information from the
standard errors by averaging the pre-MML data and the post-MML data (Donald and
 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80 75

Table 6
Robustness check for policy endogeneity by including policy leads and lags.

(1) (2)
% past-month marijuana use % past-month marijuana use

Panel A: age 12–20

MML contemporary −0.43 (0.48) −0.81 (0.63)
MML leads and lags
∼ 24-Month lead (before approval) 0.27 (0.37)
∼ 18-Month lead −0.25 (0.40)
∼ 12-Month lead 0.14 (0.93)
∼ 6-Month lead 0.57 (0.82)
∼ Approved NOT implemented 0.62 (0.69)
∼ 6-Month lag (after implementation) −0.04 (0.34)
∼ 12-Month lag −0.44 (0.68)
∼ 18-Month lag −0.30 (0.64)
∼ 24-Month lag 0.47 (0.54)

Baseline predicted mean [10.68] [10.68]

Number of observations ≈269,500 ≈269,500

Panel B: age 21+

MML contemporary 1.32** (0.58) 1.02** (0.46)
MML leads and lags
∼ 24-Month lead (before approval) 0.20 (0.71)
∼ 18-Month lead −0.36 (0.64)
∼ 12-Month lead 0.18 (0.41)
∼ 6-Month lead 0.24 (0.55)
∼ Approved NOT implemented 0.52 (0.39)
∼ 6-Month lag (after implementation) 0.73* (0.38)
∼ 12-Month lag 0.41 (0.34)
∼ 18-Month lag 0.04 (0.47)
∼ 24-Month lag 0.11 (0.64)

Baseline predicted mean [9.33] [9.33]

Number of observations ≈323,900 ≈323,900

Note:
Standard errors in parentheses are clustered at the state level.
Baseline predicted mean in square brackets is calculated as the average of predicted probabilities/counts when setting MMLst to 0 and leaving the other covariates as the
observed values.
*
Significant at the 10 percent level.
**
Significant at the 5 percent level.
***Significant at the 1 percent level.

The previously highlighted policy effects on youth marijuana use
initiation, as well as on adult past-month marijuana use, marijuana
almost daily/daily use, marijuana abuse/dependence, past-month
binge drinking, and simultaneous use of marijuana and alcohol
remain significant with similar effect size in these state-level esti-
mates (Table 7).
4. Discussion
Three main pieces of evidence from our study inform the pol-
icy discussions of MMLs. First, we find a significant effect of MML
implementation on increasing marijuana use. Estimates suggest
that the populations responsive to MMLs are adolescents and young
adults aged 12–20 who experimented with marijuana for the first
time and adults aged 21 or above who tried marijuana prior to
the introduction of the law. This latter group also has an increased
risk of progression to almost daily/daily marijuana use and mari-
juana abuse/dependence.30 We caution that even if we assume the
increases in any marijuana use and regular use come from those
who use the drug for legitimate medical purposes, there may still
be possibility that marijuana abuse/dependence would increase as
a result of MML implementation. The effect of MML implementa-
tion on marijuana abuse/dependence constitutes a potential public
health concern similar to that of prescription drug abuse epidemic
in the U.S. (CDC, 2012).
Second, among those aged 21 or above, we find a spillover
effect of MML implementation on the increasing frequency of binge
drinking, possibly through increased use of the two substances
simultaneously. The complementarity between marijuana use and
binge drinking among adults of legal drinking age could magnify
the expected harms of an MML. As Pacula and Sevigny (2014) com-
mented, “even if consumption (of marijuana) were assumed to rise
by 100 percent, the savings of liberalization policies would dwarf
the known health costs associated with using marijuana. However,
all potential savings . . . could be entirely erased, and tremendous
losses incurred, if alcohol and marijuana turn out to be economic
complements.” The 10 percent increase in the frequency of binge
drinking and the 18–22 percent increase in the probability of

Lang, 2007). We followed a two-step procedure described in Bertrand et al. (2004,

p. 267) to accommodate staggered adoption of the MMLs across states. As a result,
the data were collapsed into pre- and post-MML two periods across 7 MML states.
The standard errors were adjusted to take into account the smaller number of MML
states (Donald and Lang, 2007). abuse and marijuana dependence to be valid psychiatric disorders, and marijuana
30
A diagnosis of substance abuse/dependence, by definition, indicates that an abuse/dependence as experienced in clinical population and general population
individual is experiencing a cluster of psychological, physical, cognitive, and behav- appears very similar to other substance abuse/dependence disorders (Budney et al.,
ioral symptoms associated with substance use. The DSM-IV considers marijuana 2007).
76 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80

Table 7
Robustness check for serial correlation by examining state-aggregated data.

State-aggregated rates Age 12–20 Age 21+

(1) (2) (3) (4)

State-cluster 2-Period panelsa State-cluster 2-Period panelsa

Marijuana use outcomes

% past-month marijuana use −0.63 (0.65) −0.33 (0.50) 1.34** (0.52) 1.17** (0.56)
[11.81] [10.86] [9.40] [8.61]
% marijuana almost −0.22 (0.26) −0.10 (0.24) 0.56** (0.22) 0.51** (0.29)
daily/daily use [3.57] [3.30] [3.81] [3.45]
* **
% marijuana use initiation 0.28 (0.17) 0.29 (0.12) 0.14 (0.09) 0.11 (0.07)
[6.85] [6.28] [0.94] [0.86]

Alcohol use outcomes

# past-month drinks per capita 0.05 (0.79) −0.06 (0.74) 0.69 (0.65) 0.72 (0.82)
[8.28] [7.62] [19.02] [18.38]
** ***
# binge drinking days per capita 0.01 (0.03) 0.02 (0.05) 0.18 (0.08) 0.12 (0.04)
[0.72] [0.63] [1.54] [1.37]
% past-month marijuana use −0.54 (0.37) −0.65* (0.39) 1.22*** (0.38) 1.29*** (0.45)
and binge drinking [7.45] [6.65] [6.51] [5.92]
% marijuana use while −0.24 (0.43) −0.15 (0.25) 0.63** (0.31) 0.61* (0.35)
drinking [4.65] [4.12] [4.45] [3.52]

Other downstream outcomes

% marijuana abuse/dependence −0.26 (0.45) −0.16 (0.39) 0.35** (0.18) 0.41** (0.20)
(1-Year lag) [4.89] [4.61] [2.27] [2.15]
% marijuana abuse/dependence −0.15 (0.44) −0.02 (0.36) 0.34* (0.20) 0.26 (0.17)
(2-Year lag) [4.89] [4.60] [2.29] [2.16]
% alcohol abuse/dependence −0.10 (0.47) −0.04 (0.44) 0.67 (0.51) 0.49 (0.41)
(1-Year lag) [8.24] [8.11] [11.02] [10.73]
% alcohol abuse/dependence −0.26 (0.46) −0.28 (0.48) 0.21 (0.53) 0.28 (0.57)
(2-Year lag) [8.22] [8.10] [11.03] [10.74]
% prescription painkiller misuse 0.04 (0.50) −0.04 (0.17) −0.08 (0.36) −0.10 (0.26)
(1-Year lag) [8.75] [8.51] [6.76] [6.47]
% crack/cocaine use −0.01 (0.20) 0.01 (0.15) 0.02 (0.16) 0.01 (0.21)
(1-Year lag) [2.84] [2.58] [3.22] [2.96]
% heroin use −0.01 (0.09) −0.02 (0.08) −0.008 (0.11) 0.005 (0.10)
(1-Year lag) [2.84] [2.58] [3.22] [2.96]

Number of observations 459 20 459 20

Note:
Standard errors in parentheses are clustered at the state level.
Baseline predicted mean in square brackets is calculated as the average of predicted probabilities/counts when setting MMLst to 0 and leaving the other covariates as the
observed values.
a
We average the pre-MML data and the post-MML data (Donald and Lang, 2007) following a two-step procedure described in Bertrand et al. (2004, p. 267). The second-step
equation is estimated based on pre- and post-MML two-period panels of 10 “MML states”. The standard errors are adjusted to take into account the small number of “MML
states” (Donald and Lang, 2007).
*
Significant at the 10 percent level.
**
Significant at the 5 percent level.
***
Significant at the 1 percent level.

simultaneous marijuana and alcohol use31 that we estimated may
result in considerable economic and social costs from downstream
health care expenditures and productivity loss (Naimi et al., 2003).
It is worth noting that this implied complementarity between
marijuana use induced by an MML and binge drinking does not
necessarily contradict a conclusion made by Anderson et al. (2011,
2012) that the implementation of an MML results in reduced traf-
fic fatalities, and that the reduction is more pronounced in those
involving alcohol. A possible interpretation that may reconcile our
findings with theirs is that MML implementation may lead to a
shift of alcohol consumption from public places such as restaurants
and bars to one’s own home. Thus, we may see a reduction in the
traffic fatalities, even if the implementation of an MML, in effect,
increases binge drinking and simultaneous use of both alcohol and
marijuana. The reduced traffic fatalities may result from the fact
that those potential high-risk drivers are now more likely to stay
at home and less likely to engage in driving.
31
The interaction between marijuana and alcohol may magnify the risks posed by
the two substances individually (Liguori et al., 2002; Medina et al., 2007).
Third, neither underage drinking among those aged 12–20 nor
other substance use (i.e., non-medical use of prescription pain med-
ication, heroin use and cocaine use) in both age groups is affected
by MML implementation. In this regard, the often-voiced concerns
about the potential gateway effect of marijuana is not supported by
our findings. We caution that our study is not intended to refute the
gateway hypothesis. Rather it suggests that the gateway effect is
not likely to occur in the context of an MML: for those who respond
to MML implementation and use marijuana, their marijuana use is
not likely to act as a gateway to more dangerous substance use
through the pharmacological properties of marijuana.32 On the
other hand, our findings do not lend support to an area of potential
benefits of the law either, which is to benefit those who misuse
opioid pain medication by helping them ease opiate withdrawal
32
Nonetheless marijuana may still be a gateway drug for other marijuana users
through other pathways. For instance, those who use marijuana regardless of the
laws or those who use marijuana in response to decriminalization may progress to
hard drug use because marijuana introduces them to a shared market or subculture
of hard drugs.
 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80
symptoms and achieve success in early recovery. However, NSDUH
only includes questions about “non-medical use” of pain medica-
tion, so we cannot examine the effect of MML implementation on
patients who use pain medication according to the prescription.
The previously documented beneficial effect of an MML on reduc-
ing opioid overdose mortality may primarily come from this group
of legitimate pain patients.33 An MML may benefit these patients
by allowing them to start with medical marijuana treatment in lieu
of opioid pain medication or to switch partially or entirely from
opioids to marijuana. Whether and to what extent the legitimate
pain patients may benefit from MML implementation merit further
investigation, but are beyond the scope of our study.
Substance dependence
A maladaptive pattern of substance use leading to clinically significant
impairment or distress, as manifested by 3 or more of the following
occurring at any time in the same 12-month period:
1. Tolerance or markedly increased amounts of the substance to
achieve intoxication or desired effect or markedly diminished effect
with continued use of the same amount of substance.
2. Characteristic withdrawal symptoms or the use of certain
substances to relieve or avoid withdrawal symptoms.
3. Use of a substance in larger amounts or over a longer period than
was intended.
4. Persistent desire or unsuccessful efforts to cut down or control
substance use.
5. Involvement in chronic behavior to obtain or use the substance, or
recover from its effects.
6. Important social, occupational or recreational activities given up or
reduced due to substance use.
7. Continued substance use despite knowledge of a persistent or
recurrent physical or psychological problem that is likely to have been
caused or exacerbated by the substance.
33
More than 60 percent of the opioid pain medication users receive and take the
drug according to the prescription (Bachhuber et al., 2014).
77
Taken together, our study findings provide evidence for a sig-
nificant effect of MML implementation on increasing marijuana
use, and a spillover effect among adults of legal drinking age from
increased marijuana use to increased binge drinking. The findings
do not, however, provide evidence to support other types of sub-
stance use spillovers such as underage drinking, pain medication
misuse, and hard drug use.
Appendix 1. DSM-IV criteria for substance abuse and
substance dependence
Substance abuse
A maladaptive pattern of substance use leading to clinically significant
impairment or distress, as manifested by 1 or more of the following
occurring at any time in the same 12-month period:
1. Recurrent substance use resulting in a failure to fulfill major role
obligations at work, school, or home (e.g., repeated absences or poor
work performance related to substance use; substance-related
absences, suspensions, or expulsions from school; neglect of children
or household).
2. Recurrent substance use in physically hazardous situations (e.g.,
driving an automobile or operating a machine when impaired by
substance use).
3. Recurrent substance-related legal problems (e.g. arrests for
obtaining or using the substance, substance-related disorderly
conduct).
4. Continued substance use despite persistent or recurrent social and
interpersonal problems caused or exacerbated by the substance (e.g.,
arguments with spouse about consequences of intoxication, physical
fights).
78 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80

Appendix 2. Robustness check for alternative classification of control states

Main outcomes Age 12–20 Age 21+

(1) (2) (3) (4) (5) (6)

MML states vs. ALL MML states vs. MML states vs. MML states vs. ALL MML states vs. MML states vs.
other states always MML states no MML states other states always MML states no MML states

% past-month −0.43 (0.48) −0.52 (0.48) −0.34 (0.43) 1.32** (0.58) 1.22** (0.60) 1.37** (0.59)
marijuana use [10.68] [12.75] [10.39] [9.33] [10.57] [8.90]
% marijuana almost −0.25 (0.17) −0.28 (0.18) −0.21 (0.15) 0.58 **
(0.26) 0.50 **
(0.24) 0.62 **
(0.26)
daily/daily use [3.52] [4.29] [3.40] [3.78] [4.43] [3.56]
Cond. # marijuana use −0.28 (0.45) −0.28 (0.48) −0.23 (0.47) 0.17 (0.64) 0.08 (0.61) 0.29 (0.65)
days [12.29] [12.37] [12.25] [14.15] [14.36] [14.02]
% marijuana use 0.32** (0.16) 0.34** (0.13) 0.31** (0.15) 0.15 (0.23) 0.18 (0.28) 0.15 (0.22)
initiation [6.47] [7.14] [6.32] [0.92] [1.14] [0.95]
# past-month total −0.03 (1.57) −0.06 (1.31) −0.01 (1.46) 0.95 (1.18) 0.90 (1.11) 0.99 (1.20)
drinks [7.76] [7.70] [7.83] [18.69] [18.62] [18.75]
# binge drinking days 0.04 (0.18) 0.01 (0.14) 0.05 (0.18) 0.16** (0.08) 0.14** (0.06) 0.16* (0.09)
[0.66] [0.67] [0.66] [1.52] [1.48] [1.53]
% past-month marijuana −0.63 (0.39) −0.73 (0.46) −0.57 (0.36) 1.44*** (0.35) 1.41*** (0.38) 1.50*** (0.37)
use and binge drinking [6.41] [7.49] [6.30] [6.44] [7.09] [6.21]
% marijuana use while −0.38 (0.49) −0.42 (0.61) −0.34 (0.47) 0.82* (0.45) 0.71 (0.53) 0.84* (0.46)
drinking [4.10] [5.03] [4.01] [4.45] [5.17] [4.24]

Number of observations ≈269,500 ≈86,700 ≈229,400 ≈323,900 ≈104,400 ≈275,800

Note:
Standard errors in parentheses are clustered at the state level.
Baseline predicted mean in square brackets is calculated as the average of predicted probabilities/counts when setting MMLst to 0 and leaving the other covariates as the
observed values.
*
Significant at the 10 percent level.
**
Significant at the 5 percent level.
***
Significant at the 1 percent level.

Appendix 3. Robustness check for sampling design-based adjustment

Main outcomes Age 12–20 Age 21+

(1) (2) (3) (4)

Unweighted state-cluster Weighted sampling-design Unweighted state-cluster Weighted sampling-design

% past-month marijuana use −0.43 (0.48) −0.47 (0.65) 1.32** (0.58) 1.19*** (0.40)
[10.68] [11.28] [9.33] [8.50]
% marijuana almost daily/daily −0.25 (0.17) −0.28 (0.25) 0.58** (0.26) 0.53*** (0.13)
use [3.52] [3.76] [3.78] [3.53]
Cond. # marijuana use days −0.28 (0.45) −0.38 (0.67) 0.17 (0.64) 0.19 (0.60)
[12.29] [12.44] [14.15] [14.49]
% marijuana use initiation 0.32** (0.16) 0.33** (0.11) 0.15 (0.23) 0.14 (0.17)
[6.47] [6.58] [0.92] [0.85]
# past-month total drinks −0.03 (1.57) −0.04 (2.32) 0.95 (1.18) 0.76 (0.87)
[7.76] [8.59] [18.69] [16.32]
# binge drinking days 0.04 (0.18) 0.04 (0.17) 0.16** (0.07) 0.15*** (0.04)
[0.66] [0.74] [1.52] [1.49]
% past-month marijuana use −0.63 (0.39) −0.68* (0.37) 1.44*** (0.35) 1.28*** (0.31)
and binge drinking [6.41] [6.86] [6.44] [5.74]
% marijuana use while drinking −0.38 (0.49) −0.43 (0.37) 0.82* (0.45) 0.69* (0.38)
[4.10] [4.44] [4.45] [3.73]

Number of observations ≈269,500 ≈269,500 ≈323,900 ≈323,900

Note:
Standard errors in parentheses are clustered at the state level.
Baseline predicted mean in square brackets is calculated as the average of predicted probabilities/counts when setting MMLst to 0 and leaving the other covariates as the
observed values.
*
Significant at the 10 percent level.
**
Significant at the 5 percent level.
***
Significant at the 1 percent level.
 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80 79

Appendix 4. State heterogeneity in policy effect

Age 12–20 Age 21+

(1) (2) (3) (4) (5) (6)

% marijuana % past-month % marijuana % marijuana # binge % marijuana use
initiation marijuana use (almost)/daily abuse/dependence drinking days and binge drinking

MML implementation 0.32** (0.16) 1.32** (0.58) 0.58** (0.26) 0.25** (0.11) 0.16** (0.08) 1.44*** (0.35)
State MMLs
∼ Vermont 0.57 ***
(0.18) 1.76 ***
(0.21) 2.27 ***
(0.12) 1.19 ***
(0.12) 0.11 **
(0.03) 2.03***
(0.15)
∼ Montana 0.05 (0.20) 4.92*** (0.27) 1.26*** (0.15) 2.51*** (0.15) 0.47*** (0.05) 4.16*** (0.20)
∼ Rhode Island 0.61*** (0.16) −0.64 (0.37) −0.42* (0.23) 0.29** (0.13) 0.12*** (0.03) 0.25* (0.14)
∼ New Mexico −0.04 (0.23) −0.23 (0.17) −0.15 (0.16) 0.16 (0.11) 0.46*** (0.04) 0.08 (0.18)
∼ Michigan 0.43** (0.22) 2.37*** (0.34) 1.44*** (0.22) 0.17 (0.10) 0.22*** (0.06) 1.57*** (0.28)
∼ New Jersey 0.007 (0.19) 1.52** (0.23) 0.94** (0.15) −0.15 (0.11) −0.10 (0.07) 2.17*** (0.18)
∼ District of Columbia 0.33* (0.19) 0.59** (0.27) 0.87*** (0.21) 0.30** (0.13) 0.15** (0.06) 0.70*** (0.21)
∼ Arizona −0.02 (0.20) −0.23 (0.20) 0.47*** (0.10) −0.25 (0.16) −0.12 (0.07) 0.10 (0.12)
∼ Delaware 0.79*** (0.25) 0.39* (0.23) 0.06 (0.13) 0.21* (0.12) 0.27*** (0.04) 0.58** (0.24)
∼ Connecticut 0.21 (0.18) 1.27** (0.21) 0.85** (0.12) 0.59*** (0.11) 0.05 (0.03) 1.09*** (0.17)

Baseline predicted mean [6.47] [9.33] [3.78] [4.59] [1.52] [6.44]

Number of observations ≈323,900 ≈323,900 ≈323,900 ≈323,900 ≈323,900 ≈323,900
Note:
Standard errors in parentheses are clustered at the state level.
Baseline predicted means in square brackets are calculated as the average of predicted probabilities/counts when setting MMLst (VT) ∼ MMLst (CT) to 0 and leaving the other
covariates as the observed values.
*
Significant at the 10 percent level.
**
Significant at the 5 percent level.
***
Significant at the 1 percent level.

References
Abrams, D.I., Couey, P., Shade, S.B., Kelly, M.E., Benowitz, N.L., 2011.
Cannabinoid–opioid interaction in chronic pain. Clinical Pharmacology &
Therapeutics 90 (6), 844–851.
American Psychiatric Association (APA) (Ed.), 2000. Diagnostic and Statistical Man-
ual of Mental Disorders: DSM-IV-TR. American Psychiatric Publishing, Arlington,
VA.
Anderson, D.M., Hansen, B., Rees, D.I., 2011. Medical marijuana laws, traffic fatalities,
and alcohol consumption. Institute for the Study of Labor (IZA) Discussion Paper
Series No. 6112. http://ftp.iza.org/dp6112.pdf (accessed 10.10.13).
Anderson, D.M., Hansen, B., Rees, D.I., 2012. Medical marijuana laws and teen mari-
juana use. Institute for the Study of Labor (IZA) Discussion Paper Series No. 6592.
http://ftp.iza.org/dp6112.pdf (accessed 10.10.13).
Anderson, D.M., Hansen, B., Rees, D.I., 2013. Medical marijuana laws, traffic fatalities,
and alcohol consumption. Journal of Law and Economics 56, 333–369.
Anderson, D.M., Rees, D.I., 2014. The role of dispensaries: the devil is in the details.
Journal of Policy Analysis and Management 33 (1), 235–240.
Bachhuber, M.A., Saloner, B., Cunningham, C.O., Barry, C.L., 2014. Medical cannabis
laws and opioid analgesic overdose mortality in the United States, 1999–2010.
JAMA Internal Medicine 174 (10), 1668–1673.
Belotti, F., Deb, P., Manning, W.G., Norton, E.C., 2014. TPM: estimating two-part
models. The Stata Journal, http://econ.hunter.cuny.edu/people/economics-
faculty/pdeb/ihea-minicourse/tpm-estimating-two-part-models-working-
paper/at download/file (accessed 21.03.14; forthcoming).
Ben Amar, M., 2006. Cannabinoids in medicine: a review of their therapeutic poten-
tial. Journal of Ethnopharmacology 105 (1), 1–25.
Bertrand, M., Duflo, E., Mullainathan, S., 2004. How much should we trust
differences-in-differences estimates? The Quarterly Journal of Economics 119
(1), 249–275.
Bohnert, A.S.B., Valenstein, M., Bair, M.J., Ganoczy, D., McCarthy, J.F., Ilgen, M.A.,
Blow, F.C., 2011. Association between opioid prescribing patterns and opioid
overdose-related deaths. The Journal of the American Medical Association 305
(13), 1315–1321.
Boys, A., Marsden, J., Strang, J., 2001. Understanding reasons for drug use amongst
young people: a functional perspective. Health Education Research 16 (4),
457–469.
Budney, A.J., Roffman, R., Stephens, R.S., Walker, D., 2007. Marijuana dependence
and its treatment. Addiction Science and Clinical Practice 4 (1), 4.
Campbell, V.A., Gowran, A., 2007. Alzheimer’s disease: taking the edge off with
cannabinoids? British Journal of Pharmacology 152 (5), 655–662.
Carpenter, C., Dobkin, C., 2009. The effect of alcohol consumption on mortality:
regression discontinuity evidence from the minimum drinking age. American
Economic Journal: Applied Economics 1 (1), 164.
Center for Behavioral Health Statistics and Quality (CBHSQ), 2013. National Sur-
vey on Drug Use and Health, 2004–2011 [data files and code books]. U.S.
Dept. of Health and Human Services (HHS), Substance Abuse and Mental
Health Services Administration (SAMHSA), Center for Behavioral Health Statis-
tics and Quality (CBHSQ), Rockville, MD, https://www.datafiles.samhsa.gov,
http://www.icpsr.umich.edu/icpsrweb/SAMHDA (accessed from 26.09.13 to
21.03.14).
Centers for Disease Control and Prevention (CDC), 2012. CDC grand rounds: prescrip-
tion drug overdoses – a US epidemic. Morbidity and Mortality Weekly Report
61 (1), 10.
Chu, Y.L., 2014. The effects of medical marijuana laws on illegal marijuana use.
Journal of Health Economics 38, 43–61.
Chaloupka, F.J., Laixuthai, A., 1997. Do youths substitute alcohol and marijuana?
Some econometric evidence. Eastern Economic Journal 23 (3), 253–276.
Cohen, P.J., 2010. Medical marijuana 2010: it’s time to fix the regulatory vacuum.
The Journal of Law, Medicine and Ethics 38 (3), 654–666.
Crost, B., Guerrero, S., 2012. The effect of alcohol availability on marijuana use:
evidence from the minimum legal drinking age. Journal of Health Economics 31
(1), 112–121.
Crost, B., Rees, D.I., 2013. The minimum legal drinking age and marijuana use: new
estimates from the NLSY97. Journal of Health Economics 32 (2), 474–476.
DeSimone, J., Farrelly, M.C., 2003. Price and enforcement effects on cocaine and
marijuana demand. Economic Inquiry 41 (1), 98–115.
DiNardo, J., Lemieux, T., 2001. Alcohol, marijuana, and American youth: the unin-
tended consequences of government regulation. Journal of Health Economics
20 (6), 991–1010.
Donald, S.G., Lang, K., 2007. Inference with difference-in-differences and other panel
data. The Review of Economics and Statistics 89 (2), 221–233.
Gloss, D., Vickrey, B., 2012. Cannabinoids for epilepsy. Cochrane Database Systematic
Reviews 6.
French, M.T., Popovici, I., 2011. That instrument is lousy! In search of agreement
when using instrumental variables estimation in substance use research. Health
Economics 20 (2), 127–146.
Gostin, L.O., 2005. Medical marijuana, American federalism, and the Supreme Court.
The Journal of the American Medical Association 294 (7), 842–844.
Greene, W.H., 2011. Econometric Analysis, 5th edition. Prentice Hall, Upper Saddle
River, NJ.
Harper, S., Strumpf, E.C., Kaufman, J.S., 2012. Do medical marijuana laws increase
marijuana use? Replication study and extension. Annals of Epidemiology 22 (3),
207–212.
Hathaway, A.D., Comeau, N.C., Erickson, P.G., 2011. Cannabis normalization and
stigma: contemporary practices of moral regulation. Criminology and Criminal
Justice 11 (5), 451–469.
Heishman, S.J., Arasteh, K., Stitzer, M.L., 1997. Comparative effects of alcohol and
marijuana on mood, memory, and performance. Pharmacology Biochemistry
and Behavior 58 (1), 93–101.
Hoffmann, D.E., Weber, E., 2010. Medical marijuana and the law. New England
Journal of Medicine 362 (16), 1453–1457.
Horowitz, J.L., 2001. Should the DEA’s STRIDE data be used for economic analy-
ses of markets for illegal drugs? Journal of American Statistical Association 96,
1254–1271.
Johnson, T.P., Fendrich, M., Mackesy-Amiti, M.E., 2010. Computer literacy and the
accuracy of substance use reporting in an ACASI survey. Social Science Computer
Review 28 (4), 515–523.
Kandel, D.B., 1975. Stages in adolescent involvement in drug use. Science 190 (4217),
912–914.
Kandel, D.B. (Ed.), 2002. Stages and Pathways of Drug Involvement: Examining the
Gateway Hypothesis. Cambridge University Press, Cambridge, UK.
80 H. Wen et al. / Journal of Health Economics 42 (2015) 64–80
King, A.C., de Wit, H., McNamara, P.J., Cao, D., 2011. Rewarding, stimulant, and seda-
tive alcohol responses and relationship to future binge drinking. Archives of
General Psychiatry 68 (4), 389–399.
Krishnan, S., Cairns, R., Howard, R., 2009. Cannabinoids for the treatment of demen-
tia. Cochrane Database Systematic Reviews 2.
Liguori, A., Gatto, C.P., Jarrett, D.B., 2002. Separate and combined effects of marijuana
and alcohol on mood, equilibrium and simulated driving. Psychopharmacology
163 (3/4), 399–405.
Lukas, S.E., Orozco, S., 2001. Ethanol increases plasma delta-9-tetrahydrocannabinol
(THC) levels and subjective effects after marihuana smoking in human volun-
teers. Drug and Alcohol Dependence 64 (2), 143–149.
Lynch, M.E., Campbell, F., 2011. Cannabinoids for treatment of chronic non-cancer
pain: a systematic review of randomized trials. British Journal of Clinical Phar-
macology 72 (5), 735–744.
Lynne-Landsman, S.D., Livingston, M.D., Wagenaar, A.C., 2013. Effects of state med-
ical marijuana laws on adolescent marijuana use. American Journal of Public
Health 103 (8), 1500–1506.
Maldonado, R., Valverde, O., Berrendero, F., 2006. Involvement of the endocannabi-
noid system in drug addiction. Trends in Neurosciences 29 (4), 225–232.
Medina, K.L., Schweinsburg, A.D., Cohen-Zion, M., Nagel, B.J., Tapert, S.F., 2007.
Effects of alcohol and combined marijuana and alcohol use during adolescence
on hippocampal volume and asymmetry. Neurotoxicology and Teratology 29
(1), 141–152.
Moore, S.C., 2010. Substitution and complementarity in the face of alcohol-specific
policy interventions. Alcohol and Alcoholism 45 (5), 403–408.
Morral, A.R., McCaffrey, D.F., Paddock, S.M., 2002. Reassessing the marijuana gate-
way effect. Addiction 97 (12), 1493–1504.
Naimi, T.S., Brewer, R.D., Mokdad, A., Denny, C., Serdula, M.K., Marks, J.S., 2003. Binge
drinking among US adults. The Journal of the American Medical Association 289
(1), 70–75.
Pacula, R.L., 1998. Does increasing the beer tax reduce marijuana consumption?
Journal of Health Economics 17 (5), 557–585.
Pacula, R.L., Kilmer, B., Grossman, M., Chaloupka, F.J., 2010. Risks and prices: the role
of user sanctions in marijuana markets. The BE Journal of Economic Analysis &
Policy 10 (1), 1–36.
Pacula, R.L., Powell, D., Heaton, P., Sevigny, E.L., 2013. Assessing the effects of medical
marijuana laws on marijuana and alcohol use: the devil is in the details. National
Bureau of Economic Research (NBER) Working Paper No. w19302.
Pacula, R.L., Sevigny, E.L., 2014. Marijuana liberalization policies: why we can’t learn
much from policy still in motion. Journal of Policy Analysis and Management 33
(1), 212–221.
Pertwee, R.G., 2012. Targeting the endocannabinoid system with cannabinoid
receptor agonists: pharmacological strategies and therapeutic possibilities.
Philosophical Transactions of the Royal Society B: Biological Sciences 367 (1607),
3353–3363.
Saffer, H., Chaloupka, F., 1999. The demand for illicit drugs. Economic Inquiry 37 (3),
401–411.
Scavone, J.L., Sterling, R.C., Van Bockstaele, E.J., 2013. Cannabinoid and opioid inter-
actions: implications for opiate dependence and withdrawal. Neuroscience 248,
637–654.
Sekhon, V., 2009. Highly uncertain times: an analysis of the executive branch’s
decision to not investigate or prosecute individuals in compliance with
state medical marijuana laws. Hastings Constitutional Law Quarterly 37 (3),
553–564.
Solon, G., Haider, S.J., Wooldridge, J., 2013. What are we weighting for? National
Bureau of Economic Research (NBER) Working Paper No. w18859.
Substance Abuse and Mental Health Services Administration (SAMHSA), 2013.
Results from the 2012 National Survey on Drug Use and Health: Summary of
National Findings. U.S. Dept. of Health and Human Services (HHS), Substance
Abuse and Mental Health Services Administration (SAMHSA), Rockville,
MD http://www.samhsa.gov/data/nsduh/2k11results/nsduhresults2011.htm
(accessed 10.10.13).
Wagner, F.A., Anthony, J.C., 2002a. Into the world of illegal drug use: exposure oppor-
tunity and other mechanisms linking the use of alcohol, tobacco, marijuana, and
cocaine. American Journal of Epidemiology 155 (10), 918–925.
Wagner, F.A., Anthony, J.C., 2002b. From first drug use to drug dependence: devel-
opmental periods of risk for dependence upon marijuana, cocaine, and alcohol.
Neuropsychopharmacology 26, 479–488.
Wang, P., 2003. A bivariate zero-inflated negative binomial regression model for
count data with excess zeros. Economics Letters 78 (3), 373–378.
Wechsler, H., Dowdall, G.W., Davenport, A., Rimm, E.B., 1995. A gender-specific
measure of binge drinking among college students. American Journal of Public
Health 85 (7), 982–985.
Yörük, B.K., Yörük, C.E., 2011. The impact of minimum legal drinking age laws on
alcohol consumption, smoking, and marijuana use: evidence from a regression
discontinuity design using exact date of birth. Journal of Health Economics 30
(4), 740–752.
Yörük, B.K., Yörük, C.E., 2013. The impact of minimum legal drinking age laws on
alcohol consumption, smoking, and marijuana use revisited. Journal of Health
Economics 32 (2), 477–479.
 Journal of Health Economics 42 (2015) 81–89

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

The interaction of direct and indirect risk selection

Normann Lorenz ∗
Universität Trier, Universitätsring 15, 54286 Trier, Germany

a r t i c l e i n f o a b s t r a c t

Article history: This paper analyzes the interaction of direct and indirect risk selection in health insurance markets. It is
Received 15 April 2014 shown that direct risk selection – using measures unrelated to the benefit package like selective adver-
Received in revised form 6 September 2014 tising or ‘losing’ applications of high risk individuals – nevertheless has an influence on the distortions of
Accepted 9 December 2014
the benefit package caused by indirect risk selection. Direct risk selection (DRS) may either increase or
Available online 17 December 2014
decrease these distortions, depending on the type of equilibrium (pooling or separating), the type of DRS
(positive or negative) and the type of cost for DRS (individual-specific or not). Regulators who succeed in
JEL classification:
reducing DRS by, e.g., banning excessive advertising or implementing fines for ‘losing’ applications, may
I13
I18
therefore (unintendedly) mitigate or exacerbate the distortions of the benefit package caused by indirect
L13 risk selection. It is shown that the interaction of direct and indirect risk selection also alters the formula
for optimal risk adjustment.
Keywords: © 2015 Elsevier B.V. All rights reserved.
Risk selection
Risk adjustment
Health insurance
Discrete choice
Imperfect competition

1. Introduction advertising or ‘losing’ applications of high risk individuals.3 It has

Risk selection is considered to be one of the main problems in
regulated health insurance markets. If there is community rating,
so that insurers are not allowed to charge premiums according to
risk, they will make profits with some individuals and losses with
others. Insurers who act on these incentives to attract profitable and
repel unprofitable individuals are said to engage in risk selection.1
Two forms of risk selection can be distinguished: direct risk
selection (DRS) and indirect risk selection (IRS).2 With DRS, insur-
ers know that a particular individual or group of individuals is
characterized by non-average risk. DRS is therefore targeted at an
individual the insurer has identified to either be a high or a low
risk type (like, e.g., a hypochondriac) or at a group of individuals
the insurer knows to have non-average expected cost (like, e.g., a
certain age group or individuals living in a high cost area). Usu-
ally, the measures taken for DRS are not related to the benefit
package (i.e. the medical services) offered; examples are selective
been shown that potential profits associated with successful DRS
can be substantial.4
With IRS, insurers do not know which particular individual is
of high or low risk (or are prevented from using this information),
and so only act on their knowledge that there are different risk
types in the population. The measures taken to engage in IRS usually
consist of distorting the benefit package, so that it is attractive for
low risks, but not for high risks. Several studies have shown that
the incentives for IRS can be severe, and that insurers do indeed act
on these incentives.5
A regulator can counteract the incentives for both DRS and IRS by
implementing a risk adjustment scheme, setting transfers to (and
from) insurers depending on signals which are informative about
individuals’ expected cost. In almost all risk adjustment schemes,
the formula used to calculate these transfers is based on a regres-
sion of actual health care expenditures on a set of explanatory
variables like age, gender and morbidity. Most of the literature on

3
van de Ven and van Vliet (1992) provide an extensive list of measures insur-
∗ Tel.: +49 651 2012624. ers may use for risk selection; for differential treatment of low and high risks’
E-mail address: lorenzn@uni-trier.de applications see Bauhoff (2012).
1 4
See van de Ven and Ellis (2000). See Shen and Ellis (2002).
2 5
See Breyer et al. (2011). See Frank et al. (2000), Cao and McGuire (2003) and Ellis and McGuire (2007).

http://dx.doi.org/10.1016/j.jhealeco.2014.12.003
0167-6296/© 2015 Elsevier B.V. All rights reserved.
82 N. Lorenz / Journal of Health Economics 42 (2015) 81–89
risk adjustment has been concerned with improving this underly-
ing regression by, e.g., including additional variables or altering the
grouping algorithm for diagnoses in morbidity based risk adjust-
ment, so that a larger part of the variance of actual expenditures is
explained. The larger the explained part of the variance, the closer
transfers are to actual cost, and the lower the incentives for risk
selection should be.
Initiated by the very influential study of Glazer and McGuire
(2000), there has developed a small literature that departs from this
statistical approach and instead explicitly models insurers’ incen-
tives for risk selection. One study in this literature has shown that
conventional, i.e. regression-based, risk adjustment may decrease
welfare if there is imperfect competition, another, that it may even
increase the extent of risk selection.6 These undesirable effects of
conventional risk adjustment exemplify the need for what Glazer
and McGuire (2000) have termed optimal risk adjustment. They
have shown that a regulator can increase the effectiveness of a risk
adjustment scheme by distorting the payments as calculated from
a regression: there has to be overpayment for signals which are
correlated with high risks and underpayment for signals which are
correlated with low risks. If the over- and underpayments are cho-
sen optimally, incentives for IRS can be eliminated completely.7
Optimal risk adjustment has also been derived for a setting where
individuals differ in their elasticity to switch insurers or where
insurers are allowed to vary their premium in some dimension,
as is the case in the insurance exchanges in the US.8
A concern, already raised by Glazer and McGuire (2000) them-
selves, is that such over- and underpayments create incentives
for DRS regarding the signal, but so far it has not been analyzed
whether this has an influence on optimal risk adjustment. In fact,
we are not aware of any theoretical study that explicitly models
the interaction of direct and indirect risk selection, even in the
absence of risk adjustment.9 In this study we therefore develop
such a model and show that in general (the degree of) DRS has
an influence on the distortions of the benefit package caused by
IRS and that this alters the formula for optimal risk adjustment.
DRS may either increase or decrease the distortions caused by
IRS, depending on whether insurers try to attract the low risks
(positive DRS) or to repel the high risks (negative DRS), whether a
pooling or a separating equilibrium emerges, and whether the cost
for DRS is individual-specific or not.
If insurers’ expenditures for DRS are at least to some degree
individual-specific (and not just a fixed cost), they affect risk-type-
specific cost: Positive DRS increases the cost per low risk, negative
DRS the cost per high risk. In the first case, the cost difference
between the risk types is reduced, in the second, it is increased.
In the pooling equilibrium where both risk types pay the same pre-
mium, positive DRS therefore reduces the incentives for IRS, while
negative DRS increases it.
In the separating equilibrium, insurers’ expenditures for posi-
tive DRS translate into a higher premium for the contract offered
for the low risks; this makes this contract less attractive for the high
risks, so the distortion of the benefit package necessary to repel the
6
See Lorenz (2013) and Brown et al. (2011), respectively. In the empirical part of
their study, Brown et al. (2011) find such an increase in the extent of risk selection
for the Medicare Advantage program in the U.S.; however, there has been some
disagreement on this finding, see Newhouse et al. (2012).
7
See also Glazer and McGuire (2002) and Jack (2006).
8
For the first setting, see Bijlsma et al. (2011), and for the second, McGuire et al.
(2013) and Shi (2013).
9
Eggleston (2000) derives the optimal mix of supply and demand side cost shar-
ing for a setting with a single (semi-altruistic) HMO that can influence the level of
medical services (according to the outcome of a patient–provider bargaining pro-
cess) and can dump a share of the high risks at some cost; however, there is no
competition as there is only one provider.
Table 1
Effect of DRS with individual-specific cost on the distortion of the benefit package.
Type of equilibrium Positive DRS Negative DRS
Pooling equilibrium Distortion decreases Distortion increases
Separating equilibrium Distortion decreases Distortion decreases
high risks is reduced. Negative DRS on the other hand creates a ‘sub-
stitution effect’: If insurers are (somewhat) successful in repelling
the high risks by DRS, they can reduce the degree of IRS. For an
overview of these results see, Table 1.
A regulator who succeeds in reducing negative DRS (by, e.g.,
charging a fine for ‘losing’ applications of high risk individuals) will
therefore simultaneously reduce IRS in the pooling equilibrium, but
unintendedly increase the distortion of the benefit package in the
separating equilibrium. The distortions caused by IRS will also be
increased if he succeeds in reducing positive DRS (by, e.g., banning
excessive advertising), in this case for both the pooling and the
separating equilibrium.
In three of the four cases, optimal risk adjustment then becomes
even more important. We therefore derive the impact of DRS
on the formula for optimal risk adjustment developed by Glazer
and McGuire (2000). We show that the overpayment for a signal
that indicates a high risk has to be increased exactly by insurers’
expenditures on positive DRS; likewise, the underpayment has to
be reduced by the expenditures on negative DRS. With this modi-
fication, their formula can eliminate the incentives for IRS even in
the presence of DRS.
In the literature on optimal risk adjustment, some of the results
regarding the distortions caused by IRS have been derived under
perfect competition, but DRS seems incompatible with such a
setting where individuals are perfectly informed about all ben-
efit packages and premiums and always choose the insurer that
offers the best benefit package-premium combination. We there-
fore derive our results within a discrete choice model, which can
easily capture different levels of competition. To keep the model
simple, we assume that the benefit package is one-dimensional, but
the model can be extended to a multi-dimensional benefit package.
Also, to simplify the notation when deriving the results, we first
consider the case of two risk types, but then show that the results
also hold for an arbitrary number of risk types.
The remainder of this paper is organized as follows: In Section
2, we introduce the basic discrete choice model and show how DRS
can be incorporated in such a model. We analyze the pooling equi-
librium in Section 3 and the separating equilibrium in Section 4.
Section 5 concludes.
2. The model
2.1. Basic model without DRS
Individual preferences regarding the benefit-premium bundle
are given by
u = pr v(m) − R, (1)
where R denotes the premium and m the level of medical services
(measured in monetary terms). pr is the probability of becoming
ill, and there are two risk types r = H, L, with pH > pL ; the share of
the low risks is . The utility of receiving medical treatment, v(m),
is increasing at a decreasing rate, i.e. v (m) > 0 and v (m) < 0. The
efficient level of medical services is implicitly defined by v (mFB ) =
1.
There are n insurers j, each offering a benefit-premium bundle
{mj , Rj }. The individual’s decision of which insurer to choose may,
however, not only depend on these benefit-premium bundles, but
 N. Lorenz / Journal of Health Economics 42 (2015) 81–89
also on some other factors, like perceived friendliness of personnel,
location, or which insurer was recommended by family or friends.
In a discrete choice model, these other factors are captured by aug-
menting the individual’s utility as given in (1) by an individual- and
j
insurer-specific utility component εi ; the utility of an individual i
(being of risk type r) when choosing an insurer j therefore is
j
ui (mj , Rj ) = pr v(mj ) − Rj + εi . (2)
j
If εi is assumed to be i.i.d. extreme value, the logit model with its
analytically tractable choice probabilities arises. Denoting risk type
r’s utility of the benefit-premium bundle offered by insurer j by
j
Vr = pr v(mj ) − Rj ,
j j
and specifying the variance of εi as Var(εi ) =  2 2 /6, the probabil-
ity of individual i (being of risk type r) choosing a particular insurer
k is10
Prob(i chooses k) = Prob(Vr k + εki > Vrl + εli ∀ l =
/ k)
k
eVr /
= n j
. (3)
j=1
eVr /
Denote this probability by Prk ; it is also insurer k’s market share
among the individuals of risk type r. Prk is increasing in Vrk : a higher
share of individuals of risk type r will choose insurer k, if this insurer
offers a higher level of medical services or charges a lower premium.
j
The variance of the additional utility component, Var(εi ) =
2 2
  /6, is a measure of the level of competition in this health insur-
j memberships or special counseling services. In this case, total
ance market. If  is small, all the εi are very similar and therefore
only play a minor role in which insurer is chosen: Offering an only
somewhat higher utility level than all the other insurers will, in this
case, already attract a large share of all individuals; this implies a
high level of competition. If, on the other hand,  is large, the other
factors besides the benefit level and the premium – captured by
j generates some cost, but decreases the probability of being chosen
large positive and large negative εi – are rather important, so that
insurers, when increasing their premium (or reducing their benefit
level), only lose a small share of their insured; a large level of 
therefore corresponds to a low level of competition.
As shown by Lorenz (2013) who has analyzed this basic model
without DRS, the level of competition determines which type of
equilibrium emerges: If the level of competition is low (i.e.,  is
large), there will be a pooling equilibrium: all insurers offer the
same benefit-premium bundle, so each individual i chooses the
j tising’ might be an example, where an insurer informs about some
insurer j for which his εi is maximal. If the level of competition
is high (i.e.,  is small), a separating equilibrium similar (but not
identical) to the Rothschild–Stiglitz equilibrium under perfect com-
petition arises.11 Some of the n insurers offer a benefit-premium
bundle designated for the low risks, the remaining insurers a con-
tract designated for the high risks (with a larger benefit package
and a higher premium). All the low risks choose the insurer with
j additional (unnecessary) paper work or involve the high risk indi-
the highest εi among the first set of insurers, but only most of the
j
high risks choose the insurer with the highest εi among the second
set. Because a small share of the high risks chooses one of the insur-
ers offering the contract designated for the low risks, the separation
of risk types is not perfect.12
10
See Train (2009, p. 40).
11
See Zweifel et al. (2009), chapter 7, for the Rothschild–Stiglitz equilibrium in
this setting.
12
We explain why this occurs in Section 4.1.
83
2.2. The model with positive and negative DRS
We consider positive DRS to be an activity each insurer is
engaged in which generates some cost and increases the proba-
bility of being chosen by the individual (or group of individuals)
the activity is targeted at. We model this increase in the probability
of being chosen to stem from an increase in the utility the individ-
ual receives, which may either be real (as, e.g., with a discount for
a fitness club membership) or just perceived (as with advertising).
We denote the cost by aj and the increase in utility by g(aj ), where
g(aj ) is increasing and concave.13 With positive DRS, the (perceived)
utility of individual i choosing an insurer j therefore is
j
ui (mj , Rj ) = pr v(mj ) − Rj + g(aj ) + εi , (4)
so that insurer k’s market share among risk type r is given by14
k k ))/
e(Vr +g(a
Prk =  j +g(aj ))/
. (5)
j
e(Vr
Two cases regarding the cost aj can be distinguished:
non-individual-specific and individual-specific cost. With non-
individual-specific cost, total cost for DRS of an insurer j is
independent of the number of individuals choosing this insurer.
The prime example for this case is selective advertising, where
cost does not increase if an additional individual chooses insurer
j. With individual-specific cost, total cost for DRS of an insurer
j increases in the number of individuals choosing this insurer.
One example here are additional benefits which the regulator (or
society) considers not to be part of a ‘normal’ basic benefit package
insurers are supposed to provide, like discounts for fitness club
cost of DRS increases if an additional individual chooses insurer j.
It seems reasonable to assume that most risk selection activities
entail both non-individual-specific and individual-specific (i.e.
fixed and variable) cost.
Like positive DRS, we model negative DRS as an activity that
by a particular individual (or group of individuals). We denote the
cost of negative DRS by bj and the utility decrease by f(bj ), where
f(bj ) is increasing and concave. With negative DRS, the (perceived)
utility of individual i and insurer k’s market share are as given in
(4) and (5), but with +g(a) replaced by −f(b).
Unlike with positive DRS, it is difficult to imagine some activity
where the cost an insurer incurs for negative DRS is independent of
the number of individuals choosing this insurer. ‘Negative adver-
undesirable feature of its offer, like scrupulous utilization reviews,
but this and similar examples may seem rather far-fetched. We
think it is more realistic to consider negative DRS to be an activity
insurers are engaged in during the application process, so that cost
depends on the number of individuals applying at the insurer.15
Activities which fall into this category are that insurers require
viduals in lengthy phone calls in which they try to convince (or
13
Since we are interested in a setting where insurers are engaged in DRS, we
assume lim g  (a) → ∞ to guarantee an interior solution.
a→0+
14
To simplify the notation, we do not introduce different symbols for Pjk for the
case of no, positive or negative DRS; we will, however, always make clear to which
case we refer.
15
In Section 5 we argue that the main effects should be similar if negative DRS
does not occur during the application process, but is targeted at individuals who
already hold a contract with the insurer.
84 N. Lorenz / Journal of Health Economics 42 (2015) 81–89
urge) these individuals to choose a different insurer.16 In the work-
ing paper version of this study we explicitly model the application
process and show that the main results are the same regardless of
whether the number of individuals applying at an insurer is equal
to or larger than the number of individuals eventually choosing
the insurer.17 We will therefore refer to both cases as individual-
specific cost.
3. The pooling equilibrium
In this section we analyze the pooling equilibrium which occurs
if the level of competition is low enough. We begin by briefly deriv-
ing the equilibrium without DRS in Section 3.1. We consider the
equilibrium with positive DRS in Section 3.2 and analyze the impact
of positive DRS on risk adjustment in Section 3.3. We then show
that the results also hold for an arbitrary number of risk types in
Section 3.4. The effects of negative DRS, which are just the opposite
of positive DRS, are summarized in Section 3.5.
3.1. The pooling equilibrium without DRS
Normalizing the mass of individuals to one and assuming profit
maximization, the objective of insurer k is
max k = PLk Lk + (1 − )PHk H
k
, (6)
mk ,Rk
where rk = Rk − pr mk denotes insurer k’s profit per individual of
risk type r. The solution to this objective yields the following two
conditions (see Appendix A.1): First,
n
Lk + (1 − )H
k
= . (7)
n−1
Average profit per insured decreases in n and increases in : a
higher level of competition (a larger number of insurers n or a
smaller level of ) decreases profit. Secondly, the condition deter-
mining the distortion of the benefit level is given by
 2
 waste part of this profit on their expenditures for DRS; this reduces
(1 − )(pH − pL )
1− n mk v (mk ) = 1, (8)
n−1
p
where p = pL + (1 − )pH . Because the fraction is positive, it is
immediately apparent that v (mk ) > 1, so that mk is distorted below
the efficient level mFB . As is to be expected, the distortion increases
in the difference pH − pL and in the level of competition (captured
by n/(n − 1)).
3.2. The pooling equilibrium with positive DRS
3.2.1. Positive DRS of risk type
With positive DRS of the risk type itself and individual specific
cost, insurer k’s objective reads as
k = PLk (Lk − ak ) + (1 − )PHk H
k
, (9)
where now PLk contains g(ak ) and g(aj ) as given in (5).18
The solution to this objective yields a positive equilibrium
level of insurers’ expenditures on DRS, implicitly defined by
16
After a German sickness fund operating mainly in high cost areas went bankrupt
in 2011, members of this fund who then applied at other funds received phone calls
in which some of the insurers told them that they could not continue their drug
therapy or disease management program if they did not choose a different insurer;
see, e.g., Spiegel (2011).
17
See Lorenz (2014). There, we also illustrate all results with an example.
18
We assume Hk < 0, so that positive DRS is targeted only at the low risk. For a
very low level of competition (very high ), even the high risks would entail a profit
and positive DRS would be targeted at both risk types (albeit at different levels).
Table 2
Shares rs of the four types; positive correlation of H and O for ı > 0.
pL pH
Y LY =  + ı HY = (1 − ) − ı 
O LO = (1 − ) − ı HO = (1 − )(1 − ) + ı 1−
 1−
g  (ak )(Lk − ak ) = n/(n − 1).19 The larger g (a), i.e., the more effec-
tive insurers’ expenditures on risk selection are, the larger, c.p.,
the equilibrium level of ak will be. These expenditures affect the
distortion of the benefit level m, which is now given by
 2

(1 − )(pH − pL ) (1 − )(pH − pL ) k  k
1− n mk + n a v (m ) = 1.
n−1
p n−1
p
(10)
Because the last term in the brackets [·] is positive, compared with
(8), the condition if there is no DRS, v (mk ) has to decrease: mk
increases, so the distortion is reduced. More generally, the larger
the equilibrium level of ak , the larger mk .
Result 1. In the pooling equilibrium, the distortion of the benefit
level decreases in the level of positive DRS if cost for DRS is individual-
specific.
The incentive to distort the benefit level (with or without DRS)
arises because profit per high risk is lower than profit per low
risk, where the degree of the distortion depends on the difference
between these two profits, which, in the case without DRS, is given
by20 Lk − Hk = (pH − pL )mk .
With positive DRS, profit per low risk decreases because insurers
the difference between net profits (including ak ), and thereby the
incentive to distort the benefit level.
This is different for the case of non-individual-specific cost.21
Although in this case, expenditures for risk selection decrease total
profit, they do not specifically decrease profit per low risk, so the
difference between the risk-type-specific profits remains the same:
Therefore, positive DRS has no influence on the distortion of the
benefit level if cost is non-individual-specific. In the following, we
will therefore only consider individual-specific cost.
3.2.2. Positive DRS of a signal that is correlated with risk type
We now analyze the (probably more realistic) case that DRS is
not targeted at the risk type itself, but at a signal that is (less than
perfectly) correlated with risk type. We assume that there are two
signal types s = Y, O, young and old. The shares of the four types of
individuals, rs , are given in Table 2, where ı > 0 captures the case
of a positive correlation of high risk and old age.
This formulation of a positive correlation has the advantage that
increasing ı increases the level of correlation without altering the
shares of the two risk types,  and (1 − ), and the shares of the two
signal-types,  and (1 − ).
19
See Appendix A.2, where also condition (10) is derived.
20
See the second term in the brackets in condition (8), which contains (pH − pL )mk .
21
In this case, insurer k’s objective is k = (PLk Lk − ak ) + (1 − )PHk Hk .
 N. Lorenz / Journal of Health Economics 42 (2015) 81–89 85

With positive DRS of the young and individual-specific cost, so that the distortion of the benefit level is eliminated for
insurer k’s objective reads as22 (1 − ) H
  RAO = (p − pL )m. (15)
k k k k k ı
 = rs Prs r − rY PrY a , (11)
r s r
If there is perfect correlation, the share of the low risks equals the
share of the young, so  = ; in addition, the mass of individuals
k contains g(ak ) and g(aj ). The distortion of the benefit
where only PrY in the lower left and the upper right corner in Table 2 has to be
level is now determined by23 zero, which requires ı = (1 − ) = (1 − ). Replacing in (15) shows
 2
 that with perfect correlation the efficient benefit level is therefore
(1 − )(pH − pL ) (pH − pL ) k  k implemented for RAO = (pH − pL )m, which is just the cost difference
1− n mk + n ıa v (m ) = 1. (12)
n−1
p n−1
p between the two risk types. With less than perfect correlation,
ı < (1 − ), so there is overpayment, and the lower the level of
Because the last term in the brackets is positive, compared with no correlation (i.e., the lower ı), the larger this overpayment has to
DRS, v (mk ) has to decrease, so the distortion is reduced. However, be.
for a given level of ak , the reduction of the distortion is of course not
as large as with DRS against the risk type itself, since ı < (1 − ). 3.3.2. Optimal risk adjustment with positive DRS
If there is less than perfect correlation, expenditures on DRS of the With optimal risk adjustment there is overpayment for the old,
young not only increase the cost of the low risks, but also of the so this is the group positive DRS will be targeted at.25 Insurer k’s
high risks and therefore translate into a smaller reduction of the objective in this case is given by (13) with RAO replaced by RAO − ak
cost difference between the risk types. As is to be expected, for a and with PrO k containing g(ak ) and g(aj ) as given in (5). The posi-
given level of ak the reduction of the distortion increases in the level tive equilibrium level of ak then enters condition (14), where again
of correlation (i.e. in ı). RAO has to be replaced by (RAO − ak ).26 The efficient benefit level is
Result 2. In the pooling equilibrium, the distortion of the benefit therefore implemented with
level decreases in the level of positive DRS of a signal that is correlated (1 − ) H
with low risk if cost for DRS is individual-specific. A higher level of RAO = (p − pL )m + ak . (16)
ı
correlation (a higher ı) increases the effect of a given level of DRS on
the distortion of the benefit level. Because part of the overpayment for the old is spent on positive
DRS, the optimal overpayment for the old has to be raised by exactly
these expenditures, so that the net difference in payments (includ-
3.3. Implications of DRS on optimal risk adjustment in the ing ak ) equals the amount necessary to eliminate IRS as given in
pooling equilibrium (15).

We now discuss the implications of the interaction of direct and Result 3. With individual-specific cost, if there is positive DRS
indirect risk selection for optimal risk adjustment. As shown by regarding a signal that is used for risk adjustment and that is correlated
Glazer and McGuire (2000), if a regulator does not observe indi- with high risk, the optimal overpayment of that signal to eliminate IRS
viduals’ risk type, but only a signal that is correlated with risk type has to be increased by the expenditures for DRS.
(like age), it is still feasible to eliminate the distortion of the benefit
This result shows that the formula for optimal risk adjustment
package by overpaying for a signal that indicates a high risk, and
derived by Glazer and McGuire (2000) is not invalidated by DRS:
underpaying for a signal that indicates a low risk. In Section 3.3.1,
there is still overpayment for a signal that is correlated with high
we first show how to implement optimal risk adjustment in our
risk, and underpayment for a signal correlated with low risk. Also,
setting with two risk types and two age groups if there is no DRS;
DRS does not invalidate their claim that optimal risk adjustment
we then determine how the optimal payments have to be modified
can implement the efficient benefit level. However, the formula to
if there is DRS in Section 3.3.2.
implement the efficient benefit level has to be modified and include
insurers’ expenditures on DRS if cost is individual specific. Whether
3.3.1. Optimal risk adjustment without DRS these expenditures are negligible or significant is an empirical mat-
With risk adjustment, each insurer receives a payment of RAO ter, but, e.g., the findings of Starc (2014), who reports that insurers
for each insured that is old; these payments are financed by a risk spend a large part of potential profits on marketing and insurance
adjustment fee RAF which each insurer has to pay for each insured brokers, indicate that these expenditures may be substantial.
(including the old). The balanced budget constraint requires this fee
to be RAF = (1 − )RAO . The insurer’s objective with risk adjustment 3.4. More than two risk types
is then given by
  To keep the notation simple when deriving the results, so far
k = k
rY PrY (rk − RAF) + k
rO PrO (rk − RAF + RAO ). (13) we have considered the case of only two risk types. The results,
r r however, also hold for an arbitrary number of risk types. Let the
number of risk types be r and denote the probability of risk type r
Simplifying the optimality conditions now yields24
by pr and its share by r .
 2
 In the conditions determining the distortion of the benefit level,
(1 − )(pH − pL ) (pH − pL )
1− n mk + n ıRAO v (mk ) = 1, (14) conditions (8), (12) and (14), the term (1 − )(pH − pL ) then has to
2

n−1
p n−1
p r 2
be replaced by  (p − p) > 0; see Appendix B.1. This is just
r=1 r r
the variance of the illness probabilities, and the larger this variance,
the larger the distortion of m.
22
Like in Section 3.2.1, we assume that positive DRS is profitable for only one of
the two signal types; see footnote 18.
23 25
See Appendix A.3. As before we assume that DRS is only targeted at one of the signal types.
24 26
See Appendix A.4. See Appendix A.5.
86 N. Lorenz / Journal of Health Economics 42 (2015) 81–89

For this general case of an arbitrary number of risk types, we

define the shares of individuals (which were given in Table 2 for
the case of two risk types) by rY = r  + ır and rO = r (1 − ) − ır ,
see Table B.1 in Appendix B.2. It is then straightforward toshow that
there is a positive correlation of risk type and old age for r ır pr < 0;
ır has to be (mainly) positive for small illness probabilities and neg-
H L
ative for large illness probabilities. The term(p − p )ı in the last
fraction of (12) then has to be replaced by − r ır pr , so positive DRS
again reduces the distortion caused by IRS (see Appendix B.3).
To eliminate the incentives for IRS with optimal risk adjust-
ment, as before there has to be overpayment (see Appendix B.4)
and this overpayment has to be increased by insurers’ expenditures
for positive DRS (see Appendix B.5).

3.5. The pooling equilibrium with negative DRS

The discussion of negative DRS can be very brief because for all
the settings we considered, negative DRS creates just the opposite
effects of positive DRS.27 Negative DRS increases the loss associated
with the high risks, which increases the cost difference between the
two risk types and thereby increases the incentives for IRS.
Negative DRS regarding a signal that is used for risk adjustment
and for which there is underpayment increases the loss for this sig-
nal, which allows keeping the underpayment less pronounced than
without DRS. We can therefore summarize the results for negative
DRS as follows:
Result 4. In the pooling equilibrium and with individual specific cost,
the distortion of the benefit level increases in the level of negative
DRS against the high risk type or against a signal that is correlated
with low risk. If there is negative DRS regarding a signal that is used
for risk adjustment and that is correlated with low risk, the optimal
underpayment of that signal to eliminate IRS has to be reduced by the
expenditures for DRS.
4. DRS in the separating equilibrium
4.1. The separating equilibrium without DRS
In Lorenz (2013), it is shown that the separating equilibrium
arises for a high level of competition (i.e. a low level of ) and that
it is similar, but not identical, to the Rothschild–Stiglitz equilibrium
under perfect competition.28 Both equilibria can be found in Fig. 1.
Under perfect competition, the separating equilibrium consists
of contract B, chosen by the high risks, and contract A1 , chosen by
the low risks, where A1 is located at the intersection of the low
B distribution function can be depicted by a shaded area around the
risks’ iso-profit line pL and the high risks’ indifference curve I VH
associated with contract B (so that the incentive compatibility con-
straint is satisfied); in addition, both iso-profit lines pass through
the origin and are in this case zero-profit lines.29 The number of
insurers offering the contract for the low risks (insurers of type A)
and for the high risks (insurers of type B), nA and nB respectively,
are indeterminate under perfect competition.30
This is different under imperfect competition, where nA and nB
are implicitly defined by the profit equality condition A = B : total
profit for an insurer of type A has to be the same as for an insurer of
27
This is straightforward to show by replacing ak by bk and +g(a) by −f(b) in all
the optimization problems and the derivations we derived in the appendix; see also
Lorenz (2014).
28
The separating equilibrium for a low level of  exists under the same condition
as the Rothschild–Stiglitz equilibrium: the share of low risks must not be too large.
29
See Zweifel et al. (2009), chapter 7.
30
However, Olivella and Vera-Hernandez (2010) have shown that the smallest
number of insurers supporting this equilibrium is n = 3, with nB = 1 and nA = 2.
Fig. 1. Separating equilibrium: contracts B and A3 are offered.
type B. If, e.g., B > A , it would be profitable for an insurer of type
A to become of type B; increasing nB decreases the market share of
each insurer of type B, which decreases B . In equilibrium, profit
per high risk equals nB /(nB − 1), so the iso-profit line associated
with contract B starts at that level, see Fig. 1. However, while B
is indeed one of the equilibrium contracts, A1 is not. To simplify
the exposition of which contract is offered for the low risks, in the
following we will simply speak of ‘insurer A’ and ‘insurer B’, i.e.,
present the explanation as if there was only one insurer of type A
and one insurer of type B.
B
Because B and A1 are both located on the indifference curve I VH ,
these two benefit-premium bundles provide the same utility to the
high risk, i.e., VHB = VHA1 ; therefore, all the high risks with εAi > εBi
will choose insurer A. From the perspective of insurer A who tries
to avoid being chosen by the high risks by offering a contract that
satisfies the incentive compatibility constraint, what is important
B
is not only the indifference curve I VH , but also the additional utility
j
components εi , in particular, the difference εAi − εBi . For some of the
high risks, this difference will be negative, for others somewhat,
or even considerably, larger than zero. There is therefore no single
incentive compatibility constraint (ICC), but an individual-specific
ICC for each of the high risks depending on this difference.
One way to capture these individual-specific ICCs is by a dis-
tribution function denoting the share of the high risks for which
the ICC is violated; formulated in this way, this distribution func-
tion coincides with insurer A’s market share among the high risks
as a function of VHA given VHB , i.e., PHA = PHA (VHA |VHB ). Graphically, this
B
indifference curve I VH , representing the density of this distribution
(with the darkness of the shaded area as a measure of this density,
which is bell-shaped). Above the shaded area, the ICC is satisfied for
all the high risks, so PHA = 0. Within the shaded area, PHA increases
in VHA according to the density ∂PHA /∂VHA = PHA (1 − PHA )/. Below the
shaded area, PHA = 1, because there the ICC is violated for all the
high risks.
Insurer A, when offering A2 , would therefore not be chosen by
any of the high risks. However, he can still increase his profit by
moving along the indifference curve of the low risks to the right,
which would have two effects31 : on the one hand, it increases profit
per low risk because the indifference curve has a larger slope than
pL ; on the other hand, it also increases the number of the high risks
31
We consider a movement along the indifference curve of the low risks because
with more than one insurer of type A, this would leave this insurer’s market share
among the low risks unaffected.
 N. Lorenz / Journal of Health Economics 42 (2015) 81–89
Fig. 2. Separating equilibrium with (a) positive or (b) negative DRS.
choosing this contract. Because at the boundary of the shaded area
this second effect is of second order (since the density is about zero
if PHA ≈ 0), the two effects balance somewhere inside the shaded
area, represented by contract A3 .32 Because a small share of the
high risks choose contract A3 , this contract is somewhat above the
iso-profit line pL (which would apply if only the low risks choose
insurer A).
Due to the ‘stochastic nature’ of the incentive compatibility con-
straint, the separating equilibrium under imperfect competition is
therefore not perfectly separating. Instead, a small share of the high
risks chooses the contract designated for the low risks, but none of
the low risks choose the contract designated for the high risks.
4.2. The separating equilibrium with DRS
4.2.1. The separating equilibrium with positive DRS
For positive DRS, we have to distinguish whether the utility
increase g(a) will only (or at least primarily) be experienced by the
low risks (as, e.g., might be the case with a discount for a fitness
club membership) or equally by both risk types (as, e.g., might be
the case with advertising). We begin with the case that the utility
increase can only be experienced by the low risks. Positive DRS will
than only be performed by insurers of type A.33
The objective of insurers of type A with positive DRS of the low
risks equals the objective as given in (9). The solution to this objec-
tive determines a positive equilibrium level of expenditures on
DRS, which increases the premium charged by these insurers.34
In Fig. 2(a), this increase in RA can be shown by an upward shift of
the corresponding iso-profit line. As is apparent, insurers of type
A can then increase their benefit level before attracting the same
share of the high risks as without DRS. Because a higher bene-
fit level (accompanied by the according increase in the premium,
pL mA ) increases the utility of the low risks, insurers will offer this
higher benefit level (if there is some competition), so that the new
equilibrium will be a contract like A4 . Since positive DRS reduces
the attractiveness of the contract offered by insurers of type A for
the high risks in the premium-dimension, they can increase the
attractiveness of their contract in the benefit level-dimension.
32
See Appendix C.1; for a more detailed derivation of this equilibrium, see Lorenz
(2013).
VA
33
Because contract B is far above the indifference curve I L associated with con-
tract A3 , for the low risks there is a huge difference in utility between A3 and B. Any
moderate increase in the (perceived) utility of contract B due to positive DRS of insur-
ers of type B will reduce this difference only to a small degree and not induce any
of the low risks to choose this contract. Because insurers of type B cannot increase
their market share among the low risks (which remains at zero), they will refrain
from positive DRS.
34
See Appendix C.2.
87
Result 5. In the separating equilibrium, the distortion of the bene-
fit level is reduced if there is positive DRS of the low risk and cost is
individual-specific.
We now turn to the case that both risk types can experience the util-
ity increase g(a). In this case, DRS will be performed by all insurers:
insurers of type A engage in DRS of the low risks and insurers of type
B in DRS of the high risks (since high risks are profitable to insurers
of type B). Because the equilibrium level of expenditures on DRS in
this case is the same for both types of insurers,35 all insurers raise
their premium by the same amount. In Fig. 2(a), both iso-profit lines
are then shifted upwards by the same distance, so there is no effect
on the benefit levels mA and mB .
However, it seems reasonable to assume that positive DRS is
at least somewhat more effective when targeted at the low risks
than when targeted at the high risks. In equilibrium, insurers of
type A will spend more on DRS and raise their premium by a higher
amount than insurers of type B. The larger upward shift of their
iso-profit line then allows insurers of type A to increase their ben-
efit level (just as in the case where the iso-profit line of insurers
of type B is not shifted at all). Therefore, as long as positive DRS
reduces the attractiveness of the contract offered by insurers of
type A relative to the contract offered by insurers of type B in the
premium-dimension, insurers of type A can increase the attractive-
ness of their contract in the benefit level-dimension.
4.2.2. The separating equilibrium with negative DRS
Negative DRS will only be performed by insurers of type A,
because only these insurers try to avoid being chosen by the high
risks. As in Section 4.1, we explain the effect of negative DRS
as if there was only one insurer of type A and one insurer of
type B.
With negative DRS, a high risk chooses insurer A if VHA − f (bA ) +
εAi > VHB + εBi . Because negative DRS reduces the utility as perceived
by the high risks by f(bA ), VHA can be increased by exactly that
amount without altering the number of the high risks choosing
insurer A. This allows insurer A to increase his benefit level, see
Fig. 2(b). There, AL5 denotes a contract offered by insurer A as
perceived by the low risks, while AH 5
denotes the same contract as
perceived by the high risks. Compared to AL5 , AH 5
is shifted upwards
by f(bA ), the utility decrease of negative DRS (measured in mone-
tary terms). The larger bA , the larger f(bA ), and therefore the larger
mA can be without increasing the share of high risks choosing
insurer A. Because this effect occurs regardless of whether cost
for DRS is individual-specific or not, there exists one case where
35
This can be seen by replacing Hk by (Hk − ak ) in (6) for insurers of type B and
comparing the respective optimality condition with the one for insurers of type A.
88 N. Lorenz / Journal of Health Economics 42 (2015) 81–89

DRS influences the distortion caused by IRS even if cost is non- Appendix A.
individual-specific.
A.1. The pooling equilibrium without DRS
Result 6. In the separating equilibrium, negative DRS against the
high risks reduces the distortion of the benefit level of the low risks Using the property that the derivative of Prk with respect to Vrk
∂Prk P k (1−P k )
regardless of whether cost for negative DRS is individual-specific or can be expressed in terms of Prk itself as = r  r , the FOCs to
∂Vrk
not.
objective (6) are given by

The main mechanism in this last case is therefore different from all
   
the other cases we have considered. Here, IRS is substituted by DRS, ∂ k P k (1 − PLk ) k P k (1 − PHk ) k
while in all the other cases, IRS affects risk type specific profits and = − L L + PLk + (1 − ) − H H + PHk
∂ Rk  
thereby the gains from IRS.
=0 (17)

4.3. Optimal risk adjustment in the separating equilibrium  

∂ k PLk (1 − PLk )  k
= pL v L − PLk
With optimal risk adjustment, profit per high risk equals ∂m k 
profit per low risk, so insurers’ incentives to distort the benefit  
package are eliminated and the separating equilibrium turns into PHk (1 − PHk )
+ (1 − )pH v Hk − PHk = 0. (18)
a pooling equilibrium where all insurers offer the same benefit 
premium bundle. The impact of DRS on optimal risk adjustment
j
in the pooling equilibrium has, however, already been derived in Using the fact that in equilibrium Pr = (1/n) ∀ j, condition (17)
Section 3.3.2. yields (7). Solving (17) for Rk = (n/(n − 1)) + pmk and inserting
in (18) yields (8).

5. Conclusion A.2. Positive DRS of risk type in the pooling equilibrium

In this paper, the interaction of direct and indirect risk selec- The FOC of (9) with respect to ak is
tion (DRS and IRS) has been analyzed. It has been shown that DRS,  
∂ k PLk (1 − PLk )  k
using measures unrelated to the benefit package, nevertheless has = g (a )(Lk − ak ) − PLk = 0. (19)
an influence on the distortions of the benefit package caused by IRS. ∂ak 
If cost for DRS is (at least to some degree) individual-specific, DRS
j
selectively reduces the profit per individual of the risk type it is tar- With Pr = (1/n) ∀ j, this can be simplified to g  (ak )(Lk − ak ) =
geted at. Positive DRS therefore reduces and negative DRS increases n/(n − 1), which implicitly defines a positive level of ak , because
the difference in profits between the low and the high risks. Because lim g  (a) → ∞. Replacing Lk by (Lk − ak ) in (17) and (18), solving
in the pooling equilibrium the degree of the distortion depends a→0+

on the difference between these profits, positive DRS reduces the
distortion of the benefit level, while negative DRS increases it. In
the separating equilibrium, DRS can act as a substitute for IRS and
thereby reduce the distortion of the benefit level. In addition, it
has been shown that the effects of DRS on type- and signal-specific
profits also have an influence on the formula for optimal risk adjust-
ment: The over- and underpayments have to be inflated by insurers’
expenditures on positive DRS and reduced by their expenditures on
negative DRS.
We have derived these results for a setting where negative DRS
occurs during the application process, but the main mechanisms
should also hold if it is targeted at the high risks who already hold
a contract with the insurer. If an insurer attracts a larger share of
the high risks, this will increase the cost of negative DRS, even if the
activity of risk selection and the cost associated with it occur only
later (when the insurer tries to induce the high risks to switch to
another insurer). With negative DRS, high risks are more expensive
than without DRS; anticipating these additional cost, insurers will
therefore not make their contract as attractive for this risk group
as they would without DRS. Therefore, negative DRS will increase
the distortion in such a setting as well.
(17) for Rk and substituting in (18) then yields (10).
A.3. Positive DRS of a signal that is correlated with risk type
The FOC of (11) with respect to ak defines a positive equilibrium
level of ak . Solving the FOC with respect to Rk for Rk = (n/(n −
1)) + pmk + ak and inserting in the FOC with respect to mk yields
(12).
A.4. Optimal risk adjustment without DRS
Solving the FOC of (13) with respect to Rk for Rk = (n/(n − 1)) +
pmk + RAF − (1 − )RAO and substituting in the FOC with respect to
mk (using the definitions of rs as given in Table 2) yields (14).
A.5. Optimal risk adjustment with DRS
The derivation of condition (16) is identical to the derivation of
condition (14) described in Appendix A.4; the only difference is that
RAO has to be replaced by (RAO − ak ).

Appendix B and C. Supplementary data

Acknowledgements
Supplementary data associated with this article can be found, in
I thank Friedrich Breyer and Esther Schuch for helpful comments the online version, at http://dx.doi.org/10.1016/j.jhealeco.2014.12.
and suggestions. 003
 N. Lorenz / Journal of Health Economics 42 (2015) 81–89
References
Bauhoff, S., 2012. Do health plans risk-select? An audit study on Germany’s
Social Health Insurance. Journal of Public Economics 96 (9–10),
750–759.
Bijlsma, M., Boone, J., Zwart, G., 2011. Competition leverage: how the demand side
affects optimal risk adjustment. TILEC Discussion Paper, 2011-039.
Breyer, F., Bundorf, M.K., Pauly, M.V., 2011. Health care spending risk,
health insurance, and payment to health plans. In: Pauly, M.V., McGuire,
T.G., Barros, P.P. (Eds.), Handbook of Health Economics, vol. 2. Elsevier,
Amsterdam, pp. 691–762.
Brown, J., Duggan, M., Kuziemko, I., Woolston, W., 2011. How does risk selection
respond to risk adjustment? Evidence from the Medicare Advantage Program.
NBER Working Paper 16977.
Cao, Z., McGuire, T.G., 2003. Service-level selection by HMOs in Medicare. Journal of
Health Economics 22 (6), 915–931.
Eggleston, K., 2000. Risk selection and optimal health insurance-provider payment
systems. Journal of Risk and Insurance 67 (2), 173–196.
Ellis, R.P., McGuire, T.G., 2007. Predictability and predictiveness of health care spend-
ing. Journal of Health Economics 26 (1), 25–48.
Frank, R.G., Glazer, J., McGuire, T.G., 2000. Measuring adverse selection in managed
health care. Journal of Health Economics 19 (6), 829–854.
Glazer, J., McGuire, T.G., 2000. Optimal risk adjustment in markets with adverse
selection: an application to managed care. American Economic Review 90 (4),
1055–1071.
Glazer, J., McGuire, T.G., 2002. Setting health plan premiums to ensure efficient qual-
ity in health care: minimum variance optimal risk adjustment. Journal of Public
Economics 84 (2), 153–173.
Jack, W., 2006. Optimal risk adjustment with adverse selection and spatial compe-
tition. Journal of Health Economics 25 (5), 908–926.
Lorenz, N., 2013. Adverse selection and risk adjustment under imperfect competi-
tion. Trier University Working Paper 5/13. http://ideas.repec.org/p/trr/wpaper/
201305.html
89
Lorenz, N., 2014. The interaction of direct and indirect risk selection. Trier University
Working Paper 12/14. http://ideas.repec.org/p/trr/wpaper/201412.html
McGuire, T.G., Glazer, J., Newhouse, J.P., Normand, S.-L., Shi, J., 2013. Integrating risk
adjustment and enrollee premiums in health plan payment. Journal of Health
Economics 32 (6), 1263–1277.
Newhouse, J.P., Price, M., Huang, J., McWilliams, J.M., Hsu, J., 2012. Steps to reduce
favorable risk selection in Medicare Advantage largely succeeded, boding well
for Health Insurance Exchanges. Health Affairs 31 (12), 2618–2628.
Olivella, P., Vera-Hernandez, M., 2010. How complex are the contracts offered by
health plans? SERIEs 1 (3), 305–323.
Rothschild, M., Stiglitz, J., 1976. Equilibrium in competitive insurance markets: an
essay on the economics of imperfect information. Quarterly Journal of Eco-
nomics 90 (4), 629–649.
Shen, Y., Ellis, R.P., 2002. How profitable is risk selection? A comparison of four risk
adjustment models. Health Economics 11 (2), 165–174.
Shi, J., 2013. Efficiency in plan choice with risk adjustment and premium discrimi-
nation in Health Insurance Exchanges. Unpublished.
Spiegel, 2011. Fragwürdige Beratung: Krankenversicherer wimmelt Senioren ab.
Spiegel, May 9, 2011. http://www.spiegel.de/wirtschaft/soziales/fragwuerdige-
beratung-krankenversicherer-wimmelt-senioren-ab-a-761384.html
Starc, A., 2014. Insurer pricing and consumer welfare: evidence from Medigap. RAND
Journal of Economics 45 (1), 198–220.
Train, K.E., 2009. Discrete Choice Methods with Simulation, 2nd ed. Cambridge Uni-
versity Press, New York.
van de Ven, W.P., Ellis, R., 2000. Risk adjustment in competitive health plan markets.
In: Culyer, A.J., Newhouse, J.P. (Eds.), Handbook of Health Economics. Elsevier,
Amsterdam, pp. 755–845.
van de Ven, W.P., van Vliet, R.C., 1992. How can we prevent cream skimming in
a competitive health insurance market? The great challenge for the 90’s. In:
Zweifel, P., Frech, H. (Eds.), Healtheconomics Worldwide. Kluwer Academic Pub-
lishers, Dordrecht, Boston, London, pp. 23–46.
Zweifel, P., Breyer, F., Kifmann, M., 2009. Health Economics, 2nd ed. Springer, Berlin,
Heidelberg.
 Journal of Health Economics 42 (2015) 90–103

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Environmental regulations on air pollution in China and their impact

on infant mortality
Shinsuke Tanaka ∗
Tufts University, United States

a r t i c l e i n f o a b s t r a c t

Article history: This study explores the impact of environmental regulations in China on infant mortality. In 1998, the
Received 27 January 2014 Chinese government imposed stringent air pollution regulations, in one of the first large-scale regulatory
Received in revised form 24 February 2015 attempts in a developing country. We find that the infant mortality rate fell by 20 percent in the treatment
Accepted 26 February 2015
cities designated as “Two Control Zones.” The greatest reduction in mortality occurred during the neonatal
Available online 6 March 2015
period, highlighting an important pathophysiologic mechanism, and was largest among infants born
to mothers with low levels of education. The finding is robust to various alternative hypotheses and
JEL classification:
specifications. Further, a falsification test using deaths from causes unrelated to air pollution supports
Q56
I18
these findings.
Q53 © 2015 Elsevier B.V. All rights reserved.
I12
O13

Keywords:
Infant mortality
Air pollution
Environmental regulation
China

1. Introduction themselves to a legal framework for reducing pollution emissions.

There is little disagreement that air pollution poses a major envi-
ronmental risk to human health. Improved air quality worldwide
is correlated with the amelioration of numerous health problems,
including respiratory infections, cardiovascular diseases, and lung
cancer.1 Developing countries rank highest in air pollution world-
wide, and children under age five in these countries are considered
to be the most vulnerable population. Elevated air pollution is
generally beyond the scope of individual control and falls to the
public sector. However, environmental regulations on air pollution
are extremely contentious. This was evident at the 2009 United
Nations Climate Change Conference, also known as COP15, held in
Copenhagen, where many developing countries refused to commit
∗ Correspondence to: The Fletcher School, Tufts University, 160 Packard Avenue,
Medford, MA 02155, United States. Tel.: +1 6176274619.
E-mail address: Shinsuke.Tanaka@tufts.edu
1
A sizable literature documents health risks associated with air pollution expo-
sure. See, for example, Schwartz et al. (1996), Levy et al. (2000), Samet et al. (2000),
Chay and Greenstone (2003a), Currie and Neidell (2005), and Currie et al. (2009),
and Arceo et al. (2012).
Their opposition is largely due to the prevailing concern that the
economic costs associated with pollution abatement may outweigh
the health benefits. Accordingly, air pollution regulations are still
rare in developing countries, and whether, and to what extent,
environmental regulations on air pollution lead to health benefits
remains an important question yet to be answered.
In this paper, we examine the effect of environmental regula-
tions pertaining to air pollution on infant mortality in China. As
China’s economy continued to grow at unprecedented rates for the
last several decades, ambient air quality deteriorated to one of the
worst levels in the world due to its heavy reliance on coal-fired
energy generation. In 1998, the Chinese government imposed strin-
gent regulations on pollutant emissions from power plants, in one
of the first attempts on such a large scale in developing countries.
This so-called the Two Control Zone (TCZ) policy designated nearly
175 prefectures exceeding the nationally mandated pollution
standards as the TCZ. In these areas, the power industry, which con-
tributed more than 90 percent of air pollution, was forced to reduce
emissions and install new pollution control technologies, while also
shutting down a massive number of small, inefficient power plants.
For our purpose, the TCZ policy provides a quasi-experimental

http://dx.doi.org/10.1016/j.jhealeco.2015.02.004
0167-6296/© 2015 Elsevier B.V. All rights reserved.
 S. Tanaka / Journal of Health Economics 42 (2015) 90–103
environment, wherein the intensity of exposure to the regulations
can be defined by the TCZ regulatory status, and we are able to com-
pare changes in infant mortality rate (IMR) before and after the pol-
icy reform, between the cities assigned and not assigned as the TCZ.
To implement the analysis, we draw IMR data from the Chi-
nese Disease Surveillance Points (DSP) system that collected birth
and death registrations for 145 nationally representative sites from
1991 through 2000. IMR, defined as the number of infant deaths
under age one per 1000 live births in a given year, is available for
each DSP site by year level, linked with detailed information on
birth characteristics and parental attributes. We match this dataset
to the TCZ regulatory status assigned to individual cities, based on
the governmental report, and thereby estimate the treatment effect
of the regulations.
We find that the air pollution regulations led to significant
reductions in infant mortality. The difference-in-differences esti-
mates suggest that the regulations have led to 3.29 fewer infant
deaths per 1000 live births than would have occurred in the absence
of the regulations. This corresponds to a 20 percent reduction
in IMR. 63 percent of the reduction in infant mortality occurred
during the neonatal period, highlighting an important pathophysi-
ologic mechanism, and the greatest reduction of mortality occurred
among children born to mothers with low educational levels.
A major methodological challenge, however, is that the TCZ des-
ignation rule may not be orthogonal to unobserved characteristics
that contribute to reductions in air pollution and infant mortal-
ity. The present study conducts a number of robustness checks
and a falsification test to address this issue. First, we confirm that
the TCZ status has little association with changes in observable
covariates, assuring that there is no systematic difference in con-
current trends in observable characteristics between the TCZ and
non-TCZ cities. Although this is not a direct test of the exclusion
restriction, since it requires that TCZ status not be correlated with
trends in unobservable factors, this result leads us to believe that
the treatment effect is less likely to be confounded by differential
trends in unobservable factors as well (Altonji et al., 2005). Second,
the regression is also directly adjusted for differential pre-existing
trends in mortality, yet the estimates are essentially unaffected.
Lastly, the policy had no impact on infant deaths due to acciden-
tal causes. The absence of a causal mechanism linking air pollution
to these causes of death serves as falsification evidence, suggest-
ing that differences in access to or quality of medical services and
technologies cannot be the sources of bias. Overall, there is no evi-
dence that the estimates are driven by inappropriate identification
assumptions, leading us to believe that the treatment effect based
on the TCZ status is indeed not spurious but causal.
This study makes three major contributions to the existing liter-
ature. First, by exploiting regulation-induced changes in air quality,
it addresses a policy-relevant question: to what extent do envi-
ronmental regulations in developing countries lead to reductions
in infant mortality? Several prior studies have focused on vari-
ation in air quality induced by recession (Chay and Greenstone,
2003a), weekly fluctuations (Currie and Neidell, 2005), wildfires
(Jayachandran, 2009), or wind directions (Luechinger, 2014). Chay
and Greenstone (2003b) provide compelling evidence for the link-
age between the Clean Air Act of 1970 and infant mortality in the
U.S. It remains to be determined, however, whether, and how effec-
tively, environmental regulations can improve human health in
developing countries.2 A recent study by Greenstone and Hanna
2
Extrapolating evidence in developed countries to developing countries is dif-
ficult for a number of reasons. First and foremost importantly, we do not know
whether environmental regulations in developing countries had any impact in
reducing pollution due to weaker implemental mechanisms and an enforcement
91
(2011) examines regulations on air pollution and water pollution
in India since 1987. They find these regulations efficacious in reduc-
ing air pollution, but such reductions led to modest and statistically
insignificant reductions in infant mortality. Our study provides
an interesting contrast, finding that infant mortality significantly
responded to the environmental regulation. Further, the regulation
we focus on targeted coal for energy generation, which is the major
contributor to air pollution in many other developing countries as
well, whereas Greenstone and Hanna (2011) focus on vehicular
pollution.3,4 The findings in this study accordingly present relevant
estimates for the effect of environmental regulations in develop-
ing countries implementing similar policies on coal in the power
industry.
Second, the present study contributes to our understanding
of the relationship between air pollution and infant mortality at
greater concentration levels. Previous evidence is predominantly
derived from the United States or other developed countries, where
pollution is relatively low.5 Since we know little about the shape of
the dose–response relationship, it is consequently difficult to pre-
dict the marginal impact of pollution reduction in the presence of
non-linear relationship.6 Air pollution in China is one of the highest
in the world. Its total suspended particulates (TSP) level in 1995 was
four times higher than the WHO standards, and four times higher
than the level in the United States in 1970, when the Clean Air
Act was amended, as examined in Chay and Greenstone (2003b).
Thus, estimates in China provide compelling evidence applicable to
the distinctive context of developing countries where air pollution
levels are relatively high.
Third, there is extensive literature showing differential patterns
according to socioeconomic status, yet it is still an open question as
to whether air pollution also exhibits differential impact on infant
mortality (Currie and Hyson, 1999; Case et al., 2002; Jayachandran,
2009). While infants in poor countries are considered to be the
most susceptible to the effect of pollution, not only because of
high pollution levels but also because families lack the resources
or knowledge necessary to avoid exposure, the impacts may be
small if air pollution does not have first-order effect on them.7 Thus,
the present study helps identify vulnerable population in designing
policies.
The current research design has several advantages over previ-
ous studies. First, this study focuses on infants, not only because
system. Second, even if pollution is successfully reduced to some extent, infant mor-
tality may not fall if a concave relationship between mortality and pollution level
leads to low marginal pollution effect at high concentration levels. Third, magni-
tudes of impacts in reducing air pollution should be greater if people have limited
access to medical services, initially have lower health status, and/or have limited
knowledge in avoiding pollution.
3
Since the most heavily affected industry is the power industry, which was a
driving force behind China’s rapid economic growth, the findings in this study
accordingly highlight the important tradeoffs among economic growth, environ-
mental quality, and human health. See Tanaka et al. (2014) for the impact of the
environmental regulation on industrial performance.
4
A relatively smaller-scale air quality regulatory regime, targeting a different
industry, in another developing country, can be found in the Indian transportation
sector, which was mandated to use compressed national gas vehicles in Delhi during
working hours. Kumar and Foster (2007) show its effect on respiratory health.
5
Examining the pollution effect at low levels, especially levels lower than what
is often considered to be the standard, is also an interesting question in itself.
Currie and Neidell (2005) find that CO has significant impact on infant mortality
in California over the 1990s at relatively low levels.
6
For example, Arceo et al. (2012) show a non-linear relationship between CO
and health in Mexico. Evidence in developed countries may understate the impact
of pollution reduction in developing countries if the marginal impact of pollution is
higher at greater concentration levels.
7
For example, people with poor health tend to stay indoor with little exposure
to pollution. Children from rich households, who tend to have better health, may be
exposed more to pollution if they are more likely to have outside activities.
92 S. Tanaka / Journal of Health Economics 42 (2015) 90–103
they are particularly vulnerable to air pollution due to their weak
respiratory system, but because focusing on infant mortality mit-
igates complicating factors associated with adult mortality. For
example, adult deaths correlate more closely to chronic disease
conditions than to acute changes in air quality. In addition, adults
may migrate into less polluted areas. Addressing infant mortality
circumvents these issues, if not completely, because it is relatively
less difficult to identify causes of death during the first year of
birth, and because migration rates are low for pregnant women and
infants. Lastly, China is not only one of the most polluted countries
but also one of the first developing countries to regulate air pollu-
tion on such a large-scale. It is evident that China serves as a rare
research environment in which to assess the impact of environ-
mental regulations at greater concentration levels.
The rest of the paper is organized as follows. Section 2 provides
the historical background on air pollution and national air pollution
regulations in China. Section 3 describes the data and the descrip-
tive statistics. Section 4 presents the econometric framework and
its validity, and Section 5 presents empirical results. Section 6 con-
cludes.
2. Background on air pollution and regulations in China
2.1. Brief history
China is infamous for its air pollution, due to emissions from a
power sector that relies heavily on coal to generate electric power.
As the world’s largest coal producer, China possesses abundant
and relatively cheap coal, which constitutes the country’s primary
energy resource endowment, accounting for 75.5 percent of total
energy production in 1995 (National Bureau of Statistics of China,
2006). However, coal generally emits more pollutants than other
fossil fuels. As China underwent rapid economic growth, total SO2
emissions increased from 18.4 million tons in 1990 to 23.7 million
tons in 1995, and the ambient air pollution rose to levels detri-
mental to human health (State Environmental Protection Agency
[SEPA], 1996).
Fig. 1 illustrates the world distribution of TSP (Panel A) and SO2
(Panel B) concentration levels in 1995. The TSP level in Beijing, the
capital city of China, was 377 ␮g/m3 , almost four times higher than
the WHO guideline of 90 ␮g/m3 , and its SO2 concentration level was
90 ␮g/m3 , almost double the WHO guideline of 50 ␮g/m3 (WHO,
2002). SO2 is also an important precursor of acid rain. From the
1980s to the mid-1990s, the area of China experiencing acid rain
expanded by more than 1 million km2 (Yang and Schreifels, 2003).
During that decade, elevated air pollution gave rise to increas-
ing public concern about adverse impacts on human health.8
In response, the Chinese government formulated a series of
8
It is generally known that the smaller a particulate, the more detrimental it is
to health. For example, PM10 or PM2.5 , the particles with a diameter of 10 or 2.5
micrometers or less, respectively, or toxic gas, such as SO2 , are considered to be
the most hazardous because, when inhaled, these particulate matters or gas can
penetrate deep into the lungs and interfere with internal gas exchange. Further,
Laden et al. (2000) find that fine particles emitted from combustion sources (i.e.,
motor vehicles or coal combustion) have a stronger association with mortality than
those from non-combustion sources. Alternatively, SO2 becomes sulfuric acid when
it interacts with water, which is the main component of acid rain that may have a
direct or indirect impact on health. Yet, epidemiological evidence of the impact of
SO2 on mortality in developed countries is somewhat mixed. While Mendelsohn and
Orcutt (1979) show close associations between the two, SO2 is also considered a less
important determinant of mortality (Schwartz and Marcus, 1990; Nielsen and Ho,
2007). Hedley et al. (2002) is one of the few intervention studies that investigates
changes in SO2 caused by an overnight restriction on all power plants and road
vehicles in Hong Kong using fuel oil with a sulfur content of more than 0.5 percent.
They found that the intervention resulted in an immediate reduction in ambient
SO2 concentrations and a reduction in death rates particularly due to respiratory
environmental regulatory policies, the first version of which
was enacted in 1987, known as the Air Pollution Prevention and
Control Law (APPCL). This original law, however, failed to reduce
air pollution, mainly because it excluded the power sector, the
major contributor of SO2 emissions (Qian and Zhang, 1998). Even
worse, SO2 emissions continued to surge, and areas affected by
acid rain expanded.
APPCL was consequently amended in 1995. The major part of
the amendment was to include a section to regulate pollutant emis-
sions and coal combustion, particularly regarding the usage of high
sulfur-content coal, at power plants (Hao et al., 2007). Although the
1995 APPCL still had a weak enforcement mechanism and limited
efficacy, a prominent feature of the amendment was to propose a
future regional strategy, which would identify priority regions to
improve air quality and prevent the spread of acid rain.9,10
This was officially approved and implemented as the Two Con-
trol Zone (TCZ) policy in January 1998 (State Council, 1998). This
legislation designated prefectures exceeding nationally mandated
thresholds as either acid rain control zone or SO2 pollution control
zone.11 Based on the records in preceding years12 , prefectures were
designated as SO2 pollution control zone if;
• Average annual ambient SO2 concentrations exceeded the Class
II standard,13
• Daily average concentrations exceeded the Class III standard, or
• High SO2 emissions were recorded.14
Alternatively, prefectures were designated as acid rain control
zone if
• Average annual pH values for precipitation were less than or
equal to 4.5,
• Sulfate deposition was greater than the critical load, or
• High SO2 emissions were recorded.
In total, 175 prefectures out of 333 prefectures across 27
provinces were designated as TCZs. They accounted for 40.6 per-
cent of its population, 62.4 percent of GDP, and 58.9 percent of
the total SO2 emissions in 1995 (Hao et al., 2001). The SO2 pollu-
tion control zone was concentrated in the north due to high SO2
emissions for heating,15 whereas the acid rain control zones were
primarily in the south, where heat, humidity, and solar radiation
combine to create high atmospheric acidity. Hence, acid rain in the
and cardiovascular reasons. See also Aunan and Pan (2004) and Matus et al. (2012)
for health effects of air pollution in China.
9
Article 27 of the 1995 APPCL stipulates: “The environmental protection depart-
ment under the State Council together with relevant departments under the State
Council may, in light of the meteorological, topographical, soil and other natural
conditions, delimit the areas where acid rain has occurred or will probably occur
and areas that are seriously polluted by sulfur dioxide as acid rain control areas and
sulfur dioxide pollution control areas, subject to approval by the State Council.”
10
It is a standard practice of policy experimentation in China to implement strate-
gies in a particular region or for a set of time period, attempting to demonstrate their
effectiveness before expanding their implementation to the entire nation.
11
In this sense, the legislation can be considered to be parallel to the attainment
and nonattainment county designation by the Clean Air Acts in the United States.
12
The original document does not specify exactly which years of records they refer
to.
13
According to the Chinese National Ambient Air Quality Standards (CNAAQS)
for SO2 , Class I standard designates an annual average concentration level not
exceeding 20␮g/m3 , Class II ranges 20 ␮g/m3 < SO2 < 60 ␮g/m3 , and Class III ranges
60 ␮g/m3 < SO2 < 100 ␮g/m3 . Cities should meet Class II, which is considered to be
less harmful.
14
The original document does not specify the levels of SO2 emissions that are
considered to be “high.”
15
See Almond et al. (2009) for the impact of heating policy, which created a
discrepancy in air quality north and south of the Huai River.
 S. Tanaka / Journal of Health Economics 42 (2015) 90–103 93

Fig. 1. Air pollution across countries Notes: These figures present the world distribution of the TSP in Panel A and SO2 in Panel B in 1995.
Source: World Bank (1998).

south cannot necessarily be attributed to SO2 emissions traveling
down from the north, but is rather due to local emissions. This is
even more evident because acid deposition is the greatest in the
summer, when wind direction is generally south to north.
The TCZ status enforced more stringent regulations mandating
the use of less high-sulfur coal and the development of clean coal
technology. For example;
• All new and renovated power plants are required to use coal with
less than 1 percent sulfur content.
• Existing power plants using coal with sulfur content above 1
percent are required to install flue gas desulfurization (FGD)
equipment.
2.2. Effectiveness of TCZ policy

• No new coal mines producing coal with a sulfur content higher
than 3-percent can be established, and existing mines that pro-
duce such coal must gradually be shut down or reduce output.
• Construction of any new coal-burning thermal power plants in
large and medium-sized prefectures is prohibited.
Various studies have documented the effectiveness of TCZ reg-
ulatory actions in reducing pollutant emissions and improving air
quality. For example at the national level, SO2 emissions fell from
23.67 million tons in 1995 to 19.95 million tons in 2000, and the
percentage of prefectures exceeding the Class II standard fell from
94 S. Tanaka / Journal of Health Economics 42 (2015) 90–103

54 percent in 1995 to 20.7 percent in 2000. By the end of 1999, Table 1

Descriptive statistics of baseline sample.
collieries producing more than 50 million tons of high-sulfur coal
had been closed (Hao et al., 2001). By the end of 2000, the total Mean Std. Dev. Min. Max. Obs.
power capacity with FGD equipment exceeded 10,000 MW, and Panel A: Birth characteristics
small thermal power plants, with output capacity below 50 MW, IMR from death record 16.61 16.79 0 202.53 556
were actively shut down because they were relatively less efficient, IMR from birth record 9.62 13.43 0 133.18 550
had high coal consumption rates, and emitted massive amounts of % of Boy 0.55 0.04 0.33 0.71 550
% of January births 0.11 0.07 0 1 548
pollutants. This reduced raw coal consumption by10 million-tons
Birth order 1.40 0.28 1 3.39 551
and SO2 emissions by 0.4 million tons (Yang et al., 2002).
Panel B: Household characteristics
Importantly for our identification strategy, reductions in air
Mother’s age 25.37 1.35 22.18 30.45 552
pollution were more remarkable in TCZ cities. Among TCZs, SO2 Maternal education ≥H.S. 0.22 0.30 0 1 547
emissions fell by about 3 million tons, and about 71 percent of all % of Han 0.89 0.25 0 1 553
factories producing over 100 tons of emissions per year reduced
Panel C: District characteristics
their SO2 emissions to the standard between 1998 and 2000 (He Total population 106,551 53,062 5267 251,707 556
et al., 2002). Between 1998 and 2005, the number of prefectures Number of infants 1,706 1,213 44 6,747 556
in the SO2 pollution control zones meeting the Class II standard % of city and above 0.46 0.50 0 1 553
rose by 12.3 percent, those meeting the Class III standard increased Notes: Each column presents the variable mean, standard deviation, the minimum
by 4.2 percent, and those not meeting the Class III standard fell value, the maximum value, and the number of observations. The sample is at the
by 16.5 percent. Within the acid rain control zone, the number DSP site by year level for the period prior to 1995. All mean values are weighted by
the number of population.
of prefectures meeting the Class II standard rose by 3.3 percent,
those meeting the Class III standard increased by 7.9 percent, and
those not meeting Class III decreased by 11.2 percent (United death without information on birth date or family attributes. Fur-
Nations Environment Programme, 2009). Online Appendix III pro- ther details on how these 145 sites were selected, and how the data
vides available evidence on the effect of TCZ regulation on air were processed at each site can be found in Online Appendix I.
quality. TCZ regulatory status. The TCZ regulatory status is reported in
the document “Official Reply to the State Council Concerning Acid
3. Data sources and descriptive statistics Rain Control Areas and Sulfur Dioxide Pollution Control Areas,”
published by the State Council in 1998. The assignments are pri-
3.1. Data sources marily made at the prefecture level, while those in municipalities
are at the county level. The document lists the names of all pre-
Infant mortality. The micro-level data on infant mortality come fectures that are designated as within the Acid Rain Control Zone
from the Chinese Disease Surveillance Points (DSP) system. The or the SO2 Control Zone. Hence, we can merge this information to
DSP covers 145 sites, primarily at the county-level,16 established the DSP sites, and those sites that are in the TCZ prefectures com-
on the representative sample of the national population. It reports prise the treatment group in the analysis.18 In total, 61 of 145 DSP
the censuses of death and birth registrations for the sample sites are in the TCZ prefectures and thus comprise the treatment
population of 10 million residences (approximately 1 percent of group, and 84 sites are in the non-TCZ prefectures, forming the
the national population) in various geographic areas across 31 control group. More detailed rules on matching DSP sites and TCZ
provinces, autonomous regions, and municipalities in China. regulatory status are described in Online Appendix I.
Overall, the original data record approximately 500,000 deaths
(for all ages) and 1,000,000 births from 1991 through 2000, from 3.2. Descriptive statistics
which the dataset we obtained was aggregated to the DSP site
by year level.17 The birth record reports whether or not infants Table 1 presents the descriptive statistics of the baseline sample
died within the calendar year, and if they did, the cause of death, in the pre-reform period (prior to 1995).19 The weighted average
using the International Classification of Disease, 9th Revision (ICD- means of observations at the DSP site by year level are calculated
9) codes. For our purpose, we use the birth registration to measure using population as the weight. The table shows that the mean IMR
the infant mortality rates, because it additionally contains variables is 16.61 per 1000 live births, using the death record.20 As expected,
relevant for the analysis, i.e., the infant’s characteristics such as gen- the birth record yields a much lower level of IMR, 9.62 per 1000
der, the date of birth, birthweight, length of gestation period, and live births, because it reports the occurrence of infant deaths only
birth order as well as maternal demographics such as age, educa- within a calendar year.21 It is worth emphasizing that since the
tion, and race. On the other hand, the death record is less useful in
our analysis because it only reports the age in years and cause of
18
Because the regulatory actions were the same between the SO2 and acid rain
control zones, we consider both areas as comprising one treatment group.
16 19
There are four main administrative divisions in China: the province, the prefec- We use pre-1995 as the baseline observations in case the APPCL amendment
ture, the county, and the township. The “county” level refers to districts, county-level had any contribution to pollution reductions. We address this issue by separating
cities, or counties. 17 out of 145 sites are at the prefecture level, while the rest are controlling for this period up to when the TCZ policy was implemented (i.e., years
at one of the county-level divisions. between 1995 and 1997).
17 20
Because the latest DSP data available at the city level end in 2000 and infor- By 2003, there was a concern that the DSP system might not comprise a rep-
mation in the subsequent years are unfortunately not accessible, the analysis resentative sample of China, being more concentrated in developed and populous
necessarily ends then. Some evidence indicates that pollution has increased during areas than rural ones. Indeed, according to the estimates by the World Bank, the
the twenty-first century, after the coverage of our dataset, to bolster rapid domestic average IMR over the same period was 38.2 per 1000 live births. Adjustments have
economic growth, which almost negated the effect of TCZ policy. Thus, the effect been made to the DSP system by extending the number of sites to 160 after our study
of TCZ policy was indeed concentrated in the first several years, which our dataset period. For the purpose of our study, the issue is less important, as we estimate the
covers but likely to have dissipated afterwards. Subsequent policies such as subsi- within-city impacts.
21
dizing sulfur scrubber or other incentive programs to control SO2 emissions have Note that the definition of IMR will differ according to whether one is using
been implemented since 2006. The effects of these policies are beyond the scope of the death record or the birth record. Using the death record, the number of infants
our paper. who did not make it to age one per 1000 live births is measured, whereas using
 S. Tanaka / Journal of Health Economics 42 (2015) 90–103 95

main analysis focuses on the relative changes in IMR, a question of

whether the levels of IMR in the birth record are underestimated
or not does not invalidate the use of the birth record, unless the
data were intentionally manipulated to respond to the TCZ regu-
lation, which is unlikely due to the quality control enforced by the
institution that has little to do with SEPA. If there is any concern
with using the birth data, it is that the impact may be understated
if reductions in infant deaths are truncated at zero. As the table
shows, several DSP sites report no occurrence of infant deaths in
both death and birth records. This is not surprising, given that they
have a small number of births. To address this issue, we repeat the
main analysis using the death record as a robustness check.

4. Empirical framework

4.1. Basic specification

The main objective of this study is to assess the effect of air pol-
lution regulations on infant mortality. In an ideal research setting,
the TCZ status is randomly assigned across cities, creating varia-
tion uncorrelated with baseline characteristics. In the absence of
a randomized controlled trial, we first use a simple difference-in-
differences (DID) approach, based on the TCZ regulatory status;
Fig. 2. Trends in infant mortality rate. Notes: This figure plots the trend of infant
mortality rate due to internal causes between the TCZ and non-TCZ cities. The
annual mean is calculated using the population as the weight. The solid vertical line
indicates the timing of TCZ policy implementation in January 1998. Because each
observation represents the annual average value, the solid vertical line is located
between 1997 and 1998 to clarify the timing of their implementations.

Yjt = ˛ + 1 (Tj × Postt ) + ı1 Xjt + t + j + εjt (1)

where Yjt is IMR in city j in year t, Tj is an indicator variable that
takes on the value one if city j was assigned as a TCZ in 1998,22 and
Postt is an indicator variable that takes on the value one for years
in or after 1998.23 The city fixed effects, j , control for the perma-
nent heterogeneity across cities, whereas the year fixed effects, t ,
control for year-specific shocks that are common to both TCZ and
non-TCZ cities. Xjt controls for an additional set of covariates that
capture birth, parental, and city characteristics at the city by year
level. All standard errors are clustered at the city level, allowing for
an arbitrary correlation within cities over time.
The parameter of interest is 1 , the reduced-form impact of air
pollution regulations on infant mortality, capturing the difference
the birth record, the number of infants who died within the same calendar year in
which they were born is measured. For example, if an infant is born on January 1st,
there is almost one year during which to record her death during the first year after
birth, whereas there is only a day in which to record the death of an infant born on
December 31st.
22
Note that the non-TCZ cities are not equivalent to “non-affected” cities. It is
more appropriate to say that the regulations were more stringent in the TCZ cities,
relative to those in the non-TCZ cities. In such a context, it may be ideal for Tj to
measure a continuous intensity of exposure to the regulations, measured by either
per-capita amount of high-sulfur coal used or by per-capita amount of SO2 emis-
sions in the baseline years. A similar strategy is used in Qian (2008), where she uses
the amount of tea planted in each county in China to measure the impacts of relative
female income on sex ratio, where women have comparative advantage in picking
tea, and the price of tea was dramatically increased by the post-Mao reforms. Also,
Bleakley (2007) uses the pre-treatment hookworm infection rate to measure the
benefits of hookworm eradication on school enrollment. Unfortunately, data limi-
tation makes this infeasible, and we decided to use a discrete choice of being in the
TCZ cities or not. This is still valid because the large amount of high-sulfur coal and
SO2 emissions was produced in the TCZ cities. The TCZ policy may also have reduced
pollution beyond TCZ cities, especially when non-TCZ cities are located near TCZ
cities, either directly through the policy effect on even non-TCZ cities or indirectly
through reducing pollution that travels to non-TCZ cities. (Note that lifetime and
travel distance of pollutants may be highly dependent on conditions, such as type
of pollutants, temperature, weather, wind, etc.) However, reductions in pollution in
non-designated areas, if any, would only understate our finding. Restricting the con-
trol group to distance places would generate a concern about comparability across
the places. Instead, it is part of our important finding that the treatment effects exist
even among geographically similar sites.
23
We later discuss the theoretical and empirical rationales of using 1998 as a cut-
off year for the post-reform period, rather than1995 when the APPCL was amended.
We find that the results are robust to using 1995 instead.
in the changes in IMR before and after the regulations, between
the cities that were and were not designated as a TCZ in 1998. If
the air pollution regulations contributed to significant reductions
in IMR in the TCZ cities relative to non-TCZ cities, 1 is expected to
be negative.
4.2. Validity of the identification assumptions
The key identification assumption for Eq. (1) to provide a causal
inference is that the non-TCZ cities provide valid counterfactual
changes in infant mortality for the TCZ cities, had they not been
treated, conditional on covariates. Two potential hypotheses may
violate this assumption: (1) there is a systematic difference in pre-
existing trends in mortality reductions, and/or (2) the TCZ status
is not orthogonal to factors explaining the reductions in infant
mortality in the post-treatment period.
To examine the pre-existing trend, we plot the evolution of IMR
over time between the TCZ and the non-TCZ cities in Fig. 2. The
solid vertical line indicates the timing of TCZ policy implementa-
tion in January 1998.24 The figure provides graphical support that
IMR trends were similar in the pre-intervention period; infant mor-
tality fell between 1991 and 1993, stayed somewhat constant until
1996, and fell again in 1997 in both sets of the cities. A trend break
appears around 1998, when IMR continued to drop only among the
TCZ cities, while IMR stayed somewhat constant within the non-
TCZ cities. A similar drop in IMR is illustrated in an even-study
type analysis in Fig. 3. Notably, although the IMR was continuously
higher in the TCZ cities than those in the non-TCZ cities before 1998,
the reverse is true after 1998, which is commensurate with the tim-
ing of the TCZ policy. A similar pre-trend between the two sets of
the cities suggests that the post-trend would have been similar in
the absence of the regulations.
However, it may not be fully convincing that TCZ cities and
non-TCZ cities would be similarly trended in the absence of
the treatment. In light of this, we extend Eq. (1) by additionally
24
Because each observation represents the annual average value, the solid vertical
line is located between 1997 and 1998 (since the TCZ policy was implemented in
January, early in the year). This is to clarify the timing of the TCZ policy implemen-
tation.
96 S. Tanaka / Journal of Health Economics 42 (2015) 90–103

Table 2
Balancing test by the TCZ status.

Trend difference

(1) (2)
All Years 1996–2000

Panel A: Birth characteristics

% of Boy 0.004 −0.001
(0.009) (0.006)
% of January births 0.001 0.005
(0.006) (0.005)
Birth order 0.053* −0.007
(0.029) (0.024)

Panel B: Household characteristics

Mother’s age 0.138 0.025
(0.113) (0.100)
Mother’s education ≥H.S. 0.034 0.032
(0.045) (0.041)
% of Han 0.038* 0.035
(0.021) (0.029)

Panel C: District characteristics

Fig. 3. Event-study analysis of the effect on IMR. Notes: The figure plots the
Total population 1433 3171
coefficients and their associated 90% confidence interval based on a single regres-
(4132) (3719)
sion modifying Eq. (1) by interacting the TCZ and respective year dummies (instead
Number of infants 139.5 124.2
of a post dummy):
(138.2) (109.5)
2000
Yjt = ˛ + i (Tj × 1(year = i)) + ı1 Xjt + t + j + εjt . Panel D: Economic Characteristics
i=1992 GDP per capita 918.12
(yuan) (740.88)
where 1(*) is an indicator variable taking on the value of one if year is i.
Num. of hospitals −29.34
(in units) (32.70)
controlling for the location-specific linear trend using the following Water supply 0.109
(100 million tons) (0.114)
specification:
Electricity consumption 14.94
(in kwh) (26.87)
Yjt = ˛ + 2 (Tj × Postt ) + ı2 Xjt + t + j + j × t + εjt (2)
Notes: Each entry reports the coefficient of the interaction term of TCZ status and a
where  j × t absorbs long-term linear trend in IMR that may vary post dummy (=1 if year is 1998 or later) from a separate regression when using the
across DSP sites. The specification in Eq. (2) explicitly controls for respective variable as the dependent variable based on Eq. (1). All regressions are
any effects through differential trends across cities. Such rigorous weighted by the number of population, and the robust standard errors, clustered at
the DSP site level, are reported in the parentheses. The sample includes all obser-
regression-adjusted evidence below also shows that the treatment vations between 1991 and 2000 in column (1), and observations between 1996 and
effect is robust to DSP site-specific time trends. 2000 in column (2). Economic characteristics are available only for urban areas and
Another concern is that the regulations’ effect may be con- from 1996.
*
founded by other concurrent changes in policies or factors affecting Significant at p < 0.1 level.
infant deaths. If the central government assigned the TCZ status
solely based on the nationally mandated standards, the designation point estimates are substantially small and virtually indistinguish-
should be less likely to respond to demands from local govern- able from zero. Importantly, the TCZ status balances the trends in
ments. However, because the TCZ status is based on air pollution important predictors of infant deaths, such as mother’s age or edu-
level, and air pollution level is conditioned by various local fac- cational level as well as the percentage of male and share of January
tors, it may still be possible that the TCZ status simply proxies births. A low degree of correlation with local economic trends sug-
these other factors. For example, air pollution may be high in urban gests that underlying economic shocks are less likely to be a source
places where the number of births is decreasing due to high levels of bias.25 In column (2), the sample is restricted to a shorter period,
of women’s participation in the labor force or relatively more strin- 1996–2000, which serves as a robustness check, because a number
gent enforcement of the one child policy. In this case, the negative of policies are fixed under the 9th Five-Year Plan. In this case, none
association between the TCZ status and IMR would be confounded of the variables presents significant differences in trends, while
by the quantity-quality tradeoff. To address this concern, we inves- many of the estimates become even smaller in magnitude.
tigate whether the TCZ status has any association with changes Overall, these findings provide compelling evidence that the TCZ
in observable characteristics. Although this is not a formal test regulatory status is orthogonal to trends in observable determi-
of exclusion restrictions, as the assumption states that the treat- nants of infant deaths, strongly suggesting that the current research
ment status should not covary with unobservable characteristics, design is unlikely to be biased by changes in unobservable vari-
the absence of significant correlation with observable characteris- ables. A falsification test and robustness checks below will provide
tics suggests that there should not be significant correlations with further support.
unobservable variables either (Altonji et al., 2005).
Column (1) of Table 2 addresses potential differences in trends 5. Empirical results
of characteristics between the TCZ and non-TCZ cities, before and
after the policy change. Each entry reflects the coefficient of the Following the identification strategies outlined above, we first
interaction term, 1 , in Eq. (1) from separate regressions, when present the main impact of environmental regulations on IMR in
each characteristic is used as the dependent variable. Any signif-
icant estimate indicates a systematic difference in trend patterns
and may confound the regulation effects. Strikingly, the table sheds 25
We address channels through additional economic activities in Online Appendix
light on little systematic difference in trends. Further, most of the II.
 S. Tanaka / Journal of Health Economics 42 (2015) 90–103 97

Table 3 of the point estimate is essentially unaffected, indicating that the

Main results on IMR.
main finding cannot be explained by differential trends.
(1) (2) (3) (4) The estimate in column (4) represents approximately 20 percent
TCZ × Post −2.870 **
−3.592 ***
−3.205 ***
−3.287 reductions in infant mortality over the three years in the post-
(1.103) (1.067) (1.087) (2.128) reform period.26 As a comparison of this figure to the U.S. context,
Observations 1340 1281 1281 1281 Chay and Greenstone (2003b) find that the nonattainment status
R2 0.415 0.438 0.442 0.552 in 1972 is associated with a 3–6 percent reduction in IMR in nonat-
Year fixed effects Y Y Y Y tainment counties from the previous year. This puts our estimates
District fixed effects Y Y Y Y in about a similar magnitude in terms of average annual reduc-
HH controls N Y Y Y
tion in IMR. It is worth mentioning that these comparisons pertain
District controls N N Y Y
District trend N N N Y
to the effect of environmental regulation on infant mortality, but
do not imply differences in marginal effect of air pollution. Due to
Notes: Dependent variable is number of infant deaths per 1000 live births, estimated
from the birth record. Robust standard errors are clustered at the DSP site level. Col-
lack of reliable pollution data in our study area, it is not feasible to
umn (1) includes only year and DSP sites fixed effects without additional covariates, calculate elasticity of infant mortality in response to air pollution.27
while column (2) adds birth and parental characteristics (share of male, birth shares
in respective month, birth order, mother’s age, mother with high school degree or
more, Han), column (3) additionally includes DSP sites characteristics (number of
5.2. The biological mechanism
births, total population, rainfall), and column (4) additionally includes DSP-site spe-
cific time trends. The point estimate based on the specification in column (1) using In an effort to identify the biological mechanism through which
the same sample as in columns (2)–(4) is −3.306 (1.06). air pollution affects infant mortality, we adopt two different strate-
**
Significant at p < 0.05 level. gies. The first examines whether the regulations are associated with
***
Significant at p < 0.01 level.
birth outcomes, namely birthweight and length of gestation period.
Maternal exposure to pollution during pregnancy is considered to
retard fetal development, as absorbed pollutants limit nutrition and
the subsequent section. We then highlight the biological mech- oxygen flows to fetuses. This sometimes results in low birthweight
anism underlying these results to illustrate that the effects are or shorter gestation period, although exact channels or outcomes
concentrated during the neonatal period and induced by changes are not well-known (Dejmek et al., 1999; Perera et al., 1999).28
in internal causes of deaths, deaths potentially associated with air Column (1) of Table 4 shows that those who were born in the
pollution, and not due to external causes of deaths, ones obviously TCZ cities weighed more by 2.66 g, yet the estimate is not distin-
not associated with air pollution. Then, we present heterogeneity in guishable from zero. This result should be interpreted with caution.
effects depending on infants’ gender or mothers’ education. Finally, As shown in the table, the baseline birthweight was already high
we present a number of robustness checks that support our main in China, and parents can often intentionally adjust birthweight.
findings. Rather, we should be more concerned about children born at low
birthweight, which is often considered to have a lasting impact
on later health and socioeconomic status.29 In Column (2), hence,
5.1. Main results on infant mortality we evaluate the distributional effect on births at low birthweight;
the dependent variable is the number of births below 2500 g per
We present the DID estimates of the regulations’ effect on infant 1000 live births. The result shows that the environmental regula-
mortality in Table 3. The dependent variable is IMR per 1000 live tions correlate to a lower incidence of low birthweight births by
births for all causes of deaths. Column (1) provides the result from 3.42 per 1000 live births.30 Columns (3) and (4) further examine
Eq. (1) without any controls except city and year fixed effects, the corresponding impact on the length of gestation period and the
indicating that the TCZ status is associated with 2.87 fewer infant
deaths per 1000 live births. Column (2) controls for a number of
birth and parental characteristics, directly addressing a concern
26
that changes in infant mortality may be explained by changes Note that we use the baseline estimate of IMR based on the death record (16.61
per 1000 live births), as it captures the entire infant deaths per year.
in observable characteristics of samples that vary over time and 27
Based on back-of-envelope calculation of elasticity using changes in air pollu-
that are correlated with the TCZ status. Namely, it controls for the tion, our elasticity of changes in mortality rate relative to changes in pollution is
percentage of male births, share of births by month, birth order, about 0.9. In economics studies, evidence in developing countries is still scarce, yet
mother’s age, mother’s education, and percentage of Han. The share a recent study by Arceo et al. (2012) finds elasticity of 0.415 in Mexico, and other
studies in U.S. provide elasticities of 0.284 (Chay and Greenstone, 2003a) or 1.827
of births in every month works similarly to birth month dummies,
(Knittel et al., 2011). Thus, our estimates in China are still in the range of the lit-
if the data were at the individual level, addressing the fact that erature, though toward a higher end. Evidence on the dose–response relationship
infant deaths for those who were born in earlier of the year are between pollution and human health in the fields of medicine and epidemiology is
more likely to be in the data. With these controls, the estimated in favor of greater elasticity at greater pollution levels.
28
coefficients become larger and are statistically significant at the 1 For adults, air pollution is often linked to respiratory and cardiovascular disease,
aggravation of asthma, heart disease, lung malfunction or cancer, stroke, and possi-
percent level. Column (3) additionally controls for DSP site charac- bly carcinogenesis. For infants, high exposure to pollution in the neonatal period is
teristics, in an effort to control for changes in total number of births, likely to result in deaths by acute respiratory infections. Some recent studies show
total population, and precipitation rates. The estimated effects are strong impacts coming from maternal exposure.
29
robust to the inclusion of these covariates and remain significant See Behrman and Rosenzweig (2004), Case et al. (2005), and Almond et al. (2005)
for costs and returns of low birthweight.
at the 1 percent level. 30
In theory, it is also possible to find lower birthweight when infant morality
A major concern with any non-experimental study is that the decreases, as marginal babies who would have died before are now born alive.
treatment group and control group are not comparable. Our study This argument is particularly true in a case, for example, where improved medical
is no exception. In Column (4), we present results based on Eq. (2), technology contributes to greater infant survival at the margin (or special health-
which directly controls for linear time trends specific to the DSP care services for newborns at risk). In our case, infant mortality decreased due to
improved pre-natal environment, which is more likely to result in greater birth-
sites. While Eq. (2) represents our preferred specification through- weight. Our finding of positive effect on birthweight is indeed consistent with a
out the paper, it substantially reduces the degrees of freedom, strong association between maternal smoking (which works in a similar way as
causing the standard error to double. However, the magnitude pollution exposure) and reduced birthweight.
98 S. Tanaka / Journal of Health Economics 42 (2015) 90–103

Table 4
Identifying the biological mechanism.

(1) (2) (3) (4) (5) (6) (7)

Birthweight Low birthweight Gestation period Short gestation Deaths w/in 1day Deaths w/in 28 days Deaths w/in 6 mo.

TCZ × Post 2.66 −3.42** −0.237 −2.59 −0.85* −2.08** −2.40**

(16.32) (1.70) (0.217) (3.17) (0.44) (0.85) (0.93)
N 1249 1235 657 650 1295 1295 1295
R2 0.78 0.66 0.44 0.23 0.43 0.48 0.50
Pre-1995 mean 3313.2 15.76 39.09 26.47 2.88 6.59 8.00
[116.9] [15.47] [1.80] [118.5] [5.03] [9.35] [11.23]
Year effect Y Y Y Y Y Y Y
District effect Y Y Y Y Y Y Y
HH attributes Y Y Y Y Y Y Y
District attributes Y Y Y Y Y Y Y

Notes: Dependent variables are; birthweight in gram in column (1), the number of infants born with less than 2500 g per 1000 live births in column (2), gestation period in
week in column (3), the number of infants born in less than 32 weeks per 1000 live births in column (4), IMR within one day in column (5), IMR within 28 days in column
(6), and IMR within 6 months in column (7). The mean values of respective variable (weighted by total population) and their standard deviations in the square bracket are
provided based on the observations before 1995.
*
Significant at p < 0.1 level.
**
Significant at p < 0.05 level.

share of gestation period below 32 weeks, the lowest one percentile
level. Neither estimate is significant, indicating that the impact on
low birthweight is not driven by changes in length of gestation
period.
The results above suggest that fetal exposure to pollution affect-
ing fetal development appears to be a key. However, birthweight
and length of gestation period may not fully capture fetal devel-
opment. Hence, we now explore the effect on deaths occurring at
different time periods. Infant deaths occurring during the neona-
tal period (within 28 days after birth) are generally considered to
be associated with poor fetal development (Chay and Greenstone,
2003a,b).31 Column (5) presents the impacts on infant deaths that
occurred within one day of birth. The point estimate is small yet
marginally significant at the 10 percent level, implying that 26
percent of overall reduction in IMR occurred within one day. The
magnitude is in line with Chay and Greenstone (2003b), who esti-
mate that roughly 22 percent of overall infant deaths occurred
within one day. Yet this contrasts to the finding in Chay and
Greenstone (2003a), who attribute roughly 60 percent of overall
impact to infant deaths within one day. Column (6) reveals that
the regulations are disproportionately more associated with the
probability of death during the neonatal period, indicating that
63 percent of the effect of the regulations on infant mortality is
due to reductions in this period.32 This corresponds to the find-
ings in Chay and Greenstone (2003a) and Chay and Greenstone
(2003b), whose 73–82 percent and 80 percent of infant mortality
reductions occurred in the same period, respectively. Overall, these
findings highlight weak fetal development via maternal exposure
as an important biological mechanism.
A related concern is that the regulations are confounded by
other concurrent changes in factors contributing to the reductions
in infant mortality. For example, in response to high exposure
to pollution in the TCZ cities, the local governments may have
increased healthcare spending, leading to an improvement in qual-
ity and/or quantity of healthcare services. Then, without directly
controlling for a local health policy, a simple DID estimate would
erroneously pick up effects through a healthcare policy, causing
an upward bias. In order to rule out such a possibility, we exam-
ine the regulations’ effect on infant mortality by cause of death.
If maternal exposure is a primary channel, we expect to see larger
changes in mortality associated with prenatal disorders as opposed
to postnatal causes. There is added significance to this analysis;
the estimates are directly comparable to the falsification test that
examines the effect on external causes of deaths unrelated to air
pollution.
Given that the exact etiology and pathology of diseases caused
by fetal exposure to air pollution are not yet known, the most
important comparison is between internal and external causes
of deaths. We compute IMR due to internal deaths to include all
health-related, non-accidental causes that are potentially associ-
ated with air pollution, other than infant deaths due to external
causes of deaths, those that clearly do not pertain to air pollution:
injury and poisoning (Chay and Greenstone, 2003a,b).
As predicted, Table 5 shows statistically significant effects on
mortality from health-related causes.33 When effects are estimated
separately for four major health-related causes,34 the estimates
correspond to a reduction of 49.7 percent in mortality from ner-
vous system disorders and a reduction of 32.7 percent in circulatory
system disorders. These findings are consistent with vast literature
that also finds strong associations between these birth defects and
maternal smoking during pregnancy (which essentially works in
a similar mechanism as air pollution exposure) (see for example
Fried, 1995; Brennan et al., 1999).
On the other hand, infectious, parasitic and respiratory diseases
did not have significant impacts. Low rates of respiratory diseases
indicate that these are not typical causes of deaths for infants, par-
ticularly during the neonatal period, who spend most of their times
indoor, while these diseases have been found to be a major cause
of deaths for children or at youngest post-neonatal period (See for
example, Woodruff et al., 1997; Borja-Aburto et al., 1997; Bobak
and Leon, 1999; Woodruff et al., 2006).
Most importantly, we find no statistically significant effect on
mortality from external causes such as injury and poisoning. If

31
According to the WHO, infant deaths during the neonatal period are also asso-
ciated with preterm birth, intrapartum-related complications, and infections, and
thus we need caution in attributing it solely to fetal development, as there is also a
possibility that post-natal exposure plays a role.
32
Note that because the analysis is based on the birth record, we need to keep in
mind that deaths in a shorter period are more likely to be recorded and be cautious
about the estimates. We appreciate this discussion by an anonymous referee.
33
Over the study period between 1991 and 2000, 51 percent of infant deaths are
due to diseases of the circulatory system, 19.4 percent come from nervous systems
and sense organs, and 19 percent are due to external causes.
34
Note that this exercise of separately estimating effects for different “internal”
causes do not itself test whether the fundamental cause is due to air pollution or
others, as it is mentioned that the exact pathway is not known. Indeed, Chay and
Greenstone (2003a,b) do not disentangle effects among internal causes for this rea-
son. We do this simply to highlight variation across disease types given our data
capacity to do so.
 S. Tanaka / Journal of Health Economics 42 (2015) 90–103 99

Table 5
Effects on IMR by cause of death.

(1) (2) (3)

Pre-1995 mean mortality rate Estimated coefficients % in mortality rate

All non-accidental causes 7.61 −2.82*** −37.07

[11.28] (0.98)

All non-accidental causes disaggregated

Nervous system disorders 1.63 −0.81*** −49.69
[3.85] (0.31)
Circulatory system disorders 4.56 −1.49* −32.68
[7.18] (0.76)
Infectious and parasitic diseases 1.15 −0.38
[2.49] (0.29)
Respiratory system disorders 0.13 −0.03
[0.77] (0.13)

Accidental causes
Injury and poisoning 2.01 −0.39
[7.21] (0.48)

Notes: All specifications include year fixed effects, DSP site fixed effects, household attributes, and district attributes. The number of observations is 1281. Standard deviations
are reported in the square brackets, and robust standard errors are reported in the parentheses.
*
Significant at p < 0.1 level.
***
Significant at p < 0.01 level.

households had had improved access to healthcare services, there Table 6

Heterogeneity in effect.
would likely have been reductions in infant deaths by external
causes as well. Dependent variable
While this does not directly rule out the possibility that the reg-
IMR Low birthweight Short pregnancy
ulations spuriously pick up the effect through unobserved changes
Boys −3.12*** −2.06 −0.677
that are correlated with internal causes of deaths but not with
(1.20) (1.52) (3.47)
external ones, the finding suggests that the main results are not Girls −3.41*** −4.98** −3.11
driven by any other channels, leading us to believe that the relation- (1.18) (2.29) (3.60)
ship is causal. We provide a number of further robustness checks Mother’s education < H.S. −2.21 −3.85* −3.43
in Section 5.4. (2.15) (2.16) (3.65)
Mother’s education ≥ H.S. 4.45 −4.94** −1.52
(2.98) (2.29) (1.65)

5.3. Heterogeneity in the regulations effect
This part tests the hypothesis that the regulations on air pol-
lution may have a heterogeneous impact on infant mortality
across various subsamples. We first search for heterogeneity in
the treatment effects between boys and girls due to biologically
based gender differences given that, in the literature, male fetuses
are considered to be more physiologically sensitive than female
fetuses to environmental changes. On the other hand, heteroge-
Notes: Each cell reports the coefficient of interests and standard errors in the paren-
theses from separate regressions for respective subsample. The dependent variables
are IMR, the number of infants born at less than 2500 g per 1000 live births, and the
number of infants born in less than 32 weeks per 1000 live births, respectively.
Low maternal education is defined as educational attainment less than high school,
and high maternal education is equal to or above high school completion. Note that
the analysis based on mother’s education inevitably drops deaths associated with
missing mother’s education.
*
Significant at p < 0.1 level.
**
Significant at p < 0.05 level.
***
Significant at p < 0.01 level.

neous effects may also reflect gender discrimination, particularly

when infant deaths occur after birth. For example, if boys were
initially more likely to be protected from exposure to pollution or
were more likely to receive medical treatments for health prob-
lems caused by pollution, then the effect of pollution reductions
would be more pronounced for girls. The first and second rows
in Table 6 report the estimated effects for boys and girls, respec-
tively. The coefficients suggest that the effect on infant mortality is
larger among girls than boys, yet the difference is not statistically
significant. This indicates that pollution reduction effect on infant
mortality was similar for boys and girls. In contrast, the impact
on birthweight is substantially higher for girls than for boys. The
interpretation of the results requires a caution. On one hand, the
evidence is consistent with the gender bias hypothesis in that girls
are more sensitive to changes in environment on a condition that
parents had a means to identify the gender of the fetus, which
changed their behaviors. On the other hand, the finding is also
consistent with the literature that female fetuses have a higher
threshold at which pollution leads to mortality than boys do, and
thus only girls were affected when pollution decreased somewhat
from initially high levels to lower levels, which were still above the
threshold for male fetuses. We do not find any significant effects
on the length of gestation period.
Next we explore whether the effect differs according to the
level of mother’s education. Such heterogeneity is likely to arise,
because behavior is known to be an important determinant in
health production function when it comes to pollution (Zivin and
Neidell, 2009; Moretti and Neidell, 2011; Deschenes et al., 2012).35
The effect may be amplified among households with low lev-
els of maternal education for many reasons. For example, infants
and fetuses of poorly educated mothers tend to have a lower ini-
tial health endowment, making them more liable to be adversely
impacted by air pollution initially. Alternatively, mothers with
greater educational attainment are more likely to know how to pro-
tect their children from being exposed to pollution outside and/or
35
In empirical studies, Jayachandran (2009) finds greater wild fire air pollution
effect on low-income households in Indonesia. Arceo et al. (2012) finds that the
estimated marginal effect of carbon monoxide is larger in Mexico compared to the
U.S. estimates, while the effect of PM10 is similar between the two countries.
100 S. Tanaka / Journal of Health Economics 42 (2015) 90–103

have more access to health services to treat their children. Fami- Table 7
Robustness checks.
lies with high socioeconomic status also have a greater degree of
mobility to better areas, while the poor may continue to be exposed Specifications Coeff.
to greater pollution. On the other hand, the effect may be smaller Use death record −3.45*
among poor households, if long-term exposure to pollution in poor (1.84)
areas allowed them to be more adept at keeping infants indoors. Use 1996–2000 −3.57**
The third row reports the estimated effects for the sample (1.44)
Only districts and cities −3.62**
of households where maternal education was less than a high
(1.64)
school degree, and the fourth row for the sample of households Common support −3.24***
where mothers attained at least a high school education. We find (1.12)
that the regulations’ effect is substantially higher among house- Eliminate outliers −2.73***
(0.95)
holds with low maternal education.36 The finding suggests that
Include province × year effects −3.50***
the regulations’ effect on infant mortality should be stronger for (1.27)
the low-socioeconomic families that are more vulnerable to the Control for 1995–1997 −4.08***
effects of air pollution. (1.56)
Weight by number of birth −2.92**
(1.26)
5.4. Additional robustness checks Use 1995 cut-off −2.92*
(1.61)
The findings above leave little room for the scope of confound- Cluster at prefecture level −3.205***
(1.076)
ing factors. First, little association between TCZ status and trends Only northern China −3.138*
in observable characteristics limits the possibility that the main (1.657)
results erroneously reflect time trends that vary systematically Only southern China −3.258**
between the TCZ and the non-TCZ cities. Second, the consistency (1.497)
Use 1991–1997 sample −1.967
of the regulations’ effect in both magnitude and statistical signif-
(1.901)
icance when controlling for the set of key determinants of infant
Notes: The table provides robustness checks of the main results to various other
mortality suggests that the estimated effects are robust to compar-
hypotheses and specifications. See the text for their explanations.
isons with similar characteristics. Third, the absence of treatment *
Significant at p < 0.1 level.
effect in the falsification exercise on external infant mortality pro- **
Significant at p < 0.05 level.
vides strong evidence that health care system reform or medical ***
Significant at p < 0.01 level.
technology advancement cannot be a source of bias.
In this subsection, we extensively explore additional robust-
ness checks to rule out other possible scenarios. First, we examine
whether the finding is robust when using an alternative dataset. restrict the sample to only districts and county-level cities, mostly
As discussed above, using the birth record, which reports the urban areas, in the third row. The estimate is larger and remains sig-
occurrence of deaths only within the calendar year, may result in nificant at the 5 percent level, indicating that the treatment effect is
understating the effect, if the number of infant deaths is truncated not driven by simple comparisons between urban and rural areas.
at zero. The death record allows us to compute IMR for all deaths A major concern with any non-experimental studies is the pos-
occurring before age of one. As expected, the first row of Table 7 sibility that omitted heterogeneity may give rise to a spurious
shows that the size of the estimate becomes larger, though not sub- relationship. Controlling for the attributes, as in columns (2) and
stantially different from the main result, indicating that the effect (3) in Table 3, may not solve this issue when we cannot compare
is not sensitive to using the death record, while the main analysis the distribution of attributes across TCZ and non-TCZ cities. We
may understate the overall impact, if any. address this issue in two ways. In the fourth row, we limit to the
Second, we confirm that the treatment effect is not driven by observations under common support using the propensity score,
other national or local policy changes. In the second row, we restrict which potentially restricts the sample to DSP sites that have similar
the sample to the years between 1996 and 2000, which corresponds observed characteristics. In doing so, we first compute propensity
to the period of the 9th Five-Year Plan. The fact that the time period scores of being TCZ based on households and DSP attributes used
is shorter and falls within one policy regime helps reduce a set of in the main analysis. Then, we re-estimate the treatment effects
potential confounders in the pre- and post-natal health environ- using observations only under the common support. Alternatively,
ment other than pollution. Also, Table 2 shows that all observed in the fifth row, we examine whether the main findings are robust
characteristics had balanced trends during this short period. The to eliminating outliers. Specifically, we eliminate DSP sites whose
estimated effect is similar, suggesting that various other national IMR are above 99th percentile. The orders of the both magnitudes
or local policy changes should not confound the effect. are similar.
Third, despite evidence that the treatment effect is robust to het- Another concern is that there may be unobserved policy changes
erogeneous city-specific trends in IMR, as predicted in Table 3, there that affected infant mortality. In addressing this, we control for
may still be a concern that the TCZ status may be correlated with province times year fixed effects in the sixth row. In this specifica-
administrative divisions. For example, urban districts and county- tion, the effect of the regulations on infant mortality is identified
level cities may be more likely to be treated, whereas poor counties using variation in regulatory stringency before and after 1998
may be less likely to be assigned as TCZ. To address this issue, we within the province, thus purging any potential effects resulting
from any other policy changes at the provincial level. Further, in
the seventh row, we control for the years between 1995 and 1997,
the intermediate period after the APPCL and before the TCZ pol-
36
Note that, although neither of these estimates in themselves nor the differences icy was in effect. Note that this additional variable controls for the
between them are statistically significant, when we restrict the IMR to internal
causes, the coefficient to low maternal education level is −4.40 and significant
immediate impacts of APPCL, if any, but does not directly rule out
at the 5 percent level, whereas that to high maternal education is −0.34 and not delayed impacts of APPCL after 1998. Both of the estimates are again
statistically significant. unchanged.
 S. Tanaka / Journal of Health Economics 42 (2015) 90–103
Next, we re-estimate the main analysis using different spec-
ifications. Namely, we weight the regressions using the number
of population aged 0 in the eighth row;37 we use 1995 as a cut-
off year instead of 1998 to incorporate all years after APPCL was
enacted in the ninth row;38 we cluster the standard error at the
prefecture level in the tenth row;39 we use only northern China in
the eleventh row; we use only southern China in the twelfth row.40
All these estimates effects remain similar, showing robustness to
alternative specifications.
Lastly, we repeat the main analysis using samples only in the
pre-reform period in the last row. We use 1995 as a placebo cut-
off year, and thus years between 1995 and 1997 are defined as
“post”-reform observations. The point estimate is lower and not
statistically indistinguishable from zero, reassuring that the trends
between the TCZ and non-TCZ cites are similar in the pre-reform
period.
Taken all above together, there is no evidence to indicate that
the main results are driven by inappropriate identification assump-
tions, leading us to believe that the relationship is causal. The
collection of these robustness checks substantially limits the scope
of omitted variables, leading us to believe that the main find-
ings substantiate the causal impact of environmental protection
on infant mortality.
6. Conclusions
China suffers from notoriously bad air pollution, the health
effects of which have been of increasing public concern. In 1998,
37
This intends to adjust the regression weighted by the number of birth, where
we do not have information on actual numbers of birth.
38
The main analysis uses 1998 as a cut-off year for the post-reform period, rather
than 1995 when the APPCL was amended, based on both theoretical and empir-
ical rationales. For theoretical purposes, it is plausible to take two to three years
before the regulations are carried out to the full extent. Chay and Greenstone (2005)
are based on a similar argument when they use 1975 nonattainment status as an
instrumental variable in estimating the effect of 1970 Clean Air Act on housing
prices between 1970 and 1980. In their context, the nonattainment status changes
every year. By using the mid-decade regulation, they also take into account a two-
to three-year lag before the policy was fully executed. This is also relevant in our
context because the regulations required power plants to alter the energy sources
and install costly technology (such as FGD). Informal conversations with officials at
local power plants provide anecdotal evidence to support this assertion. Time-lags
are also likely in China because it is common for the government to set policy targets
or guidelines, often very ambitious ones, without specifying the critical details until
later, thereby largely leaving implementation up to the local governments or indi-
vidual firms. Further, the 1995 amendment had a weak implementation mechanism,
and more drastic actions (such as shutting down numerous inefficient power plants
and enforcing stringent air pollution regulations) were enforced only after the TCZ
policy in 1998. The empirical rationale for the cut-off date is based on the finding
that the interaction term between the TCZ status and the post-1998 period better
balances important determinants of infant mortality, compared with using 1995 as
a cut-off year, suggesting that the former is less likely to be confounded. Further,
using a 1998 cut-off year enables us to restrict the sample to observations between
1996 and 2000, where the interaction term is not correlated with any observable
variables. Therefore, using 1998 better averts omitted variable bias.
39
The main results are clustered at DSP level for two reasons. (1) It is conventional
to cluster the standard errors at the treatment-site level, and in our case, the treat-
ment status varies at the DSP level, not the prefecture level. There are cases that one
prefecture accommodates multiple DSP sites that have different treatment status,
i.e., one is a district or county-level city, while the other is a county. (2) Indeed,
most DSP sites are in different prefecture. Out of 145 DSP sites, there are only 12
cases that we observe multiple sites in a single prefecture (two of which are three
DSP sites in one prefecture, and the rest of the cases include two DSP sites in one
prefecture). This robustness check clustering at the prefecture level should thus be
seen as accounting for serial correlations within prefecture.
40
The northern and southern China is defined as prefectures accommodating SO2
control zones (north) or acid rain control zone (south) (or more so if accommo-
dating both). We eliminated Tibet Autonomous Region, Qinghai Province, Xinjiang
Autonomous Region, as they are not typically perceived as either northern or south-
ern China.
101
the TCZ policy, which went into effect in 1998, was one of the
largest-scale air pollution regulatory schemes ever implemented
in a developing country, imposing stringent regulations on pollut-
ant emissions from power plants in cities exceeding the nationally
mandated standards.
The major objective of this paper is to test the hypothesis that
these regulations led to reductions in infant mortality within the
TCZ cities subjected to particularly stringent regulations. Using the
difference-in-differences approach, comparing changes in infant
mortality between the cities assigned and not assigned as the TCZ,
before and after the policy reform, we find substantial impacts:
infant mortality decreased by 20 percent; a large fraction of the
reduction occurred in the neonatal period, which can be attributed
to fetal exposure; and infants of mothers with low levels of educa-
tion benefited the most.
The set of falsification tests and robustness checks limits the
role of omitted variables in biasing these estimates, leading us
to believe that the linkage between the air pollution regulations
and infant mortality reduction is causal. First, the estimates are
robust to DSP site-specific trends in IMR. Second, we confirm that
the treatment effect is absent for infant deaths caused by external
causes, ruling out a potential mechanism through improved local
healthcare system. Lastly, the estimates are robust to various alter-
native hypotheses, i.e., using the death record, controlling for other
policies at national or local levels, and limiting to only urban areas.
The findings in this study have important implications for pol-
icy. First, the question of whether, and to what extent, air pollution
regulations in developing countries can lead to reducing infant
mortality remains unanswered. This study highlights a significant
reduction in infant mortality correlating to air pollution reductions,
in contrast to Greenstone and Hanna (2011) who find that air pol-
lution reductions in India had only modest and insignificant impact
on infant mortality. Our results substantiate compelling evidence
to support air pollution regulations in countries that suffer from
high levels of air pollution. The size of the benefits itself may not
be enough to justify environmental protection without consider-
ing their costs. In the United States, the Clean Air Act Amendments
are found to have caused distortions on productivity (Gollop and
Roberts, 1983; Barbera and McConnell, 1990; Greenstone et al.,
2012), firm’s location decisions (Henderson, 1996; Becker and
Henderson, 2000; List et al., 2003); employment (Greenstone,
2002; Deschenes, 2010; Walker, 2011, 2013); and foreign direct
investment inflows and outflows (Eskeland and Harrison, 2003;
Keller and Levinson, 2002; Hanna, 2010). On the other hand, Tanaka
et al. (2014) provide compelling evidence that polluting firms in the
TCZ cities substantially improved economic performance through
increased market dynamics via the entry of more efficient firms
and the exit of less efficient ones. In addition, costs of reducing
pollution are likely to be smaller under convex marginal cost func-
tions. Therefore, our finding is likely to pass cost-benefit analysis
and suggest further implementation of air pollution regulations in
similarly polluted countries.
Second, while the precise mechanisms through which air pol-
lution reductions lead to health benefits is not known, our findings
highlight substantial reductions in infant mortality during the
neonatal period, shedding light on maternal exposure to pollu-
tion as a potential pathophysiologic mechanism. This necessitates
additional policy interventions to protect pregnant women against
environmental risks.
Third, our findings identify that children in households with
low maternal education are particularly susceptible to fluctua-
tions in air quality. Although all individuals are potentially exposed
to ambient pollution, the evidence indicates that socioeconomic
status cushions the effect of air pollution, either through behav-
ioral factors in avoiding pollution or socioeconomic factors such
102 S. Tanaka / Journal of Health Economics 42 (2015) 90–103
as increased access to medical care. As such, our findings provide
justifications to interventions targeting low-income households,
including information provision about air pollution effect.
This study has clear policy implications for developing countries
in general: namely, while climate change does not currently appear
to be a sufficiently strong motivation for these countries to embark
on more aggressive air pollution regulations, our findings leave
little doubt that protecting the environment is vital to improving
domestic public health.
Acknowledgements
I am indebted to Daniele Paserman, Dilip Mookherjee, Tavneet
Suri, and Wesley Yin for invaluable advice and feedback. I am also
grateful to Lucas Davis, Esther Duflo, Michael Greenstone, Hsueh-
Ling Huynh, Kelsey Jack, Ginger Zhe Jin, Hiroaki Kaido, Kevin Lang,
Adriana Lleras-Muney, Michael Manove, Kenneth A. Rahn, Leena
Rudanko, Marc Rysman, Johannes Schmieder, Jeremy Smith, three
anonymous reviewers, and seminar participants at Boston Univer-
sity, Hiroshima University, Loyola Marymount University, National
University of Singapore, Tufts University, University of Washing-
ton, 2010 NEUDC, and the 2011 Royal Economic Society for their
comments and suggestions. I also thank the Chinese Center for Dis-
ease Control and Prevention for sharing the data. Financial support
from the Institute for Economic Development at Boston Univer-
sity, as well as Hewlett/IIE Dissertation Fellowship in Population,
Reproductive Health and Economics, is gratefully acknowledged.
Jieshuang He provided excellent research assistance. All remaining
errors are my own.
References
Almond, D., Chay, K.Y., Lee, D.S., 2005. The costs of low birth weight. Quarterly
Journal of Economics 120 (3), 1031–1083.
Almond, D., Chen, Y., Greenstone, M., Li, H., 2009. Winter heating or clean air? Unin-
tended impacts of China’s Huai River policy. American Economic Review Papers
and Proceedings 99 (2), 184–190.
Altonji, J.G., Elder, T.E., Taber, C.R., 2005. Selection on observed and unobserved
variables: assessing the effectiveness of catholic schools. Journal of Political
Economy 113 (1), 151–184.
Arceo, E., Hanna, R., Oliva, P., 2012. Does the effect of pollution on infant mortality
differ between developing and developed countries? Evidence from Mexico City.
NBER Working Paper No. 18349.
Aunan, K., Pan, X.C., 2004. Exposure-response functions for health effects of ambi-
ent air pollution applicable for China: a meta-analysis. Science of the Total
Environment 329, 3–16.
Barbera, A.J., McConnell, V.D., 1990. The impact of environmental regulations on
industry productivity: direct and indirect Effects. Journal of Environmental Eco-
nomics and Management 18 (1), 50–65.
Becker, R., Henderson, V., 2000. Effects of air quality regulations on polluting indus-
tries. Journal of Political Economy 108 (2), 379–421.
Behrman, J.R., Rosenzweig, M.R., 2004. Returns to birthweight. Review of Economics
and Statistics 86 (2), 586–601.
Bleakley, H., 2007. Disease and development: evidence from hookworm eradication
in the American South. Quarterly Journal of Economics 122 (1), 73–117.
Bobak, M., Leon, D.A., 1999. The effect of air pollution on infant mortality appears
specific for respiratory causes in the postneonatal period. Epidemiology 10 (6),
666–670.
Borja-Aburto, V.H., Loomis, D.P., Bangdiwala, S.I., Shy, C.M., Rascon-Pacheco, R.A.,
1997. Ozone, suspended particulates, and daily mortality in Mexico City. Amer-
ican Journal of Epidemiology 145 (3), 258–268.
Brennan, P.A., Grekin, E.R., Mednick, S.A., 1999. Maternal smoking during preg-
nancy and adult male criminal outcomes. Archives of General Psychiatry 56
(3), 215–219.
Case, A., Fertig, A., Paxson, C., 2005. The lasting impact of childhood health and
circumstance. Journal of Health Economics 24 (2), 365–389.
Case, A., Lubotsky, D., Paxson, C., 2002. Economic status and health in childhood:
the origins of the gradient. American Economic Review 92 (5), 1308–1334.
Chay, K.Y., Greenstone, M., 2003a. The impacts of air pollution on infant mortality:
evidence from geographic variation in pollution shocks induced by a recession.
Quarterly Journal of Economics 118 (3), 1121–1167.
Chay, K.Y., Greenstone, M., 2003b. Air quality, infant mortality, and the Clean Air Act
of 1970. NBER Working Papers No. 10053.
Chay, K.Y., Greenstone, M., 2005. Does air quality matter? Evidence from the housing
market. Journal of Political Economy 113 (2), 376–424.
Currie, J., Hyson, R., 1999. Is the impact of health shocks cushioned by socioeco-
nomic status? The case of low birthweight. American Economic Review 89 (2),
245–250.
Currie, J., Neidell, M., 2005. Air pollution and infant health: what can we learn
from California’s recent experience? Quarterly Journal of Economics 120 (3),
1003–1030.
Currie, J., Neidell, M.J., Schmieder, J., 2009. Air pollution and infant health: lessons
from New Jersey. Journal of Health Economics 28 (3), 688–703.
Dejmek, J., Selevan, S.G., Benes, L., Solansky, I., Sram, R.J., 1999. Fetal growth and
maternal exposure to particulate matter during pregnancy. Environmental
Health Perspectives 107 (6), 475–480.
Deschenes, O., 2010. Climate policy and labor markets. NBER Working Paper No.
16111.
Deschenes, O., Greenstone, M., Shapiro, J.S., 2012. Defensive investments and the
demand for air quality: evidence from the NOX budget program and ozone
reductions. NBER Working Paper No. 18267.
Eskeland, G.S., Harrison, A.E., 2003. Moving to greener pastures? Multinationals and
the pollution haven hypothesis. Journal of Development Economics 70 (1), 1–23.
Fried, P.A., 1995. Prenatal exposure to marihuana and tobacco during infant infancy,
early and middle childhood: effects and an attempt at synthesis. Archives of
Toxicology Supplement 17, 233–260.
Gollop, F.M., Roberts, M.J., 1983. Environmental regulations and productivity
growth: the case of fossil-fueled electric power generation. Journal of Political
Economy 91 (4), 654–674.
Greenstone, M., 2002. The impacts of environmental regulations in industrial activ-
ity: evidence from the 1970 and 1977 Clean Air Act Amendments and the census
of manufactures. Journal of Political Economy 110 (6), 1175–1219.
Greenstone, M., List, J.A., Syverson, C., 2012. The effects of environmental regulation
on the competitiveness of U.S. manufacturing. NBER Working Paper 18392.
Greenstone, M., Hanna, R., 2011. Environmental regulations, air and water pollution,
and infant mortality in India. NBER Working Paper 17210.
Hanna, R., 2010. US environmental regulation and FDI: evidence from a panel of
US-based multinational firms. American Economic Journal: Applied Economics
2 (3), 158–189.
Hao, J., He, K., Duan, L., Li, J., Wang, L., 2007. Air pollution and its control in
China. Frontiers of Environmental Science and Engineering in China 1 (2),
129–142.
Hao, J., Wang, S., Liu, B., He, K., 2001. Plotting of acid rain and sulfur dioxide pollution
control zones and integrated control planning in China. Water, Air, and Soil
Pollution 130, 259–264.
He, K., Huo, H., Zhang, Q., 2002. Urban air pollution in China: current status, char-
acteristics, and progress. Annual Review of Energy and the Environment 27,
397–431.
Hedley, A.J., Wong, C.M., Thach, T.Q., Ma, S., Lam, T.H., Anderson, H.R., 2002. Car-
diorespiratory and all-cause mortality after restrictions on sulphur content fuel
in Hong Kong: an intervention study. Lancet 360 (9346), 1646–1652.
Henderson, V., 1996. Effects of air quality regulation. American Economy Review 86
(4), 789–813.
Jayachandran, S., 2009. Air quality and early-life mortality: evidence from
Indonesia’s wildfires. Journal of Human Resources 44 (4), 916–954.
Keller, W., Levinson, A., 2002. Pollution abatement costs and foreign direct invest-
ment inflows to U.S. states. Review of Economics and Statistics 84 (4),
691–703.
Knittel, C., Miller, D.L., Sanders, N.J., 2011. “Caution, Drivers! Children Present: Traf-
fic, Pollution, and Infant Health,” NBER Working Paper No. 17222.
Kumar, N., Foster, A., 2007. Respiratory Health Effects of Air Pollution in Delhi and
its Neighboring Areas. Mimeo, India.
Laden, F., Neas, L.M., Dokery, D.W., Schwartz, J., 2000. Association of fine particulate
matter from different sources with daily mortality in six U.S. cities. Environ-
mental Perspectives 108 (10), 941–947.
Levy, J.I., Hammitt, J.K., Spengler, J.D., 2000. Estimating the mortality impacts of
particulate matter: what can be learned from between-study variability? Envi-
ronmental Health Perspectives 108 (2), 109–117.
List, J.A., Millimet, D.L., Fredriksson, P.G., McHone, W.W., 2003. Effects of envi-
ronmental regulations on manufacturing plant births: evidence from a
propensity score matching estimator. Review of Economics and Statistics 85 (4),
944–952.
Luechinger, S., 2014. Air pollution and infant mortality: a natural experiment from
power plant desulfurization. Journal of Health Economics 37, 219–231.
Matus, K., Nam, K.M., Selin, N.E., Lamsal, L.N., Reilly, J.M., Paltsev, S., 2012. Health
damages from air pollution in China. Global Environmental Change 22 (1),
55–66.
Mendelsohn, R., Orcutt, G., 1979. An empirical analysis of air pollution
dose–response curves. Journal of Environmental Economics and Management
6 (2), 85–106.
Moretti, E., Neidell, M., 2011. Pollution, health, and avoidance behavior: evidence
from the ports of Los Angeles. Journal of Human Resources 46 (1), 154–175.
National Bureau of Statistics of China, 2006. http://www.stats.gov.cn/
Nielsen, C.P., Ho, H.S., 2007. Air pollution and health damages in China: an intro-
duction and reviews. In: Ho, Mun, S., Chris, P., Nielsen (Eds.), Clearing the Air:
The Health and Economic Damages of Air Pollution in China. The MIT Press,
Cambridge.
Perera, F.P., Jedrychowski, W., Rauh, V., Whyatt, R.M., 1999. Molecular epidemi-
ological research on the effects of environmental pollutants on the fetus.
Environmental Health Perspectives Supplements 107 (3), 451–460.
 S. Tanaka / Journal of Health Economics 42 (2015) 90–103
Qian, J., Zhang, K., 1998. China’s desulfurization potential. Energy Policy 26 (4),
345–351.
Qian, N., 2008. Missing women and the price of tea in China: the effect of sex-
specific earnings on sex imbalance. Quarterly Journal of Economics 123 (3),
1251–1285.
Samet, J.M., Zeger, S.L., Dominici, F., Curriero, F., Coursac, I., Douglas, W., Dockery, J.S.,
Zanobetti, A., 2000. The National Morbidity, Mortality, and Air Pollution Study.
Part II. Morbidity, Mortality, and Air Pollution in the United States. Health Effects
Institute, Boston.
Schwartz, J., Dockery, D.W., Neas, L.M., 1996. Is daily mortality associated specifically
with fine particles? Journal of the Air and Waste management Association 46
(10), 927–939.
Schwartz, J., Marcus, A., 1990. Mortality and air pollution in London: a time series
analysis. American Journal of Epidemiology 131, 185–194.
State Council, 1998. Official Reply to the State Council Concerning Acid Rain Control
Areas and Sulfur Dioxide Pollution Control Areas.
State Environmental Protection Agency (SEPA), 1996. The Report on Environmental
Quality in 1991–1995. SEPA.
Tanaka, S., Yin, W., Jefferson, G., 2014. Environmental Regulation and Industrial
Performance: Evidence from China. Mimeo.
United Nations Environment Programme, 2009. Two Control Zone Plan and Pro-
gram to Control Sulfur Pollution, Available at: http://www.ekh.unep.org/files/
GP-2.pdf
103
Walker, R.W., 2011. Environmental regulation and labor reallocation: evidence from
the clean air act. American Economics Review Papers & Proceedings 101 (3),
442–447.
Walker, R.W., 2013. The transitional costs of sectoral reallocation: evidence from
the Clean Air Act and the workforce. Quarterly Journal of Economics 128 (4),
1787–1835.
Woodruff, T.J., Grillo, J., Schoendorf, K.C., 1997. The relationship between selected
causes of postneonatal infant mortality and particulate air pollution in the
United States. Environmental Health Perspectives 105 (6), 608–612.
Woodruff, T.J., Parker, J.D., Schoendorf, K.C., 2006. Fine particulate matter (PM2.5)
air pollution and selected causes of postneonatal infant mortality in California.
Environmental Health Perspectives 114 (5), 786–790.
World Bank, 1998. World Development Indicators. The World Bank, Washington,
D.C.
World Health Organization (WHO), 2002. Air Quality Guidelines for Europe, 2nd ed,
Copenhagen.
Yang, J., Cao, D., Ge, C., Gao, S., 2002. Air pollution control strategy for China’s power
sector. In: Chinese Academy for Environmental Planning, Beijing.
Yang, J., Schreifels, J., 2003. Implementing SO2 emissions in China. In: Presented in
OECD Global Forum on Sustainable Development. Emissions Trading, Paris.
Zivin, J.G., Neidell, M., 2009. Day of haze: environmental information disclosure
and intertemporal avoidance behavior. Journal of Environmental Economics and
Management 58 (2), 119–128.
 Journal of Health Economics 42 (2015) 104–114

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

How do health insurer market concentration and bargaining power

with hospitals affect health insurance premiums?
Erin E. Trish a,∗ , Bradley J. Herring b,1
a
Leonard D. Schaeffer Center for Health Policy and Economics, University of Southern California, and Department of Health Policy and Management,
University of California, Los Angeles Verna and Peter Dauterive Hall 301-3 635 Downey Way, Los Angeles, CA 90089-3333, United States
b
Johns Hopkins Bloomberg School of Public Health, Department of Health Policy and Management, 624 North Broadway, Room 408, Baltimore, MD 21205,
United States

a r t i c l e i n f o a b s t r a c t

Article history: The US health insurance industry is highly concentrated, and health insurance premiums are high and
Received 17 April 2014 rising rapidly. Policymakers have focused on the possible link between the two, leading to ACA pro-
Received in revised form 5 November 2014 visions to increase insurer competition. However, while market power may enable insurers to include
Accepted 28 March 2015
higher profit margins in their premiums, it may also result in stronger bargaining leverage with hospitals
Available online 8 April 2015
to negotiate lower payment rates to partially offset these higher premiums. We empirically examine the
relationship between employer-sponsored fully-insured health insurance premiums and the level of con-
JEL classification:
centration in local insurer and hospital markets using the nationally-representative 2006–2011 KFF/HRET
I11
L11
Employer Health Benefits Survey. We exploit a unique feature of employer-sponsored insurance, in which
L41 self-insured employers purchase only administrative services from managed care organizations, to dis-
D4 entangle these different effects on insurer concentration by constructing one concentration measure
representing fully-insured plans’ transactions with employers and the other concentration measure rep-
Keywords: resenting insurers’ bargaining with hospitals. As expected, we find that premiums are indeed higher for
Insurance
plans sold in markets with higher levels of concentration relevant to insurer transactions with employers,
Competition
lower for plans in markets with higher levels of insurer concentration relevant to insurer bargaining with
Hospitals
Premiums hospitals, and higher for plans in markets with higher levels of hospital market concentration.
Bargaining power © 2015 Elsevier B.V. All rights reserved.

1. Introduction by the Department of Justice (DOJ) and Federal Trade Commission

The US healthcare industry has become increasingly consoli-
dated. While the wave of hospital mergers in the 1990s gave way
to numerous studies of the implications of hospital consolidation,
newfound attention in recent years has focused on consolidation
in the US health insurance industry. Robinson (2004) documents
the increasing concentration of these markets over the first half of
the 2000s, as well as the predominance of insurance markets dom-
inated by a small number of large, nationwide insurers. Similarly,
a report from the American Medical Association (2013) highlights
the preponderance of health insurance markets across the country
that are highly concentrated, as defined by the standards set forth
∗ Corresponding author. Tel.: +1 213 821 6178.
E-mail addresses: etrish@healthpolicy.usc.edu (E.E. Trish), herring@jhu.edu
(B.J. Herring).
1
Tel.: +1 410 614 5967.
(FTC) in their Horizontal Merger Guidelines (2010).
This increased level of interest in insurer concentration is war-
ranted for several reasons. Understanding the effects of these high
levels of market concentration and their implications for premi-
ums is valuable generally, but particularly so for an industry facing
such high and rapidly rising premiums. Further, there are a num-
ber of policy provisions included in the 2010 Patient Protection
and Affordable Care Act (ACA) that have important implications for
the level of competition in the US health insurance industry. The
creation of health insurance exchanges and the inclusion of vari-
ant forms of health insurers (such as CO-OP plans and nonprofit
plans directed by Office of Personnel Management) as competitors
alongside more traditional insurers are examples of ACA provi-
sions targeted toward increasing competition in the private health
insurance industry.
However, the ultimate effect of the level of health insurance
concentration on health insurance premiums is not straightfor-
ward, because there are potentially offsetting effects of the level
of insurer competition on premiums. On one hand, higher levels

http://dx.doi.org/10.1016/j.jhealeco.2015.03.009
0167-6296/© 2015 Elsevier B.V. All rights reserved.
 E.E. Trish, B.J. Herring / Journal of Health Economics 42 (2015) 104–114
of insurer concentration should lead to increased insurer market
power in the markets where insurance is sold (to employers and
individuals), likely resulting in relatively higher premiums due to
higher plan profit margins, all else equal. On the other hand, insur-
ers also engage in bilateral bargaining over transaction prices with
providers, one of the key drivers of insurer costs. Thus, higher levels
of insurer market concentration may yield stronger insurer bar-
gaining leverage with local providers, thereby enabling them to
negotiate lower provider prices, which may partly be passed on
to insurance purchasers in the form of lower premiums. This pur-
chasing power effect is particularly important, given the recent
movement toward increased consolidation among provider mar-
kets driven by the ACA and other trends (Cutler and Scott Morton,
2013).
Moreover, the effects of insurer market power may depend on
the amount of provider market power, and vice versa. The extent
to which insurers can use their bargaining leverage to negotiate
lower provider prices likely depends on the level of competition in
the local provider market, as these prices may already be at or near
the point at which economic profits are zero in relatively competi-
tive provider markets. Furthermore, the extent to which hospitals
can use their bargaining leverage likely depends on local insur-
ance market conditions. A better understanding of the extent to
which higher prices resulting from concentrated provider markets
are passed through to consumers in the form of higher premiums
(rather than simply representing a transfer of rents from insurers
to providers) is particularly relevant for antitrust enforcement in
terms of evaluating the extent to which hospital market consolida-
tion ultimately harms consumers.2
1.1. Our empirical contribution
In this paper, we empirically analyze the relationships between
insurer concentration, hospital concentration, and employer-
sponsored health insurance premiums. Our primary empirical
contribution is that we identify a way to disentangle insurer con-
centration’s differing effects on higher insurer profits and lower
provider prices. We do so by exploiting a unique feature of the mar-
ket for employer-sponsored insurance whereby smaller employers
tend to purchase fully-insured coverage whereas larger employers
tend to self-insure and purchase only administrative services from
managed care plans (such as provider network assembly and claims
processing). An insurer’s market share in the fully-insured market
is mainly relevant to the plan’s profits, while an insurer’s market
share in the fully-insured and self-insured markets combined is
mainly relevant to provider prices.
More specifically, we construct two distinct measures of health
insurance market concentration to disentangle these two effects.
Both concentration measures use the HealthLeaders-InterStudy
census of private insurers to construct Herfindahl-Hirschman
Indices (HHI) of market concentration, and we consider HHIs
alternatively using Core-Based Statistical Areas (CBSA), with the
Metropolitan Divisions therein, and counties as the geographic
market boundaries. One HHI market concentration measure
focuses on the profit portion of the premium’s administrative over-
head tied to the transactions between fully-insured plans and
employers by only using HealthLeaders-InterStudy’s fully-insured
plans in its HHI’s market share calculation. We hypothesize that,
all else equal, concentration in the fully-insured market will be
associated with relatively higher health insurance premiums.
2
We thank Chris Garmon for highlighting this point.
105
The second HHI market concentration measure focuses on the
hospital price’s portion of the premium tied to the negotiations
between insurers and hospitals. While self-insured enrollment rep-
resents a distinct product that is sold to employers, the insurer’s
patient volume across the entire combined “book of business” (i.e.,
the fully-insured market and the self-insured market) represents
its market share relevant to the price negotiations with hospitals.
We therefore use these HealthLeaders-InterStudy data to mea-
sure each plan’s fully-insured and self-insured combined market
share in this HHI calculation representing insurer bargaining with
providers. We hypothesize that concentration in the fully-insured
and self-insured markets combined will be associated with rela-
tively lower health insurance premiums. (We also hypothesize that
higher hospital market concentration – derived from the American
Hospital Association’s (AHA) Annual Survey – will be associated
with relatively higher health insurance premiums.)
Using plan-level premium data from the restricted-use
Kaiser Family Foundation/Health Research and Educational Trust
(KFF/HRET) Employer Health Benefits Survey for years 2006
through 2011, we find that premiums are indeed higher among
markets with higher levels of insurer concentration representing
fully-insured coverage sold to employers (and higher among more
concentrated hospital markets), and we find that premiums are
indeed lower among markets with higher levels of insurer con-
centration representing insurer bargaining with hospitals (derived
from combined fully-insured and self-insured market shares).
Regarding the organization of the remainder of the paper, we
first summarize the relevant literature on the effects of insurer and
hospital concentration and then describe the conceptual frame-
work. We then explain our empirical model, data, and market defi-
nitions. Our results, discussion, limitations, and conclusions follow.
2. Relevant literature
The majority of studies related to competition in the US
healthcare industry over the past few decades have focused on
competition and consolidation among hospitals. Gaynor and Vogt
(2000), Vogt and Town (2006), and Gaynor and Town (2011, 2012)
provide excellent reviews of this literature. While many of these
studies yield unique findings, the results generally suggest that
increasing consolidation in the hospital industry is associated with
higher hospital prices.
The literature on the association between insurance premiums
and the level of competition in the US health insurance indus-
try, particularly within the employer-sponsored market, is more
limited, largely due to data limitations. Early studies by Wholey
et al. (1995) and Dranove et al. (2003) find that markets with
more HMO competitors are associated with lower premiums. Dafny
(2010) finds evidence of price discrimination as a consequence of
insurer market power. Using a proprietary dataset containing infor-
mation about the insurance benefits offered by large employers
between 1998 and 2005, she utilizes variation in the profitability
of these large employers to illustrate that insurers in concen-
trated insurance markets impose higher premium increases on
more profitable employers (assumed to be less price sensitive).
Dafny et al. (2012) observe a positive effect of insurer consolidation
on health insurance premiums by exploiting the 1999 merger of
nationwide insurers Aetna and Prudential as a source of differential
changes in local insurance market concentration across the coun-
try. Using this instrument and the same dataset of large employers
as above, they find a significant effect of increases in local insurance
market concentration on increases in health insurance premi-
ums. Additionally, they explore the possible effects of insurance
consolidation on bargaining power with providers, finding that
106 E.E. Trish, B.J. Herring / Journal of Health Economics 42 (2015) 104–114
increased insurance concentration is associated with a substitu-
tion of nurses for physicians. Similarly, Dafny et al. (Forthcoming)
exploit United Healthcare’s uneven impact of its non-participation
in state exchanges to conclude that more concentrated insurance
exchanges were associated with higher premiums in 2014.
Numerous recent studies document the significant inter- and
intra-market variation in negotiated provider prices, including its
association with market-level factors, such as insurer or hospital
market concentration (for example, White et al., 2013; Berenson
et al., 2012; Ginsburg, 2010; MedPAC, 2009; Massachusetts
Attorney General, 2010; US GAO, 2005). Several recent papers
examine the relationship between both insurer and hospital con-
centration and negotiated hospital prices. McKellar et al. (2013)
and Moriya et al. (2010) both find that higher levels of insurance
concentration are associated with lower hospital prices, but that
higher levels of hospital concentration are not significantly associ-
ated with higher hospital prices. However, Melnick et al. (2011) find
that higher hospital concentration is indeed associated with higher
hospital prices, and also find that hospital prices are lower in the
most concentrated health plan markets compared to more compet-
itive health plan markets. Ho and Lee (2013) find heterogeneous
effects of insurer competition on negotiated hospital prices; while
increased insurer competition actually reduces hospital prices on
average, they observe a positive and significant effect on the prices
negotiated by the most attractive hospitals.3
Similar effects have been documented among physician
markets. Schneider et al. (2008) find that higher physician concen-
tration is associated with higher prices but find no effect of insurer
concentration on prices, while Dunn and Shapiro (2012) find that
physician prices are higher in concentrated physician markets and
lower in concentrated insurance markets. Additionally, Dunn and
Shapiro (2013) find that negotiated physician prices increased as a
result of health reform in Massachusetts, including some evidence
suggesting that these price increases are at least partly attributable
to increased competition among insurers. However, the outcome
of interest in each of these studies is the prices negotiated between
insurers and providers, leaving open the question of whether and,
if so, the extent to which such prices are ultimately passed through
to consumers in the form of higher premiums.
In perhaps the most closely related paper to our study, Town
et al. (2006) analyze the effects of hospital industry consolida-
tion in the 1990s on HMO premiums. They derive a theoretical
model demonstrating the effects of horizontal mergers in upstream
markets on consumer prices in downstream markets, and apply
this model to the hospital (i.e., the upstream input to the prod-
uct of health insurance) and health insurance (i.e., the downstream
output) industries. Their theory predicts that the effects of con-
solidation in the upstream industry will have differential effects
on the price and quantity of the downstream product dependent
on the level of competition in the downstream product industry,
and their empirical findings support this theory. Specifically, they
find that the hospital mergers that occurred in the 1990s resulted
3
Several other papers also focus on the effect of the type of hospital with
respect to negotiation between hospitals and insurers (which we do not consider
in our empirical analysis). Ho (2009) develops a sophisticated model of the insurer-
hospital bargaining game, estimating the expected division of profits between
insurers and hospitals. She finds that specific hospital features have important
effects on the outcome of this bargaining game – that “star” and capacity constrained
hospitals have stronger bargaining leverage with insurers and higher profits. This
result is also documented by Berenson et al. (2012) who, using data from qualitative
interviews with hospital and insurance executives from the Community Track-
ing Study, find that “must-have hospital systems. . .can exert considerable market
power to obtain steep payment rates from insurers.” Lewis and Pflum (2014) also
find that multi-market participation by a hospital system may increase bargaining
leverage.
in higher HMO premiums and reductions in insurance coverage,
and that these effects were strongest among competitive insurance
markets.
3. Conceptual framework for premiums
Premiums set by insurers for a given employer represent a com-
bination of expected medical spending covered by the insurer and a
loading factor. The loading factor reflects the insurer’s administra-
tive costs (such as marketing and paying claims) and any possible
mark-up in the profit margin resulting from the insurer exercising
market power in selling the insurance policy. Expected medical
spending is a function of prices and quantities of medical care to
be consumed. Prices generally represent the outcome of negotia-
tions between insurers and providers, and the expected quantity of
healthcare consumed generally reflects the generosity of the plan
and the health status and other features of the group covered.4
As noted above, an increase in the level of concentration in
the insurance market likely has offsetting effects on premiums
as market concentration may differentially affect the loading and
expected spending components of the premium. Regarding the
loading component of the premium, the most straightforward
effect of increasing insurance market concentration is the likely
positive effect on loading as the insurer gains more market power
and attains higher profit margins on policies sold to employers.5
However, higher levels of insurer market concentration may poten-
tially yield efficiencies in certain administrative costs such as lower
advertising costs and an increased ability to spread certain fixed
costs over a larger population. Would an insurer with increased
market power ever pass any portion of these saving in adminis-
trative costs along to consumers in the form of relatively-lower
premiums? Consider the extreme case of one monopolist insurer
setting the price of the premium such that its marginal revenue
(from selling an additional policy) equals its marginal cost. Unless
the aggregate demand for insurance is completely inelastic, any
decrease in the marginal cost (from administrative efficiencies via
larger market share) implies a partial decrease in the premium (to
thus reduce marginal revenue in equilibrium). That said, the over-
all effect of increased market concentration would seem to likely
increase premiums, with the partial effect of increased profits on
higher premiums exceeding the partial effect of reduced adminis-
trative costs on lower premiums, unless the aggregate demand for
insurance is highly elastic.
Regarding the expected spending components of the premium,
increased insurance market concentration may also result in lower
healthcare spending, as the insurer gains stronger bargaining
leverage with hospitals and is able to negotiate lower payment
rates. However, the extent to which these lower provider prices
attained by an insurer are passed along as savings to consumers
in the form of lower premiums is also unclear. Similar to the
4
Quantity consumed is also a function of the price; however, here we are focusing
on prices in terms of total transaction price negotiated with the hospital by the
insurer. Given the presence of insurance coverage, the portion of this price faced
by the consumer seeking medical care is likely to be considerably smaller than this
negotiated transaction price, so the price effects would likely reflect the change in
consumer cost-sharing, rather than the change in overall price. In a similar paper
which disentangles the price and quantity effects on physician services consumed,
Dunn and Shapiro (2012) find very small price effects on quantity consumed in this
state of insurance coverage. Additionally, McKellar et al. (2013) find that, despite
an inverse correlation between market-level private prices and utilization, overall
the price effect dominates, resulting in a positive relationship between prices and
spending.
5
Competitive pressures on insurers could also lead to improvements in quality
for the insurance plan, holding spending constant, although Scanlon et al. (2008)
find no evidence to support competition’s effect on quality.
 E.E. Trish, B.J. Herring / Journal of Health Economics 42 (2015) 104–114
above consideration of reduced administrative costs, a reduction
in negotiated provider prices is essentially a downward shift in
the insurer’s marginal cost curve. If the price elasticity of demand
for insurance is completely inelastic, all of the savings from lower
provider prices paid by the insurer would be retained by the
insurer as higher profits. Otherwise (unless the price elasticity
is extremely high), a portion of the savings from lower provider
prices would likely be passed on to consumers as lower premiums
(tied with the desire to sell more policies)6 while a portion of
the savings from lower provider prices would be retained by the
insurer as a higher profit margin.
Conversely, as hospital markets become more concentrated,
hospitals may gain stronger leverage in the bargaining game with
insurers, resulting instead in higher premiums via increased spend-
ing due to higher negotiated payment rates to hospitals. Moreover,
as hospital markets become more competitive, the effect of insurer
concentration may have a negligible impact on hospital prices if
those prices cannot be negotiated downwards any further by insur-
ers due to hospital solvency constraints.
As a result, the relative magnitudes of these potentially off-
setting effects of increasing insurance concentration on health
insurance premiums are not clear. Our study therefore aims to
empirically isolate some of these potentially countervailing effects
of health insurance concentration, and their interaction with local
hospital market concentration, on health insurance premiums in
the employer-sponsored insurance market.
4. Empirical model and data
4.1. Empirical overview
We run plan-level OLS regressions to test the relationship
between insurer and hospital market concentration and employer-
sponsored fully-insured premiums from the KFF/HRET Employer
Health Benefits Survey from 2006 through 2011. These models
use the logged single-employee’s total annual premium (i.e., the
employer and employee shares combined) as the dependent vari-
able of interest and include plan, firm, industry, and market-level
controls for premiums. Our model uses continuous HHI measures
for insurer and hospital market concentration and, as noted above,
incorporates two separate measures of insurance market concen-
tration to disentangle the effects of insurer market concentration
in the market for selling fully-insured coverage to employers from
those effects of insurer market concentration in the market for
bargaining over service prices with hospitals.
An important limitation of our analysis is that we ultimately
rely on cross-sectional geographic variation in these market con-
centration measures for both insurers and hospitals, and thus the
endogeneity of these market concentration measures is a potential
concern. A good instrument for cross-sectional variation in market
concentration is simply not apparent to us. Many studies therefore
use variation over time in these market concentration measures,
but we think that firm decisions to merge with one another are also
likely endogenous to market characteristics themselves. Regard-
less, there is very little within-market variation over time in either
the insurer HHIs or the hospital HHIs during this 2006 through 2011
time period for our premium data. (Despite this lack of variation
over time, we tested models including market-level fixed-effects
6
In the presence of adverse selection, insurers may also pass on savings in the
form of lower premiums in an effort to attract a healthier risk pool. For example,
Starc (2014) shows that medical spending is positively associated with premiums in
the Medigap market and that adverse selection in this market somewhat restrains
insurer premium markups despite insurer market power.
107
models anyway, but these models yielded insignificant results.)
Our prior is that exogenously-high insurer profits would increase
insurer competition and exogenously-high hospital prices would
increase hospital competition, leading to a bias against observing
our hypothesized positive effects of HHIs on premiums.
4.2. Data
We obtain data on employer-sponsored health insurance pre-
miums from a restricted-use version of the annual KFF/HRET
Employer Health Benefits Survey for 2006–2011. (The public-use
version of this dataset does not have geographic identifiers.) The
KFF/HRET survey provides nationally representative data regard-
ing employers’ health benefits offerings for roughly 2000 firms
per year. The data include plan-level information on the largest
plan of each type of plan (i.e., HMO, PPO, etc.) offered by the
employer in the year. We obtain plan-level premiums, type of
plan, and generosity factors such as deductible and out-of-pocket
maximum information from these data and focus our regression
analyses on single coverage. We also include firm-level control
variables from these data including firm industry, size, unioniza-
tion, and workforce characteristics. We restrict our analysis to
employers purchasing fully-insured coverage by excluding self-
insured employers. We also exclude rural employers from our
analyses, as we ultimately link these data to market concentration
measures constructed for urban markets. Finally, we exclude obser-
vations with premiums in the highest and lowest one percentile of
the distribution of the data.
We also include time-variant market-level control variables in
these premium regressions. These include mean per capita income
at the CBSA-level, which we obtain from the Bureau of Economic
Analysis, and the age, sex, and race-adjusted mean annual Medicare
hospital reimbursement per enrollee at the HRR-level, obtained
from the Dartmouth Atlas of Health Care, which we include to con-
trol for local variation in practice patterns that would be expected
to affect utilization and therefore premiums. Additionally, we con-
trol for state-level premium tax rates and an index of high-cost
state-mandated benefits,7 both of which may increase premiums
for fully-insured coverage. We use a one year lag for all market-
level variables, except the premium tax rates and mandated benefit
index, which are contemporaneous (though highly invariant over
the time period studied).
4.3. Market definitions
We construct HHI measures of insurance market concentra-
tion from the HealthLeaders-InterStudy census of private insurers
and subsequently merge these market concentration measures to
the KFF/HRET data.8 The HealthLeaders-InterStudy data include
enrollment at the managed care organization (MCO)-product-
county-level for each year.9 We construct two distinct measures
7
The index is constructed by summing the number of high cost benefit mandates
in effect for the given year in the state in which the policy is sold. “High cost” man-
dates are defined as those for which associated healthcare spending is estimated to
be more than 1% of overall premium by the Council for Affordable Health Insurance
(2006–2011).
8
The HHI is the sum of the squared market shares of each competitor in the
market, and is a commonly used measure of market competitiveness in horizontal
merger analyses conducted by the DOJ and FTC. The measure ranges from 0 to 10,000
with 10,000 representing a perfect monopoly. We scale this by 100 points (such that
the HHI ranges from 0 to 100) in all of our regression analyses for easier presentation
of results.
9
The InterStudy data have been criticized for work on health insurance
markets due to concerns with accuracy and consistency (see, for example,
Dafny et al., 2011). One important point to note is that earlier criticisms of the
108 E.E. Trish, B.J. Herring / Journal of Health Economics 42 (2015) 104–114
of insurance market concentration based on the market shares of
the relevant transaction.
For the market transaction in which the insurer sells fully-
insured coverage to employers, we define the product market as
all fully-insured managed care insurance products and aggregate
the MCO’s enrollment in these products within a defined geo-
graphic market (described below). We refer to this measure as the
“Insurer:Employer HHI” representing the level of competition of
the market in which the employer is purchasing a fully-insured
managed care product to provide coverage for its employees.
For the market transaction in which the insurer uses its bargain-
ing leverage to negotiate with hospitals, we define the insurance
product market by aggregating each insurer’s enrollment in a
geographic market for its entire commercial book of business
(i.e., combined enrollment in fully-insured and self-insured man-
aged care products) because that full set of commercially insured
patients represents that insurer’s purchasing power. We refer to
this measure as the “Insurer:Hospital HHI” measure. We exclude
observations in markets in which the HealthLeaders-InterStudy
data provide implausibly high, low, or variant total enrollment.
To construct the measure of market concentration for hospi-
tal services, we use data from the AHA annual survey. We include
all non-federal short-term general acute care hospitals in the US.
We define the product market as the number of private-pay inpa-
tient days aggregated to the hospital system within the geographic
market.10 This concentration measure, referred to as “Hospital HHI”
represents the relative bargaining strength of the local hospital
market with which insurers must negotiate hospital prices.
We exclude plan enrollment and hospital admissions among
three specific integrated delivery systems – namely, Kaiser Per-
manente, Geisinger Health System, and Intermountain Healthcare
– from the calculations of the Insurer:Hospital and Hospital HHI
measures, respectively, in the geographic markets where their
hospitals exclusively treat patients from the integrated insurer
and there is thus no relevant hospital price negotiation. However,
we do not remove the enrollment among these integrated delivery
InterStudy data related to the fact that they only measured enrollment in HMOs do
not apply to our study, as information on PPOs and other products was added begin-
ning in 2005 associated with combining with HealthLeaders. Nonetheless, there are
still concerns regarding the validity and volatility of enrollment. We have addressed
these concerns in several ways. In particular, we have removed some enrollment
to address the double counting issue of “rental network” enrollment, particularly
in 2007–2008, following our own analysis and discussions with database managers
at HealthLeaders-InterStudy. Additionally, we have taken several steps to address
volatility in the data. First, we have taken the average MCO-product enrollment
of the two observations per year (January and July) and used this as MCO-product
enrollment for the year. Next, we aggregate total managed care enrollment (fully-
and self-insured) in the data at the market (CBSA/Division or county) level, and com-
pare this enrollment to estimates of the under-65 population for the market, which
we obtain from the Small Area Health Insurance Estimates. We exclude from our
analyses any markets where the aggregate private enrollment in the HealthLeaders-
InterStudy data is less than 30% or greater than 100% of the total under-65 population
in the market. We believe these are conservative cutoffs, as the under-65 population
includes not only those that are privately insured, but also those that are uninsured
and those with Medicaid or another source of public coverage (such as VA, Medicare,
etc.). Additionally, we drop any market-year observations for markets in which the
HHI is more than 25% greater or less than the mean HHI of that same market across
the six-year time period included in our study (i.e., an implausibly large insurance
market one-year outlier). Overall, these restrictions result in an exclusion of about
20% of the total plan-level observations in the KFF/HRET data. These excluded mar-
kets tend to have higher levels of insurance market concentration (likely due to
mis-measurement), but are otherwise similar to the markets retained in our study.
10
We also run models with alternative definitions of hospital product market, such
as beds, total volume, total admissions, and Medicare discharges; hospital market
concentration based on these different measures are all very highly correlated and
our results are robust to these alternative definitions. Moreover, we believe that
using the system-level measures (rather than individual hospitals) more accurately
represent the bargaining nature with insurers.
systems from the Insurer:Employer HHI calculations, as these
plans compete with other plans for employer coverage.11
We consider two ways to define geographic markets: Core-
Based Statistical Areas (CBSA) and counties. The CBSA is a
geographic area defined by Office of Management and Budget to
represent an area with commuting ties to an urban center. The
11 largest CBSAs (e.g., greater New York City, greater Chicago) are
separated into smaller Metropolitan Divisions (e.g., four Divisions
within New York City, three Divisions within Chicago), and so we
use the smaller Metropolitan Division codes, when available, to
define the geographic markets within these larger CBSAs.12 While
we believe that HHIs using the CBSA as the geographic market def-
inition should reasonably characterize the market transactions for
relatively-smaller employers purchasing coverage among compet-
ing private health insurers and reasonably characterize the market
transactions between private health insurers and hospital systems,
we also construct HHI measures using counties as the geographic
market. Accordingly, we run a regression model using the CBSA for
these three HHIs and then a separate regression model using the
county for these three HHIs and report results for both measures.13
The joint distribution of the Insurer:Employer and Insurer:
Hospital CBSA-based HHI measures is shown in Fig. 1A. While the
two measures are strongly correlated across markets (i.e., the cor-
relation coefficient is 0.83), there is actually a considerable level
of differences between the Insurer:Employer and Insurer:Hospital
HHI measures, so that we are able to disentangle these opposing
effects of higher profits and lower hospital prices on premiums.
The joint distribution of relative bargaining leverage (i.e.,
Insurer:Hospital HHI and Hospital HHI for CBSAs) is shown in
Fig. 1B. The correlation coefficient is 0.22, indicating that there is
a mix of markets where the insurers have more bargaining power
than hospitals, insurers have less bargaining power than hospitals,
and insurers and hospitals have comparable bargaining power. The
DOJ/FTC Horizontal Merger Guidelines provide particular HHI cut-
offs as one way to categorize the level of competition in a market;
by these standards, markets with an HHI between 1500 and 2500
are considered moderately concentrated, and markets with an HHI
greater than 2500 are considered highly concentrated (US DOJ/FTC,
2010).
For the Insurer:Employer HHI measure using the CBSA for the
market definition, 2.6% of plans are in un-concentrated markets,
39.3% are in moderately concentrated markets, and 58.1% are in
highly concentrated markets. For the Insurer:Employer HHI mea-
sure using the smaller county for the market definition, 1.8% of
plans are in un-concentrated markets, 35.1% are in moderately con-
centrated markets, and 63.1% are in highly concentrated markets.
For the Hospital HHI measure using the CBSA for the market def-
inition, 27.2% of plans are in un-concentrated markets, 29.8% are
in moderately concentrated markets, and 43.1% are in highly con-
centrated markets. For the Hospital HHI measure using the county
for the market definition, 10.8% of plans are in un-concentrated
markets, 20.1% are in moderately concentrated markets, and 69.1%
are in highly concentrated markets. Only 1.0% of plans are in
11
The results are qualitatively unchanged if we either do not exclude this enroll-
ment and/or if we simply drop observations in markets with integrated delivery
systems present in the market.
12
The results are also qualitatively unchanged if we use CBSAs to define geographic
markets without using the smaller Metropolitan Divisions within these 11 largest
CBSAs.
13
We also run models using the Dartmouth Atlas’ Hospital Referral Region (HRR)
as the geographic market for both Insurer:Hospital HHIs and Hospital HHIs, as the
HRR has been frequently been used as a geographic market for healthcare. HRRs are
generally larger geographic areas than CBSAs, so CBSA-level markets are typically
more highly concentrated than HRR-level markets.
 E.E. Trish, B.J. Herring / Journal of Health Economics 42 (2015) 104–114 109

100
mandated benefit index. (We lag the income and Medicare utiliza-
tion values because they are essentially t − 1 forecasts by an insurer
while the tax rates and benefit mandates are known in advance.)
 t is a year-indicator variable and εpt is the random error. We use
80

cluster-corrected robust standard errors at the market-year level.

We first estimate the model for the entire sample. We then con-
Ins:Emp HHI
60

duct sensitivity analyses by systematically excluding one or two

of the three market concentration measures to allow us to exam-
40

ine our ability to disentangle the effects of insurer concentration

through these measures. We then also estimate separate models for
different subsamples, by first stratifying the sample by insurance
20

market concentration and then stratifying the sample by hospital

market concentration. These stratified models allow us to examine
whether the effects of insurer and hospital market concentration
0

0 20 40 60 80 100 on premiums appear to vary across markets. We use an HHI of 2500

Ins:Hosp HHI
(A) for stratifying these samples, as this is the FTC/DOJ cutoff for a high
level of concentration and it splits the sample roughly in half for
the CBSA-defined markets.
100

5. Results and discussion

80

The first two columns of Table 1 include the weighted means and
Hospital HHI
60

standard deviations (where applicable) of the variables used in our

analyses. The weighted mean annual single-employee premium in
our sample is $4567; this, as well as the insurer and hospital mar-
40

ket concentration measures discussed below are similar to overall

population-weighted national measures over this time period and
20

thus do not appear to be idiosyncratic to the KFF/HRET sampling

(which is designed to be nationally representative) or our exclusion
criteria.
0

0 20 40 60 80 100
Ins:Hosp HHI
(B)
Fig. 1. Comparison of the two insurance market concentration measures and the
joint distribution with hospital market concentration using CBSA market definitions.
Notes: The scatterplots depict the joint distribution of the Insurer:Employer and
Insurer:Hospital CBSA-based HHI measures of insurance market concentration in
Panel (A) and the joint distribution of Insurer:Hospital and Hospital CBSA-based
HHI measures for market concentration in Panel (B). Each dot represents a plan
from the 2006–2011 KFF/HRET Employer Health Benefits Survey. HHI is Herfindahl-
Hirschman Index.
CBSA markets where both insurer and hospital markets are un-
concentrated. In contrast, 31.6% of plans are in CBSA markets where
both insurer and hospital markets are highly concentrated.
4.4. Empirical model
We estimate parameters from the following OLS plan-level pre-
mium regression:
ln Ppt = ˛ + ˇIns:Emp HHImt−1 + ϕIns:Hosp HHImt−1
+ Hosp HHImt−1 + Xp + ıFf + Mmt−1 + t + εpt
where the indices are plan p, firm f, market m, and year t. The
Ins:Emp HHI term in this equation is the one-year lagged HHI of
the market in which insurers sell fully-insured policies to employ-
ers, the Ins:Hosp HHI term is the one-year lagged HHI of the market
in which insurers bargain with hospitals, and the Hosp HHI term
is the one-year lagged HHI of the hospital market. The Xp and Ff
covariates are plan-level and firm-level control variables, while the
Mmt−1 are market-level controls, including the one-year lagged and
logged CBSA-level per capita income, the lagged and logged mean
HRR-level Medicare hospital reimbursement values, the contem-
poraneous state premium tax rate, and the contemporaneous state
The next three columns of Table 1 present the full results from
the OLS regression model for the annual premium shown in Eq.
(1). Before discussing the results for insurer and hospital market
concentration, we note that the results for the control variables
generally appear as expected, indicating that the overall data and
model is well specified. For instance, plans with higher deductibles
have lower premiums, unionized firms have higher premiums, and
smaller firms have higher premiums.
In this model using the CBSA as the geographic market, the
coefficient for a 100 point increase in the Insurer:Employer HHI
is 0.0021 and the coefficient for a 100 point increase in the Hospi-
tal HHI is 0.0019. These findings, which are statistically significant
at the 5% and 1% levels, respectively, support our hypothesis that
higher levels of both Insurer:Employer and Hospital concentration
are associated with higher employer-sponsored health insurance
premiums. To put a relative magnitude on these coefficients, we
consider their effect size in the commonly-used example of a stan-
dard “five-to-four” merger – a market in which two of five equally
sized firms merge, resulting in an 800 point increase in HHI (i.e., an
HHI increase from 2000 to 2800). These coefficient estimates imply
that a simulated five-to-four merger in the Insurer:Employer mar-
ket is associated with 1.7% ($78) increase in premiums, and the
(1)
same increase in the Hospital HHI is associated with a 1.5% ($67)
increase in premiums.
The coefficient for a 100 point increase in the Insurer:Hospital
HHI is −0.0024, also statistically significant at the 1% level. A simu-
lated five-to-four merger in this insurer bargaining leverage market
is associated with a 1.9% ($90) decrease in predicted premiums. This
finding of a positive coefficient on the Insurer:Employer HHI term
and a negative coefficient on the Insurer:Hospital HHI term pro-
vides support that there are indeed offsetting effects of increases in
insurer concentration in terms of market power in selling insurance
to employers (increasing premiums) versus negotiating leverage
with hospitals (decreasing premiums).
110 E.E. Trish, B.J. Herring / Journal of Health Economics 42 (2015) 104–114

Table 1
Summary statistics and premium regression results for insurance and hospital market concentration.

Dependent variable Mean Std Dev Coeff Std Err p-Value

ln(Premium) 8.43 0.30 3.9720 0.4987 0.000

Market concentration variables
Insurer:Employer HHI (CBSA) 31.07 13.44 0.0021 0.0010 0.029
Insurer:Hospital HHI (CBSA) 27.39 12.65 −0.0024 0.0009 0.006
Hospital HHI (CBSA) 30.78 23.79 0.0019 0.0003 0.000
Plan-level controls
HMO 30.9% Ref
PPO 49.1% 0.1062 0.0154 0.000
POS 20.0% 0.0949 0.0197 0.000
Annual deductible ($000s) 0.44 0.72 −0.0581 0.0099 0.000
OOP Max < $1500 16.5% 0.0080 0.0156 0.607
Firm-level controls
Unionized workers 21.2% 0.0810 0.0156 0.000
Percent low income (<$23,000) 13.0% 20.1% −0.0564 0.0315 0.073
Percent part time 12.0% 16.5% 0.0344 0.0340 0.312
Firm size: 500+ 26.7% Ref
Firm size: 100–499 25.1% 0.0211 0.0152 0.166
Firm size: 25–99 23.0% 0.0232 0.0173 0.179
Firm size: 3–24 25.2% 0.0407 0.0200 0.042
Construction 7.0% Ref
Manufacturing 9.1% 0.0398 0.0306 0.193
Mining 0.4% 0.0105 0.0770 0.891
Transportation 7.4% 0.1016 0.0391 0.009
Wholesale 4.7% 0.0409 0.0335 0.223
Retail 7.2% 0.0512 0.0363 0.159
Finance 8.9% 0.1222 0.0322 0.000
Service 40.8% 0.1277 0.0291 0.000
Government 7.1% 0.2106 0.0334 0.000
Healthcare 7.3% 0.1425 0.0364 0.000
Market-level controls
ln(Per capita income) 10.62 0.20 0.2835 0.0417 0.000
ln(Medicare hospital payments) 8.38 0.20 0.1300 0.0387 0.001
State premium tax rate 0.99% 1.03% 0.8981 0.7783 0.249
State Mandated Benefit Index 9.49 4.20 −0.0004 0.0019 0.814
Regional controls
South 29.8% Ref
Northeast 20.9% 0.0855 0.0195 0.000
Midwest 16.1% 0.0793 0.0216 0.000
West 33.2% −0.0181 0.0192 0.345
Year controls
Year 2006 15.9% Ref
Year 2007 17.1% −0.0034 0.0227 0.880
Year 2008 16.5% 0.0618 0.0239 0.010
Year 2009 18.9% 0.1039 0.0223 0.000
Year 2010 16.9% 0.1439 0.0241 0.000
Year 2011 14.6% 0.1924 0.0260 0.000

Notes: The left-hand side of the table shows enrollment-weighted means and standard deviations from the 2006–2011 KFF/HRET Employer Health Benefits Survey with
geographic markets defined as CBSAs. The right-had side of the table shows enrollment-weighted OLS regression results from a plan-level regression of log annual premium
on the insurance and hospital market concentration. N = 5270; F(32, 1288) = 20.49 (p = 0.000); R2 = 0.2176. Standard errors are robust cluster-corrected at the market-year
level. The insurance and hospital market concentration measures and other market controls are lagged by one year and HHIs are scaled by 100. Percentages may not sum to
100% due to rounding. The coefficient, standard error, and p-value included in the first row are for the intercept.

The results in Table 2 illustrate how the inclusion of the two
distinct and potentially offsetting measures of insurance con-
centration (in the employer and hospital markets) appear to be
necessary to disentangle the different effects of market power on
the seller and buyer side. Table 2A’s second column (labeled Model
1) repeats our main results including all three HHI measures using
the CBSA market definition, and the next six columns show the
results from additional separate regressions to show the possi-
ble permutations of including/excluding these three concentration
measures. Our ability to capture these related but offsetting effects
of insurer concentration is supported by the fact that the magni-
tude and statistical significance of the coefficients on these terms
are diminished when only one of them is included in the regression
specification. This is illustrated in the other columns of Table 2A;
when only one of these insurance market concentration mea-
sures is included without the other (i.e., Models 2 and 3 with the
hospital HHI excluded, Models 6 and 7 with the hospital HHI
included), the portion of the offsetting effect on premiums that is
picked up by the measure limits its magnitude and significance in
the direction that we expect. This provides support for the fact that
these two insurance market concentration terms are indeed con-
tributing unique information regarding the structure of the market
in which the plan is sold and that in which the insurer bargains
over hospital prices, and their relevant association with premiums.
Finally, Model 4 indicates that our findings for insurance market
concentration are not sensitive to the exclusion of hospital market
concentration, and Model 5 indicates that our finding for the associ-
ation between higher premiums and hospital market concentration
is not sensitive to the exclusion of insurer market concentration.
Table 2B presents the results from this same set of regressions
but instead using counties as the geographic market. While the
mean concentration measures for the insurer and hospital markets
are higher for county-defined markets compared to CBSA-defined
markets (especially for hospital markets), overall the same pattern
of regression results holds for the models using county as the mar-
ket. In the model including all three concentration measures with
 E.E. Trish, B.J. Herring / Journal of Health Economics 42 (2015) 104–114 111

Table 2
Premium regression results for insurance and hospital market concentration measures excluding each measure, using (A) CBSA and (B) county market definitions.

ln(Premium) Mean Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7

(A)
Ins:Emp HHI 3107 0.0021 0.0005 0.0028 0.0001
SD/SE 1344 0.0010 0.0005 0.0010 0.0005
p-Value 0.029 0.308 0.004 0.768

Ins:Hosp HHI 2739 −0.0024 −0.0004 −0.0028 −0.0006

SD/SE 1265 0.0009 0.0004 0.0009 0.0004
p-Value 0.006 0.328 0.002 0.139

Hospital HHI 3078 0.0019 0.0019 0.0019 0.0020

SD/SE 2379 0.0003 0.0003 0.0003 0.0003
p-Value 0.000 0.000 0.000 0.000

N 5270 5270 5270 5270 5270 5270 5270

R2 0.2595 0.2623 0.2624 0.2617 0.2599 0.2600 0.2599

(B)
Ins:Emp HHI 3225 0.0025 0.0005 0.0029 0.0004
SD/SE 1351 0.0009 0.0005 0.0009 0.0005
p-Value 0.008 0.351 0.002 0.416

Ins:Hosp HHI 2844 −0.0025 −0.0006 −0.0030 −0.0005

SD/SE 1269 0.0008 0.0004 0.0008 0.0004
p-Value 0.003 0.181 0.000 0.283

Hospital HHI 4293 0.0009 0.0011 0.0011 0.0011

SD/SE 2757 0.0003 0.0003 0.0003 0.0003
p-Value 0.001 0.000 0.000 0.000

N 5210 5210 5.210 5210 5210 5210 5210

R2 0.2608 0.2624 0.2624 0.2616 0.2614 0.2614 0.2613

Notes: These seven models show selected enrollment-weighted OLS regression results from a plan-level regression of log annual premium from the KFF/HRET survey on the
insurance and hospital market concentration measures using the CBSA as the geographic market in Panel (A) and the county as the geographic market in Panel (B). Standard
errors (SE) are robust cluster-corrected at the market-year level. The three market concentration measures in the regression models are lagged by one year and scaled by
100 (i.e., the coefficients represent the effect of an HHI that is 100 points higher).

markets defined at the county-level, the coefficient for a 100 point Table 3
Premium regression results for insurance and hospital market concentration mea-
increase in the Insurer:Employer HHI is 0.0025 and the coefficient
sures, stratified by insurance market concentration.
for a 100 point increase in the Hospital HHI is 0.0009, both signif-
icant at the 1% level. Additionally, the coefficient for a 100 point ln(Premium) 1 2 3
Full sample Ins:Emp HHI Ins:Emp HHI
increase in the Insurer:Hospital HHI is −0.0025, also significant at
≤2500 >2500
the 1% level.
Ins:Emp HHI 0.0021 0.0061 0.0026
SE 0.0010 0.0039 0.0010
5.1. Stratified analyses p-Value 0.029 0.115 0.026

Ins:Hosp HHI −0.0024 −0.0004 −0.0023

Based on our discussion in Section 3, as well as the results from SE 0.0009 0.0029 0.0010
Town et al. (2006), we hypothesize that the effects of insurance and p-Value 0.006 0.879 0.026
hospital market concentration may be more (or less) pronounced Hospital HHI 0.0019 0.0020 0.0018
depending on the relative concentration of the other market. SE 0.0003 0.0006 0.0004
Specifically, the Town et al. theoretical and empirical results sug- p-Value 0.000 0.001 0.000
gest that the positive effects of hospital concentration on insurance N 5270 2207 3063
premiums will be more pronounced among competitive insurance R2 0.2595 0.2427 0.2659
markets. To test this hypothesis, we re-ran the regressions on two Notes: The first column repeats the regression results in Table 2A’s Model 1, while
subsamples of the plans, stratified by the level of competition in the next two columns show enrollment-weighted OLS regression results for premiums
Insurer:Employer market; i.e., an HHI either below or above 2500. from the KFF/HRET survey stratified by level of insurance market concentration
(We also stratify our models based on hospital concentration fur- above or below an HHI of 2500. All markets are geographically defined using CBSAs.
The three market concentration measures are lagged by one year and scaled by 100.
ther below.) We use this stratified-sample approach, rather than
Standard errors are robust cluster-corrected at the market-year level.
an interaction term, so that we can observe the baseline effects in
more competitive and more concentrated markets independently.
The findings from this analysis using CBSAs as the geographic association between hospital concentration and premiums is sig-
market are presented in Table 3.14 While we find that the nificant among plans sold in both relatively more competitive
downstream insurance markets and relatively more concentrated
downstream insurance markets, we do not find evidence that
14
Repeating these stratified analyses using the county-based HHI measures with
the same HHI cutoff above and below 2500 yields slightly different results. How-
ever, this appears to be due, at least in part, to the difference in the magnitudes
of HHIs for markets using the county-defined markets as compared to the CBSA- the HHI threshold cutoff (e.g., from 2500 to 3500 and thus roughly evenly-split
defined markets (i.e., the mean HHIs for the smaller county-defined markets are subsample sizes) for stratifying more competitive versus more concentrated mar-
higher). We instead observe a similar pattern of results using county-defined mar- kets. Additional results from these analyses are available from the authors upon
kets to those presented in Tables 3 and 4 using CBSA-defined markets if we increase request.
112 E.E. Trish, B.J. Herring / Journal of Health Economics 42 (2015) 104–114

the relationship is stronger among more competitive down- Table 4

Premium regression results for insurance and hospital market concentration mea-
stream insurance markets. The coefficient on premiums for a
sures, stratified by hospital market concentration.
100 point increase in Hospital HHI is 0.0020 in more competi-
tive Insurer:Employer markets and 0.0018 in more concentrated ln(Premium) 1 2 3
Full sample Hospital HHI Hospital HHI
Insurer:Employer markets, both significant at the 1% level.
≤2500 >2500
However, we do observe that the negative association between
the Insurer:Hospital concentration measure and premiums is Ins:Emp HHI 0.0021 0.0013 0.0030
SE 0.0010 0.0015 0.0013
stronger among more concentrated insurance markets. Specifi-
p-Value 0.029 0.378 0.023
cally, the coefficient on premiums for a 100 point increase in the
Ins:Hosp HHI −0.0024 −0.0012 −0.0034
Insurer:Hospital HHI measure is −0.0023 (p < 0.05) for plans sold
SE 0.0009 0.0012 0.0012
in more concentrated insurance markets (Column 3), and statis- p-Value 0.006 0.343 0.005
tically insignificant (−0.0004) for plans sold in more competitive
Hospital HHI 0.0019 0.0012 0.0015
insurance markets (Column 2). In addition, this stratification by
SE 0.0003 0.0014 0.0004
level of concentration in the insurance market reveals that, while p-Value 0.000 0.372 0.000
the relationship between Insurer:Employer competition and pre-
N 5270 3001 2269
miums among more competitive insurance markets is statistically R2 0.2595 0.2487 0.2632
indistinguishable from zero, the extent of Insurer:Employer com-
Notes: The first column repeats the regression results in Table 2A’s Model 1, while
petition among more concentrated insurance markets is important.
next two columns show enrollment-weighted OLS regression results for premiums
These findings suggest that the effects of the levels of con- from the KFF/HRET survey stratified by level of hospital market concentration above
centration in healthcare markets on premiums vary with the or below an HHI of 2500. All markets are geographically defined using CBSAs. The
overall market characteristics. They provide evidence that increas- three market concentration measures are lagged by one year and scaled by 100.
ing consolidation among hospital markets not only results in Standard errors are robust cluster-corrected at the market-year level.

higher negotiated prices, but that these higher prices are ulti-
mately passed-through to consumers in the form of higher
premiums, regardless of downstream insurer market structure.
The finding that the positive and negative associations between
Insurer:Employer and Insurer:Hospital market concentration and
premiums are strongest among the more concentrated insurance
markets may suggest that the association between higher levels
of insurance concentration on the ability to charge higher premi-
ums to employers and to negotiate lower prices with providers
may be particularly important among relatively more concentrated
insurance markets.
Next we examined whether a similar pattern of results would
emerge from stratifying the observations based on the level of
hospital market concentration. Here we hypothesize that the asso-
ciation of the level of competition in the Insurer:Hospital market
and premiums would be strongest among the more concentrated
hospital markets, as the effects of increased insurer bargaining
leverage may be less pronounced if hospital prices are already rela-
tively low due to hospital market competition alone. Analogous to
the above analysis, we stratified the observations into two groups
according to level of concentration in the hospital market; i.e., Hos-
pital HHI below or above 2500 using the CBSA-defined geographic
markets.
The findings from this analysis stratified by hospital market con-
centration are presented in Table 4, where Column 1 again simply
repeats the results from Table 2A’s Column 1, and Columns 2 and
3 show the results among plans sold in markets with more or less
competitive hospital markets. We find that the statistical signifi-
cance of the positive and negative associations between premiums
and Insurer:Employer HHI and Insurer:Hospital HHI, respectively,
hold only among the more concentrated hospital markets. We
also find that the statistical significance of the positive association
between Hospital HHI and premiums holds only among the more
concentrated hospital markets. These findings provide support for
our hypothesis that, while hospital prices may already be rela-
tively lower among more competitive hospital markets, in more
concentrated hospital markets, more concentrated insurers may
leverage their stronger bargaining power to negotiate lower prices
among concentrated hospital markets and use their market power
in selling insurance to employers to increase their profit margins.
Taken together, these results (from Tables 1 and 2) suggest that
the levels of competition in the health insurance and hospital mar-
kets are significantly associated with employer-sponsored health
insurance premiums. Further, these results (from Tables 3 and 4)
suggest that the negative relationship between premiums and
insurer bargaining power with hospitals may be particularly pro-
nounced among more highly concentrated insurer and hospital
markets. Generally, they support the suggestion that the relative
balance of insurer and hospital concentration also has important
implications for insurance premiums, reflecting the underlying
market structure that insurers must bargain with hospitals to set
transaction prices, and thus the level of concentration in both
insurer and hospital markets and their relative bargaining leverage
jointly impact these negotiated prices. Importantly, they provide
empirical evidence that these higher provider prices are often
ultimately passed-through to consumers in the form of higher pre-
miums, and that this pass-through also depends on relative market
conditions. In general, we observe higher premiums among plans
sold in markets with higher levels of hospital and Insurer:Employer
concentration and lower premiums among plans sold in markets
with higher levels of Insurer:Hospital concentration, especially
among more highly concentrated markets.
5.2. Limitations
As noted earlier, the cross-sectional design of this study limits
our ability to infer the causal relationship between the concen-
tration of insurance and hospital markets and health insurance
premiums. It is difficult to construct instruments that would rep-
resent an exogenous source of variation in market concentration
that would be unrelated to premiums, particularly instruments
that would be uniquely predictive of the two different insurer and
the hospital concentration measures. One approach (for examining
provider prices) is to instrument for insurer concentration using the
underlying distribution of firms, but it is doubtful that this would
be unrelated, independently, to employer-sponsored premiums.
Another common approach is to exploit mergers as a source of
variation in concentration over time, but there was little consol-
idation activity for insurers and hospitals during this time period
for which we have these rich KFF/HRET data. Moreover, it is not
clear that the merger itself would necessarily be exogenous. Yet
another commonly-used approach is to construct measures of mar-
ket concentration based on predicted, rather than actual patient
flows, but we are limited by data and analytical resources to com-
plete this exercise at the national level for hospitals and unsure how
 E.E. Trish, B.J. Herring / Journal of Health Economics 42 (2015) 104–114
one would apply this consumer-flow approach to insurer market
shares. We have tried to alleviate these endogeneity concerns by
lagging the market concentration and market control variables by
one year; this at least implies a temporal relationship that is con-
sistent with the hypothesis that the level of market concentration
affects health insurance premiums. Additionally, if higher premi-
ums in fact encourage market entry by other insurers, this would
result in a more competitive health insurance market, which would
bias our results downward. Nonetheless, we interpret our results
as associations between market concentration and premiums and
not necessarily a causal relationship.
Another limitation is that our model also measures the asso-
ciation between aggregate market-level measures of insurer and
hospital concentration and the premium of a specific insurance
plan purchased by an employer in that market. The KFF/HRET data
do not allow for the identification of which insurer sold the pol-
icy to the employer, so we therefore are also unable to link the
specific market share of that insurer to the observation. Similarly,
we do not know anything about which hospitals are included in
a given plan’s network, nor about the insurer-hospital contracts.
Thus, due to data limitations, we are unable to model the bargain-
ing between insurers and hospitals and to consider constructs such
as “Option Demand/Willingness to Pay” (Capps et al., 2003) mea-
sures for inclusion of certain hospital systems and the effects that
this may have on premiums. Nonetheless, while the market mea-
sures may not reflect the specific insurer and associated network
from which the plan is purchased, they do represent the overall
market conditions within which the employer is choosing a policy.
Thus, we believe that they provide important information regard-
ing the relationship between market conditions and policies sold in
those markets. Additionally, while we control for some plan gen-
erosity features such as plan type, deductible, and out of pocket
maximum, premium variation may also reflect differences in qual-
ity and plan generosity that are not accounted for by these control
variables (including the plan’s network) which could conceivably
be correlated with the extent of market concentration.
Additionally, our market concentration measures are reliant
on how we have chosen to define the markets. While we believe
that CBSAs represent a reasonable geographic market for employ-
ers purchasing insurance and a reasonable geographic market for
hospital care, the extent to which our measures accurately rep-
resent the true level of competition in these markets depends on
the degree to which they accurately reflect the markets in which
insurance is purchased and hospital network inclusion negotiations
occur, respectively. However, given that CBSAs are constructed to
represent person-flows for commuting to employment, we believe
that they represent reasonable choices for the markets we are try-
ing to measure. We are also reassured by the observation that our
results are similar when using counties instead of CBSAs.
Further, given that these KFF/HRET data are of plans actually
purchased by employers, our analysis does not explicitly model the
employers’ option to not offer coverage or to self-insure as an alter-
native choice.15 However, given that all the plans in our model are
indeed purchased, they reflect choices that employers have made
dependent on the local market conditions that do in fact exist,
and thus we feel that they represent an appropriate association
between market conditions and premiums.
15
We have also examined firm-level decisions to self-insure using these KFF/HRET
data, with a primary focus on examining the influence of state community rating
rules for low-risk versus high-risk industries among firms with 25–100 workers
(Trish and Herring, 2014). In those analyses, there were no statistically significant
associations between insurer market concentration nor hospital market concentra-
tion (included as control variables) and a firm’s decision to self insure.
113
Overall, we believe that our ability to construct two distinct
insurance market concentration measures using the variation in
fully-insured and self-insured enrollment represents an improve-
ment in depicting these markets and their unique association with
health insurance premiums. Nonetheless, further work on these
open questions is warranted.
6. Conclusion
The US health insurance industry is highly concentrated and
health insurance premiums are high and rising rapidly. Our data
demonstrate that less than 3% of the markets in which employers
purchase fully-insured coverage are considered un-concentrated
by the guidelines set forth by the DOJ/FTC. Similarly, more than
half of these markets are considered highly concentrated. Provi-
sions included in the ACA are focused on increasing competition
in these markets, with the expectation that increased competi-
tion within the health insurance industry would help to lower
premiums. Though focused on the individual rather than the
employer-sponsored market, early evidence suggests that not all
state insurance markets are, in fact, becoming more competitive
(Cox et al., 2014) but that more competitive exchanges have lower
premiums (Dafny et al., Forthcoming).
However, health insurers operate in a complex bilateral
oligopoly, whereby they must negotiate service prices with hos-
pitals and other providers, and higher levels of market power may
in fact result in stronger bargaining leverage with these providers
to drive down prices, which could then be partially passed through
in the form of lower premiums. Thus, the ultimate impact of the
level of competition in the health insurance industry on premiums
is unclear – but it seems likely that the underlying goal of reduc-
tions in insurer administrative overhead associated with increased
insurer competition can generally not be achieved without the
unintended consequence of higher provider prices associated with
decreased bargaining power with providers. The analyses pre-
sented herein suggest that the effects of increasing competition in
health insurance markets on health insurance premiums are likely
to depend on the level of competition in local hospital markets,
as well as the relative competitiveness of the fully-insured and
self-insured markets and insurers’ overall bargaining leverage with
hospitals and other local providers.
We find that employer-sponsored insurance premiums among a
nationally representative sample of firms purchasing fully-insured
products are higher in markets where insurance and/or hospital
markets are highly concentrated, as compared to those in which
they are more competitive. Further, we find that higher levels of
concentration among the market in which insurance is sold to
employers are associated with higher premiums, whereas higher
levels of concentration among the market in which insurers bargain
with hospitals are associated with lower premiums. Importantly,
we find that higher levels of concentration in hospital markets are
also associated with higher premiums – providing evidence that
the well-documented higher prices resulting from consolidation
among hospitals do in fact affect consumers in the form of higher
premiums, and that local market conditions affect the extent of this
pass-through.
However, our findings, along with recent literature suggesting
that hospital prices are lower among more concentrated insur-
ance markets, suggest that higher levels of insurer bargaining
leverage with hospitals may lead to lower health insurance pre-
miums via lower negotiated hospital prices, as long as there
is sufficient competition in the market for selling insurance to
small employers that these lower prices get passed through to
employers in the form of lower premiums. Recent policy changes,
114 E.E. Trish, B.J. Herring / Journal of Health Economics 42 (2015) 104–114
such as the introduction of minimum medical loss ratios, may
help to ensure that such savings are passed through, even in the
absence of higher levels of competition in this market. Addition-
ally, our results suggest that this important negative relationship
between insurer bargaining power and premiums is particularly
pronounced among more highly concentrated markets. Taken
together, these findings suggest that ACA provisions to increase
competition in health insurance markets may be unsuccessful by
not also considering the level of concentration in local hospital mar-
kets, particularly if they dilute insurers’ overall bargaining leverage
with local hospital systems. This may be particularly problematic
due to recent provider consolidation trends and the strong incen-
tives for such provider consolidation included in the ACA, such
as increased horizontal and vertical integration resulting from the
formation of Accountable Care Organizations. Therefore, efforts tar-
geted toward reducing health insurance premiums may be better
directed toward insurer–provider negotiations and rate regula-
tion, or efforts to simultaneously reduce the level of concentration
among insurers and providers.
Acknowledgements
This research was supported by Grant No. 69070 from the Robert
Wood Johnson Foundation’s Changes in Healthcare Financing and
Organization (HCFO) initiative. We would like to thank Anthony
Damico, Matthew Rae, and Gary Claxton of the Kaiser Family Foun-
dation for their assistance with the KFF/HRET Employer Health
Benefits Survey data and Adele Shartzer for her assistance with the
AHA data. We are grateful for helpful comments by two anonymous
reviewers, Chris Garmon, Jeff Stensland, Abe Dunn, Vivian Wu,
Rich Lindrooth, Lisa Dubay, and seminar participants at CBO, NCHS,
Urban Institute, Emory University, Mathematica Policy Research,
Analysis Group, University of Massachusetts-Amherst, RAND Cor-
poration, University of Wisconsin, Colorado School of Public Health,
Virginia Commonwealth University, an RWJF HCFO grantee brief-
ing, AcademyHealth’s Annual Research Meeting, and the American
Society of Health Economists Conference.
References
American Medical Association, 2013. Competition in Health Insurance. A Compre-
hensive Study of US Markets: 2013 Update.
Berenson, R., Ginsburg, P., Christianson, J., Yee, T., 2012. The growing power of
some providers to win steep payment increases from insurers suggests policy
remedies may be needed. Health Affairs (Millwood) 31 (5), 973–981.
Capps, C., Dranove, D., Satterthwaite, M., 2003. Competition and market power in
option demand markets. RAND Journal of Economics 34 (4), 737–763.
Council for Affordable Health Insurance, 2006–2011. Health Insurance Mandates in
the States 2006–11. The Council for Affordable Health Insurance.
Cox, C., Ma, R., Claxton, G., Levitt, L., 2014. Sizing Up Exchange Market Competition.
Kaiser Family Foundation Issue Brief.
Cutler, D.M., Scott Morton, F., 2013. Hospitals, market share, and consolidation.
Journal of the American Medical Association 310 (18), 1964–1970.
Dafny, L., 2010. Are health insurance markets competitive? American Economic
Review 100 (4), 1399–1431.
Dafny, L., Dranove, D., Limbrock, F., Morton, F.S., 2011. Data impediments to empir-
ical work on health insurance markets. The B.E. Journal of Economic Analysis &
Policy 11 (2), Article 8.
Dafny, L., Duggan, M., Ramanarayanan, S., 2012. Paying a premium on your pre-
mium? Consolidation in the US Health Insurance Industry. American Economic
Review 102 (2), 1161–1185.
Dafny, L., Gruber, J., Ody, C., 2015. More insurers lower premiums: evidence from
initial pricing in the health insurance marketplaces. American Journal of Health
Economics 1 (1), 53–81.
Dartmouth Atlas of Health Care, 2012. Data by Region, Available from: http://www.
dartmouthatlas.org/data/region/
Dranove, D., Gron, A., Mazzeo, M., 2003. Differentiation and competition in HMO
markets. Journal of Industrial Economics 51 (4), 433–454.
Dunn, A., Shapiro, A., BEA Working Papers 0084 2012. Physician Market Power and
Medical-Care Expenditures. Bureau of Economic Analysis.
Dunn, A., Shapiro, A., Working Paper 2013-36 2013. The Impact of Health Care
Reform on Physician Payments: Evidence from Massachusetts. Federal Reserve
Bank of San Francisco.
Gaynor, M., Town, R., 2011. Competition in health care markets. In: Pauly, M.,
Mcguire, T., Barros, P. (Eds.), Handbook of Health Economics, vol. 2. Elsevier,
Amsterdam, pp. 499–637.
Gaynor, M., Town, R., 2012, June. The Impact of Hospital Consolidation – Update.
Robert Wood Johnson Foundation Synthesis Report.
Gaynor, M., Vogt, W., 2000. Antitrust and competition in health care markets. In:
Culyer, A., Newhouse, J. (Eds.), Handbook of Health Economics, vol. 1B. Elsevier,
Amsterdam, pp. 1405–1487.
Ginsburg, P., 2010. Wide Variation in Hospital and Physician Payment Rates Evidence
of Provider Market Power. HSC Research Brief No. 16, Center for Studying Health
System Change.
Government Accountability Office (GAO), 2005. Report to the Honorable Paul Ryan:
House of Representatives: Federal Employees Health Benefits Program – Com-
petition and Other Factors Linked to Wide Variation in Health Care Prices,
Washington, DC.
Ho, K., 2009. Insurer–provider networks in the medical care market. American Eco-
nomic Review 99 (1), 393–430.
Ho, K., Lee, R.S., NBER Working Paper #19401 2013. Insurer Competition and Nego-
tiated Hospital Prices. National Bureau of Economic Research.
Lewis, M.S., Pflum, K.E., 2015. Diagnosing hospital system bargaining power in
managed care networks. American Economic Journal: Economic Policy 7 (1),
243–274.
Massachusetts Attorney General, 2010. Examination of Health Care Cost Trends
and Cost Drivers. Report for Annual Public Hearing, Office of Attorney General
Martha Coakley, Boston, MA.
McKellar, M.R., Naimer, S., Landrum, M.B., Gibson, T.B., Chandra, A., Chernew, M.,
2013. Insurer market structure and variation in commercial health care spend-
ing. Health Services Research (Epub ahead of print: 5 December 2013).
Medicare Payment Advisory Commission (MedPAC), 2009. Report to the Congress:
Medicare Payment Policy, Washington, DC.
Melnick, G., Shen, Y., Wu, V., 2011. The increased concentration of health plan
markets can benefit consumers through lower hospital prices. Health Affairs
(Millwood) 30 (9), 1728–1732.
Moriya, A., Vogt, W., Gaynor, M., 2010. Hospital prices and market structure in the
hospital and insurance industries. Health Economics, Policy and Law 5, 459–479.
Robinson, J., 2004. Consolidation and the transformation of competition in health
insurance. Health Affairs (Millwood) 23 (6), 11–24.
Scanlon, D.P., Swaminathan, S., Lee, W., Chernew, M., 2008. Does competition
improve health care quality? Health Services Research 43 (6), 1931–1951.
Schneider, J., Li, P., Klepser, D., Peterson, N., Brown, T., Scheffler, R., 2008. The effect of
physician and health plan market concentration on prices in commercial health
insurance markets. International Journal of Health Care Finance and Economics
8, 13–26.
Starc, A., 2014. Insurer pricing and consumer welfare: evidence from Medigap. RAND
Journal of Economics 45 (1), 198–220.
Town, R., Wholey, D., Feldman, R., Burns, L., NBER Working Paper #12244 2006.
The welfare consequences of hospital mergers. National Bureau of Economic
Research.
Trish, E., Herring, B., Working Paper 2014. Does Small Group Market Community
Rating Affect Firm Self-Insurance?
US Department of Justice and the Federal Trade Commission, 2010. Hori-
zontal Merger Guidelines, Available from: http://www.justice.gov/atr/public/
guidelines/hmg-2010.pdf
Vogt, W.B., Town, R., 2006. How has Hospital Consolidation Affected the Price and
Quality of Hospital Care? Research Synthesis Report No. 9, Robert Wood Johnson
Foundation.
White, C., Bond, A.M., Reschovsky, J.D., 2013. High and Varying Prices for Privately
Insured Patients Underscore Hospital Market Power. HSC Research Brief No. 27,
Center for Studying Health System Change.
Wholey, D., Feldman, R., Christianson, J., 1995. The effect of market structure on
HMO premiums. Journal of Health Economics 14, 81–105.
 Journal of Health Economics 42 (2015) 115–124

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Does women’s education affect breast cancer risk and survival?

Evidence from a population based social experiment in education
Mårten Palme a,∗ , Emilia Simeonova b
a
Department of Economics, Stockholm University, SE-106 91 Stockholm, Sweden
b
Johns Hopkins University and NBER, 100 International Drive, Baltimore, MD 21202, United States

a r t i c l e i n f o a b s t r a c t

Article history: Breast cancer is a notable exception to the well documented positive education gradient in health. A
Received 12 July 2013 number of studies have found that highly educated women are more likely to be diagnosed with the dis-
Received in revised form 30 October 2014 ease. Breast cancer is therefore often labeled as a “welfare disease”. However, it has not been established
Accepted 5 November 2014
whether the strong positive correlation holds up when education is exogenously determined. We esti-
Available online 22 November 2014
mate the causal effect of education on the probability of being diagnosed with breast cancer by exploiting
an education reform that extended compulsory schooling and was implemented as a social experiment.
JEL classification:
We find that the incidence of breast cancer increased for those exposed to the reform.
I12
I18 © 2014 Elsevier B.V. All rights reserved.

Keywords:
Education gradient in health
Schooling reform
Breast cancer

1. Introduction et al., 2008). This is in stark contrast with the well-documented

Worldwide, breast cancer is the most common cancer and the
leading cause of cancer death among women. In the year 2008
alone, 2.6 women were diagnosed with breast cancer every minute
across the globe. More than 52 women died of breast cancer every
hour in the same year (Ferlay et al., 2010). These aggregate num-
bers mask large differences in trends in breast cancer incidence and
mortality across the developed and developing economies. Histor-
ically, western societies have experienced a heavier burden of the
disease, however in the last couple of decades the incidence and
mortality from breast cancer has been on the rise in developing
countries (Althius et al., 2005). While it is plausible that this rise
is due to increased screening and better medical and vital records
keeping, some have argued that more affluent societies and west-
ernization also contribute to these recent trends (ibid).
Breast cancer in women is one of the rare health conditions
that exhibit a positive incidence gradient with socio-economic
status (SES), and in particular with attained education (see e.g.
Hemminki and Li, 2003, 2004; Lund and Jacobsen, 1991; Hussain
negative association between education and all-cause mortality
and with the positive effects of education on health-promoting
behaviors (Cutler and Lleras-Muney, 2006, 2011 for reviews of the
literature). Frequently diagnosed cancers of the female reproduc-
tive organs, such as cervical cancer, show the opposite, negative
association with education in correlational studies (Baquet et al.,
1990). Part of the observed positive correlation between educa-
tion and breast cancer could be due to more frequent screening
and more adequate response to risk factors among the better edu-
cated (Lange, 2011). Still, environmental and social factors could
also affect breast cancer risk and survival. A recent report by the
Interagency Breast Cancer and Environmental Research Coordinat-
ing Committee (IBCERCC) in the US forcefully argues that research
on the causes of increased breast cancer risk and consequently on
increased prevention is of first order importance in designing public
health strategies to contain the disease.1
A key question on the etiological background to the link between
education and the incidence of breast cancer is whether the rela-
tion is made up by life style factors, such as delayed childbearing,

∗ Corresponding author. Tel.: +46 8 16 33 07; fax: +46 8 15 94 82. 1

“Prioritizing Prevention” Summary of Recommendations of the Interagency
E-mail addresses: Marten.Palme@ne.su.se (M. Palme), Breast Cancer and Environmental Research Coordinating Committee (http://www.
Emilia.Simeonova@gmail.com (E. Simeonova). niehs.nih.gov/about/assets/docs/ibcercc full.pdf).

http://dx.doi.org/10.1016/j.jhealeco.2014.11.001
0167-6296/© 2014 Elsevier B.V. All rights reserved.
116 M. Palme, E. Simeonova / Journal of Health Economics 42 (2015) 115–124
that may be acquired along with prolonged education, or if it can
be attributed to factors and individual characteristics correlated
with both educational attainments and the probability to get breast
cancer. The most common research strategy used in epidemiolog-
ical studies is to add confounders that are known to be associated
with educational attainments and potentially etiologically related
to breast cancer, such as delayed childbearing in a regression frame-
work, and to investigate if the correlation remains (see e.g. Braaten
et al., 2004; Danø et al., 2004; Heck and Pamuk, 1997). There are
at least two problems with this strategy. First, there is an identi-
fication problem. Most confounders, such as fertility behavior, are
likely to be endogenous to educational attainment. This means that
it is still not clear if including them in the regression makes up for a
causal relation with education, or if they just proxy individual char-
acteristics correlated with educational attainments. Second, adding
independent variables in a regression would in most cases aggra-
vate the downward bias from measurement errors (see e.g. Greene,
2003).
An alternative strategy to analyze this research question is to
use exogenous variations in educational attainments created by
natural experiments. A number of influential studies have used
this research strategy to study the relationship between education
and measures of general health. Lleras-Muney (2005), Oreopoulos
(2006) and Clark and Royer (2012) use variation induced by
changes in compulsory schooling legislations in the US and the
UK as a source of exogenous variation in education. Spasojevic
(2010), Meghir et al. (2012) as well as Lager and Torssander (2012)
investigate the health consequences of the introduction of compre-
hensive school reform in Sweden. An interesting related question
is whether the health effects of education vary by gender2 and
diagnosis.
In this paper we investigate whether there is a causal effect of
education on the incidence and mortality from breast cancer in the
population of women born in Sweden between 1940 and 1957 who
survived until at least 1985. We make use of a compulsory schooling
reform that increased the number of compulsory years of education
from 7 or 8 depending on municipality to 9 years nationwide. We
also compile a unique nationally representative dataset from vari-
ous Swedish national data registries, including the Swedish Cancer
Registry.
The Swedish setting is particularly well suited to study how edu-
cation affects the incidence of a “welfare disease” such as breast
cancer in women for several reasons. First, Sweden is ethnically
and racially homogenous, especially in the cohorts under study.
This reduces potential omitted confounders that could correlate
both with the hereditary genetic make-up and the SES of some eth-
nic or racial subgroups. Second, health care is free at the point of
access and the Swedish government provides free universal health
insurance. Disparities arising from differential access to care due to
financial constraints are unlikely to play a role in the Swedish set-
ting. Breast cancer screening covers the entire female population
in the critical ages and is free of charge. The screening program
was adopted nation-wide in 1986 after the first results from the
Swedish mammography trials became available (Tabar et al., 1985).
The take up rate of this screening program after the first invita-
tion to screen is about 80 percent (see e.g. Hussain et al., 2008).
Third, the Swedish Cancer Registry is the oldest cancer registry and
one of the best in terms of data quality and accuracy in the world
today.
2 3
Clark and Royer (2012) as well as Meghir et al. (2012) investigate for differential
effects of education by gender and find inconclusive evidence. Gathman et al. (2012)
analyze a number of compulsory schooling reforms in Europe and find diverging
effects of education on mortality by gender.
The closest study we are aware of is by Glied and Lleras-
Muney (2008) who use the Surveillance, Epidemiology and End
Results Program (SEER) data to estimate the effects of technological
progress on cancer deaths by education, relying on US compulsory
schooling laws for exogenous variation in educational attainment.
They find that conditional on technological progress, extra educa-
tion reduces overall cancer mortality in men, but not in women.
Excluding cancers of the reproductive system, inclusive of breast
cancer, makes the estimated effects for men and women consis-
tent. The authors do not specifically test for the effects of education
on survival from reproductive system cancers in women, relying
on the findings in the medical literature we discuss above.
This study finds that attaining higher levels of education
increases the risk of being diagnosed with breast cancer in women,
confirming the results obtained from purely correlational studies.
However, we also find that this heightened probability of diagnosis
is later followed by an elevated probability of death from breast
cancer among better educated women. Further, we investigate the
potential role of fertility decisions, which has been pointed out as
the mechanism linking education and the incidence of breast can-
cer. We find no convincing evidence in favor of this hypothesis.
The curious association between education and the most common
cancer diagnosis in women appears to be affected by qualities,
behaviors, and risk factors acquired in the process of obtaining more
education, rather than pre-existing characteristics that predispose
some women to both get more education and be diagnosed with
the disease.
2. The comprehensive school reform
2.1. The Swedish school system before and after the reform
Sweden implemented a compulsory schooling reform as a social
experiment between 1949 and 1962. Prior to the implementa-
tion of the reform, pupils attended a common basic compulsory
school (folkskolan) until grade six. After the sixth grade pupils were
selected to continue either for one or, in mainly urban areas, two
years in the basic compulsory school, or to attend the three year
junior secondary school (realskolan). The selection of pupils into the
two different school tracks was based on their past academic per-
formance, measured by grades. The pre-reform compulsory school
was in most cases administered at the municipality level. The junior
secondary school was a prerequisite for the subsequent upper sec-
ondary school, which was itself required for higher education.
In 1948 a parliamentary committee proposed a school reform
that implemented a new nine-year compulsory comprehensive
school.3 The reform had three main elements:
1. An extension of the number of years of compulsory schooling to
9 years in the entire country.
2. Abolition of early selection and tracking based on academic per-
formance. Although pupils in the comprehensive schools were
able to choose between three tracks after the sixth grade – one
track including vocational training, a general track, and an aca-
demic level preparing for later upper secondary school – they
were kept in common schools and classes until the ninth grade.
3. Introduction of a national curriculum. The new curriculum
replaced the pre-existing curriculum which varied between
municipalities.
We offer a brief description of the main parts of the Swedish comprehensive
school reform. The school reform and its development are described in Meghir
and Palme (2003, 2005), and Holmlund (2007). For more detailed reference on the
reform, see Marklund (1981).
 M. Palme, E. Simeonova / Journal of Health Economics 42 (2015) 115–124
2.2. The social experiment
The social experiment with the new comprehensive nine-year
compulsory school started during an assessment period between
1949 and 1962, when the final curriculum was decided.4 The pro-
posed new comprehensive school system, as described above, was
introduced in municipalities or parts of city communities, which
in 1952 numbered 1055. The cohorts included in our empirical
analysis, born between 1940 and 1957, cover the entire period
of implementation of the comprehensive school. In 1962 it was
decided that the new comprehensive school would become the
standard education in Sweden. The last class that graduated from
the old schooling system did so in 1970.
The selection of municipalities into the new comprehensive
school was not based on random assignment. Still, the decision
to select the areas was based on an attempt to choose locations
that were representative for the entire country, both in terms
of demographics as well as geographically. At first the National
Board of Education contacted the municipalities, or sometimes they
themselves applied to participate. From this pool of applicants a
“representative” sample of municipalities was chosen. Municipal-
ities could elect to implement the comprehensive school starting
with first or fifth grade cohorts. Once the grade of implementation
was fixed, all individuals from the cohort immediately affected and
all subsequent cohorts went to comprehensive school. The older
cohorts continued in the per-reform school.
Meghir and Palme (2005) and Holmlund (2007) study the effect
of the comprehensive school reform on educational attainments.5
The Meghir and Palme (2005) estimates for their entire sample
are 0.252 additional years for males and 0.339 years for females;
for low SES persons the estimates are 0.3 extra years for males
and 0.512 for females.6 Holmlund has estimates in the range
0.21–0.61 additional years of schooling for men and 0.13–0.44 for
women.
3. Data
This is a population-level study. We match data from the
Swedish Cancer Register to Swedish population register data, the
1990 Swedish Education register, and the Cause of Death regis-
ter. The population register contains information on the parish of
birth for all individuals born in Sweden in the 1940–1957 cohorts.
We use this register to assign municipality of birth for all women
in the cohorts affected by the schooling reform. The municipal-
ity of birth is then used to assign the year in which the reform
was implemented in that locality, and the reform treatment status
to different cohorts of women who were born in the municipal-
ity. Note this means that all estimated effects are of the “intention
to treat” type, but we avoid potential bias coming from selective
migration. Holmlund (2007) offers a detailed exposition of the exact
matching algorithm used.7
The Cause of Death register contains information on the date
of death and the principal cause of death. The Census data provide
information on the date of birth and the number of children born
to the women in the 1940–1957 cohorts. We use this informa-
tion to assign age at first childbearing and the total completed
fertility per woman. The multi-generational register is used to
4
The official evaluation was mainly of administrative nature. Details on this eval-
uation are also described in Marklund (1981).
5
Holmlund (2007) does not have individual treatment status and imputes it from
municipality of residence in 1960.
6 9
Note that Meghir and Palme (2005) use the exact reform assignment from the
school registries for a random subset of the cohorts born in 1948 and 1953. Their
estimates are free of measurement error in the reform assignment variable.
7
We are grateful to Helena Holmlund for sharing her algorithm with us.
117
link women in the sample to their parents. We then use the
Education Register for the parents to determine the level of
education of each woman’s father. Fathers who had more than
the basic required (7 years) education are considered highly
educated.
All women who died of breast cancer as a primary cause of
death were found in the Cancer Register as having been previ-
ously diagnosed with the disease. We record all diagnoses and
deaths until 2006. The Swedish Cancer Register is the oldest Can-
cer Register in the world and contains detailed information on all
incidences of cancer diagnosis in Sweden. It is compiled from com-
pulsory cancer diagnosis registrations by physicians, cytologists
and pathologists and covers close to 100% of all cancer diagnoses
in Sweden (Swedish National Board of Health and Welfare, 2006).
Studies of the accuracy of the Cancer Register have shown that cases
of breast cancer are the most reliably reported cancer diagnosis
in the Register, with under-reporting rates of less than 1.1% of all
cases diagnosed within the reporting year (Barlow et al., 2009).
Importantly, the exact date of every diagnosis is recorded, and the
data can be linked to the population registers through a unique
person ID.
In the empirical analysis we use the population of all women
born in Sweden between 1940 and 1957 and surviving until at least
1985. We exclude 414,214 women with missing parental educa-
tion background and use the remaining sample of 562,814 women.
Of those, 19,736 women were diagnosed with breast cancer after
1984. Of those who were diagnosed, 2370 women died, and breast
cancer was noted as the cause of death on their death certificate.
Another 401 of the women diagnosed with breast cancer after 1984
died from a different main diagnosis.8
Table 1 summarizes the main explanatory and control variables
used in the analysis. The mortality data start in 1985 and include
the exact date of death and the main cause of death as recorded in
the death certificate. We restrict the time of first diagnosis to be
after 1984 in order to avoid selection of women who were diag-
nosed previously and survived until the period after 1984. The
women in our sample were aged between 28 and 45 in 1985 and
(those surviving) between 49 and 66 in 2006. As a percent of total
female mortality, breast cancer mortality peaks between ages 40
and 60 at around 15% of total deaths in the age group. This implies
that we are capturing the interval in women’s lives during which
they are most likely to be affected by breast cancer (as opposed
to another lethal disease). Aggregate mortality in Sweden is very
low at ages below 45 at 6 per 1000 (from data), and breast can-
cer mortality is even lower at 1 per 1000 (from data). A back of
the envelope calculation suggests that we are potentially missing
at most 100 deaths from breast cancer that may have occurred
in our study population before 1985.9 This is a very small part
of the total number of breast cancer deaths in the sample – less
than 5%.
Several differences in the raw means between women of high
and low SES family backgrounds are worth discussing. Unsur-
prisingly, on average women of higher SES obtained more years
of education. They are less likely to have had any children and
the average age at first childbearing in this group is about two
years higher. High SES women are also more likely to have been
8
The distribution of the main causes of death among those women is: 61 women
died from ovarian cancer; 45 from lung cancer; 22 from AMIs; 15 from pancreatic
cancer; 12 died from colon cancer; 9 from melanoma; 15 from unknown causes.
The remaining 222 deaths are distributed across more than one hundred different
causes.
Assuming mortality at missing ages is equivalent to mortality at those ages
among observed cohorts; assuming also that cohort sizes at different ages are sim-
ilar over time, which gives an upper bound estimate since demographic trends led
to steady cohort size increase between 1940 and 1957.
118 M. Palme, E. Simeonova / Journal of Health Economics 42 (2015) 115–124

Table 1
Main explanatory and outcome variables in interest. Standard deviations are reported in square brackets under the mean. P-values of tests of differences in means between
high and low SES background women are also reported.

Father’s education Low FE High FE P-value diff

Obs Mean Obs Mean

Variable
Years of education 360,240 11.155 180,612 12.823 0
[2.805] [3.063]
No children (nulliparous) 372,894 0.11 187,702 0.134 0
[0.315] [0.34]
Age at first childbearing 331,164 26.5 162,587 28.1 0
[5.4] [5.5]
Age at diagnosis 12,723 50.196 6654 49.513 0.088
[6.581] [6.511]
Death year – year of diagnosis 1538 4.83 724 5.3 0
[4.55] [4.63]
Diagnosed with breast cancer 372,894 0.035 187,702 0.036 0.012
[0.183] [0.186]
Died from breast cancer 372,894 0.004 187,702 0.004 0.16
[0.065] [0.063]

diagnosed with breast cancer, to have received the diagnosis at an 4. The relation between educational attainment and breast
earlier age and, conditional on dying from breast cancer, to have cancer incidence and mortality
lived longer between their initial diagnosis and the time of death.
There is no significant difference in the probability of death from We start the analysis by documenting correlations between
breast cancer by SES background. These facts suggest that either the years of attained education and socio-economic background
(1) higher SES women are more likely to have been diagnosed ear- and the probability of diagnosis and death from breast cancer in
lier or that (2) higher SES women received better treatment, or Sweden. Table 2 presents the estimates. We use all available obser-
both. The differences in age at diagnosis appear in favor of the first vations to maximize power. Coefficients and standard errors are
hypothesis, but we cannot draw any firm conclusions based on this multiplied by 1000 for better presentation. Women with an extra
evidence. year of education are 3 percent (evaluated at the mean incidence of
Here it is important to consider the importance of breast can- breast cancer diagnosis in the population) more likely to have been
cer screening for early diagnosis and treatment. The large clinical diagnosed with breast cancer than their less educated peers. The
trials that produced evidence on the beneficial effects of mam- correlation with high socio-economic family background is larger
mography, were done in Sweden in the 1970s and 1980s (see – even if we control for years of attained education, women who
Tabar et al., 1985). Thus, policy makers in Sweden were quite were born in better-off families are 7 percent more likely to receive
aware of the importance of breast cancer screening at the time a breast cancer diagnosis.
our study period begins. After the first results of the random- The effect of education on deaths from breast cancer is not as
ized trials came out, the National Board of Health and Welfare clear. None of the education coefficients attain statistical signif-
issued guidelines in 1986 recommending that the county councils icance at the 10% level even though the coefficient on years of
invite women ages 40–54 years to screening every 18 months and education implies a negative correlation both with and without
women ages 55–74 years every second year. Thus, national ser-
vice screening with mammography was initiated in 1986. Local Table 2
health administrations are in charge of running the screening Correlations between years of educational attainment and diagnosis/death from
programs. All women of eligible ages receive a letter giving a spe- breast cancer in women.
cific date and time for a mammography examination. Failure to Panel A (1) (2) (3)
attend the scheduled examination or re-schedule the appointment
Diagnosis
results in a second invitation up to six consecutive invitations. A
Years of schooling 0.96* 0.89*
regional case study from Uppsala reports that of the 46,041 eligible coef * 1000 (0.09) (0.08)
women only 5.6% never attended after six attempted appoint- High SES 2.99* 1.92*
ments. Non-attenders tend to be older (over the age of 60), foreign coef * 1000 (0.54) (0.56)
born and single. Note that all foreign-born women residing in Observations 541,135 560,596 541,135
Mean incidence per 1000 34.1 34.1 34.1
Sweden are excluded from our sample by construction. Interest- R-squared 0.006 0.007 0.010
ingly, the relationship between education and the probability of Empirical model Linear Prob Linear Prob Linear Prob
non-attendance is u-shaped, with women finishing high school,
some college, and college more likely to attend than those with Panel B (1) (2) (3)
professional education or high school drop outs (Lagerlund et al., Death from breast cancer
2002). Years of schooling −0.01 −0.02
Breast cancer is a common killer in our sample. In this rel- coef * 1000 (0.01) (0.01)
High SES −0.09 0.12
atively young population, 15% of all deaths are due to breast
(0.18) (0.09)
cancer. Cardio-vascular diseases account for an extra 13.5% of Mean deaths per 1000 4.1 4.1 4.1
total mortality. Deaths from other cancers are responsible for Observations 541,135 560,596 541,135
another 33%. In total, cardio-vascular and cancer-related mortality Empirical model Linear Prob Linear Prob Linear Prob
account for close to two-thirds of all female deaths in the sample Note: Robust standard errors in parentheses; *Significant at 1%; SE clustered on the
cohorts. municipality of birth level.
 M. Palme, E. Simeonova / Journal of Health Economics 42 (2015) 115–124
controls for parental SES background. Coupled with the evidence
on the higher incidence of diagnoses among the more educated
women, this suggests that conditional on being diagnosed with
breast cancer, more educated women are more likely to survive.
This is consistent both with evidence that educated people are more
adept at using new medical technologies (Glied and Lleras-Muney,
2008; Lichtenberg and Lleras-Muney, 2005) and with earlier diag-
nosis and earlier treatment in higher SES background women. The
table of means shows supportive evidence for the latter expla-
nation. Even though the mortality point estimate suggests that
education has a positive effect on survival, the precision of the
estimates is not high enough to make any strong conclusions.
The corresponding Cox proportional hazard estimates are: in
the full sample one year of schooling reduces the probability of
death from breast cancer by a statistically insignificant 1.6% relative
to the mean (SE 0.0197), which is a larger estimate than the LP
estimate evaluated at the mean (0.03%); the Cox estimate of high
SES in model 2 is a statistically insignificant decrease of 2.4% relative
to the mean, not too far from the LP estimate of 2.2% relative to the
mean.
5. Empirical specification
We use two main types of outcomes in the empirical analy-
sis. When we consider breast cancer mortality, we use the binary
mortality outcome and the time to death as the outcome vari-
ables. When we study the incidence of breast cancer, we use a
binary outcome variable equal to one if the woman was ever diag-
nosed with breast cancer after 1984 and zero otherwise. We use
the same identification strategy for the effect of the reform for
both types of outcomes. If the reform would have been randomly
distributed among Sweden’s 1000 or so municipalities we could
have simply compared the outcomes in the treated and non-treated
municipalities conditional on year of birth. However, as has been
discussed in previous studies (see e.g. Meghir et al., 2012), this was
not the case. Therefore, we will control for both birth cohort and
municipality of birth. We start with the following latent variable
specification:
∗
yi,m,t = ˛ + ˇ1 Ri,m,t + 1 Ti + 2 Mi + εi.m.t ,
where i, m and t are sub-indices for individual, municipality and
birth cohort, respectively; y* is a latent variable for health status; T
is a vector of dummy variables for year of birth; M is a corresponding
vector of dummy variables for municipality of birth; finally, ε is an
individual random disturbance.
The key identifying assumption is that the distribution f(·) of
ε does not depend on the assignment to reform treatment, con-
ditional on cohort and municipality. In practice we impose the
stronger assumption that the distribution of ε is independent of
all right hand side variables. It is important to note that the reform
assignment in this analysis depends on the municipality of birth,
rather than the municipality of schooling. On the one hand, this
means that the estimates are of the “intention-to-treat” type. On
the other hand we avoid selection issues coming from differential
(and potentially endogenous) mobility.10
For the binomial outcome breast cancer diagnosis, we use
linear probability models. The reason for using a linear proba-
bility model, rather than e.g. logit and probit, which restrict the
probabilities in the [0, 1] interval and relax the linearity assump-
tion, is computational convenience, since all models include about
10
Meghir and Palme (2005) show, however, that 90.1 percent have the same
reform assignment based on predictions from their municipality of birth as their
municipality of schooling; 5.3 percent moved from reform to non-reform munici-
palities; 4.6 moved in the other direction.
119
1000 municipality indicator variables in addition to the 17 birth
cohort dummies. For relatively small treatment effects, when both
approaches have been used in a similar context, the results are
almost identical.11
We also use linear probability models as one of two methods
of estimating the probability of death from breast cancer. The lin-
ear probability model is handy because it can efficiently estimate a
large number of dummy coefficients in specifications where we
also include municipality-specific time trends. We complement
the linear probability estimates with estimates from Cox semi-
parametric proportional hazard models. For the time to death from
breast cancer outcomes we use Cox proportional hazard models of
this type:
I1,i,m,t (r|Ri,m,t , Ti , Mi ) = I0 (r) exp{˛ + ˇ1 Ri,m,t + 1 Ti + 2 Mi }, (2)
where r is exposure time and I0 (r) is the baseline hazard. This model
is semi-parametric in the sense that no functional form assumption
is imposed on the base line hazard. Importantly, when we consider
the hazard of death from breast cancer, we consider only deaths
from breast cancer as terminal event. Thus, all women who died
from causes other than breast cancer are considered still living
at the end of the observation window. Prior research has found
that the compulsory schooling reform did not significantly affect
life expectancy for (high and low SES) Swedish women (Meghir
et al., 2012). Moreover, the age at first diagnosis in this sample
is fairly young. These two facts suggest that a competing risks
phenomenon is an unlikely explanation for our estimates. Never-
theless, as Honoré and Lleras-Muney (2006) show that decreasing
cardio-vascular disease mortality in the US contributed to a steady
(non-declining) cancer mortality rate between the 1970s and 2000s
we construct Peterson bounds on our estimates taking into account
the association between cardio-vascular and cancer mortality risks.
The assumption here would be that obtaining extra education,
while reducing the likelihood of cardio-vascular mortality, indi-
rectly increases the likelihood of breast cancer mortality. To control
for unobserved differential trends that might affect municipalities
differently depending on the timing of the education reform, we
(1) include linear trends by year of reform implementation. All munic-
ipalities that implemented the reform in the same year are assigned
the same linear trend. The empirical results section reports the
results from these preferred specifications.
Table 3 demonstrates the effects of being exposed to the school-
ing reform on the number of years of attained education for Swedish
women of affected cohorts. We first show the effects on the entire
sample and then split the sample according to the education level
of the father. We expect that the education reform affected chil-
dren from low SES families more as they were more likely to drop
out of school earlier. The results confirm that women from low SES
backgrounds increased their education by more than those from
high SES backgrounds. The reform resulted in an average increase
of 1.8 months of schooling for girls coming from relatively disad-
vantaged backgrounds. The corresponding estimate for the high
SES group is about one third of the size and does not attain sta-
tistical significance at the 10% level. We present estimates from
models including linear time trends grouped by year of imple-
mentation and from specifications including municipality-specific
linear time trends. The estimated coefficients are very similar,
which is reassuring that unobserved municipality-specific changes
coincidental with reform implementation are unlikely to bias our
results.
11
See for example Meghir et al. (2011).
120 M. Palme, E. Simeonova / Journal of Health Economics 42 (2015) 115–124

Table 3
The effect of education reform on women’s educational attainment in years of education.

Father’s education All Low FE High FE

(1) (2) (3) (4) (5) (6)

Reform 0.119* 0.106* 0.149* 0.14* 0.055 0.043

(0.029) (0.023) (0.022) (0.021) (0.035) (0.033)
Mean years of education 11.7 11.7 11.1 11.1 12.8 12.8
Linear trends by year of reform implementation Yes Yes Yes
Municipality trends Yes Yes Yes
Observations 540,852 540,852 360,240 360,240 180,612 180,612
R-squared 0.041 0.045 0.041 0.045 0.024 0.031

Note: Robust standard errors in parentheses; SE clustered on the municipality of birth level. + Significant at 10%; **significant at 5%; *significant at 1%.

Table 4
Educational reform and the risk of diagnosis and death from breast cancer.

Father’s education All Low High

Diagnosis
Reform coef * 1000 1.5** 1.71** 1.6 1.92+ 1 1.1
(0.78) (0.775) (1.1) (1.025) (1.5) (1.47)
Mean dep var * 1000 34.1 33.7 35
Linear trends by year of reform implementation Yes Yes Yes
Municipality trend Yes Yes Yes
Observations 560,596 560,596 372,894 372,894 187,702 187,702
R-squared 0.006 0.0075 0.006 0.009 0.010 0.015

Death from breast cancer

Reform coef * 1000 0.63** 0.66** 0.68+ 0.74+ 0.65 0.6
(0.3) (0.33) (0.4) (0.44) (0.9) (0.87)
Mean dep var * 1000 4.1 4.1 4.2 4.2 3.9 3.9
Linear trends by year of reform implementation Yes Yes Yes
Municipality trend Yes Yes Yes
Observations 560,596 560,596 372,894 372,894 187,702 187,702
R-squared 0.002 0.004 0.003 0.0061 0.006 0.010

Note: Municipality fixed effects included in all specifications; birth cohort dummies included in all specifications; robust standard errors in parentheses; standard errors
clustered on the municipality of birth level; + significant at 10%; **significant at 5%; *significant at 1%.

6. Results
6.1. Main findings
We next turn to the effects of the compulsory education reform
on breast cancer incidence and death. Since we know from previ-
ous research (see e.g. Meghir and Palme, 2005) that the reform had
very different effects on later-life economic wellbeing depending
on parental SES, we run separate regressions by women’s fam-
ily SES background. It is important to note that since the reform
was not limited to simply increasing the number of compulsory
years of education, but had additional elements, the results that
follow are not directly comparable with the education correla-
tions presented in Table 2. The results on how reform treatment
affected the probability of diagnosed breast cancer are shown
in the top panel of Table 4 and on mortality from the disease
in the bottom panel. We present estimates with year of imple-
mentation specific linear trends for easy comparison with the
semiparametric Cox estimates, as well as results from linear prob-
ability models including municipality-specific linear trends. We
multiply all coefficients and standard errors by 1000 for ease of
presentation.
There is a significantly positive effect of reform assignment
on the probability of being diagnosed with breast cancer in the
full sample. Although the precision in the estimate in the low
SES subsample is somewhat inferior, it is obvious that the effect
is attributable to the group originating from low SES families,
who experienced the largest effect of the education reform. The
point estimate of the magnitude of the effect suggests a 1.5
percentage point elevated risk, which is somewhat more than
the correlation estimate corresponding to one year extra year of
education although the precision is not sufficient to make any def-
inite conclusion.12
The linear probability estimation results in Table 4 also show
that the reform causes a significant increased risk of mortality with
breast cancer as a primary cause of death.13 That is, the expected
improvement in the effect of better response to cancer treatment
from more education was not sufficient to offset the increased risk
of being diagnosed with breast cancer.
In addition to the linear probability models we obtained Cox
PH mortality estimates stratified at the municipality level. The
Cox semi-parametric model imposes fewer restrictions on the
estimates, however it suffers from severe incidental parameters
problems with a large number of dummy variables, such as would
be included in a specification including municipality specific lin-
ear trends. That is why we ran the Cox estimations with linear
trends by year of implementation only, so these estimates are com-
parable to the linear probability coefficients reported in columns
(1), (3), and (5). The hazard is stratified by municipality of birth,
allowing for potentially different underlying breast cancer mortal-
ity hazards by municipality. The Cox estimates are as follows: full
12
If we consider only the reform’s effect on schooling attainment, we would mul-
tiply the reform estimates from Table 4 by 1/(estimated change in years of education
from Table 2), resulting in much larger estimates of the effect of an additional year of
education than what is obtained in the correlations reported in Table 3. We empha-
size, however, that using the reform as an IV for years of attained education is
most likely flawed, as the reform contained additional elements that challenge the
exclusion restrictions.
13
In the analyses we exclude all women who have received a diagnosis of breast
cancer pre-1985 to avoid selection bias. This is because our mortality data start in
1985 and survival following a breast cancer diagnosis could be related to the reform.
 M. Palme, E. Simeonova / Journal of Health Economics 42 (2015) 115–124
sample coefficient 0.18+ (SE 0.1), which is very similar to the LP
estimate evaluated at the mean (15% increase); low father’s edu-
cation sample estimate 0.2 (SE 0.15) – again very similar to the LP
coefficient estimated at the mean – a 16% increase.
6.2. Competing risks
A potentially important concern in analyzing mortality by dif-
ferent causes has been raised by Honoré and Lleras-Muney (2006).
Technological progress in medicine or any other factor that affects
the treatment or detection of certain diseases would affect the
probability of death from related diagnoses but also the prob-
ability of death from other conditions, which pose “competing
risks”. In essence, failure to die from one condition at a given age
increases the probability of death from another condition. Honoré
and Lleras-Muney (2006) show in particular that cardio-vascular
(CVD) and cancer deaths in the US are related in this manner.
Improvements in the treatment of CVD led to decreased mortality
from CVD but also to increased mortality from cancer compared to
the counterfactual. This is important in our setting because educa-
tion may have affected the early detection and proper treatment
of CVD, leading to a reduction in the probability of death from
CVD. Through the competing risks channel, this reduction may have
increased the probability of death from breast cancer. To exam-
ine this hypothesis, we first estimate the effects of being exposed
to the reform on CVD mortality and compute bounds for our
estimates.
The probability of death from CVD is reduced by reform treat-
ment by 0.53% (Cox estimate 0.99472, CI 0.8353–1.1845) in the
full sample. In the subsample of low SES background women, the
reform treatment leads to a 2.7% decrease in CVD mortality (Cox
estimate 0.97304, CI 0.80374–1.178). Assuming that everyone who
did not die from CVD died from breast cancer, we compute a lower
bound on our breast cancer mortality estimates. The education
reform increases the probability of death from breast cancer or CVD
by 9.8% (hazard ratio 1.098, CI from 0.997 to 1.209). Thus even if
the entire reduction in CVD mortality is translated into breast can-
cer mortality, we still find a positive effect of the reform on the
(combined) mortality rate, even though it is about half the size
of the effect we obtain when we assume the risks are unrelated
(18%).14
6.3. Parallel trends assumption
Our difference-in-differences analysis relies on the assumption
of parallel trends in the incidence of diagnosed breast cancers
before and after the cohort affected by the reform in each munici-
pality. We implement two different tests of this assumption. First,
Fig. 1 plots the conditional marginal effects of exposure in the
6 cohorts pre-implementation to 6 cohorts post-implementation
14
A separate issue emerges if we consider testing for the effect of the reform on
deaths from breast cancer as one of a series of multiple mortality tests we could
perform, including the reform effect on death from CVD and death from other
causes. We performed an adjustment procedure to calculate the q-value, which is
the P-value of the test adjusted for the false discovery rate. This methodology was
developed by Storey and co-authors and software was created by Dabney and Storey
(Storey, 2002; Storey and Tibshirani, 2003). We picked a 0 of 1 and an FDR threshold
of 0.05. The q-value on the reform coefficient in the first linear probability regression
of breast cancer mortality is 0.096 (the P-value is 0.032); the q-value on the reform
coefficient in a linear probability regression with binary outcome “death from any
other condition” is 0.141 (P-value 0.95); the q-value on the reform coefficient in a
linear probability regression with binary outcome “death from CVD” 0.79 (P-value
0.79). While a P-value threshold of 0.1 implies that 1 in every 10 tests will be a false
positive, a q-value threshold of 0.1 implies that one in every 10 positive tests will
be a false positive. Even after a conservative adjustment for multiple hypotheses
testing, we obtain a reform effect that is still significant at the 10% level.
121
sample. The reference cohort is the one born 2 years before the first
treated cohort. The regressions control for municipality and year of
birth fixed effects, as well as municipality group by year of imple-
mentation linear trends. As the figures demonstrate the conditional
probability of diagnosis and death is not significantly different from
zero in cohorts born pre-implementation. There is however a sharp
increase in the probability that starts with the cohort right before
the first fully treated cohort and levels off at a new and increased
level with the second fully treated cohort (1 year after year zero of
the implementation in the figures below).
Second, we performed a number of placebo tests in which we
pretend that the reform was implemented earlier or later than
the actual implementation year. The placebo treatment groups are
defined by falsely assigning treatment to women born 6, 4 and 2
years before the first fully treated cohort, as well as 2, 4 and 6 years
after the first cohort. In the first arrangement women who were not
treated receive false treatment status. In the latter arrangement we
pretend that women who were (actually) treated and were born 2,
4, and 6 years from the first treated cohort were not treated. Thus,
in this set-up treated women receive false untreated status. We
present all these tests together in Fig. 2.
Every estimate is obtained from a separate regression includ-
ing cohort and municipality fixed effects, as well as year of
implementation linear trends. The regressions assigning treat-
ment to untreated cohorts include only women from untreated
cohorts. Similarly, the regressions assigning non-treatment to
treated cohorts include only treated women. As the figures
demonstrate, the largest in absolute value and only statistically
significant effects are estimated when we assign the correct treat-
ment values. Further, there are no particular discernible patterns,
suggesting that there is nothing that systematically biases our
estimates.
6.4. Changes in fertility behavior as a possible mechanism behind
the results
The causal estimates confirm the positive correlations between
education and the probability of breast cancer diagnosis. Medical
studies have pointed to several channels that might contribute to
these findings (see Nechuta et al., 2010 for a recent review). For two
of these – the inverse relation between educational attainments
and completed fertility as well as the positive relation between
education and age at first birth – we have information in our data
set allowing us to analyze how these two outcomes were affected
by the schooling reform.
As a background to this analysis, Table 5 shows associations
between attained education and women’s fertility behavior in
Sweden using the same population we analyzed in the mortality
regressions. Column (1) reports the correlation between year of
schooling and the probability of having no children; column (2)
displays the relation between year of schooling and number of chil-
dren; finally, column (3) The shows the correlation between years
of schooling and age at first child.
As can be seen in Table 5, there is a statistically significant rela-
tion between years of schooling and each of the three outcomes
under study. The point estimates suggest that one additional year
of schooling is associated with a 0.003 increase in the probability of
having no children; 0.013, or an about 0.8 percent, fewer children;
and, finally, almost half a year older age at first birth.
In Table 6 we turn to analyzing the effect of schooling reform
on the same set of outcomes as those analyzed in Table 5. None
of the point estimates attain statistical significance. Comparing
the estimates for the effect of the reform with the correlations
shown in Table 5, it is evident that the precision in the reform
effect estimates for the probability of having no children as well
as the total number of children is too low to enable us to reject
122 M. Palme, E. Simeonova / Journal of Health Economics 42 (2015) 115–124

Fig. 1. Probability of breast cancer diagnosis and death from breast cancer among cohorts of women born close to the first cohort affected by the reform. Note: Conditional
marginal effects plotted in solid line, 95% confidence intervals in dashed lines. The omitted category is women born 2 years before the first cohort that was affected by the
reform. Cohort and municipality fixed effects included in the regressions, as well as municipality group by year of implementation-specific linear trends.

Educaon Death from breast cancer

0,2 2

0,15 1,5
1
0,1
0,5
0,05
0
-6 -4 -2 0 2 4 6
0 -0,5
-6 -4 -2 0 2 4 6
-0,05 -1
-1,5
-0,1
-2

Diagnosis
0,006

0,004

0,002

0
-6 -4 -2 0 2 4 6
-0,002

-0,004

-0,006

Fig. 2. Placebo tests assigning treatment status to untreated or untreated status to treated cohorts.

the hypothesis that the effects are the same as for one additional
year of schooling. However, for the age at first child outcome,
the point estimate is very different and the precision sufficient
to allow us to reject that the effect is as large as the almost 0.5
years as suggested by the result in Table 5. The upper confidence
limit for a 95 percent confidence interval for the reform effect
is as small as 0.025 for age at first birth, and so we can con-
clude that it is unlikely that the mechanism behind our result
of the reform effect on cancer diagnosis incidence is through
delayed childbearing among those who had children. For the
other two outcomes, the precision is too low for any definite
conclusions.

Table 5
Education and women’s fertility behavior.

Father’s education All All All

Outcome No children Total fertility (number of children) Age at first childbearing
(1) (2) (3)

Years of schooling 0.0034* −0.0134* 0.4608*

(0.0002) (0.0010) (0.0038)
Mean outcome variable 0.11 1.7 27
Empirical model Linear Prob OLS OLS
Observations 541,135 541,135 478,946
R-squared 0.0057 0.0057 0.0868

Note: + Significant at 10%; **significant at 5%; *significant at 1%; robust standard errors in parentheses clustered at the municipality of birth; linear trends by year of reform
implementation included in all specifications; birth cohort dummies included in all specifications.
 M. Palme, E. Simeonova / Journal of Health Economics 42 (2015) 115–124 123

Table 6 were sufficient for excluding it as a major mechanism behind our

Education reform and women’s fertility behavior.
results on breast cancer incidence and mortality.
Father’s education All Low education High education Epidemiological studies of the incidence of breast cancer have
(1) (2) (3) discussed the possibility that breast feeding, breast feeding dura-
Probability of not bearing a child tion and the duration of oral contraceptive use may affect the
Reform 0.0018 0.0011 0.0033 probability of breast cancer. Since we have not been able to find
(0.0018) (0.0021) (0.0028) large enough dataset including these outcomes, we have not been
Mean outcome variable 0.11 0.11 0.134
able to explore their potentials as a possible mechanism behind our
Empirical model OLS OLS OLS
Observations 560,596 372,894 187,702 results and have to leave this for further research.
R-squared 0.0050 0.0057 0.0082

Total fertility (number of children)

Reform 0.0028 0.0034 −0.0018 References
(0.0061) (0.0073) (0.0107)
Observations 560,596 372,894 187,702
Althius, M., et al., 2005. Global trends in breast cancer incidence and mortality
R-squared 0.0045 0.0055 0.0077
1973–1997. International Journal of Epidemiology 34 (2), 405–412.
Age at first childbearing Baquet, C., Horm, J., Gibbs, T., Greenwald, P., 1990. Socio-economic factors and cancer
Reform −0.0445 −0.0464 −0.0426 incidence among Blacks and Whites. Journal of the National Cancer Institute 83
(0.0350) (0.0364) (0.0572) (8), 551–557.
Barlow, L., et al., 2009. The completeness of the Swedish Cancer Register – a sample
Mean outcome variable 27 26.5 28.1
survey for year 1998. Acta Oncologica 48 (1), 27–33.
Empirical model OLS OLS OLS
Braaten, T., Weiderpass, E., Kumle, M., Adami, H.-O., Lund, E., 2004. Education and
Observations 493,751 331,164 162,587 risk of breast cancer in the Norwegian-Swedish women’s lifestyle and health
R-squared 0.0269 0.0225 0.0319 cohort study. International Journal of Cancer 110 (4), 579–583.
Note: + Significant at 10%; **significant at 5%; *significant at 1%; robust standard Center for Epidemiology, 2006. Cancer Incidence in Sweden 2004. Center for Epi-
errors in parentheses clustered at the municipality of birth; linear trends by year demiology, National Board of Health and Welfare, Stockholm.
Clark, D., Royer, H., 2012. The effect of education on adult mortality and health:
of reform implementation included in all specifications; birth cohort dummies
evidence from Britain. American Economic Review 103 (6), 2087–2120.
included in all specifications.
Cutler, D., Lleras-Muney, A., 2006. Education and health: evaluating theories and
evidence. In: House, J.S., Schoeni, R.F., Kaplan, G.A., Pollack, H. (Eds.), The Health
Effects of Social and Economic Policy. Russell Sage Foundation, New York.
Cutler, D., Lleras-Muney, A., 2011. Education and Health (Unpublished manuscript).
Danø, H., Hansen, D.K., Jensen, P., Pedersen, J.H., Jacobsen, R., Ewertz, M., Lynge,
E., 2004. Fertility pattern does not explain social gradient in breast cancer in
7. Concluding remarks
Denmark. International Journal of Cancer 111 (3), 451–456.
Ferlay, J., Shin, H.R., Bray, F., Forman, D., Mathers, C., Parkin, D.M., 2010. Estimates
Numerous studies have shown that higher educational attain- of worldwide burden of cancer in 2008: GLOBOCAN 2008. International Journal
ment is conducive to better health in the affected cohorts and their of Cancer 127 (12), 2893–2917.
Gathman, C., Jürges, H., Reinhold, S., 2012. Compulsory Schooling Reforms, Education
offspring (see review by Cutler and Lleras-Muney, 2011). Breast and Mortality in Twentieth Century Europe. University of Wuppertal (mimeo).
cancer is an exception to this rule in the sense that the incidence Glied, S., Lleras-Muney, A., 2008. Health Inequality, Education and Medical Innova-
of diagnosed cases increases with education and it has therefore tion. Demography 45 (3), 741–761.
Greene, W., 2003. Econometric Analysis. Prentice Hall, New Jersey.
been labeled a “welfare disease”. We show that this relation holds Heck, K., Pamuk, E., 1997. Explaining the relation between education and post-
also as a response to an exogenous policy change that induced an menopausal breast cancer. American Journal of Epidemiology 145 (4), 366–372.
increase in compulsory schooling. This result suggests that the rela- Hemminki, K., Li, X.J., 2003. Level of education and the risk of cancer in Sweden.
Cancer Epidemiology Biomarkers Preview 12, 796–802.
tion between women’s educational attainments and breast cancer Hemminki, K., Li, X.J., 2004. University and medical education and the risk of cancer
is likely due to some characteristic or risk factor that is acquired in Sweden. European Journal of Cancer Prevention 13, 199–205.
as additional education is obtained, rather than some innate qual- Holmlund, H., 2007. A Researcher’s Guide to the Swedish Compulsory School Reform.
Working paper 9/2007, Swedish Institute for Social Research, Stockholm Uni-
ity that is correlated with educational attainments. Many social
versity.
and health behaviors fit these categories, such as the use of hor- Honoré, Bo E., Lleras-Muney, A., 2006. Bounds in competing risks models and the
monal therapies and oral contraceptives, which have been linked war on cancer. Econometrica 74 (6), 1675–1698.
Hussain, S.K., Altieri, A., Sundquist, J., Hemminki, K., 2008. Influence of education
to increased probability of breast cancer.
level on breast cancer risk and survival in Sweden between 1990 and 2004.
The reform may have made women more willing to participate International Journal of Cancer 122 (1), 165–169.
in screening programs. Meghir et al. (2013) shows that cognitive Lager, A., Torssander, J., 2012. Causal effect of education on mortality in a quasi-
skills were improved as a result of the reform, which, in turn, experiment on 1.2 million Swedes. Proceedings of the National Academy of
Sciences 109 (22), 8461–8466.
may make people more adequately aware of different risk fac- Lagerlund, M., et al., 2002. Sociodemographic predictors of non-attendance at invi-
tors (see e.g. Cutler and Lleras-Muney, 2006). Since participation in tational mammography screening – a population-based register study. Cancer
screening programs is not included in our data, we are not able to Causes and Control 13 (1), 73–82.
Lange, F., 2011. Education and allocative efficiency. Evidence from cancer screening.
estimate this effect separately. However, given that about 80 per- Journal of Health Economics 30, 43–54.
cent of Swedish females participate in the nationwide screening Lichtenberg, F., Lleras-Muney, A., 2005. The effect of education on medical technol-
program it is not likely that improved participation in the pro- ogy adoption: are the more educated more likely to use new drugs? Special issue
of the Annales d’Economie et Statistique in memory of Zvi Griliches, No 79/80.
gram makes up the entire effect. Further, superior screening among Lleras-Muney, A., 2005. The relationship between education and adult mortality in
higher SES women by itself would not explain the elevated risks of the U.S. Review of Economic Studies 72 (1), 189–221.
death that we document. The findings in this study imply that the Lund, E., Jacobsen, B.K., 1991. Education and breast-cancer mortality – experience
from a Large Norwegian Cohort Study. Cancer Causes and Control 2, 235–238.
quality and availability of breast cancer screening and preventive
Marklund, S., 1981. Skolsverige 1950–1975: Försöksverksamheten. Liber Utbild-
health care must keep pace with improving educational opportu- ningsförlaget, Stockholm.
nities for women world-wide. Meghir, C., Palme, M., 2003. Ability Parental Background and Educational Policy:
Empirical Evidence from a Social Experiment. Institute for Fiscal Studies, IFS
We find no convincing evidence that fertility behaviors, often
Working Papers: W03/05.
cited as a potential mechanism behind the higher incidence of Meghir, C., Palme, M., 2005. Educational reform, ability, and family background.
breast cancer in educated women, were significantly affected by American Economic Review 95 (1), 414–424.
the reform and we could therefore not conclude that it is driving Meghir, C., Palme, M., Schnabel, M., 2011. The Effect of Education Policy on Crime:
An Intergenerational Perspective. NBER Working Paper 18145.
the result of elevated risk of breast cancer caused by the educa- Meghir, C., Palme, M., Simeonova, E., 2012. Education, Health and Mortality: Evi-
tion reform. For delayed childbearing the precision in our estimates dence from a Social Experiment. NBER Working Paper 17932.
124 M. Palme, E. Simeonova / Journal of Health Economics 42 (2015) 115–124

Meghir, C., Palme, M., Simeonova, E., 2013. Education, Cognition and Health: Evi-
dence from a Social Experiment. NBER Working Paper 19002.
Nechuta, S., Paneth, N., Velie, E., 2010. Pregnancy characteristics and maternal breast
cancer risk: a review of the epidemiologic literature. Cancer Causes and Control
21, 967–989.
Oreopoulos, P., 2006. Estimating average and local average treatment effects of edu-
cation when compulsory school laws really matter. American Economic Review
96 (1), 152–175.
Spasojevic, J., 2010. Effects of education on adult health in Sweden: results from a
natural experiment. Contributions to Economic Analysis 290, 179–199.
Storey, J., 2002. A direct approach to false discovery rates. Journal of the Royal
Statistical Society: Series B 64, 479–498.
Storey, J., Tibshirani, R., 2003. Statistical significance for genome-wide experiments.
Proceeding of the National Academy of Sciences 100, 9440–9445.
Tabar, L., et al., 1985. Reduction in mortality from breast cancer after mass screening
with mammography. The Lancet 325 (8433), 829–832.
 Journal of Health Economics 42 (2015) 125–138

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Peer effects, fast food consumption and adolescent weight gain夽

Bernard Fortin a,∗ , Myra Yazbeck b
a
CIRPÉE, IZA, CIRANO and Department of Economics, Université Laval, Canada
b
School of Economics, University of Queensland, Australia

a r t i c l e i n f o a b s t r a c t

Article history: This paper aims at opening the black box of peer effects in adolescent weight gain. Using Add Health
Received 1 February 2015 data on secondary schools in the U.S., we investigate whether these effects partly flow through the eating
Accepted 12 March 2015 habits channel. Adolescents are assumed to interact through a friendship social network. We propose a
Available online 28 March 2015
two-equation model. The first equation provides a social interaction model of fast food consumption. To
estimate this equation we use a quasi maximum likelihood approach that allows us to control for common
JEL classification:
environment at the network level and to solve the simultaneity (reflection) problem. Our second equation
C31 I10
is a panel dynamic weight production function relating an individual’s Body Mass Index z-score (zBMI)
I12
to his fast food consumption and his lagged zBMI, and allowing for irregular intervals in the data. Results
Keywords: show that there are positive but small peer effects in fast food consumption among adolescents belonging
Obesity to a same friendship school network. Based on our preferred specification, the estimated social multiplier
Overweight is 1.15. Our results also suggest that, in the long run, an extra day of weekly fast food restaurant visits
Peer effects increases zBMI by 4.45% when ignoring peer effects and by 5.11%, when they are taken into account.
Social interactions
© 2015 Elsevier B.V. All rights reserved.
Fast food

1. Introduction
For the past few years, obesity has been one of the major con-
cerns of health policy makers in the U.S. It has also been one of the
principal sources of increased health care costs. In fact, the increas-
ing trend in children’s and adolescents’ obesity (Ogden et al., 2012)
has raised the annual obesity-related medical costs to $147 billion
夽 We wish to thank Christopher Auld, Charles Bellemare, Luc Bissonnette, Vincent
Boucher, Paul Frijters, Guy Lacroix, Paul Makdissi, Daniel L. Millimet, Kevin Moran,
Bruce Shearer for useful comments and Yann Bramoullé, Badi Baltagi, Rokhaya Dieye,
Habiba Djebbari, Tue Gorgens, Bob Gregory, Louis Hotte, Linda Khalaf, Lung fei
Lee, Xin Meng and Rabee Tourky for useful discussions. The suggestions of two
anonymous referees have substantially improved the paper. We are grateful to
Rokhaya Dieye for outstanding assistance research. The usual disclaimer applies.
Financial support from the Canada Research Chair in the Economics of Social Policies
and Human Resources and le Centre interuniversitaire sur le risque, les politiques
économiques et l’emploi is gratefully acknowledged. This research uses data from
Add Health, a program project directed by Kathleen Mullan Harris and designed by
J. Richard Udry, Peter S. Bearman, and Kathleen Mullan Harris at the University of
North Carolina at Chapel Hill, and funded by grant P01-HD31921 from the Eunice
Kennedy Shriver National Institute of Child Health and Human Development, with
cooperative funding from 23 other federal agencies and foundations.
∗ Corresponding author. Tel.: +1 418 656 5678.
E-mail addresses: Bernard.Fortin@ecn.ulaval.ca (B. Fortin),
m.yazbeck@uq.edu.au (M. Yazbeck).
per year (Finkelstein et al., 2009). Obesity is also associated with
increased risk of reduced life expectancy as well as with serious
health problems such as type 2 diabetes (Maggio and Pi-Sunyer,
2003), heart disease (Calabr et al., 2009) and certain cancers (Calle,
2007), making obesity a real public health challenge.
Recently, a growing body of the health economics literature has
tried to look into the obesity problem from a new perspective using
a social interaction framework. An important part of the evidence
suggests the presence of peer effects in weight gain. On one hand,
Christakis and Fowler (2007), Trogdon et al. (2008), Renna et al.
(2008) and Yakusheva et al. (2014) are pointing to the social multi-
plier as an important element in the obesity epidemics. As long as it
is strictly larger than one, a social multiplier amplifies, at the aggre-
gate level, the impact of any shock (such as the reduction in relative
price of junk food) that may affect obesity at the individual level.
This is so because the aggregate effect incorporates, in addition to
the sum of the individual direct effects, positive indirect peer effects
stemming from social interactions. On the other hand, Cohen-Cole
and Fletcher (2008b) found that there is no evidence of peer effects
in weight gain. Also, results from a placebo test performed by
the same authors (Cohen-Cole and Fletcher, 2008a) indicate that
there are peer effects in acne (!) in the Add Health data when one
applies the Christakis and Fowler (2007) method discussed later
on.

http://dx.doi.org/10.1016/j.jhealeco.2015.03.005
0167-6296/© 2015 Elsevier B.V. All rights reserved.
126 B. Fortin, M. Yazbeck / Journal of Health Economics 42 (2015) 125–138
While the presence (or not) of peer effects in weight has been
widely researched,1 the literature on the mechanisms by which
peer effects flows is still scarce. Indeed, most of the relevant lit-
erature attempts to estimate the relationship between variables
such as an individual’s Body Mass Index (BMI) and his average
peers’ BMI, without exploring the channels at source of this poten-
tial linkage.2 The aim of this paper is to go beyond the black box
approach of peer effects in weight gain and try to identify one
potentially important mechanism through which peer effects in
adolescence overweight may flow: eating habits (as proxied by fast
food consumption).
Three reasons justify our interest in eating habits in analyzing
the impact of peer effects on teenage weight. First of all, there is
an important literature that points to eating habits as an important
component in weight gain (e.g., Niemeier et al., 2006; Rosenheck,
2008).3 Second, one suspects that peer effects in eating habits are
likely to be important in adolescence. Indeed, at this age, youngsters
have increased independence in general and more freedom as far
as their food choices are concerned. Usually vulnerable, they often
compare themselves to their friends and may alter their choices
to conform to the behaviour of their peers. Therefore, unless we
scientifically prove that obesity is a virus, it is counter intuitive to
think that one can gain weight by simply interacting with an obese
person.4 This is why we are inclined to think that the presence of
real peer effects in weight gain can be estimated using behavioural
channels such as eating habits. Third, our interest in peer effects in
youths’ eating habits is policy driven. There has been much discus-
sion on implementing tax policies to address the problem of obesity
(e.g., Caraher and Cowburn, 2007; Powell et al., 2013). As long as
peer effects in fast food consumption is a source of externality that
may stimulate overweight among adolescents, it may be justified
to introduce a consumption tax on fast food. The optimal level of
this tax will depend, among other things, on the social multiplier
of eating habits, and on the causal effect of fast food consumption
on adolescent weight.
In order to analyze the impact of peer effects in eating habits on
weight gain, we propose a two-equation model. The first linear-in-
means equation relates an individual’s fast food consumption to his
individual characteristics, his reference group’s mean fast food con-
sumption (endogenous peer effect), and his reference group’s mean
characteristics (contextual peer effects). The endogenous peer effect
reflects the possibility that eating behaviour of his friends influ-
ences a teenager’s own behaviour. For instance, one reason why an
adolescent may want to go to a fast food restaurant is to be with
his friends during the lunch. Contextual effects, such as the average
level of education of his friends’ mother, may also affect a teenager’s
eating habits. Thus, mothers with higher education may encourage
not only their children but also their children’s friends to develop
accurate eating bebaviour.
The second equation is a panel dynamic production function that
relates an individual’s BMI adjusted for age (BMI z-score or zBMI) to
his fast food consumption, his lagged zBMI and other control vari-
1
For a complete review see Fletcher et al. (2011) who conducted a systematic
review of literature that shows that school friends are similar as far as body weight
and weight related behaviours.
2
One recent exception is Yakusheva et al. (2011) and Yakusheva et al. (2014)
who look at peer effects in overweight and in weight management behaviours such
as eating and physical exercise, using randomly assigned pairs of roommates in
freshman year.
3
An indirect evidence of the relationship between eating habits and weight gain
come from the literature on the (negative) effect of fast food prices on adolescents’
BMI (see Auld and Powell, 2009; Powell, 2009; Powell and Bao, 2009). See also Cutler
et al. (2003) which relates the declining relative price of fast food and the increase
in fast food restaurant availability over time to increasing obesity in the U.S.
4 6
Of course, having obese peers may influence an individual’s tolerance for being
obese and therefore his weight management behaviours.
ables. The system of equations thus allows us to evaluate the impact
of an eating habits’ exogenous shock on an adolescent’s weight,
when peer effects on fast food consumption are taken into account.
To estimate our two-equation model, we use three waves of the
National Longitudinal Study of Adolescent Health (Add Health), that
is, Wave II (1996), Wave III (2001) and Wave IV (2008).5 We define
peers as the nominated group of individuals reported as friends
within the same school. The consumption behaviour is depicted
through the reported frequency (in days) of fast food restaurant
visits in the past week.
Estimating our system of equations raises serious econometric
problems. It is well known that the identification of peer effects
(first equation) is a challenging task. These identification issues
were first pointed out by Manski (1993) and discussed among oth-
ers by Bramoullé et al. (2009) and Blume et al. (2015). On one
hand, (endogenous + contextual) peer effects must be identified
from correlated (or confounding) factors. For instance, students in a
same friendship group may have similar eating habits because they
share similar characteristics (i.e., homophily) or face a common
environment (e.g., same school). On the other hand, simultane-
ity between an adolescent’s and his peers’ behaviour (referred
to as the reflection problem by Manski) may make it difficult to
identify separately the endogenous peer effect and the contextual
effects.
We use a new approach based on Bramoullé et al. (2009) and
Lee et al. (2010), and extended by Blume et al. (2015) to address
these identification problems and to estimate the peer effects
equation. First, we assume that in their fast food consumption
decisions, adolescents interact through a friendship network. Each
school is assumed to form a network. School fixed effects are
introduced to capture correlated factors associated with network
invariant unobserved variables (e.g., similar preferences due to self-
selection in schools, same school nutrition policies, distance from
fastfood restaurants). The structure of friendship links within a
network is allowed to be stochastic but conditional on the school
fixed effects and observable individual and contextual variables,
is strictly exogenous. The possibility that friends select each other
using unobservable traits that may be correlated with their fast
food consumption decisions is an important issue and is discussed
later on.
To solve the reflection problem, we exploit results by Bramoullé
et al. (2009) who show that if there are at least two agents who
are separated by a link of distance 3 within a network (i.e., there
are two adolescents in a school who are not friends but are linked
by two friends), both endogenous and contextual peer effects are
identified. Finally, we exploit the similarity between the linear-
in-means model and the spatial autoregressive (SAR) model with
or without autoregressive spatial errors.6 The model is estimated
using a quasi maximum likelihood (QML) approach as in Lee et al.
(2010). The QML is appropriate when the estimator is derived
from a normal likelihood but the error terms in the model are not
truly normally distributed. We also estimate the model using gen-
eralized spatial two-stage least square proposed in Kelejian and
Prucha (1998) and refined in Lee (2003), which is less efficient than
QML.
The estimation of the production function (second equation)
also raises serious econometric issues. First, fast food consumption
is likely to be an endogenous variable correlated with the individ-
ual error term. Moreover, the short and the long term impacts of
fast food consumption on zBMI may be different, suggesting the
5
Note that for the first equation we use wave II and for the second equation we
use the three waves.
Our approach is more general than the SAR model as the latter usually ignores
contextual effects and spatial fixed effects.
 B. Fortin, M. Yazbeck / Journal of Health Economics 42 (2015) 125–138
introduction of lagged zBMI as an explanatory variable. Finally, Add
Health data waves are collected at irregular intervals. As a con-
sequence, estimators obtained from standard dynamic panel data
models are inconsistent. In order to deal with these problems, we
use a nonlinear instrumental approach developed by Millimet and
McDonough (2013).
Results suggest that there is a positive but small endogenous
peer effect in fast food consumption among adolescents in gen-
eral. Based on our QML specification, the estimated social multiplier
is 1.15. Moreover, the production function estimates indicate that
there is a positive significant impact of fast food consumption on
zBMI. Combining these results, we find that, in the long run, an extra
day of weekly fast food restaurant visits increases zBMI by 4.45%
when ignoring peer effects and by 5.11%, when they are taken into
account.
The remaining parts of this paper will be laid out as follows. Sec-
tion 2 provides a survey of the literature on peer effects in obesity as
well as its decomposition into the impact of peer effects on fast food
consumption and the impact of fast food consumption on obesity.
Section 3 presents the specification of our fast food equation with
peer effects. Section 4 is devoted to our weight production function.
In Section 5, we give an overview of the Add Health Survey and we
provide descriptive statistics of the data we use. In Section 6, we
discuss estimation results. Section 7 concludes.
2. Previous literature
In recent years, a number of studies found strong “social network
effects” in weight outcomes. In a widely debated article, Christakis
and Fowler (2007) found that an individual’s probability of becom-
ing obese increased by 57% if he or she had a friend who became
obese in a given interval.7 However, their analysis has been crit-
icized for suffering from a number of limitations (see Cohen-Cole
and Fletcher, 2008b; Lyons, 2011; Shalizi and Thomas, 2011).8 In
particular, it ignores potential spurious correlations between two
friends’ BMI resulting from the fact that they are exposed to the
same environment. Both Shalizi and Thomas (2011) and Lyons
(2011) show that relying on link asymmetries does not rule out
shared environment as it claims. Also, the simultaneity problem
between these two outcomes is not directly addressed by allowing
the peer’s obesity to be endogenous.
In the same spirit, Trogdon et al. (2008) investigate the presence
of peer effects in obesity using Add Health data. They include school
fixed effects to account for the fact that students in a same school
share a same surrounding. The authors also estimate their BMI peer
model with an instrumental variable approach using information
on friends’ parents’ obesity and health and friends’ birth weight
as instruments for peers’ BMI. They find that a one point increase
in peers’ average BMI increases own BMI by 0.52 point. Based
on a similar approach and using Add Health dataset, Renna et al.
(2008) also find positive peer effects. These effects are significant
for females only (=0.25 point).9 These analyses raise a number of
concerns though. In particular, they assume no contextual variables
reflecting peers’ mean characteristics. This rules out the reflec-
tion problem by introducing non-tested restriction exclusions. In
our approach, we introduce school fixed effects and, for each
7
They used a 32-year panel dataset on adults from Framingham, Massachusetts
and a logit specification.
8 10
For a response to these criticisms and others, see Fowler and Christakis (2008),
VanderWeele (2011) and Christakis and Fowler (2013).
9 11
Also Ali et al. (2011) and Ali et al. (2011) provide evidence that there are peer cor-
relations in weight related behaviours and peer influence in weight misperception
respectively.
127
individual variable, the corresponding contextual variable at the
reference group level.
Using the same dataset as Trogdon et al. (2008) and Renna et al.
(2008), Cohen-Cole and Fletcher (2008b) exploit panel information
(wave II in 1996 and wave III in 2001) for adolescents for whom at
least one of same-sex friend is also observed over time. Compared
with Christakis and Fowler’s approach, their analysis introduces
time invariant and time dependent environmental variables (at
the school level). Friendship selection is controlled for by individ-
ual fixed effects. The authors find that peer effects are no longer
significant with this specification.
All the studies discussed up to this point focus on peer effects
in weight outcomes without analyzing quantitatively the mecha-
nisms by which they may occur. The general issue addressed in this
paper is whether the peer effects in weight gain among adolescents
partly flow through the eating habits channel. This raises in turn two
basic issues: (a) are there peer effects in fast food consumption?,
and (b) is there a link between weight gain (or obesity) and fast
food consumption? In this paper, we address both issues.
The literature on peer effects in eating habits (first issue) is
recent and quite limited. In a recent paper, in which the formation
of the network is randomized, Yakusheva et al. (2011) estimate peer
effects in explaining weight gain among freshman girls using a simi-
lar set up but in school dormitories. In their paper, they test whether
some of the student’s weight management behaviours (i.e., eating
habits, physical exercise, use of weight loss supplements) can be
predicted by her randomly assigned roommate’s behaviours. Their
results provide evidence of the presence of negative peer effects
in weight gain. Their results also suggest positive peer effects in
eating habits, exercise and use of weight loss supplements. In a
subsequent paper, Yakusheva et al. (2014) investigate the pres-
ence of peer effects in weight gain exploiting random assignment of
roommates during first year of college. The authors find evidence
that suggests that peer effects in weight gain are predominantly
significant among females.10
Our paper finds its basis in this literature as well as the literature
on peer effects and obesity discussed above. However, while works
by Yakusheva et al. (2011) and Yakusheva et al. (2014) rely upon
experimental data, we use observational non-experimental data.
Peers are considered to have social interactions within a school
network. This allows for the construction of a social interaction
matrix that reflects how social interaction between adolescents in
schools occurs in a more realistic setting (as in Trogdon et al., 2008;
Renna et al., 2008). An additional originality of our paper lies in the
fact that it relies upon a linear-in-means approach when relating
an adolescent’s behaviour to that of his peers. Also, the analogy
between the forms of the linear-in-means model and the spatial
autoregressive (SAR) model allows us to exploit the particularities
of this latter model, in particular the natural instruments that are
derived from its structural form.
Regarding the second issue, i.e., the relationship between weight
(or obesity) and fast food consumption, it is an empirical ques-
tion that is still on the debate table.11 There is no clear evidence
in support of a causal link between fast food consumption and
obesity. Nevertheless, most of the literature in epidemiology
finds evidence of a positive correlation between fast food con-
sumption and obesity (see, Anderson et al., 2011; Rosenheck,
2008).
In contrast, De la Haye et al. (2010) provide evidence that close adolescent male
friends tend to be similar in their consumption of high-calorie food.
The literature on the impact of physical activity on obesity is also inconclu-
sive. For instance, Berentzen et al. (2008) provide evidence that decreased physical
activity in adults does not lead to obesity.
128 B. Fortin, M. Yazbeck / Journal of Health Economics 42 (2015) 125–138
The economic literature tends to be conservative with respect
to this question. It focuses on the impact of “exposure” to fast
food on obesity. Dunn et al. (2012), using an instrumental variable
approach, investigates the relationship between fast food avail-
ability and obesity. They finds that an increase in the number of
fast food restaurants has a positive effect on the BMI among non-
whites. Alviola et al. (2014), using a similar approach, provides
evidence that the number of fast food restaurants has a significant
impact on school obesity rates. Similarly, Currie et al. (2010) find
evidence that proximity to fast food restaurants has a significant
effect on obesity for 9th graders. Also, Anderson and Matsa (2011),
exploiting the placement of Interstate Highways in rural areas to
obtain exogenous variations in the effective price of restaurants,
did not find any causal link between restaurant consumption and
obesity. More generally, Cutler et al. (2003) and Bleich et al. (2008)
argue that the increased calorie intake (i.e., eating habits) plays a
major role in explaining current obesity rates. Importantly, weight
prior to adulthood sets the stage for weight in adulthood. While
most of the economics literature analyses the relationship between
adolescents’ fast food consumption and their weight using an indi-
rect approach (i.e, effect to fast food exposure), we adopt a direct
approach linking weight as a function of fast food consumption,
lagged weight and control variables.
In the next two sections, we present our two-equation model of
weight production function with peer effects in fast food consump-
tion. We first propose a linear-in-means social interaction equation
of fast food consumption (first equation) and discuss the economet-
ric methods we use to estimate it. We then present our econometric
weight production function which relates the adolescent’s zBMI
level to his fast food consumption (second equation).
3. Social interactions equation of fast food consumption
We assume a set of N adolescents i that are partitioned in a set
of L networks. A network is defined as a structure (e.g., school) in
which adolescents are potentially tied by a friendship link. Each
adolescent i in his network has a set of nominated friends Ni of size
ni that constitute his reference group (or peers). We assume that i is
excluded from his reference group. Since peers are defined as nom-
inated friends, the number of peers will not be the same for every
network member. Let Gl (l = 1, . . ., L) be the social interaction matrix
for a network l. Its element glij takes a value of 1/ni when i is friend
with j, and zero otherwise. Therefore, assuming no isolated individ-
uals, the Gl matrix is row normalized. We define yli as the fast food
consumed by adolescent i in network l, xli represents the adolescent
i’s observable characteristics, yl the vector of fast food consump-
tion in network l, and xl is the corresponding vector for individual
characteristics. To simplify our presentation, we look at only one
characteristic (e.g., adolescent’s mother education).12 The network
invariant unobservable variables are captured through fixed net-
work effects (the ˛l ’s). They take into account unobserved factors
such as preferences of school, school nutrition policies, or pres-
ence of fast food restaurants around the school. The εli ’s are the
idiosyncratic error terms. They capture i’s unobservable character-
istics that are not invariant within the network. Formally, one can
write the linear-in-means equation for adolescent i as follows:
 
y
j∈Ni lj
x
j∈Ni lj
yli = ˛l + ˇ + xli + ı + εli ,
ni ni
12
Later on, in Section 3.1.1, we will generalize the equation to account for many
characteristics.
 
j∈N
ylj
j∈N
xlj
i i
where ni and ni are respectively his peers’ mean fast
food consumed and characteristics. In the context of our paper,
ˇ is the endogenous peer effect. It reflects how the adolescent’s
  mean fast food
consumption of fast food is affected by his peers’
consumption. It is standard to assume that ˇ < 1. The contex-
tual peer effect is represented by the parameter ı.13 It captures
the impact of his peers’ mean characteristic on his fast food con-
sumption. It is important to note that the Gl matrix and the xl ’
vector are allowed to be stochastic but are assumed strictly exoge-
nous conditional on ˛l , that is, E(εli |xl , Gl , ˛l ) = 0. This assumption
is flexible enough to allow for correlation between the network’s
unobserved common characteristics (e.g., school’s cafeteria quality)
and observed characteristics (e.g., mother’s education). Neverthe-
less, once we condition on these common characteristics, mother’s
education is assumed to be independent of the idiosyncratic error
terms. Let Il be the identity matrix for a network l and l the corre-
sponding vector of ones, Eq. (1) for network l can be rewritten in
matrix notation as follows:
yl = ˛l l + ˇGl yl + xl + ıGl xl + εl , forl = 1, ..., L. (2)
Note that Eq. (2) is similar to a SAR model (e.g., Cliff and Ord, 1981)
generalized to allow for contextual and fixed effects (hereinafter
referred to as the GSAR model). Since ˇ < 1, (Il − ˇGl ) is invertible.
Therefore, in matrix notation, the reduced form of Eq. (2) can be
written as:
−1 −1
yl = ˛l /(1 − ˇ)l + (Il − ˇGl ) (Il + ıGl )xl + (Il − ˇGl ) εl , (3)
−1 ∞
where we use the result that (I−ˇGl ) = k=0
ˇk Gkl , so that
the vector of intercepts is ˛l /(1 − ˇ)l , assuming no isolated
adolescents.14
Eq. (3) allows us to evaluate the impact of a marginal shock
in ˛l (i.e., a common exogenous change in fast food consumption
within the network) on an adolescent i’s fast food consumption,
when the endogenous peer effect is taken into account. One has
∂(E(yli | ·)/∂˛l = 1/(1 − ˇ). This expression is defined as the social
multiplier in our model. When ˇ > 0 (strategic complementarity in
fast food consumption), the social multiplier is larger than 1. In this
case, the impact of the shock is amplified by social interactions as
more fast food consumption by his peers induces an adolescent to
adopt a similar behaviour.
We then perform a panel-like within transformation. More pre-
cisely, we average Eq. (3) over all students in network l and subtract
it from i’s equation. This transformation allows us to address prob-
lems that arise from the fact that adolescents are sharing the
same environment or preferences. Let Kl = Il − Hl be the matrix that
obtains the deviation from network l mean with Hl = n1 (l l ). The
l
network within transformation will eliminate the network fixed
effect ˛l . Pre-multiplying (3) by Kl yields the reduced form of the
model for network l, in deviation:
−1 −1
Kl yl = Kl (Il − ˇGl ) (Il + ıGl )xl + Kl (Il − ˇGl ) εl . (4)
3.1. Identification
Our peer effects structural Eq. (2) raises two basic identification
problems.
(1)
13
It is standard to assume the presence of a contextual effect for each individual
characteristic influencing the outcome. Otherwise, the model may impose ad hoc
exclusion restrictions which generate invalid instruments and inconsistent estima-
tors.
14
When an adolescent is isolated, his intercept is ˛l .
 B. Fortin, M. Yazbeck / Journal of Health Economics 42 (2015) 125–138
3.1.1. Simultaneity
Simultaneity between individual and peer behaviour (the reflec-
tion problem) may prevent separating contextual effects from
endogenous effects. This problem has been analyzed by Bramoullé
et al. (2009) when individuals interact through social networks.
They show that the conditions of identification depend on both the
values of parameters and the structure of the network. More explic-
itly, let us first assume throughout that ˇ + ı = / 0. Then define G
the block-diagonal matrix with the Gl ’s on its diagonal. Assume first
the absence of fixed network effects (i.e., ˛l = ˛ for all l). In this case,
Bramoullé et al. (2009) show that the structural parameters of Eq.
(2) are identified if the matrices I, G, G2 are linearly independent.
This condition is satisfied when there are at least two adolescents
who are separated by a link of distance 2 within a network. This
means that they are not friends but have a common friend.15 The
intuition is that this provides exclusion restrictions in the model.
More precisely, the friends’ friends mean characteristic can serve as
instrument for the mean friends’ fast food consumption. Of course,
when fixed network effects are allowed, the identification condi-
tions are more restrictive. Bramoullé et al. (2009) show that, in this
case, the structural parameters are identified if the matrices I, G, G2
and G3 are linearly independent. This condition is satisfied when at
least two adolescents are separated by a link of distance 3 within a
network, i.e., we can find two adolescents who are not friends but
3 > 0 while g 2 = g = 0.
are linked by two friends. In this case, glij ij ij
Hence, no linear relation of the form G3 = 0 I + 1 G + 2 G2 can exist.
This condition holds in most friendship networks and, in particular,
in the data we use.16
3.1.2. Correlated effects
The presence of confounding unobservable variables affect-
ing fast food consumption and correlated with the explanatory
variables raises difficult identification problems. First, since ado-
lescents are not randomly assigned into schools, endogenous
self-selection through networks may be the source of potentially
serious biases in estimating (endogenous + contextual) peer effects.
Indeed, if the variables that drive this process of selection are
not fully observable, correlations between unobserved network-
specific factors and the regressors are potentially important sources
of bias. In our approach, we assume that network fixed effects cap-
ture these factors. This is consistent with two-step models of link
formation. Each adolescent joins a school in a first step, and forms
friendship links with others in his school in a second step. In the
first step, adolescents self-select into different schools with selec-
tion bias due to specific school characteristics. In a second step,
link formation takes place within schools randomly or based on
observable individual characteristics only. Recall also that network
fixed effects take into account common unobservable variables at
the school level that may influence fast food consumption (e.g.,
availability of fast food restaurants).
Of course, one limitation of using network fixed effects is that
it ignores the possibility that the links formation within a network
depend on omitted variables. The matrix G may be endogenous
even when controlling for the network fixed effects and observable
characteristics. Thus friends may select each other using unob-
servable traits that may be correlated with fast food consumption
15
More generally, Eq. (2) is identified when individuals do not interact in groups
or interact in groups with at least three different sizes (see Bramoullé et al., 2009).
16
Identification fails, however, for a number of non trivial networks. This is notably
the case for complete bipartite networks. In these graphs, the population of students
is divided in two groups such that all students in one group are friends with all
students in the other group, and there is no friendship links within groups. These
include star networks, where one student, at the centre, is friend with all other
students, who are all friends only with him.
129
(e.g., impulsivity, a specific taste for sugar- and fat-rich food).
Recently, some researchers (e.g., Hsieh and Lee, 2011; Goldsmith-
Pinkham and Imbens, 2013; Liu et al., 2013; Badev, 2013) have
made attempts to develop econometric models allowing for the
joint estimation of network formation and network interactions.
However, empirical results using Add Health data and focusing on
outcomes such as smoking, sleeping behaviour, and scholar perfor-
mance, do not seem to detect much difference in peer effects when
networks are assumed exogenous and when they are allowed to be
endogenous.17
One specification of our peer effects equation also allows the
error terms to be (first-order) autocorrelated within networks.
Therefore its structure becomes analogous to that of a general-
ized spatial autoregressive model with network autoregressive
disturbances (denoted as the GSARAR model). This model implies
that in addition to the endogenous and contextual effects,
some unobserved characteristics of the friends are also inter-
dependent. In this case, the error terms in (2) can be written
as:
εl = Gl εl +  l , (5)
where the innovations,  l , are assumed to be i.i.d. (0,  2 Il ) and || < 1.
Given these assumptions, we can write:
εl = (Il − Gl )−1 l . (6)
Allowing for many characteristics and performing a Cochrane-
Orcutt-like transformation on the structural equation (4) in
deviation, the latter is given by the following structural form:
Kl Ml yl = ˇKl Ml Gl yl + Kl Ml Xl  + Kl Ml Gl Xl ı + l , (7)
where Xl is the matrix of adolescents’ characteristics18
in the lth
network, Ml = (I − Gl ) and l = Kl Ml  l .
Following Lee et al. (2010), we propose two approaches to esti-
mate the peer effects equation (7): a quasi maximum likelihood
approach (QML) and a generalized spatial two stage least squares
(GS-2SLS) approach. The QML estimators are estimated assum-
ing that the disturbances are normally distributed. However, we
do allow the log-likelihood function to be partially misspecified,
as standard errors are computed to be robust to non-normal dis-
turbances (using a sandwich formula). Assuming that the error
terms are i.i.d. and under a number of regularity assumptions (see
Lee et al., 2010, p. 152), QML estimators are consistent but not
asymptotically efficient. On the other hand, GS-2SLS estimators also
assume that the error terms are i.i.d. but impose less regularity con-
ditions than QML estimators. QML estimators are asymptotically
more efficient than GS-2SLS estimators.19
4. Weight production function
In this section, we propose a dynamic (AR(1)) weight production
function that relates an individual’s zBMI in period t (assumed a
year) to his lagged zBMI, his fast food consumption and his own
characteristics in period t. Let yitb be an individual i’s zBMI level in
f
period t, and yit be the individual’s fast food consumption in period
17
One reason may be that this data base includes a very large number of observable
characteristics, some of them being used in the regressions. Another explanation
is that, although statistically significant, the explanatory power of the individual
characteristics on the probability that two individuals are friends is extremely small
(Boucher, 2014).
18
Following the linear-in-means model, we allow the peers’ mean characteristic
corresponding to each individual’s characteristic to have a potential effect on his
fast food consumption. Therefore, we do not impose ad hoc (identifying) exclusion
restrictions to the structural peer effects equation.
19
The derivation of the QML and GS-2SLS estimators are presented in the
Appendix.
130 B. Fortin, M. Yazbeck / Journal of Health Economics 42 (2015) 125–138

t. Then, for a given vector of characteristics x̃it , the data generating and serially uncorrelated. Also, while the estimators are inconsis-
process (DGP) of the weight production function can be formally tent when T is fixed and N→ ∞, Monte–Carlo simulations by MM
expressed as follows (for notational simplicity we suppress l): suggest that this approach has superior small sample properties
compared to other dynamic panel data estimators.
f
yitb = 1 yi,t−1
b
+ 2 yit + 3 x̃it + i +
it , (8) Second, some covariates (in particular, the individual’s fast food
consumption, yf ) are likely to be correlated with the unobserved
where 1 is the autoregressive parameter (|1 | < 1), i is the indi-
effect and/or to be serially correlated. In the first case, Everaert
vidual i’s time-invariant error component (fixed effect) and
it , his
(2013) suggests to use Hausman and Taylor (1981) type instru-
idiosyncratic error that may change across t. One difficult problem
ments for these covariates, that is, deviations from individual
with (8) is that the Add Health data set, the waves are irregu-
sample means (e.g., ÿf ). Also, in the presence of serially correlated
larly spaced. This means that the successive periods of observed
covariates, one solution suggested by MM is to impute data for the
data (that is, for 1996, 2001 and 2008) do not conform to suc-
missing periods. For instance, we can use current value of covariates
cessive (yearly) periods as defined by our underlying DGP. In that
to approximate missing covariates between periods m and m − 1.21
case, standard methods to estimate a dynamic panel model with
Therefore, in Eq. (10), we can write:
endogenous variables (e.g., Anderson and Hsiao, 1981; Arellano
and Bond, 1991) yield inconsistent estimators. To address this 
gm −1
f j f
g
1 − 1m
point, we follow Millimet and McDonough (2013) (hereinafter MM) (2 yi,t(m)−j + 3 x̃i,t(m)−j )1 ≈ (2 yi,m + 3 x̃i,m ) .
1 − 1
approach. From repeated substitution in Eq. (8) we rewrite Eq. (8) j=1
defined over the observed periods m = 1, 2, 3, one obtains: (11)

b f
yim = 1gm yi,m−1
b
+ 2 yim + 3 x̃im +  i +
˜ im , (9)
In this paper, we estimate the weight production function given
where gm is the gap size or the number of years between observed by Eqs. (9)–(11) using a nonlinear instrumental approach and based
period m and m − 1, (which, in our case, are equal to g1 = 1, g2 = 5, on current values of covariates to approximate missing data for the
g3 = 7)20 ;  = (1 − 1gm )/(1 − 1 ), and missing periods. Following MM, we denote this estimator: E-NLS-
IV-C. We also present a GMM version of this estimator using a two-

gm −1
f j

gm −1
j step approach to obtain an optimal weighting matrix (clustered at

˜ im = (2 yi,t(m)−j + 3 x̃i,t(m)−j )1 + 1

i,t(m)−j , (10)
the individual level).
j=1 j=0
As discussed earlier, our interest in this production function goes
where t(m) is the actual period reflected by the observed period m: beyond a mere association between fast food consumption and
t(1) = 1 ; t(2) = 6 ; t(3) = 13. weight. We are particularly interested to analyze the magnitude of
Eq. (9) shows that when data are irregularly spaced, (1) the coef- a change in zBMI resulting from a common exogenous shock on fast
ficient on the lagged dependent variable is not constant but equal to food consumption within the network, when peer effects are taken
1gm ; (2) the error term
˜ im contains the covariates and the idiosyn- into account. Our two equation model allows us to compute this
f
cratic errors from the missing periods between m and m − 1, and the result. Partially differentiating (8) with respect to yi,t−1 and using
current error; (3) the unobserved fixed effect has a period-specific the social multiplier [= 1/(1 − ˇ)] yields the magnitude of a short
b
run change in zBMI (i.e., for yi,t−1
factor loading, . The first point raises the following difficulty: the given) resulting from a common
equation is now nonlinear in 1 , which suggests the use of a nonlin- 2
marginal shock on fast food consumption: ∂E(yitb | · )/∂˛l = 1−ˇ
. This
ear in parameters approach. More importantly, unequally spaced expression entails two components: the impact of the fast food con-
data relegate missing covariates into the error term (point 2). This 1
sumption on zBMI (= 2 ) and the multiplier effect (= 1−ˇ ). In the
is a serious source of concerns as long as some contemporary
long run, at the new stationary state, the impact of the shock on
covariates are serially correlated and therefore become mechan- 2
zBMI is given by (1−ˇ)(1− .
ically endogenous. Finally, one cannot eliminate the fixed effect i ) 1

using standard first-differencing or mean-differencing transforma-

tions since the factor loading parameter  varies from one observed 5. Data and descriptive statistics
period to another (point 3).
To estimate such an equation, MM suggests the use of a non- The Add Health survey is a longitudinal study that is nationally
linear instrumental approach extending Everaert (2013) technique representative of American adolescents in grades 7 through 12. It is
for estimating dynamic panel data models. It consists first in instru- one of the most comprehensive health surveys that contains fairly
menting the lagged zBMI (= yi,m−1 b b
) with the OLS residual of yi,m−1 exhaustive social, economic, psychological and physical well-being
b b m−1 variables along with contextual data on the family, neighbourhood,
regressed on its backward mean yi,m−1 , where yi,m−1 = m 1
yb .
s=0 i,s community, school, friendships, peer groups, romantic relation-
The intuition here is that the residual (which reflects the part of
ships, etc. In wave I (September 1994 to April 1995), all students
the lagged zBMI not explained by its backward mean) is likely
(around 90 000) attending the randomly selected high schools
to be highly correlated with the lagged zBMI. However, it should
were asked to answer a short questionnaire. An in-home sample
be uncorrelated with the fixed effect reflecting the time-invariant
(core sample) of approximately 20 000 students was then randomly
unobserved part of the individual’s zBMI. Also it should be uncorre-
drawn from each school. These adolescents were asked to partic-
lated with the contemporary idiosyncratic error term as long as the
ipate in a more extensive questionnaire where detailed questions
latter is i.i.d. (0, 2 ). Therefore, the residual is a good candidate as
were asked. Information on (but not limited to) health, nutrition,
an instrument for the lagged zBMI. More explicitly, MM shows that
expectations, parents’ health, parent-adolescent relationship and
a nonlinear IV version using of Everaert (2013) technique (referred
friends nomination was gathered.22 This cohort was then followed
to as E-NLS-IV) to account for irregular spacing yields consistent
estimators, when T→ ∞ and the covariates are strictly exogenous

21
No approximation is needed for variables such age, for which we have perfect
information at each period.
20 22
One has g(1) = 1 since Wave I from Add Health data (which corresponds to m = 0) Adolescents were asked to nominate up to five female friends and five male
was collected in 1995. friends.
 B. Fortin, M. Yazbeck / Journal of Health Economics 42 (2015) 125–138 131

in-home in the subsequent waves in 1996 (wave II), 2001 (wave Table 1
Descriptive statistics.
III) and 2008–2009 (wave IV). The extensive questionnaire was
also used to construct the saturation sample that focuses on 16 Variable Mean SD
selected schools (about 3000 students). Every student attending Fast food consumptiona 2.33 1.74
these selected schools answered the detailed questionnaire. There
Female 0.50 0.50
are two large schools and 14 other small schools. All schools are
Age 16.36 1.44
racially mixed and are located in major metropolitan areas except
one large school that has a high concentration of white adolescents White 0.57 0.49
Black 0.15 0.34
and is located in a rural area. Consequently, fast food consump- Asian 0.01 0.09
tion may be subject to downward bias if one accepts the argument Native 0.13 0.33
that the fast food consumption among white adolescents is usually Other 0.14 0.35
lower than that of black adolescents. Mother present 0.85 0.35
In this paper we use the saturation sample of wave II in-home
Mother education
survey to investigate the presence of peer effects in fast food
No high school degree 0.15 0.35
consumption.23 One of the innovative aspects of this wave is the High school/GED/Vocational Instead of high school 0.36 0.48
introduction of the nutrition section. It reports among other things Some college/vocational after high school 0.21 0.39
food consumption variables (e.g., fast food, soft drinks, desserts, College 0.18 0.38
Advanced degree 0.06 0.24
etc.). This allows us to depict food consumption patterns of each
Don’t know 0.04 0.20
adolescent and relate it to that of his peer group. In addition, the
availability of friend nomination allows us to retrace school friends Father education
No high school degree 0.16 0.36
and thus construct friendship networks. To estimate the weight
High school/GED/vocational instead of high school 0.33 0.47
production function, we considered information from wave I, wave Some college/vocational after high school 0.17 0.37
II, wave III and wave IV. College 0.18 0.38
We exploit friends nominations to construct the network of Advanced degree 0.08 0.26
Don’t know 0.06 0.24
friends. Thus, we consider all nominated friends as network
Missing 0.02 0.16
members regardless of the reciprocity of the nomination. If an
adolescent nominates a friend then a link is assigned between Grade 7–8 0.11 0.32
Grade 9–10 0.27 0.44
these two adolescents (directed network with non symmetric
Grade 11–12 0.62 0.48
links).
Allowance per week 8.28 11.65

5.1. Descriptive statistics Observations 2355

a
Frequency (in days) of fast food restaurant visits in the past week.
In our social interactions equation, the dependent variable of
interest is fast food consumption, as approximated by the reported
frequency (in days) of fast food restaurant visits in the past 7 days.
Table 1 reports respectively the mean and the standard deviation
of the endogenous variable, the covariates used and other relevant the fast food variables for wave II, III and IV are detailed in
characteristics. We note that on average, adolescents’ fast food con- Table 4.25
sumption is within the range of 2.33 times/week. This is consistent
with the frequency reported by the Economic Research Service of
the United States Department of Agriculture. Around 62% of the
5.2. The construction of the graph matrix
adolescents consumed fast food twice or more in the past week
and 44% of the adolescents who had consumed fast food did so 3
We construct a sub-matrix of graph for each school separately
times in the past week.
(matrice Gl ) and then we include all these sub-matrices in the block-
The covariates of the fast food peer effect equation include
diagonal matrix G.26 As we have no prior information about how
the adolescent’s personal characteristics, family characteristics
social interaction takes place, we assume, as in most studies, that
as well as the corresponding contextual social effects. The per-
an adolescent is equally influenced by his nominated friends. In
sonal characteristics are gender, age, ethnicity (white or other)
each school we eliminate adolescents for which we have missing
and grade. We observe that 50% of the sample are females,
values. As mentioned earlier, Bramoullé et al. (2009) show that
that the mean age is 16.3 years and that 57% are white. Fam-
the structural parameters are identified if the matrices I, G, G2
ily characteristics are dummies for mother and father education.
and G3 are linearly independent. This condition is verified with
We observe that around 45% of mothers and fathers have at
our data. We also compute the Belsley, Kuh, and Welsch condi-
least some college education. To control further for parents’
tion index to check for the presence of collinearity between these
income we use child allowance as a proxy. An adolescent’s
matrices. If this index is below 30, then collinearity is said not to
allowance is on average 8.28$ per week, around 50% of the
be a problem and linear independence of the four matrices is veri-
adolescents in our sample have a weekly allowance. At this
fied. In our data, the reflection problem is clearly solved since I, G,
point, it is important to highlight that since we use cross sec-
G2 and G3 are linearly independent and the condition index value
tion data, we do not have to control for fast food prices as
is 2.21.
they are taken into account by network fixed effects. As for the
weight production function, the dependent variable that we use
is zBMI in waves III and IV.24 The zBMI variables for each wave,
25
It is important to note that information on fast food consumption was not col-
lected in wave I.
23 26
It includes all meals that are consumed at a fast food restaurant such as McDon- Following the previous literature and given the lack of information on this mat-
ald’s, Burger King, Pizza Hut, Taco Bell and other fast food outlets. ter, we assume that there can be social interactions within each school but no
24
To compute the backward mean we used all four waves. interactions across schools.
132 B. Fortin, M. Yazbeck / Journal of Health Economics 42 (2015) 125–138
6. Results
6.1. Baseline: OLS peer effects estimates
We first estimate a naive OLS of the peer effects equation where
we regress the fast food consumption of an adolescent on the
average fast food consumption of his peers, his individual char-
acteristics as well as the average characteristics of his peers. We
then apply a panel-like within transformation to account for corre-
lated effects (OLS w ). It is clear that the estimates of naive OLS and
OLS w are inconsistent. The former ignores both correlated effects
and simultaneity problems while the latter ignores simultaneity
problems. However, they are reported to provide a baseline for this
study.
Estimation results reported in Table 2 show that there is a posi-
tive significant peer influence in fast food consumption. According
to the naive OLS estimates, an adolescent would increase his weekly
frequency (in days) of fast food restaurant visits by 0.21 in response
to an extra day of fast food restaurant visits by his friends. On
average, this corresponds to an increase of 9% (=0.21/2.33). OLS w
estimate is slightly lower (=0.15, or 6.6%). This reduction in the
estimated effect may partly be explained by the fact that adoles-
cents in the same reference group tend to choose a similar level
of fast food consumption partly because they face a common envi-
ronment or because adolescents with similar characteristics tend
to attend the same school (homophily). As for the individual char-
acteristics, age, father education and weekly allowance positively
affect fast food consumption. Turning our attention to the contex-
tual peer effects, we notice that the latter variable decreases with
mean peers’ mother’s education and increases with mean peers’
father’s education. The former result indicates that friends’ mother
education negatively affects an adolescent’s fast food consump-
tion.
6.2. GS-2SLS and QML peer effects estimates
Next, we estimate our peer effects equation with school fixed
effects using GS-2SLS (with i.i.d. error terms and without impos-
ing autoregressive disturbances:  = 0). We then estimate this
equation using a QML approach with i.i.d. error terms. Also, we esti-
mate another plausible version of this model by allowing network
autoregressive disturbances (GSARAR model).
Estimation results are displayed in Table 3. The GS-2SLS
approach (see last two columns) assumes that the instrument for
∗ effects in eating habits due to data limitations.
G∗ y∗ is given by G∗ ŷ (see Eq. 14).27 One can check whether this
∗ As for estimated individual effects and focusing on the GSAR
instrument is weak by regressing G∗ y∗ on G∗ ŷ , X∗ and G∗ X∗ and
performing a Stock-Yogo test (see Table 3). It consists in comparing
the Cragg-Donald F statistic associated with the estimated coeffi-
∗ cation as well as positively associated with weekly allowance.
cient of G∗ ŷ (= 17.80) with its critical value when one assumes a
10% tolerance28 for the size distortion of the 5% Wald test (=16.38).
Based on this test, we reject that the instrument is weak. The
endogenous effect resulting from GS-2SLS estimation is positive
(=0.11 or 4.73%) but non significant. When using the GSAR QML
approach, estimation results show a positive endogenous effect
of 0.129 (or 5.3%). This estimate reveals to be very close to the
one obtained by GS-2SLS and is slightly smaller than the OLS ones
obtained in the previous sub-section. It is statistically significant at
the 5% level if we perform a one tail test (one-tail p value = 0.039)
but at 10% if we consider a two-tail test (two tail p value = 0.0785).
27
The star superscript indicates that the original variable has been transformed
to eliminate the problem of singular variance matrix generated by the use of the
within transformation to eliminate fixed network effects. See Appendix.
28
This level of tolerance is the smallest one that can be computed given that there
is only one excluded instrument.
While this result suggests that one has to be quite cautious when
accepting the estimate, it can be argued that one should perform
a one-tail test since one expects the endogenous peer effect to be
either positive or zero. In that case, the social multiplier associated
with an exogenous increase in an adolescent fast food consump-
1
tion is 1.15 (= 1−0.129 ) and is significantly different from 1 at the
10% level, based on a one-tail test (its standard error is 0.096 using
the delta method, with a one-tail p-value of 0.059). This reflects a
relatively low endogenous peer effect.
How can we compare these results to those obtained previ-
ously in the related literature? Although there are few studies that
investigated the presence of peer effects in fast food consumption
using the linear-in-means equation, a richer body of literature has
investigated a tangent issue: obesity. As compared with endoge-
nous effects obtained in the literature on obesity, our peer effect is
intermediate between studies that obtain no peer effects (Cohen-
Cole and Fletcher, 2008b) and the literature that provides evidence
that there are peer effects are strong, for instance, generating a
social multiplier larger than 1.5 (e.g., Christakis and Fowler, 2007;
Trogdon et al., 2008).29
To check the sensitivity of these results to the presence of SAR
disturbances, we also estimate our model using a GSARAR QML
specification. The estimated spatial autocorrelation coefficient is
negative but not significant at the 5% level. Moreover the endoge-
nous peer effect is large (=0.3655) but no longer significant even at
the 10% level (one-tail test). Also a likelihood test does not reject
the GSAR QML specification. Therefore we consider the latter as
our preferred one. This suggests a much lower endogenous peer
effect (=0.13), which can be interpreted as a lower bound to this
parameter, at least when assuming that selection on unobserv-
ables is not an important source of biases, after controlling for
network fixed effects and observable characteristics (see our dis-
cussion above).
To sum up, we can say that results in general are consistent with
the hypothesis that fast food consumption is linked to issues of
interactions with friends. However, our social multiplier estimate
does not appear to be very strong (as the endogenous effects are less
than 0.3), at least when we consider a specification which seems
reasonable. This result, despite its small magnitude, addresses the
puzzle around the behavioural channels through which peer effects
in weight gain flows. Indeed, while Yakusheva et al. (2014) in their
attempt to uncover the channels through which these effects flow
have tested for two behavioural channels exercise and eating dis-
orders (e.g., anorexia), they could not test for the presence of peer
QML specification, they follow fairly the baseline model. Fast food
consumption is positively associated with age and father’s edu-
Mother’s education seems to have a negative but non-significant
impact on fast food consumption. It is important to note that while
the general perception is that fast food is an inferior good, the
empirical evidence suggests a positive income elasticity (Aguiar
and Hurst, 2005). The positive relation between fast food consump-
tion and allowance is thus in line with the positive relation between
income and fast food consumption.
One advantage of our spatial approach is that it allows to iden-
tify both endogenous and contextual peer effects. Turning our
attention to the latter, we note in particular that an adolescent’s
fast food consumption decreases with peers’ mother’s education
29
More specifically, Cohen-Cole and Fletcher (2008b) finds a statistically insignif-
icant social multiplier of 1.03, Christakis and Fowler (2007) find a statistically
significant social multiplier of 2.63 and Trogdon et al. (2008) find a statistically
significant social multiplier of 2.08.
 B. Fortin, M. Yazbeck / Journal of Health Economics 42 (2015) 125–138 133

Table 2
Peer effects in fast food consumption.

OLS OLSw

Coef. S.E. Coef. S.E.

Endogenous peer effects 0.2078 *** 0.0331 0.1548 *** 0.0344

Individual characteristics
Female −0.0721 0.0787 −0.0847 0.0789
Age 0.1559 *** 0.0434 0.1315 *** 0.0461
White −0.1076 0.0940 −0.0602 0.1127
Mother present −0.0152 0.0997 −0.0358 0.0989

Mother no high school (omitted)

Mother high school −0.0848 0.1195 −0.0455 0.1202
Mother some college −0.0377 0.1335 −0.0210 0.1340
Mother college 0.0214 0.1421 −0.0137 0.1425
Mother advanced −0.0259 0.1875 −0.0353 0.1877
Mother don’t know −0.1714 0.2067 −0.2124 0.2059

Father no high school (omitted)

Father high school 0.2743 ** 0.2067 0.2682 ** 0.1167
Father some college 0.2117 0.2067 0.1971 0.1338
Father college 0.3115 ** 0.1375 0.2592 * 0.1381
Father advanced 0.1732 0.1752 0.1294 0.1760
Father don’t know 0.2778 0.1756 0.2393 0.1750
Father missing 0.0908 0.2338 0.0477 0.2331

Grade 7–8 (Omitted)

Grade 9–10 0.0883 0.1931 −0.0776 0.2183
Grade 11–12 0.3164 0.2265 0.1269 0.2526
Allowance per week 0.0093 *** 0.0031 0.0074 ** 0.0031

Contextual peer effects

Female −0.0898 0.1245 −0.1071 0.1285
Age −0.0321 0.0215 0.0316 0.0718
White 0.0111 0.1244 −0.0055 0.1694
Mother present 0.0773 0.1668 0.1008 0.1707

Mother no high school (omitted)

Mother high school −0.3878 ** 0.1868 −0.2977 0.1913
Mother some college −0.3947 * 0.2127 −0.3825 * 0.2168
Mother college −0.2531 0.2180 −0.2935 0.2213
Mother advanced −0.7011 ** 0.3089 −0.5954 * 0.3112
Mother don’t know −0.4337 0.3598 −0.4150 0.3610

Father no high school (omitted)

Father high school 0.2060 0.1943 0.2999 0.1914
Father some college 0.3639 * 0.2128 0.3890 * 0.2139
Father college 0.2850 0.2238 0.3068 0.2263
Father advanced 0.2760 0.2891 0.2171 0.2953
Father don’t know 0.4737 0.2995 0.5358 * 0.3001
Father missing 0.6931 0.4619 0.7692 * 0.4640

Grade 7–8 (Omitted)

Grade 9–10 −0.0769 0.2383 0.0104 0.2773
Grade 11–12 −0.0094 0.2630 −0.0396 0.3388

Allowance per week 0.0056 ** 0.0053 0.0043 0.0054

Constant −0.5199 0.6618

N = 2339
*
Significant at 10% level.
**
Significant at 5% level.
***
Significant at 1% level.

but increases with mean peers’ father’s education. While the for-
mer causal effect seems natural as mothers with higher education
may (directly and indirectly) encourage both their children and
their friends to have better eating habits, the latter effect is rather
puzzling. One partial explanation is that fathers with higher edu-
cation are more likely to be absent from home. Therefore they
have less positive influence on their children’s and friends’ eating
habits.
6.3. Weight production function estimates
Estimation results presented in the earlier sections are consis-
tent with the presence of peer effects in fast food consumption.
Nevertheless, we still need to provide evidence of the presence of
a relationship between fast food consumption and weight gain. In
this section we report estimates of the weight production function
presented earlier.
Results from the estimation of the production function are
reported in Table 5. Specification (1) shows baseline OLS estimates
of equation (8) (where t is replaced by m), specification (2) shows
the NLS estimates of Eq. (9), specification (3) shows E-NL-IV-C esti-
mation results for Eq. (9) and finally specification (4) shows GMM
version the previous estimator using a two step with an optimal
weighting matrix. All specifications are estimated using wave 3 and
wave 4, but where information from wave 1 and wave 2 are used
to construct the instruments.
134 B. Fortin, M. Yazbeck / Journal of Health Economics 42 (2015) 125–138

Table 3
Peer effects in fast food consumption GSAR, GSARAR and GS-2SLS.

Quasi maximum likelihood GS-2SLS

GSAR S.E. GSARAR S.E. GSAR S.E.

†
Endogenous peer effects 0.1293 * 0.0729 0.3511 3.3869 0.1102 0.3945
 −0.2342 3.6954

Individual characteristics
Female −0.0787 0.0589 −0.0861 0.0832 −0.0839 0.0780
Age 0.1401 *** 0.0479 0.1477 0.1386 0.1346 ** 0.0531
White −0.0622 0.0795 −0.0582 0.2083 −0.0619 0.1169
Mother present −0.0319 0.0760 −0.0278 0.1097 −0.0375 0.0973

Mother no high school (omitted)

Mother high school −0.0333 0.0510 −0.0549 0.3647 −0.0437 0.1212
Mother some college −0.0106 0.1156 −0.0258 0.2193 −0.0161 0.1401
Mother college 0.0040 0.1043 −0.0025 0.1384 −0.0143 0.1449
Mother advanced −0.0159 0.1737 −0.0555 0.5745 −0.0366 0.1873
Mother don’t know −0.2193 *** 0.0813 −0.2327 ** 0.1101 −0.2138 0.2098

Father no high school (omitted)

Father high school 0.2774 ** 0.1165 0.2680 0.4137 0.2689 ** 0.1179
Father some college 0.2027 ** 0.0889 0.1923 0.4408 0.1957 0.1370
Father college 0.2775 *** 0.0837 0.2800* 0.1540 0.2577 * 0.1364
Father advanced 0.1336 0.1821 0.1418 0.1918 0.1275 0.1785
Father don’t know 0.2512 *** 0.0917 0.2519 0.2648 0.2420 0.1742
Father missing 0.0548 0.1302 0.0721 0.1264 0.0515 0.2308

Grade 7–8 (omitted)

Grade 9–10 −0.1499 0.2569 −0.1444 0.2677 −0.0789 0.1957
Grade 11–12 0.0225 0.2706 0.0329 0.3268 0.1249 0.2300
Allowance per week 0.0076 *** 0.0016 0.0076 *** 0.0016 0.0075 ** 0.0032

Contextual peer effects

Female −0.1563 0.1049 −0.1721 0.1178 −0.1108 0.1323
Age −0.0379 ** 0.0152 0.0158 0.5929 0.0359 0.0813
White 0.0079 0.1064 0.0246 0.4373 −0.0159 0.1865
Mother present 0.0651 0.2354 0.0955 0.6757 0.1078 0.1794

Mother no high school (omitted)

Mother high school −0.3181 ** 0.1327 −0.4149 0.6549 −0.3001 0.1858
Mother some college −0.4254 *** 0.1534 −0.4448 0.2741 −0.3882 * 0.2213
Mother college −0.3443 0.2255 −0.3462 0.3121 −0.3080 0.2496
Mother advanced −0.6565 ** 0.3032 −0.6871 0.9187 −0.5775 * 0.3407
Mother don’t know −0.4664 * 0.2508 −0.5488 * 0.2901 −0.4038 0.3604

Father no high school (omitted)

Father high school 0.3197 *** 0.1211 0.3074 0.4624 0.3290 0.3128
Father some college 0.3890 *** 0.1161 0.3476 0.5849 0.4051 * 0.2440
Father college 0.3214 0.2128 0.3354 0.2238 0.3298 0.3035
Father advanced 0.1765 0.2179 0.2330 0.5672 0.2341 0.3299
Father don’t know 0.5544 *** 0.1172 0.5478 0.7167 0.5684 0.3985
Father missing 0.7770 ** 0.3058 0.7783 0.5355 0.7769 * 0.4461

Grade 7–8 (omitted)

Grade 9–10 0.1998 0.2873 0.1900 0.4396 0.0058 0.2520
Grade 11–12 0.3272 0.2311 0.384 0.3932 −0.0341 0.3195

Allowance per week 0.0025 0.0033 0.0047 0.0237 0.0048 0.0072

Log Likelihood −4488.847 −4487.55

N = 2339
Stock-Yogo test 17.80
Critical value (r = 0.10)
at sign. level of 0.05% 16.38
†
One tail test: Significant at 5% level.
*
Two tail test: Significant at 10% level.
**
Two tail test: Significant at 5%.
***
Two tail test: Significant at 1%.

In line with our expectations, the general results indicate that
lagged zBMI and current fast food consumption have positive sig-
nificant effect (which is between 0 and 1, in the case of the lagged
zBMI) on current zBMI. These results seem to be robust across dif-
ferent specifications with some differences that can be explained
by the differences in the assumptions made on the DGP. More
specifically results in specification (1), our baseline specification,
are comparable to previous findings by Niemeier et al. (2006) who
used the same data set. The impact of lagged zBMI is 0.7591 (com-
pared to 0.7600 for Niemeier et al. (2006)). As for the impact of fast
food consumption, it is 0.014 (compared to 0.020 for Niemeier et al.
(2006)). 30
30
It is important to note that Niemeier et al. (2006) used a different wave and a
different approach.
 B. Fortin, M. Yazbeck / Journal of Health Economics 42 (2015) 125–138 135

Comparing specification (1) with specification (2), we notice Table 4

Descriptive statistics-production function.
that the estimate of the fast food consumption marginal effect is
much higher in the former than in the latter case (0.0145 vs. 0.0039). Variable Mean S.D.
The basic explanation is that when estimating the parameter 2 in zBMI wave I 0.4017 1.0178
the specification (1), we are not accounting for missing data on zBMI wave II 0.4485 1.0210
f
yit in time intervals between m and m − 1. In fact, we are estimat- zBMI wave III 0.7279 1.0903
g g zBMI wave IV 1.0173 0.9712
1− m 1− m
ing 2 ( 1−1 ), with ( 1−1 ) > 1, instead of estimating 2 . Also the Fast food wave II 2.3869 1.7810
1 1
Fast food wave III 2.6206 2.0685
estimate associated with the lagged endogenous variable is smaller
Fast food wave IV 2.2602 2.0790
in specification (1) than in specification (2) (i.e., 0.7591 vs. 0.9545). Agewave II 16.5741 1.5674
The explanation is that in the OLS specification we are implicitly Age wave III 22.0200 1.5613
estimating 1gm with |1 | < 1 as we are ignoring the irregularly time Age wave IV 29.1677 1.5417
Female 0.5156 0.4999
intervals.
White 0.6250 0.4842
While specification (2) accounts for the unequally spaced inter-
vals, it does not account for the correlation between lagged zBMI Obs. 1848
and the time-invariant unobserved effect and the idiosyncratic
error term of zBMI. Further it does take into account the fact that fast
food consumption may be correlated with the time-invariant unob-
served effect. To account for these possibilities, we follow MM and that lagged zBMI has a positive significant effect on current zBMI
instrument zBMI using the OLS residuals of the regression lagged level (=0.8447). This suggests that an exogenous shock on weight
zBMI on its backward mean. As for the fast food consumption, it has a stronger effect in the long term than in the short term. Based
is instrumented using Hausman and Taylor (1981) type of instru- on specification (4), an extra day of fast food restaurant visit per
ments, that is, by taking the deviations of fast food consumption 0.0031
week increases zBMI by 0.02 (= 1−0.8447 ) zBMI points or 4.45% in
with respect to individual sample mean as an instrument. the long term. The presence of a causal link between fast food
The estimation of specification (3) shows results that are consis- consumption and zBMI does not come as a surprise since pre-
tent with previous results, with some differences in the magnitude vious findings have been pointing in this direction (e.g., Levitsky
of the parameters. The estimated parameter associated with zBMI et al., 2004; Niemeier et al., 2006; Rosenheck, 2008). Combining
is smaller than the one obtained in the second specification (0.8461 the impact of fast food on weight gain with the social multiplier,
vs. 0.9545). As for the estimated parameter for fast food con- our results suggest that an extra day of fast food restaurant visits
sumption it is marginally higher when using E-NL-IV-C (0.0040 vs. per week leads to a zBMI increase of 0.0351 zBMI points, or 5.11%
0.0039). All results remain statistically significant. To complement (=4.45 % ×1.15) on average, as compared with 4.45% with no peer
results from the E-NL-IV-C we estimate an optimal GMM version effects. These results highlight a role for peer effects in fast food con-
(with a weighted matrix clustered at the the individual level) of sumption as one transmission mechanism through which weight
the same model. Estimation results are again within the expected gain is amplified.
lines and consistent with previous estimates with some variation As for the other covariates, age reveals to have a positive signif-
in the estimates of the lagged zBMI (0.8447 vs. 0.8461) and fast food icant effect on zBMI, an additional year increasing zBMI by 0.0072
consumption parameters (0.0039 vs. 0.0031). zBMI points. Also being female and white have a negative effect on
We retain specification (4) as our preferred one as it provide zBMI (respectively 0.0176 and 0.0153). To test for the relevance of
optimal and robust estimators. Results from this specification show the instruments we use Kleibergen-Paap rank test (which is used

Table 5
Weight production function.

Spec 1 Spec 2 Spec 3 Spec 4

OLS NLS E-NL-IV-C GMM (Adj. W)

Fast food consumption 0.0145*** 0.0039*** 0.0040** 0.0031**

(0.0050) (0.0009) (0.0020) (0.0014)
zBMI (lagged) 0.7591*** 0.9545*** 0.8461*** 0.8447***
(0.0115) (0.0024) (0.0186) (0.0146)
Age 0.0164*** 0.0031*** 0.0072*** 0.0072***
(0.0013) (0.0001) (0.0008) (0.0006)
Female = 1 −0.0189 −0.0022 −0.0188* −0.0176**
(0.0187) (0.0034) (0.0108) (0.0077)
White = 1 −0.0317* −0.0030 −0.0158 −0.0153**
(0.0190) (0.0034) (0.0107) (0.0077)

Obs. 3696 3696 3696 3696

Kleibergen Paap rk Wald F statistic 210.67 210.67

Stock-Yogo test Critical value (r = 0.10) at sign. level of 0.05% 7.03 7.03
Excluded instruments
Residual No No Yes Yes
Fast food de-meaned No No Yes Yes

All S.E. are robust and clustered at individual level.

GMM weighted matrix clustered at the individual level.
1848 observations per wave.
*
Significant at 10% level.
**
Significant at 5% level.
***
Significant at 1% level.
136 B. Fortin, M. Yazbeck / Journal of Health Economics 42 (2015) 125–138
here to take into account the fact that the estimated standard errors
are panel clustered). The test rejects the null hypothesis that the
instruments are weak.31
7. Conclusion
This paper investigates whether peer effects in adolescent
weight partly flow through the eating habits channel. We first
attempt to study the presence of significant endogenous peer
effects in fast food consumption. New methods based on spatial
econometric analysis are used to identify and estimate our model,
under the assumption that individuals interact through a friendship
social network. Our results indicate that an increase in his friends’
mean fast food consumption induces an adolescent to increase his
own fast food consumption. This peer effect amplifies through a
social multiplier the impact of any exogenous shock on fast food
consumption. However, our estimated social multiplier based on
our preferred (conservative) specification is small as it is equal to
1.15.
We also estimate a dynamic weight production function which
relates the individual’s Body Mass Index to his fast food consump-
tion. Our results reveal a positive significant impact of a change
in fast food consumption on the change in zBMI. Specifically, in
the long run, a one-unit increase in the weekly frequency (in days)
of fast food consumption produces an increase in zBMI by 4.45%.
This effect reaches 5.11% when the social multiplier is taken into
account. This suggests the presence of a positive but low endoge-
nous peer effect. In short, our results are intermediate between
studies on overweight or obesity that report no peer effects (e.g.,
Cohen-Cole and Fletcher, 2008a) and others that provide evidence
of strong peer effects (e.g., Trogdon et al., 2008; Christakis and
Fowler, 2007)
Coupled with the reduction in the relative price of fast food
and the increasing availability of fast food restaurants over time,
the social multiplier could somewhat increase the prevalence of
obesity in the years to come. Conversely, this multiplier may con-
tribute to the decline of the spread of obesity and the decrease
in health care costs, as long as it is exploited by policy mak-
ers through tax and subsidy reforms encouraging adequate eating
habits among adolescents, or used to implement network based
interventions to promote healthy eating behaviours (Fletcher et al.,
2011).
There are many possible extensions to this paper. From a
policy perspective, it would be interesting to investigate the pres-
ence of peer effects in physical activity of adolescents. A recent
study by Charness and Gneezy (2009) finds that there is room
for intervention in peoples’ decisions to perform physical exercise
through financial incentives. It would be thus valuable to investi-
gate whether there is a social multiplier that can be exploited to
amplify these effects. Furthermore, in the same way, it would be
interesting to study the presence of peer effects weight percep-
tions. So far, most of the peer effects work has focused mainly on
BMI outcomes. At the methodological level, a possible extension
would be to assume a Poisson or a Negative Binomial distribution
to account for the count nature of the consumption data at hand. As
far as we know, no work has been carried out in this area. Finally,
it would be most useful to develop a general approach that would
allow same sex and opposite sex peer effects to be different for both
males and females.
31
As the model is exactly identified, it is not possible to conduct an over identifi-
cation test.
Appendix A. Quasi maximum likelihood (QML) of the peer
effects model
Let us rewrite Eq. (7) for convenience:
Kl Ml yl = ˇKl Ml Gl yl + Kl Ml Xl  + Kl Ml Gl Xl ı + l .
The elimination of fixed network effects using a within trans-
formation leads to a singular variance matrix such that E(l l |
Xl , Gl ) = Kl Kl  2 = Kl  2 . To resolve this problem of linear depend-
ency between observations, we follow a suggestion by Lee et al.
(2010) and applied by Lin (2010). Let [Ql Cl ] be the orthonor-
mal matrix of Kl , where Ql corresponds to the eigenvalues of 1
and Cl to the eigenvalues of 0. The matrix Ql has the follow-
ing properties: Ql Ql = In∗ , Ql Ql = Kl and Ql  = 0, where n∗l = nl − 1
l
with nl being the number of adolescents in the lth network. Pre-
multiplying (7) by Ql , the structural model can now be written as
follows:
M∗l y∗l = ˇM∗l G∗l y∗l + M∗l X∗l  + M∗l G∗l X∗l ı + ∗l , (12)
where M∗l = Ql Ml Ql , y∗l = Ql yl , G∗l = Ql Gl Ql , X∗l = Ql Xl , and ∗l =
Ql l . With this transformation, our problem of dependency
between the observations is solved, since we have E(∗l ∗ l
| Xl , Gl ) =
 2 In∗ .
l
Assuming that ∗l is a n∗l -dimensional i.i.d normally distributed
disturbance vector, the log-likelihood function of (12) is given by:
−n∗  L
 L
ln L = ln(2 2 ) + ln |In∗ − ˇG∗l | + ln |In∗ − M∗l |
2 l l
l=1 l=1
1  ∗ ∗
L
− l l , (13)
2 2
l=1
L
where n∗ = n∗ = N − L, and, from (12), ∗l = M∗l (y∗l − ˇG∗l y∗l −
l=1 l
∗ ∗ ∗
Xl  − Gl Xl ı). Maximizing (13) with respect to (ˇ,   , ı , , ) yields
the maximum likelihood estimators of the model. Interestingly, the
QML method is implemented after the elimination of the network
fixed effects. Therefore, the estimators are not subject to the inci-
dental parameter problem that may arise since the number of fixed
effects increases with the size of the networks sample. To compute
robust standard errors, we use a sandwich form A−1 BA−1 , where A
is minus the expectation of the Hessian matrix and B is the expec-
tation of the outer product of the gradient matrix. An advantage of
this approach is that it allows us to obtain robust standard errors
that are not driven by the normality assumption that ML imposes
on the error term.
Appendix B. Generalized spatial two stage least squares
(GS-2SLS) of the peer effects model
To estimate the model (12), we also adopt a generalized spatial
two-stage least squares procedure presented in Lee et al. (2010).
This approach provides a simple and tractable numerical method
to obtain asymptotically efficient IV estimators within the class of
IV estimators. In the case of our paper this method will consist of
a two-step estimation.32 To simply the notation, Let X∗ be a block-
diagonal matrix with X∗ l on its diagonal, G∗ be a block-diagonal
matrix with G∗ l on its diagonal, and y∗ the concatenated vector of
the yl∗ ’s over all networks.
32
Note that for this particular case we impose  = 0 and thus Ml = Il .
 B. Fortin, M. Yazbeck / Journal of Health Economics 42 (2015) 125–138
∗
Now, let us denote by X̃ the matrix of explanatory variables
∗
such that X̃ = [G∗ y∗ X∗ G∗ X∗ ]. Let P be the weighting matrix
such that P = S(S S)−1 S , and S a matrix of instruments such that
S = X∗ G∗ X∗ G∗2 X∗ . In the first step, we estimate the following
2SLS estimator:
∗ ∗ −1 ∗
ˆ 1 = (X̃ PX̃ ) X̃ Py∗ ,
where  1 is the first-step 2SLS vector of estimated parameters
 
ˆ ) of the structural model. This estimator is consistent but
(ˆ 1 , ı̂1 , ˇ
1 Currie, J., DellaVigna, S., Moretti, E., Pathania, V., 2010. The effect of fast food restau-
not asymptotically efficient within the class of IV estimators.
Now, in the second step, we estimate a 2SLS using a new matrix
of instruments Ẑ given by:
 ∗
 Journal of Economic Perspectives 17, 93–118.
Ẑ = G∗ ŷ X∗ G∗ X∗ ,
∗ Dunn, R.A., Sharkey, J.R., Horel, S., 2012. The effect of fast-food availability on
where G∗ ŷ is computed from the first-step 2SLS reduced form (pre-
multiplied by G∗ ):
∗ −1
G∗ ŷ = G∗ (I − ˇ
ˆ 1 G∗ ) (X∗ ˆ 1 + G∗ X∗ ı̂1 ).
We then estimate:
 ∗ −1
 2 = (Ẑ X̃ ) Ẑy∗ .
This estimator can be shown to be consistent and asymptoti-
cally best IV estimator. Its asymptotic variance matrix is given
∗ ∗ −1
by N[Z X̃ R−1 X̃ Z] . The matrix R is consistently estimated by
 N 2 Hausman, J.A., Taylor, W.E., 1981. Panel data and unobservable individual effects.
R̂ = s2 ẐNẐ , where s2 = N −1 i=1 ûi and ûi are the residuals from
the second step. It is important to note that, as in Kelejian and
Prucha (1998), we assume that errors are homoscedastic. The esti-
mation theory developed by Kelejian and Prucha (1998) under the
assumption of homoscedastic errors does not apply if we assume
heteroscedastic errors (Kelejian and Prucha, 2010).
References
Aguiar, M., Hurst, E., 2005. Consumption versus expenditure. Journal of Political
Economy 113 (5), 919–948.
Ali, M.M., Amialchuk, A., Heiland, F.W., 2011. Weight-related behavior among ado-
lescents: the role of peer effects. PLoS ONE 6 (6), e21179.
Ali, M.M., Amialchuk, A., Renna, F., 2011. Social network and weight misperception
among adolescents. Southern Economic Journal 77 (4), 827–842.
Alviola, I.V., Nayga Jr., P.A., Thomsen, R.M.M.R., Danforth, D., Smartt, J., 2014. The
effect of fast-food restaurants on childhood obesity: a school level analysis.
Economics & Human Biology 12, 110–119.
Anderson, B., Lyon-Callo, S., Fussman, C., Imes, G., Rafferty, A.P., 2011. Peer reviewed:
fast-food consumption and obesity among Michigan adults. Preventing chronic
disease 8 (4).
Anderson, M., Matsa, D., 2011. Are restaurants really supersizing America? American
Economic Journal: Applied Economics 3 (1), 152–188.
Anderson, T., Hsiao, C., 1981. Estimation of dynamic models with error components.
Journal of the American Statistical Association 76, 598–606.
Arellano, M., Bond, S., 1991. Some tests of specification for panel data: Monte Carlo
evidence and an application to employment equations. Review of Economic
Studies 58 (2), 277–297.
Auld, M.C., Powell, L.M., 2009. Economics of food energy density and adolescent
body weight. Economica 76 (304), 719–740.
Badev, A., 2013. Discrete games in endogenous networks: Theory and policy, Work-
ing Papers 2-1-2013. University of Pennsylvania Scholarly Commons.
Berentzen, T., Petersen, L., Schnohr, P., Sørensen, T., 2008. Physical activity in
leisure-time is not associated with 10-year changes in waist circumference.
Scandinavian Journal of Medicine & Science in Sports 18 (6), 719–727.
Bleich, S., Cutler, D., Murray, C., Adams, A., 2008. Why is the developed world obese?
Annual Review of Public health 29, 273–295.
Blume, L., Brock, W., Durlauf, S., Jayaraman, R., 2015. Linear social interactions mod-
els. Journal of Political Economy 123 (2), 444–496.
Boucher, V., 2014. Conformism and Self-Selection in Social Networks. Mimeo.
Bramoullé, Y., Djebbari, H., Fortin, B., 2009. Identification of peer effects through
social networks. Journal of Econometrics 150 (1), 41–55.
Calabr, P., Golia, E., Maddaloni, V., Malvezzi, M., Casillo, B., Marotta, C., Calabrò,
R., Golino, P., 2009. Adipose tissue-mediated inflammation: the missing link
between obesity and cardiovascular disease? Internal and Emergency Medicine
4 (1), 25–34.
Calle, E., 2007. Obesity and cancer. British Medical Journal 335 (7630), 1107–1108.
Caraher, M., Cowburn, G., 2007. Taxing food: implications for public health nutrition.
Public Health Nutrition 8 (08), 1242–1249.
137
Charness, G., Gneezy, U., 2009. Incentives to exercise. Econometrica 77 (3), 909–931.
Christakis, N.A., Fowler, J.H., 2013. Social contagion theory: examining dynamic
social networks and human behavior. Statistics in Medicine 32 (4), 556–577.
Christakis, N., Fowler, J., 2007. The spread of obesity in a large social network over
32 years. New England Journal of Medicine 357 (4), 370–379.
Cliff, A., Ord, J., 1981. Spatial Processes: Models & Applications. Pion Ltd.
Cohen-Cole, E., Fletcher, J., 2008a. Detecting implausible social network effects in
acne height and headaches: longitudinal analysis. British Medical Journal 337,
a2533.
Cohen-Cole, E., Fletcher, J., 2008b. Is obesity contagious? Social networks vs. envi-
ronmental factors in the obesity epidemic. Journal of Health Economics 27 (5),
1143–1406.
rants on obesity and weight gain. American Economic Journal: Economic Policy
2, 34–65.
Cutler, D., Glaeser, E., Sphapiro, J., 2003. Why have Americans become more obese?
De la Haye, K., Robins, G., Mohr, P., Wilson, C., 2010. Obesity-related behaviors in
adolescent friendship networks. Social Networks 32 (3), 161–167.
fast-food consumption and obesity among rural residents: an analysis by
race/ethnicity. Economics & Human Biology 10 (1), 1–13.
Everaert, G., 2013. Orthogonal to backward mean transformation for dynamic panel
data models. Econometrics Journal 16, 179–221.
(14) Finkelstein, E.A., Trogdon, J.G., Cohen, J.W., Dietz, W., 2009. Annual medical spending
attributable to obesity: payer-and service-specific estimates. Health Affairs 28
(5), w822–w831.
Fletcher, A., Bonell, C., Sorhaindo, A., 2011. You are what your friends eat: system-
atic review of social network analyses of young people’s eating behaviours and
bodyweight. Journal of Epidemiology and Community Health, jech-2010.
Fowler, J., Christakis, N., 2008. Estimating peer effects on health in social networks.
Journal of Health Economics 27 (5), 1400–1405.
Goldsmith-Pinkham, P., Imbens, G.W., 2013. Social networks and the identification
of peer effects. Journal of Business and Economic Statistics 31 (3), 253–264.
Econometrica 49 (6), 1377–1398.
Hsieh, C.-S.-C., Lee, L.-F., 2011. A social interactions model with endogenous friend-
ship formation and selectivity, Working paper. Mimeo.
Kelejian, H., Prucha, I., 1998. A generalized spatial two-stage least squares procedure
for estimating a spatial autoregressive model with autoregressive disturbances.
Journal of Real Estate Finance and Economics 17 (1), 99–121.
Kelejian, H., Prucha, I., 2010. Specification and estimation of spatial autoregres-
sive models with autoregressive and heteroskedastic disturbances. Journal of
Econometrics 157, 53–67.
Lee, L., 2003. Best spatial two-stage least squares estimators for a spatial autore-
gressive model with autoregressive disturbances. Econometric Reviews 22 (4),
307–335.
Lee, L.-F., Liu, X., Lin, X., 2010. Specification and estimation of social interaction
models with networks structure. Econometrics Journal 13 (2), 143–176.
Levitsky, D., Halbmaier, C., Mrdjenovic, G., 2004. The freshman weight gain: a model
for the study of the epidemic of obesity. International Journal of Obesity 28 (11),
1435–1442.
Lin, X., 2010. Identifying peer effects in student academic achievement by spatial
autoregressive models with group unobservables. Journal of Labor Economics
28 (4), 825–860.
Liu, X., Patacchini, E., Rainone, E., 2013. The allocation of time in sleep: a social
network model with sampled data, Working Papers w162. Center For Policy
Research, The Maxwell School.
Lyons, R., 2011. The spread of evidence-poor medicine via flawed social-network
analysis. Statistics, Politics, and Policy 2 (1), 2.
Maggio, C., Pi-Sunyer, F., 2003. Obesity and type 2 diabetes. Endocrinology and
metabolism clinics of North America 32 (4), 805–822.
Manski, C.F., 1993. Identification of endogenous social effects: the reflection prob-
lem. Review of Economic Studies 60 (3), 531–542.
Millimet, D., McDonough, I., 2013. Dynamic panel data models with irregular spac-
ing: with applications to early childhood development, Working paper.
Niemeier, H., Raynor, H., Lloyd-Richardson, E., Rogers, M., Wing, R., 2006. Fast food
consumption and breakfast skipping: predictors of weight gain from adoles-
cence to adulthood in a nationally representative sample. Journal of Adolescent
Health 39 (6), 842–849.
Ogden, C.L., Carroll, M.D., Kit, B.K., Flegal, K.M., 2012. Prevalence of obesity and trends
in body mass index among us children and adolescents, 1999–2010. Journal of
the American Medical Association 307 (5), 483–490.
Powell, L., Bao, Y., 2009. Food prices access to food outlets and child weight out-
comes: a longitudinal analysis. Economics and Human Biology 7, 64–72.
Powell, L., Chriqui, J., Khan, T., Wada, R., Chaloupka, F., 2013. Assessing the poten-
tial effectiveness of food and beverage taxes and subsidies for improving public
health: a systematic review of prices, demand and body weight outcomes. Obe-
sity Reviews 14 (2), 110–128.
Powell, L.M., 2009. Fast food costs and adolescent body mass index: evidence from
panel data. Journal of Health Economics 28 (5), 963–970.
Renna, F., Grafova, I.B., Thakur, N., 2008. The effect of friends on adolescent body
weight. Economics and Human Biology 6 (3), 377–387.
Rosenheck, R., 2008. Fast food consumption and increased caloric intake: a system-
atic review of a trajectory towards weight gain and obesity risk. Obesity Reviews
9 (6), 535–547.
138 B. Fortin, M. Yazbeck / Journal of Health Economics 42 (2015) 125–138

Shalizi, C., Thomas, A., 2011. Homophily and contagion are generically confounded
in observational social network studies. Sociological Methods & Research 40 (2),
211.
Trogdon, J.G., Nonnemaker, J., Pais, J., 2008. Peer effects in adolescent overweight.
Journal of Health Economics 27 (5), 1388–1399.
VanderWeele, T., 2011. Sensitivity analysis for contagion effects in social networks.
Sociological Methods & Research 40 (2), 240.
Yakusheva, O., Kapinos, K.A., Eisenberg, D., 2014. Estimating heterogeneous
and hierarchical peer effects on body weight using roommate assign-
ments as a natural experiment. Journal of Human Resources 49 (1),
234–261.
Yakusheva, O., Kapinos, K., Weiss, M., 2011. Peer effects and the freshman 15:
evidence from a natural experiment. Economics & Human Biology 9 (2),
119–132.
 Journal of Health Economics 42 (2015) 139–150

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Unintended effects of reimbursement schedules in mental health care

Rudy Douven a,b,∗,∗∗ , Minke Remmerswaal a , Ilaria Mosca c
a
CPB Netherlands Bureau for Economic Policy Analysis, The Hague, The Netherlands
b
Erasmus University, Rotterdam, The Netherlands
c
Ecorys, Rotterdam, The Netherlands

a r t i c l e i n f o a b s t r a c t

Article history: We evaluate the introduction of a reimbursement schedule for self-employed mental health care
Received 10 December 2014 providers in the Netherlands in 2008. The reimbursement schedule follows a discontinuous discrete
Received in revised form 20 March 2015 step function—once the provider has passed a treatment duration threshold the fee is flat until a next
Accepted 28 March 2015
threshold is reached. We use administrative mental health care data of the total Dutch population from
Available online 6 April 2015
2008 to 2010. We find an “efficiency” effect: on the flat part of the fee schedule providers reduce treatment
duration by 2 to 7% compared to a control group. However, we also find unintended effects: providers
Keywords:
treat patients longer to reach a next threshold and obtain a higher fee. The data shows gaps and bunches
Mental health care
Provider payment
in the distribution function of treatment durations, just before and after a threshold. About 11 to 13% of
Regression discontinuity design treatments are shifted over a next threshold, resulting in a cost increase of approximately 7 to 9%.
Policy evaluation © 2015 Elsevier B.V. All rights reserved.
Regulated competition
The Netherlands

JEL classification:
I11
I12
I18

1. Introduction a case mix based reimbursement that we will review in Section 22 .

Before 2008, all mental health care in the Netherlands was
organized and funded in a national insurance scheme (Exceptional
Medical Expenses Act (AWBZ)). The AWBZ was paid for by income-
differentiated premiums raised through taxes and it provided long
term and mental health care for all citizens. Mental health care
providers were mainly funded with budgets. This changed in 2008,
when the Dutch government placed a part of mental health care,
the curative and acute mental health care, under the regime of reg-
ulated competition1 . The goal of this policy change was to improve
the efficiency in the sector by letting private insurers buy care on
behalf of their enrollees. Providers no longer receive budgets, but
∗ Corresponding author at: CPB Netherlands Bureau for Economic Policy Analysis,
PO Box 80510, 2508 GM The Hague, The Netherlands. Tel.: +31 703383377.
E-mail addresses: R.Douven@cpb.nl (R. Douven), M.Remmerswaal@cpb.nl
(M. Remmerswaal), Ilaria.Mosca@ecorys.com (I. Mosca).
∗∗
When this paper was written Douven was a Harkness Fellow 2013/2014 at Har-
vard Medical School, supported by the Commonwealth Fund, a private independent
foundation based in New York City. The views presented here are those of the author
and not necessarily those of The Commonwealth Fund, its directors, officers or staff.
1 2
Managed competition in the Dutch curative care sector was introduced in 2006
(Van de Ven and Schut, 2008).
Mason and Goddard (2009) review the international literature on
reimbursing mental health care providers and argue that case mix
based funding offers incentives for a range of objectives, includ-
ing improvements in efficiency, quality of care and patient choice.
They criticize the Dutch reimbursement schedule and state: “. . .it
[. . .] therefore does not appear to encourage early discharge. . .” and
“. . .could incentivize providers. . .to deliver medically unnecessary
treatments. . .”. Dutch policymakers also recognized that the reim-
bursement schedule in mental health care might create unintended
incentives (VWS, 2010; NZa, 2010). This research aims to quantify
these possible effects.
The design of a payment system is a complicated matter,
especially in mental health care. Uncertainty and variations in
treatments are likely to be great in the mental health care market
making the response of patients and providers to financial incen-
tives larger than in other areas of health care (Frank and McGuire,
2000). A large body of the literature in health economics estab-
lishes that health care providers respond to financial incentives
The case mix refers to the mix of different types of patients that are treated by
the provider.

http://dx.doi.org/10.1016/j.jhealeco.2015.03.008
0167-6296/© 2015 Elsevier B.V. All rights reserved.
140 R. Douven et al. / Journal of Health Economics 42 (2015) 139–150
(for excellent overviews see Chandra et al., 2012; McGuire, 2000).
Most empirical evidence concerns the US and shows that fee-for-
service payment provides incentives for overtreatment. Some of
the first papers on this topic are Epstein et al. (1986), Hickson
et al. (1987) and Stearns et al. (1992). Recently, in the Netherlands,
similar behavioral responses have also been reported since the
introduction of regulated competition in the Dutch hospital market
(Douven et al., 2015) and market for general practitioners (van Dijk
et al., 2013). Less research has been done on case mix based fund-
ing in the mental health care market (Mason and Goddard, 2009). In
the US, Jennison and Ellis (1987) found an 18% increase in the rate
of visits per mental health provider per month when they shifted
from a salaried basis to a fee-for-service basis. Rosenthal (2000)
has examined the effects of risk sharing with mental health care
providers. She found that providers that received a salary reduced
their number of visits by 20 to 25% compared to providers who
were still paid for each visit. Bellows and Halpin (2008) studied
the impact of Medicaid reimbursement on mental health quality
indicators and found evidence of upcoding of quality indicators to
increase reimbursement.
This is the first study to evaluate the introduction of a new
reimbursement schedule in mental health care in the Netherlands.
The reimbursement function follows a discontinuous discrete step
function —once the provider has passed a treatment duration
threshold the fee does not increase until a next threshold is reached.
We look at two effects: efficiency and unintended effects. Our study
shows that the unintended effects – i.e. providers treat patients
longer to reach a next threshold and obtain a higher fee – out-
weigh the efficiency effect—i.e. on the flat part of the fee schedule
providers treat patients shorter and prolong treatment only if
marginal benefits to patients outweigh marginal costs. We separate
out these two effects by using regression discontinuity design (see
e.g. Lee and Lemieux, 2010)3 . Providers’ behavior around discontin-
uous fee thresholds are most likely be explained by the change in
fee, and not by other contemporary factors such as medical quality,
treatment outcome, location or other unobserved factors. We use
a quasi-experimental design in which 10% of all mental health care
providers are paid according to the new reimbursement schedule,
while 90% of providers were not subject to the reform. This latter
group serves as a control group. We find an efficiency effect: we
estimate a reduction in treatment duration by 2 to 7% and lower
costs by 3 to 6% compared to a control group. However, we also
find unintended effects: in total, about 11 to 13% of treatments
are shifted to over a next threshold, resulting in a cost increase
of approximately 7 to 9%.
The outline of our paper is as follows. Section 2 provides a con-
cise overview of the Dutch mental health care system. Section 3
describes the economic theory relating to the new reimbursement
schedule. Section 4 describes the data and Section 5 presents the
estimation methods. Section 6 presents the results and Section 7
concludes.
2. The Dutch mental health care system
Although the mental health status of the Dutch population has
been roughly stable since 1975, the number of people that use pro-
fessional mental health services has increased with about 10% per
year from 535,000 patients in 2001 to about 1 million patients in
2009 (GGZ Nederland, 2010).
Dutch mental health care distinguishes between primary and
secondary care. Patients with mild mental disorders usually go to
3
In the analysis we use cross-sectional comparisons between the two groups
instead of a difference-in-difference model because there is no data available prior
to the reform in 2008.
primary care, which is provided by a general practitioner, psy-
chologist, psychotherapist or psychiatrist4 . Patients with a more
serious condition need specialized care and are referred to sec-
ondary care. Secondary care is split into curative care and long-term
care. Long-term care patients usually remain in an institution such
as a residence or other kind of mental health facility for longer than
a year. Our study focuses on patients who receive curative care.
They can receive care in an inpatient or outpatient setting and their
treatment does not last longer than a year.
The reform to regulated competition in 2008 required many
changes for providers, health insurers and regulators. The govern-
ment decided upon a transition period between 2008 and 2010, in
which health insurers became responsible for the services of mental
health care providers. However, during the transition period insur-
ers did not incur financial risk on providing mental health care5 .
Since 2008, providers are reimbursed on their case mix, called a DBC
(Diagnosis Treatment Combination). A DBC refers to the complete
treatment episode of a patient. It starts with the initial consulta-
tion and continues until the provider ends the treatment. Consider
for example a patient with mild depression that for ten months
receives each month an individual therapy for 60 min by a psy-
chotherapist (and no other form of medication or treatment). This
patient’s treatment can be coded with the following DBC: “Depres-
sion, 250 to 800 min, no medication” (DBC Onderhoud, 2013). If
a treatment episode lasts longer than one year, the DBC is closed
automatically. After that year a new DBC is opened. With the closed
DBC a provider can receive reimbursement from his patient’s health
care insurer. The fee covers all labor and capital costs related to
the treatment episode. The reimbursement fee for a DBC was fixed
during our period of study and set prospectively by the Dutch
Healthcare Authority (NZa). Patients’ out-of pocket payments were
limited6 .
Most mental health providers worked in large regional institu-
tions in the period under consideration. These institutions can be
a regional facility for ambulatory care, but also a specialized psy-
chiatric hospital. Often, many different types of mental health care
specialists work together. Their payment was before (and after)
2008 still based on annual budgets. These budgets were based on
expected case mix and several regional budget parameters (such
as inflation, wages, capital costs etc.). Mental health care specialists
who work at a budgeted institution received a fixed salary7 . Negoti-
ations with health care insurers only took place with the dominant
health insurer in the geographical region between 2008 and 2010.
These mostly large mental health care institutions account for
about 90% of the sector (NZa, 2012). Henceforth, we will use these
‘budgeted’ or B providers in our study as a control group because
their individual salaries during 2008–2010 were not related to the
new reimbursement schedule.
About 10% of the mental health care providers choose to work
independently, e.g. private practices. Only this group of self-
employed providers, and new providers that entered the market
after January 1st of 2008, received their income according to
the new reimbursement schedule. Contrary to B providers, the
4
As of 2008, groups of practice nurses, social workers and psychologists (named
POH-GGZ) entered the market to support general practitioners.
5
Health insurers had therefore no financial incentives to control costs. The pol-
icy was that first a proper risk adjustment system should be implemented before
health insurers could bear more financial risks. In 2013, DBC-fees became subject
to negotiation between insurers and providers. To stimulate efficiency, the govern-
ment started programs to develop quality indicators in mental health care. In 2013,
a critical report (Rekenkamer, 2013) concluded that the stability and quality of most
indicators is poor and needs to be improved.
6
There was a mandatory annual deductible of 150 euro (in 2008) to 165 euro
(in 2010) for all curative services (including mental health services) except general
practitioner care and obstetrics.
7
The government made agreement with labor unions about these salaries.
 R. Douven et al. / Journal of Health Economics 42 (2015) 139–150
Fig. 1. Reimbursement schedule for the specialty ‘depression’.
self-employed had to negotiate with all health care insurers in
the period concerned. The focus of self-employed providers was
initially mostly on new, innovative segments of the mental care
market, such as addiction clinics, youth mental health, and a combi-
nation of services on work and mental health recovery (NZa, 2011).
This group of self-employed providers will be our treatment group
and, henceforth, we will call these providers ‘non-budgeted’ or NB
providers.
To obtain mental health services patients need a referral from a
general practitioner. After a referral patients are in principle free to
choose any mental health care provider; in practice however they
will often follow the advice of their general practitioner. Although
B and NB providers tend to specialize in certain mental health con-
ditions (see Section 4), we assume that for a given mental health
condition there are on average no differences between NB and B
providers in the types of patients they treat. This assumption is
likely to hold true as we analyze many mental health conditions
that are offered by both types of providers.
3. The economic theory of bunching of treatment durations
In this section we will explain in more detail the new reimburse-
ment schedule and how we separate out efficiency and unintended
effects. Treatment duration forms the basis of the size of the fee
in the new schedule, and is calculated as a weighted sum of sev-
eral components, i.e. several activities8 . Individual contact by the
provider with the patient receives the highest weight. It can be a
consult, intake or therapy session. Lower weighted components are
the time that a patient spends on other organized activities, such
as group therapy session, and the number of days that a patient
stays overnight in an institution. If for example, a patient receives
eight therapy sessions of 1 h (duration is 480 min), 10 h of group
sessions (weighted as a total duration of 150 min) and three days
in a residence (weighted as 180 min), it accounts for total treatment
duration of 810 min. Fig. 1 shows the reimbursement schedule for
DBC category ‘depression’. The X-axis shows the different classes of
treatment duration. All DBC categories, in all specialties of mental
health care, have the same treatment duration thresholds: at 250,
800, 1800, 3000, 6000 min9 . The Y-axis shows the corresponding
fees. They are unweighted averages for the years 2008–2010. Fig. 1
8
The main activities that shape a DBC are: diagnostics and treatment, daytime
activities, residence in an institution, and (medical) activities. Each one of these
activities is further split into other activities. To each one of this activity belongs a
duration in minutes and a corresponding tariff.
9
Thresholds occur also at 12,000, 18,000 and 24,000 min but we capped the
duration time at 7000 min because such long treatment durations were rare.
141
shows that the reimbursement schedule is a discrete step function,
in which fees are flat and only increase after a threshold is reached.
The fees at each duration threshold slightly differ across specialties
(e.g. depression, anxiety disorders etc.). Only the specialty ‘other
childhood disorders’ has higher fees for treatments with more than
3000 min (NZa, 2007–2009)10 .
The idea of the step function is a combination between a
prospective fee per episode of care (the flat part of the reimburse-
ment schedule) and fee-for-service (fees increase after a threshold
has been reached). Prospective fees create incentives for providers
to limit treatment duration which may lead to an efficient provi-
sion of care (McGuire, 2000). However, if the prospective fee is not
adjusted for the severity of the patient then a provider has a poten-
tial incentive to select less severe patients, i.e. patients that need
short treatment duration. The idea of the step function is to prevent
such selection incentives by setting higher fees for more severe
patients. However, setting higher fees exacerbate fee-for-service
incentives like overprovision of care around thresholds. For exam-
ple, the reimbursement fee for a treatment duration of 2900 min
is 3703 euro, while a prolongation of the treatment with 100 min
yields 6374 euro. This is a small difference in terms of treatment
duration but a large difference in financial reward. The reimburse-
ment schedule in Fig. 1 may result in overprovision or “bunching”
of treatments at thresholds.
In line with Ellis and McGuire (1986, 1990), referred to as E&M
from here on, we formulate a utility function of provider j for pro-
viding patient i with health severity  i a treatment duration xi . The
function is composed of two parts: benefits Bi to the patient and
profits i for the provider.
 
Uij = Bi xi , i + ˛j i (xi ) (1)
As in E&M, the agency parameter ˛j describes the extent to
which a provider weights the benefits to the patient relative to
its own profits. For example, an entrepreneurial provider may
attribute a higher ˛j to profits. For the benefits to the patient Bi (xi ,
 
 i ) we make the standard assumptions ∂Bi xi , i /∂xi > 0 at xi = 0
2   2
and ∂ Bi xi , i /∂ xi < 0, indicating that at the start of the treat-
ment there is a positive benefit to the patient and the marginal
benefit to the patient declines as treatment duration increases. We
model the profit function i (xi ) for a NB provider in (1) as follows:
i (xi ) = P (xi ) − cxi with P (xi ) = Pl for kl ≤ xi < kl+1 (2)
where kl represents the treatment duration thresholds, with l = 1,
. . .,5, and k1 = 250, k2 = 800, k3 = 1800, k4 = 3000, k5 = 6000 min. P(xi )
is the flat fee rate for a treatment duration xi . For example, in Fig. 1,
P(350) = 1038 and P(1000) = 2050 euro11 . Provider costs are repre-
sented by a simple linear cost function cxi and indicate production
costs as well as indirect costs such as foregone leisure time. Note
that the profit function i (xi ) is discontinuous at a threshold xi = kl
In line with E&M we assume that a provider maximizing its utility
solves the problem:
 
max xi Bi xi , i + ˛j (P (xi ) − cxi ) (3)
Thus, given a patient’s severity  i , and provider agency type ˛j ,
a provider will choose a treatment duration xi that solves the max-
imization problem in (3). Solving (3) returns that marginal benefits
equal marginal costs:
 
∂Bi xi , i
= ˛j c, with discontinuities at xi = kl . (4)
∂xi
10
No major changes occurred in the reimbursement schedule throughout the
studied period.
11
These are the reimbursement fees for depression.
142 R. Douven et al. / Journal of Health Economics 42 (2015) 139–150
Fig. 2. Bunching at treatment duration thresholds.
In Fig. 2, we illustrate that solving this optimization prob-
12
lem results in bunching at treatment duration thresholds   kl .
We plot the various marginal benefit functions ∂Bi xi , i /∂xi
and the marginal loss line ˛j c, which is discontinuous at thresh-
olds k1 and k2 . We observe a spike at both treatment duration
thresholds because reaching such a threshold implies that the
provider receives a higher reimbursement fee (or bonus). The size
of both spikes depends on the fee difference before and after the
threshold13 . When the marginal benefit function of the patient with
severity  1 crosses the marginal loss line ˛j c in Fig. 2 the provider
will not end its treatment but prolong treatment until k1 because
its utility is maximized at the threshold k1 . A similar reasoning
applies to the marginal benefit function  2 , the provider prolongs
treatment until k2 .
The result is bunching. The distribution of treatment durations
will exhibit gaps before treatment duration thresholds. These gaps
are expected to be larger for treatment durations closer to k1 and
k2 14 .
The reimbursement schedule provides  also incentives for effi-
ciency. For example, if ˛ j = 1 then ∂Bi xi , i /∂xi = c, and all dots
on the marginal loss line ˛j c in Fig. 2 correspond with socially opti-
mal treatment durations, where marginal benefit to the patient
equals marginal costs (McGuire, 2000). In the case of ˛j = 1, bunch-
ing implies overtreatment. If ˛j > 1 all dots on the marginal loss line
˛j c correspond with under treatment. Bunching implies that some
treatment durations are prolonged and become closer to the cost
efficient duration (although some overshooting may also happen).
Similarly, if ˛j < 1, there is overtreatment and bunching implies
even more overtreatment.
Important for our estimation procedure is the notion that
there is only a financial incentive to prolong, and not to shorten,
12
Bajari et al. (2011) perform a similar analyses with figures.
13
The size of the spike has to be determined empirically. Around kl , locally holds
∂Pi (xi ) /∂xi = −∞, implying an infinite spike. However, in practice the decision to
prolong treatment is more discrete in nature. For substantial shorter treatment dura-
tions than at thresholds kl , the provider has to trade off the costs associated with
treating the patient longer versus the size of the fee difference equal to Pl+1 − Pl .
14
Suppose treatment duration is at a local optimum. The farther away this treat-
ment duration is from a threshold duration k the more costly it will be for a provider
to move to the threshold k.
treatment durations. For example, a provider that hits a treatment
duration threshold will not end the treatment but will prolong
treatment as long as marginal benefits to the patient outweigh
marginal costs15 . Now, consider our comparison group, the B
providers who receive a fixed salary. Compared to NB providers, we
expect no bunching at treatment duration thresholds kl because B
providers face no particular financial consequences around these
thresholds. We make a general  assumption
 about the behavior of
B providers, namely that ∂Bi xi , i /∂xi = d, where d is a constant.
The incentive structure may differ between B and NB providers.
B and NB providers. B providers are paid a fixed amount for a
fixed period of time. Salaried providers have no incentives to
deliver unnecessary services, nor an incentive for “underprovision”
except to the degree that providers may “shirk” under salaried
arrangements. That means that they may attempt to provide fewer
services –shorten treatment duration− than under other con-
tractual arrangements (Christianson and Conrad, 2011). Under a
prospective fee shirking is less of a problem because production of
NB providers is directly related to their income. Although, keep-
ing treatment durations short allows them to treat more patients
in a given time frame, resulting in more income. Moreover, B and
NB providers may differ in how they are confronted with costs.
For example, salaried providers may put less weight on costs than
NB providers because their institution covers partly these costs. In
the extreme case, they face no costs (d = 0) and salaried providers
care only about patient benefits and not about costs. We take an
agnostic approach here and let the data decide whether treatment
duration differs between B and NB providers. There can be three dif-
ferent outcomes in Fig. 2. The marginal loss line d is located below
the largest vertical spikes of the NB providers then B providers
would treat all patients longer than NB providers. The marginal
loss line d is located above the marginal loss line ˛j c and B providers
treat patients shorter. In Fig. 3, we plotted the third possibility: the
marginal loss line d of the budgeted providers is situated below the
marginal loss line ˛j c but not below the spikes. Treatment durations
of B providers can be shorter (for example for patients with sever-
ity  2 ) and longer (for example for patients with severity  1 ) than
for NB providers. In this case we can distinguish two effects. First,
some patients may be treated shorter. We call this an “efficiency”
effect. This effect is measured by the vertical distance between the
marginal loss line d and the marginal loss line ˛j c. Second, some
patients may be treated longer. Unintended effects of bunching
around thresholds (vertical lines in Figs. 2 and 3) may result in
“overprovision” of care.
4. Descriptive statistics
We obtained our dataset from an administrative database main-
tained by the NZa and contains all registered DBCs from providers
in the secondary curative mental health care in the Netherlands for
the period 2008 to 2010. All treatments had a minimum duration of
250 min and there were only few DBCs with a very long treatment
duration, therefore we restricted our sample to DBCs with a max-
imum treatment duration of 4,000 min16 . Table 1 summarizes the
15
An exception could be a provider with (too) many patients in his practice. Such a
provider may have a financial incentive to end a treatment after hitting a threshold
because treating a new patient may be more rewarding (in terms of profits and
total patient benefits). Vice versa a provider with a shortage of patients may have
an incentive to prolong treatment duration securing his financial income. In our
analysis we assume that these are second order effects.
16
Treatment durations below 250 min belong to primary mental health care.
About 8% of the treatments had treatment duration longer than 4000 min. Note that
4000 min is well below 6000 min, so estimation errors that occur because providers
prolong treatment duration to 6000 min are likely to be small.
 R. Douven et al. / Journal of Health Economics 42 (2015) 139–150 143

Table 1
Description of data* .

Specialty Number of DBCs

2008 2009 2010 Total

Depression 80,444 78,944 76,975 236,363

Anxiety disorders 57,829 60,262 60,706 178,797
Other mental disorders and problems 49,282 50,901 49,602 149,785
Adjustment disorders 46,693 49,865 49,080 145,638
Hyperkinetic disorders 35,271 41,463 43,442 120,176
Personality disorders 39,077 39,122 39,127 117,326
Other diagnoses 26,797 28,362 30,611 85,770
Schizophrenia 24,832 27,053 28,234 80,119
Pervasive disorders 27,425 25,820 24,968 78,213
Delirium dementia and other disorders 17,796 17,680 17,617 53,093
Other substance use disorders 14,544 15,004 15,353 44,901
Alcohol use disorders 14,170 14,071 13,796 42,037
Other childhood disorders 9427 13,398 18,576 41,401
Bipolar disorders 13,228 12,423 12,349 38,000
Total 456,815 474,368 480,436 1411,619
*
The numbers in the table correspond to DBCs with treatment durations smaller than 4000 min.

and 100 (no symptoms). Providers report these GAF-scores at the

Fig. 3. Marginal profit line of budgeted providers lower than ˛jc .
data. It contains approximately 1,4 million observations in fifteen
specialties.
Table 2 distinguishes between B and NB providers. B a providers
produce the most DBCs for all categories. and some mental dis-
orders are almost exclusively treated by B providers, for example
this holds for the categories ‘delirium, dementia and other disor-
ders’, ‘alcohol use disorders’ and to a lesser extent ‘schizophrenia’.
For NB providers we observe in many cases the profession of the
therapist, we have 1302 psychologists, 431 psychiatrists and 74
providers working in institutions. For B providers we do not observe
the profession because they are all grouped together in a large
regional institution. The data contains for each DBC information
on the type of therapy (for example adult, forensic, crisis or child
care) and whether this is individual therapy, or (also) group ther-
apy or a overnight stay. Other variables are the reason for closing a
DBC (for example closed on a regular basis, or duration exceeding a
year, or patient dissatisfied with treatment), and whether providers
have prescribed drugs during a treatment. Another important vari-
able are the global assessment of functioning (GAF) scores. The
GAF-score is a quality measure for the severity of a patient’s men-
tal illness. GAF-scores range between 0 (very severe symptoms)
beginning of a treatment.
Table 2 shows that patients are unevenly distributed across
providers. To obtain enough power for our tests we narrowed
down our patient sample and considered only patients within the
following specialties: depression, anxiety disorders, adjustment
disorders, and personality disorders17 . To obtain similar patient
characteristics for comparing our treatment and control group we
only selected patients in the category “adults” that received indi-
vidual therapy sessions. Also, DBCs were closed on a regular basis
and patients received no prescribed medication. Furthermore, we
corrected the subsamples for the severity of the diseases. Based
on the GAF scores four subsamples per specialty were created. The
first subsample considers all patients that received as initial assess-
ment a GAF score between 41 and 70. The other three subsamples
are selected from this subsample, each containing only patients
with one of the following GAF scores: 41–50, 51–60, or 61–7018 .
For these subsamples patients treated by B and NB providers have
exactly the same characteristics and, thus, can be compared19 .
Table 3 summarizes and shows the number of observations for each
subsample20 . Note we also included the total sample in our estima-
tions. Patient characteristics of the total sample are very likely to
differ between B and NB providers but it provides an estimate of the
total effect of prolonging treatment durations due to the existence
of various thresholds.
5. Estimation method
Fig. 4 shows the distribution of treatment durations in the total
sample (this corresponds to ‘Total Sample’ in Table 3) for both
types of providers. The three vertical black lines correspond to three
treatment duration thresholds at 800, 1800 and 3000 min. The dis-
tribution function clearly differs between the B and NB providers.
The treatment distribution for the budgeted providers is smooth
for all treatment durations. However, in stark contrast with the B
17
We choose for these four categories because they are most prevalent treated
mental illnesses with a clear diagnosis (see Table 2).
18
The patient has some mild symptoms (e.g., depressed mood and mild insom-
nia) [GAF scale 61–70], moderate symptoms (e.g., flat affect and circumlocutory
speech, occasional panic attacks) [GAF-scale 51–60] or serious symptoms (e.g., sui-
cidal ideation, severe obsessional rituals, frequent shoplifting) [GAF-scale 41–50].
19
The age and sex distributions are very similar across subsamples.
20
The number of observations shrinks the more narrowly the subsample is defined.
Important is also that many records were not filled in completely, and therefore had
to be excluded from our subsample.
144 R. Douven et al. / Journal of Health Economics 42 (2015) 139–150

Table 2
Type of provider and number of DBCs (years 2008–2010)* .

Specialty Budgeted providers (%) Non-budgeted providers (%) Total

Depression 181,487 (77%) 54,876 (23%) 236,363

Anxiety disorders 142,747 (80%) 36,050 (20%) 178,797
Adjustment disorders 115,416 (79%) 30,222 (21%) 145,638
Personality disorders 107,545 (89%) 12,631 (11%) 120,176
Hyperkinetic disorders 91,126 (78%) 26,200 (22%) 117,326
Other diagnoses 73,216 (85%) 12,554 (15%) 85,770
Schizophrenia 75,096 (94%) 5023 (6%) 80,119
Pervasive disorders 76,633 (98%) 1580 (2%) 78,213
Delirium dementia and other disorders 52,891 (100%) 202 (0%) 53,093
Other substance use disorders 43,958 (98%) 943 (2%) 44,901
Alcohol use disorders 40,717 (97%) 1320 (3%) 42,037
Other childhood disorders 27,969 (68%) 13,432 (32%) 41,401
Bipolar disorders 34,557 (91%) 3,443 (9%) 38,000
Other mental disorders and problems 112,309 (75%) 37,476 (25%) 149,785
Total 1175,667(83%) 235,952 (17%) 1411,619
*
The numbers in the table correspond to DBCs with treatments duration smaller than 4000 min.

Table 3
Number of observations in various subsamples (years 2008–2010)** .

Budgeted providers (%) Non-budgeted providers (%) Total

1. Total sample 1175,667 (83%) 235,952 (17%) 1411,619

2. Sample depression 181,487 (77%) 54,876 (23%) 236,363
2a GAF: 41–70* 57,740 (65%) 30,508 (35%) 88,248
2b GAF: 41–50* 12,132 (71%) 4963 (29%) 17,095
2c GAF: 51–60* 32,730 (65%) 17,395 (35%) 50,125
2d GAF: 61–70* 12,878 (61%) 8150 (39%) 21,028
3 Sample anxiety disorders 142,747 (80%) 36,050 (20%) 178,797
3a GAF: 41–70* 55,505 (72%) 21,581 (28%) 77,086
3b GAF: 41–50* 10,360 (78%) 2934 (22%) 13,294
3c GAF: 51–60* 31,051 (72%) 12,086 (28%) 43,137
3d GAF: 61–70* 14,094 (68%) 6561 (32%) 20,655
4 Sample adjustment disorder 115,416 (79%) 30,222 (21%) 145,638
4a GAF: 41–70* 55,545 (72%) 21,607 (28%) 77,152
4b GAF: 41–50* 5985 (76%) 1934 (24%) 7919
4c GAF: 51–60* 30,571 (72%) 12,067 (28%) 42,638
4d GAF: 61–70* 18,989 (71%) 7606 (29%) 26,595
5 Sample personality disorder 107,545 (89%) 12,631 (11%) 120,176
5a GAF: 41–70* 39,571 (69%) 17,977 (31%) 57,548
5b GAF: 41–50* 8467 (78%) 2457 (22%) 10,924
5c GAF: 51–60* 22,102 (69%) 10,056 (31%) 32,158
5d GAF: 61–70* 9002 (62%) 5464 (38%) 14,466
*
In these samples we only consider individual adult therapies without medical prescriptions that were closed on a regular basis.
**
The numbers in the table correspond to DBCs with treatments duration smaller than 4000 min.

providers, for NB providers we observe large gaps and spikes at
thresholds. Similar figures are obtained if we plot subsamples of
our dataset.
To estimate whether B providers treat on average longer or
shorter than NB providers we use ideas from regression discon-
tinuity design (RDD)21 . However, while RDD-studies use local
linear smoothing around single thresholds to determine non-linear
responses, we have reasonably large bunches and gaps of several
thresholds that may be connected22 . Therefore, we use a global
21
RDD studies related to health care include Card et al. (2008, 2009) who study
the discontinuity of health care utilization around age 65 when US citizens become
eligible for Medicare. Sojourner et al. (2013) use RDD to study the effects of union-
ization of nursing homes. Shi (2013) finds evidence of income manipulation when
studying labor supply responses to income cutoffs of a subsidized health insurance
program in Massachusetts. Einav et al. (2013) study the response of drug expen-
diture to non-linear contracts in Medicare part D. These studies are all related to
consumer responses. Our study is about provider responses and more related to
Bajari et al. (2011) who study hospital’s responses to discontinuities in linear reim-
bursement schedules. Their identification strategy is much more complicated than
in our paper because reimbursement schedules are only discontinuous in the first
derivative, and thresholds are not fixed but may differ across hospitals.
22
For example, combining several separate local linear estimation procedures to
one distribution function may not necessarily result in a smooth function.
estimation approach which allows us to estimate in one step the
distribution functions for both types of providers.
We fit the non-linear regression equation (5) for each mental
disorder category i, and provider type j (in what follows we omit i,
j):
 
Yt = f ˇ + t with t = Bt − Gt + εt (5)
where Yt , t = 3, 3.5, 4, . . ., 39, 39.5 is the distribution function
of treatment durations defined in treatment duration classes of
50 min23 . Alike Lee and Lemieux (2010) we assume that all fac-
tors evolve “smoothly”. If there are no discontinuities (Gt = 0, Bt = 0)
in the reimbursement schedule, f(ˇ) would be a reasonable guess
for explaining Yt . This assumption is confirmed by estimates of f(ˇ)
for the distribution function of B providers.
In standard RDD applications, sudden shifts in the outcome
variable result from an exogenous change. In this study we have
the same. Bunches and gaps in treatment durations of the NB
providers are caused by exogenous changes in the fee structure, and
23
Thus, Y3 represents all treatment durations in the 300–350 min time interval and
Y39.5 in the 3950–4000 min time interval. The size of the surface of all distributions
is normalized to 1. Note that we performed our analysis also for classes of 100 min
time intervals. This yielded similar, but slightly less stable, results.
 R. Douven et al. / Journal of Health Economics 42 (2015) 139–150 145

Fig. 4. Distribution of treatment duration for B and NB providers (all categories).

not by medical outcome or other unobserved factors of individual

patients. This implies that a conscious prolongation of treatment
duration by NB providers, introduces systematic errors in Yt . In (5),
t represents both systematic and random errors. We distinguish
systematic positive errors or “bunches” after a threshold (Bt ≥ 0)
and systematic negative errors or “gaps” before a threshold (Gt ≥ 0).
Lastly, εt represents the random error term in (5).
To estimate the smooth function f(ˇ) we constructed a class
of smoothing functions that are able to describes similar shapes
as the B providers in Fig. 4. A property of this function is that it
must increase at t = 300, has a top somewhere around t = 600 min,
and monotonically declines thereafter24 . Furthermore, the func-
tion must be flexible enough to capture various shapes. Exponential Fig. 5. Estimation of distribution function and bunches and gaps.
function (3) satisfies these criteria:
To estimate ˇ, we follow a weighted non-linear least squares
minimization problem with four restrictions.


39.5
  2
min wt Yt − f ˇ with restrictions :
ˇ
t=3
(7)

10.5
   
20.5
   
32.5
   
39.5
  
Yt − f ˇ = 0, Yt − f ˇ = 0, Yt − f ˇ = 0, Yt − f ˇ =0
t=5 t=11 t=21 t=3

The first three restrictions correspond to the shift of treatment

durations: B[k] − G[k] = 0, for k = 8, 18, 3025 . We observed in the data
that bunching occurs up to 300 min after a threshold. Therefore we
fixed possible bunching to the first 300 min after a threshold in our
restrictions.
ˇ3 ˇ4 To obtain smooth convergence of our non-linear estimations,
f (ˇ) = ˇ1 + ˇ2 t + + + ˇ5 e−ˇ6 t (6) we added a fourth restriction: the total sum of the errors is zero26 .
t t
Weights wt were also introduced27 . Our global estimation strategy
with restrictions is quite powerful compared to three separate local
We have to estimate the six parameters ˇj , j = 1, . . ., 6 in (6). First,
we substitute (6) in (5). Then we estimate (5). The size of the gap
before each threshold [k] ∈ {[8], [18], [30]} should equal the size of 7.5 10.5 17.5
25
G[8] = − t , B[8] = t , G[18] = − t , B[18] =
the bunch after this threshold (see Fig. 5). This restriction reflects 20.5 4.5
29.5 8
32.5 11

our theory in Section 2: bunching after a threshold occurs through t , G[30] = − t , B[30] = t .
18 21 30
26
a shift of treatment durations from before to after a threshold. This implies all systematic shifts are explained by the three previous restrictions,
and that no treatments with duration between 300 and 500 min are shifted to over
800 min threshold, and between 3300 and 4000 min are shifted to over the 6000 min
threshold.
27
In most cases we used wt = 1, however sometimes we experimented with
somewhat higher weights to obtain smooth convergence. We performed our opti-
24
Using standard smoothing functions in econometric software programs do not mizations with the numerical non-linear global optimization function “NMinimize”
work here because these functions ‘try to explain’ the bunches and gaps as well. of the software program Mathematica. To obtain convergence we sometimes had to
146 R. Douven et al. / Journal of Health Economics 42 (2015) 139–150
RDD-estimations at each individual threshold. The global approach
allows us to connect the “bunches” and “gaps” estimates at individ-
ual thresholds and convergence of our estimation procedure will
only occur if our assumption of equal gaps and bunches is supported
by the data.
ˆ ˆ
Minimization procedure (7) generates ˇ1 , . . ., ˇ6 . This allows
 in
us to compute  ˆ
ˆ t = Yt − f ˇ . Next, we can compute our estimates
for the gaps and bunches.
In order to present the significance of our estimates for bunches
and gaps we need an estimate for our error term εt in (5). Because
our computation does not allow us to compute for each t, B̂t , Ĝt in
(5) separately, we cannot properly estimate the random error term
εt . Therefore we assume ε̂t =  ˆ Bt where 
ˆ Bt are the estimated errors
of the budgeted providers after estimating (5). Thus, we assume the
standard error of the non-budgeted providers sNB in (5) equals the
standard error of the budgeted providers sB :28
increase of 7.1 to 7.9%. The efficiency effect for the “Total Sample
1   2
NB B
S =S = Bt (8)
(74 − 6) t
We use a 68 degrees of freedom correction (see e.g. Verbeek,
2004), 74 minus 6 (parameters ˇ to estimate in (8)). After obtain-
ing these statistics we can derive additional statistics such as an
estimate of the average treatment duration, prolongation time as a
result of shifting treatments and associated costs.
6. Estimation results
In this section we present our estimation results for all the sam-
ples described in Table 3. We first show our results graphically in
Fig. 6 for the three samples 1, 2 and 2a in Table 3: “total sample”,
“depression” and the subsample “depression with similar patient
characteristics (GAF-scores 41–70)”. Fig. 6 contains for each sam-
 Fig. 6a, d and g, show Yt and
ple three panels. The first panels,
the corresponding estimate f ˇ ˆ of the B provider, from which
we will derive an estimate for our standard error. The estimates
indicate that our exponential identification in (6) can fit f(ˇ) to Yt
very well. The middle panels, Fig. 6b, e and h, indicate the unin-
tended effects. Bunches and gaps are present in all three samples.
The size of bunches and gaps are remarkably stable across subsam-
ples. Bunches and gaps are largest (and significant) at the first two
thresholds of 800 and 1800 min and positive at the threshold of
3000 min in all cases29 . Differences in treatment duration between
both providers, after controlling for the bunches and gaps, can be
seen in the right three panels, Fig. 6c, f and i. For the total sample
and depression sample (panels c and f) we observe large effects; on
average NB providers treat patients much shorter than B providers.
However, this effect almost disappears in the case of patients with
similar characteristics (panel i). Controlling for patient characteris-
tics is therefore crucial to identify possible differences in treatment
duration between B and NB providers.
The estimation results of the three subsamples are summarized
in Table 4. The first column presents the unintended effects: the
percentage of treatments that are shifted over each of the three
alter the minimization method in Mathematica (gradient-based and direct search
methods), weights and starting values.
28
We make the reasonable assumption that the random errors and corresponding
standard deviations sB and sNB are of the same order of magnitude. If there are
small systematic errors in  ˆ Bt we will overstate sNB Note that we calculate sB from
a Yt distribution that has the same number of observations as the corresponding Yt
distribution of the NB providers.
29
The effects are smaller around the 3000 min threshold; there are fewer obser-
vations and it may be the case that the marginal benefit to patients is closer to zero
(i.e. “flat of the curve”).
thresholds. In total about 11–13% of treatments are shifted to over
a next threshold. The second column in Table 4presents
average
treatment duration. The difference between f ˇ ˆ and Yt for B
providers is small, confirming the good fit and resulting in small
standard errors sB30 . ForNB providers the average treatment dura-
ˆ
tion corresponding to f ˇ is 19–24 min lower than Yt , indicating
that the increase in average treatment duration as a result of bunch-
ing is relatively small31 . Important is the large difference in average
treatment duration between B and NB providers in the “Total sam-
ple”, 22.2%, and “Total sample depression”, 24.2%, indicating that
B providers treat on average more sick patients. After control-
ling for patient characteristics (“Subsample depression, GAF scores
41–70”) treatment duration shrinks to 2.2%. In the third column of
Table 4 we present average treatment costs. The unintended effects
increase average costs per treatment by 137 to 157 euro or a cost
Depression, GAF scores 41–70” yields that on average treatments
are 3.3% (or 39 euro) more expensive for B than NB providers. This
effect is however more than offset by the unintended effects; sum-
ming both effects yields that NB providers treat on average patients
165–39 = 126 euro more expensive than B providers32 .
In addition to the three subsamples, we have also looked into
other mental illnesses (see Table 3 for the subsamples and the num-
ber of observations in each subsample). We performed the same
estimations for these sixteen subsamples. The results are reported
in Table 5. Columns (1)–(3) present the volume effects. Column
(1) represents the size of the unintended effects: the percentage
treatments that are shifted to over a next threshold. Column (2)
shows the average treatment duration for the actual distribution
Yt , and estimated distribution f ˇ ˆ , and column (3) shows the
differences in treatment duration or the “efficiency” effect; the per-
cent change in treatment duration between NB and B providers.
Columns (4)–(6) show the same effects but now for fees.Col-
umn (4) shows the average fee of a treatment for Yt and f ˇ ˆ .
Column (5) presents the unintended cost effects; the percent dif-
ference between the two variables. Finally, column (6) represents
the cost difference related to the “efficiency” effect between NB and
B providers.
The results in Table 5 confirm our previous findings. First of all,
we observe that the unintended effects (column (1)) are present in
all subsamples. The effects are fairly stable across all our subsam-
ples and vary roughly between ±11 and 13%, with some outliers33 .
This corresponds with a cost increase that varies between ±7
and 9% (column 5). The efficiency effect in column (3) shows
that B providers treat patients approximately ±2–7% longer than
NB providers with corresponding cost increases of approximately
±3–6% (column (6))34 . Thus, for almost all cases we find that the
30
For smaller subsamples the graph Yt is less smooth increasing the size of the
standard error sB .
31
The average prolongation of treatment duration for treatments that are shifted
over to a next threshold is about 200 min.
32
We have tested the significance of the efficiency effect with thenon-parametric

Kolmogorov–Smirnov test. It rejected the hypothesis of similar f ˇ ˆ distribution
functions for B and NB providers in the first two samples in Table 4. However, it does
not reject the hypothesis in the third sample. We therefore test various different
subsamples in Table 5.
33
The estimation results for the unintended effects are all significant on a 0.01
level.
34
Only for the subsample adjustment disorders GAF: 61–70 we find a 0.4 higher
average treatment duration for NB providers. The efficiency effects are not signif-
icant at a 0.05 level (see footnote 29). However, we still conclude that efficiency
effects are present in our data because we repeated our estimations many times
(see column (3)) and our data covers the complete sample.
 R. Douven et al. / Journal of Health Economics 42 (2015) 139–150 147

Fig. 6. (a–c) Total sample. (d–f) Total sample depression. (g–i) Subsample depression, GAF scores 41–70.

Table 4
Estimation results for “total sample”, “total sample depression” and “total sample depression (GAF scores 41–70)”.a

Bunches and gaps (%) Average treatment Average treatment

Unintended effect duration (min) costs (euro)
   
B̂[8] B̂[18] B̂[30] Total Yt f ˆ
ˇ dif Yt f ˆ
ˇ Unintended effect

Total sample (all specialties)

B providers 1407 1407 2279 2266
NB providers 7.6** 3.1** 0.7** 11.5** 1170 1151 19 2053 1901 152 (8.0%)
Efficiency effect 22.2% 19.8%

Total sample depression

B Providers 1345 1345 2226 2214
NB providers 8.4** 3.6** 0.5* 12.5** 1107 1083 24 2008 1847 161 (8.7%)
Efficiency effect 24.2% 20.5%

Total sample depression (GAF scores 41–70)

B Providers 1224 1224 2043 2036
NB providers 8.3** 4.0** 0.6* 12.9** 1204 1185 19 2149 1986 165 (8.2%)
Efficiency effect 3.3% 2.9%
a
The *,** in the Table indicate significance levels of 0.05 and 0.01 respectively. Average treatment costs for B providers are calculated on the premise that they are paid
according to the reimbursement schedule for NB providers.
148 R. Douven et al. / Journal of Health Economics 42 (2015) 139–150

Table 5
Estimation results for subsamples 2a–d, 3a–d, 4a–d, 5a–d (see Table 3).

Subsample (1) Bunches, (2) Avg. treatment (3) “Efficiency” (4) Avg. treatment (5) Unintended (6) “Efficiency”
gaps (%) duration (min) effect (%) (min) costs (euro) effect (euro) effect (%)

Type of provider NB NB NB (NB-B)/B NB NB NB (NB-B)/B:

         
Distribution Yt Yt f ˆ
ˇ Yt , f ˆ
ˇ Yt f ˆ
ˇ Yt , f ˆ
ˇ Yt , f ˆ
ˇ

Depression
2a. GAF: 41–70 12.9 1204 1185 −3.3 2149 1986 8.2 −2.9
2b. GAF: 41–50 14.1 1330 1300 −2.3 2353 2152 9.4 −2.4
2c. GAF: 51–60 13.1 1215 1189 −4.0 2162 1987 8.8 −3.5
2d. GAF: 61–70 11.8 1105 1083 −3.9 1996 1844 8.2 −3.2

Anxiety disorders
3a. GAF: 41–70 12.1 1186 1161 −7.9 2094 1926 8.7 −7.5
3b. GAF: 41–50 11.3 1325 1303 −7.1 2327 2146 8.4 −6.9
3c. GAF: 51–60 12.8 1201 1175 −8.4 2118 1942 9.1 −8.0
3d. GAF: 61–70 11.4 1096 1076 −7.0 1944 1800 8.0 −6.6

Adjustment disorders
4a. GAF: 41–70 10.6 1054 1039 −2.2 1761 1645 7.1 −2.1
4b. GAF: 41–50 10.6 1215 1174 −2.8 2010 1839 9.3 −2.1
4c. GAF: 51–60 10.8 1061 1042 −4.5 1771 1646 7.6 −4.3
4d. GAF: 61–70 9.5 1000 984 0.4 1682 1572 7.0 0.3

Personality disorders
5a. GAF: 41–70 11.6 1391 1372 −5.5 2435 2251 8.1 −5.1
5b. GAF: 41–50 11.9 1495 1475 −5.2 2598 2402 8.2 −5.2
5c. GAF: 51–60 12.3 1422 1402 −5.3 2489 2290 8.7 −5.3
5d. GAF: 61–70 10.4 1286 1277 −5.2 2265 2117 7.0 −4.4

marginal loss line d is situated somewhat below the line ˛j c (see
Fig. 3 in Section 3). The unintended financial effects in column (5)
are in all cases larger than the “efficiency” effects (column (6)).
To conclude, the unintended effects appear very clear in the data
and are very stable across all subsamples. The “efficiency” effects
are smaller and less certain because these effects are estimated by
comparing B and NB providers. A limitation of our measure for the
“efficiency” effects could be that there is still unobserved variation
in the treatment and control group that we do not capture ade-
quately. For example, we may have overestimated the efficiency
effects if NB providers select more low severity patients (even for
groups with similar GAF scores)35 . Another possibility is that our
“efficiency” effect captures not genuine efficiency but quality dif-
ferences in outcome between B and NB providers. In future research
we may be able to address some of these points if more information
becomes available.
7. Discussion
We have evaluated the implementation of a new reimburse-
ment schedule in Dutch mental health care. The reimbursement
schedule follows a discontinuous discrete step function: once the
provider has passed a treatment duration threshold the fee is flat
until a next threshold is reached. We find an “efficiency” effect: on
the flat part of the fee schedule providers prolong treatment only if
marginal benefits to patients outweigh marginal costs. We estimate
a reduction in treatment duration by 2 to 7% and lower costs by 3 to
6% compared to a control group. However, we also find unintended
effects: providers treat patients longer to reach a next threshold
and obtain a higher fee. The data shows gaps and bunches in the
distribution function of treatment durations, just before and after
a threshold. In total, about 11 to 13% of treatments are shifted to
over a next threshold, resulting in a cost increase of approximately
7 to 9%.
35
If this were the case then we would expect that selection effects are greater for
less severe patients (groups with high GAF-scores). However, we do not observe
that efficiency effects are stronger for those groups (see Table 5).
An important message of our study is that the unintended effects
clearly demonstrate that mental health care providers react to
financial incentives. Since financial rewards are high, one would
expect that NB providers would anticipate ex-ante on the thresh-
olds in the reimbursement schedule. Indeed, an article in a Dutch
newspaper suggests that some providers have institutionalized the
number of therapy sessions. A psychologist stated: “Our institution
has calculated that each patient should receive eight ór sixteen ses-
sions. This would be financially very attractive but not be seen as
fraud” (Effting, 2015).
Monitoring providers’ behavior is therefore an important ele-
ment for the system to function properly. In the Dutch system of
regulated competition health insurers have the role to discipline
providers. However, until 2014 health insurers lacked informa-
tion about the exact treatment duration of health care providers.
They received only global information on treatment duration of
individual providers, i.e. they received only information between
which two treatment duration thresholds the provider performed
the treatment, and not the exact treatment time. Thus, insurers had
no possibility to perform the same analysis as we carried out in this
paper. This is now gradually changing; since 2014 health insurers
obtain exact information about treatment durations and are also
becoming more financially responsible for mental health care cost
containment.
We measure an “efficiency” effect. However, we cannot be cer-
tain that we measure genuine efficiency since we cannot rule out
the possibility that patients may also have received too little care.
Our efficiency arguments do hold if we assume ˛j = 1 in our utility
function (1), which is a fairly standard assumption (McGuire, 2000).
In that case NB providers produce cost efficient on the flat part of the
reimbursement schedule and bunching corresponds to overtreat-
ment. Efficiency differences between B and NB providers could also
be related to differences in practice styles or quality of treatments
(see e.g. Chandra et al., 2012). To address these issues more properly
quality information about treatments would be necessary.
In 2014, the Dutch government decided to pay B providers also
according to the new reimbursement schedule. Our findings sug-
gest that this policy may lead to higher costs since the higher costs
associated with the unintended effects outweigh the lower costs
 R. Douven et al. / Journal of Health Economics 42 (2015) 139–150
of the efficiency effect. However, an important difference is that
B providers correspond to large mental health institutions where
doctors are still paid on a salary basis, so the unintended effects
have to be induced by the management. Furthermore, there are still
many external dynamic demand and supply factors that are diffi-
cult to assess. For example, B providers may put a lower weight
on profits (lower agency parameter ˛j in (1) for B providers) than
NB providers because the latter category of providers is of a more
entrepreneurial type. In sum, the unintended effects may therefore
be lower for B providers than NB providers. Also, insurers may be
better equipped to monitor providers’ treatment duration and, in
the longer run, quality. Another important difference is that the
Dutch government changed the flat reimbursement fees to maxi-
mum fees. Thus, health insurers can bargain with providers lower
reimbursement fees, if providers’ performances turn out to be inad-
equate.
Also the conclusion that the introduction of the new reimburse-
ment schedule for NB providers in 2008 led to higher costs is
premature. Before 2008, NB providers received a fixed fee for each
visit. A fee for each visit is similar to the reimbursement schedule
in our study but now there are thresholds after each visit of 60 min.
A fee for each visit is closer to a fee-for-service type of payment and
may also result in overtreatment. Unfortunately, we have no data
for the period before 2008 available, making a comparison between
the two regimes not possible.
An important policy question is how an optimal reimburse-
ment schedule for mental health care providers should look like.
The reimbursement schedule that we study in this paper is inter-
esting because it combines a prospective fee per episode of care
with elements of fee-for-service, to prevent selection incentives.
The drawback is that the unintended effects are quite large which
may make the schedule less attractive than salary or even fee-
for-service. One possible way to proceed would be to improve
the current reimbursement schedule. A first option would be to
diminish the unintended effects by changing the position of the
thresholds. Ideally, thresholds should be placed where the mass of
the distribution function f(ˇ) is small. If the mass before a thresh-
old is small, unintended effects will diminish because there are
only few treatments to shift over to a next threshold. Unfortu-
nately, the threshold of 800 min is placed just after the top of
the distribution function (see Fig. 4), thus exacerbating the unin-
tended effects. Moving the 800 min threshold to 500 min, just
before the top of the distribution function, would diminish the
unintended effects. Thus by taking into account provider behavior
the reimbursement could be made much more attractive. A second
option to diminish the unintended effects would be to decrease (or
even increase) the number of thresholds. Here there is a trade-off
between efficiency, equity and selection. For example, removing all
thresholds would yield a single prospective fee for the total treat-
ment. This would remove all unintended effects thereby increasing
efficiency. However, if patients’ characteristics across providers dif-
fer substantially, it could also result in a larger income variation
across providers thereby diminishing equity considerations across
providers. As a result providers might increase their incentives for
selecting more favorable patients (McGuire, 2000). Adding more
thresholds might also be an improvement since more thresholds
will diminish the financial incentives at each individual threshold.
More research is necessary to study these trade-offs.
Another possible way is to change the nature of the payment
system and to consider a mixed payment system of a prospec-
tive fee and a linear reimbursement schedule, as advocated by
Ellis and McGuire (1986). The prospective fee reimburses the “non-
contractible” activities that may vary across specialties while the
linear reimbursement fee should be set less than marginal (and
average) costs. This type of fee schedule is likely to diminish the
unintended effects as well. Lastly, quality should be integrated into
149
tariffs and providers should ultimately get paid taking into account
the patient’s wellbeing as well. This is a long shot and much more
research is needed to integrate quality aspects into the payment
system.
In this study we rely on providers that register their own DBCs.
We assume that providers register their treatment duration cor-
rectly and honestly in their administration. However, literature
indicates that fraudulent behavior may also occur in payment sys-
tems based on DRGs in the US, or DBCs in the Netherlands. This
fraudulent behavior is often referred to as ‘upcoding’ (Steinbusch
et al., 2007). The Dutch reimbursement system may be vulnerable
to this ‘upcoding’ because Dutch providers code DBCs themselves.
They could tamper with the data. Especially, in mental health care
the risk for fraud may even be greater than for less discretionary
treatments, as hip or knee replacements. Third parties, such as
health insurers, also might find it particularly difficult to verify and
dispute mental health diagnoses.
Acknowledgement
We would like to thank the Dutch Healthcare Authority (NZa)
for providing the data. The data are not publicly available. We are
grateful to the NZa and DBC-Onderhoud for explaining the data. We
would like to thank seminar participants at BU/Harvard/MIT Health
Economics seminar in Boston at April 4, 2014, at NZa-seminar in
Utrecht at May 8, 2014, Academy of Health in San Diego, June
8–10, 2014, IHEA in Dublin in July 13–16, 2014, CPB-seminar in
The Hague, October 14, 2014, ESE-seminar at Erasmus University,
January 15, 2015, ICMPE at Venice, March 28, 2015 for comments.
Furthermore, we are grateful to two anonymous referees, Pieter
van Baal, Pieter Bakx, Leon Bettendorf, Aaron Maras, Tom McGuire,
Jan van Ours, Bastian Ravesteijn, Ingrid Seinen, Harry van Til and
Gert Jan Verhoeven for providing comments on earlier versions of
this paper.
References
Bajari, P., Hong, H., Park, M., Town, R., 2011. Regression discontinuity designs with
an endogenous forcing variable and an application to contracting in health care.
In: NBER Working Paper No. 17643.
Bellows, N.M., Halpin, H.A., 2008. Impact of Medicaid reimbursement on mental
health quality indicators. Health Serv. Res. 43, 582–597.
Card, D., Dobkin, C., Maestas, N., 2008. The impact of nearly universal insurance
coverage on health care utilization: evidence from medicare. Am. Econ. Rev. 98
(5), 597–636.
Card, D., Dobkin, C., Maestas, N., 2009. Does medicare saves lives? Q. J. Econ. 123 (1),
597–636.
Chandra, A., Cutler, D., Song, Z., 2012. Who ordered that? The economics of treat-
ment choices in medical care. In: Pauly, M.V., McGuire, T.G., Barros, P.P. (Eds.),
Handbook of Health Economics, vol. II. Elsevier, Amsterdam, pp. 397–432.
Christianson, J.B., Conrad, D., 2011. Provider payment and incentives. In: Glied, S.A.,
Smith, P.C. (Eds.), The Oxford Handbook of Health Economics. Oxford University
Press, Oxford, pp. 624–628.
Douven, R., Mocking, R., Mosca, I., 2015. The effect of physician remuneration
on regional variation in hospital treatments. Int. J. Health Econ. Manage.,
10.1007/s10754-015-9164-2.
van Dijk, C.E., van den Berg, B., Verheij, R.A., Spreeuwenberg, P., Groenewegen, P.P.,
de Bakker, D.H., 2013. Moral hazard and supplier-induced demand: empirical
evidence in general practice. Health Econ. 22 (3), 340–352.
DBC Onderhoud, 2013. Spelregels, DBC-Registratie GGZ, Versie RG13a. DBC Onder-
houd, Utrecht.
Einav, L., Finkelstein, A., Schrimpf, P., 2013. The response of drug expenditure to
non-linear contract design: evidence from Medicare Part D. In: MIT Working
Paper.
Effting, M., 2015. Tjak, tjak, volgende patient: de ontsporing van de ggz. De Volk-
skrant. (Dutch Newspaper).
Ellis, R.P., McGuire, T.G., 1986. Provider behavior under prospective reimbursement.
Cost sharing and supply. J. Health Econ. 5 (1986), 129–151.
Ellis, R.P., McGuire, T.G., 1990. Optimal payment systems for health services. J. Health
Econ. 9 (4), 375–396.
Epstein, A.M., et al., 1986. The use of ambulatory testing in prepaid and fee-for-
service group practices: relation to perceived profitability. N. Engl. J. Med. 314,
1089–1093.
150 R. Douven et al. / Journal of Health Economics 42 (2015) 139–150
Frank, R.G., McGuire, T.G., 2000. Economics and mental health. In: Culyer, A.J., New-
house, J.P. (Eds.), Handbook of Health Economics, vol. 1B. Elsevier, Amsterdam,
pp. 893–954.
Nederland, G.G.Z., 2010. Zorg op waarde geschat, update. In: Sectorrapport ggz 2010,
Amersfoort. (in Dutch).
Hickson, G.B., et al., 1987. Physician reimbursement by salary or fee-for-service:
effect on a physician’s practice behavior in a randomized prospective study.
Pediatrics 80, 744–750.
Jennison, K., Ellis, R.P., 1987. Comparison of psychiatric service utilization in a single
group practice. In: McGuire, Scheffler (Eds.), The Economics of Mental Health
Services: Advances in Health Economics and Health Services Research, vol. 8.
JAI Press, Greenwich, USA, pp. 175–194.
Lee, D.S., Lemieux, T., 2010. Regression discontinuity designs in economics. J. Econ.
Lit. 48, 281–355.
Mason, A., Goddard, M., 2009. Payment by Results in Mental Health: A Review of the
International Literature and an Economic Assessment of the Approach in the
English NHS, Research Paper 50. Centre for Health Economics, The University of
York.
McGuire, T.G., 2000. Physician Agency. In: Culyer, A.J., Newhouse, J.P.
(Eds.), Handbook of Health Economics, vol. 1A. Elsevier, Amsterdam,
pp. 461–536.
NZa, 2007. Tariefbeschikking DBC GGZ 2008. Nederlandse Zorgautoriteit, Utrecht
(in Dutch).
NZa, 2008. Tariefbeschikking DBC GGZ 2009. Nederlandse Zorgautoriteit, Utrecht
(in Dutch).
NZa, 2009. Tariefbeschikking DBC GGZ 2010. Nederlandse Zorgautoriteit, Utrecht (in
Dutch).NZa (2010). In: Invoering Prestatiebekostiging Curatieve GGZ: Advies op
Hoofdlijnen. Duth Healthcare Authority, Utrecht (in Dutch).
NZa, 2010. De curatieve GGZ in 2009: Ontwikkelingen in aanbod en volume. In:
Monitor. Nederlandse Zorgautoriteit, Utrecht (in Dutch).
NZa, 2011. Curatieve GGZ 2010 Een sector in ontwikkeling. In: Monitor. Nederlandse
Zorgautoriteit, Utrecht (in Dutch).
NZa, 2012. Marktscan Geestelijke Gezondheidszorg. Weergave van de markt
2008–2011. Nederlandse Zorgautoriteit, Utrecht (in Dutch).
Rekenkamer, 2013. Indicatoren voor kwaliteit in de zorg. In: Algemene Rekenkamer,
March 28. Algemene Rekenkamer, The Netherlands, The Hague.
Rosenthal, M.B., 2000. Risk sharing and the supply of mental health services. J. Health
Econ. 19 (6), 1047–1065.
Shi, J., 2013. Labor supply response to income cutoffs of health insurance in the
Massachusetts reform. In: Working Paper. Boston University.
Sojourner, A.J., Grabowski, D.C., Town, R.J., Chen, M.C., Frandsen, B.R., 2013. impacts
of unionization on quality and productivity: regression discontinuity evi-
dence from nursing homes. In: Working Paper, https://economics.byu.edu/
frandsen/Documents/Nursing Home Unions.pdf.
Stearns, S., Wolfe, B., Kindig, D., 1992. Physician responses to fee-for-service and
capitation payment. Inquiry 29, 416–425.
Steinbusch, P.J.M., Oostenbrink, J.B., Zuurbier, J.J., Schaepkens, F.J.M., 2007. The risk
of upcoding in casemix systems: a comparative study. Health Policy 81 (2–3),
289–299.
Van de Ven, W.P.M.M., Schut, F.T., 2008. Universal mandatory health insurance in
The Netherlands: a model for the United States? Health Affairs 27 (3), 771–781.
Verbeek, M., 2004. A Guide To Modern Econometrics, second ed. Wiley, New York,
NY.
VWS, 2010. Interdepartementaal beleidsonderzoek curatieve GGZ. In: Attachment
by the Report: Heroverweging curatieve zorg. Ministry of Health, Welfare and
Sport, The Hague (in Dutch).
 Journal of Health Economics 42 (2015) 151–164

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Health and agricultural productivity: Evidence from Zambia

Günther Fink a,∗ , Felix Masiye b
a
Harvard School of Public Health, USA
b
Department of Economics, University of Zambia, Zambia

a r t i c l e i n f o a b s t r a c t

Article history: We evaluate the productivity effects of investment in preventive health technology through a random-
Received 19 August 2014 ized controlled trial in rural Zambia. In the experiment, access to subsidized bed nets was randomly
Received in revised form 12 April 2015 assigned at the community level; 516 farmers were followed over a one-year farming period. We find
Accepted 21 April 2015
large positive effects of preventative health investment on productivity: among farmers provided with
Available online 28 April 2015
access to free nets, harvest value increased by US$ 76, corresponding to about 14.7% of the average output
value. While only limited information was collected on farming inputs, shifts in the extensive and the
JEL classification:
intensive margins of labor supply appear to be the most likely mechanism underlying the productivity
I15
J24
improvements observed.
J43 © 2015 Elsevier B.V. All rights reserved.

Keywords:
Investment
Health
Productivity
Agriculture
Malaria

1. Introduction
Despite the rapid speed of urbanization over the past decades,
rural small-scale farming remains the primary source of food and
income for a majority of the population in developing countries
(World Bank, 2007). In most settings, the degree of agricultural
mechanization is limited, so that agricultural production remains
primarily dependent on the availability and productivity of human
labor. While labor is abundant in principle in most developing
countries (Pitt and Rosenzweig, 1986), labor inputs can be com-
promised by episodes of ill health and can result in output losses if
absent labor cannot be replaced immediately.
In this paper we investigate the economic impact of short-term
morbidity on agricultural output in the context of small-scale farm-
ing in Zambia. The study setting is representative of many rural
areas in the developing world both in terms of the general lack of
advanced farming technology and in terms of the dominant role
of farming as source of nutrition and income. With farming land
available free of charge in most communities, a large majority of
the working-age population engages in agriculture, while formal
sector jobs are generally scarce. Despite major government efforts
to reduce the burden of the disease in recent years (NMCC, 2010;
Zambia Ministry of Health, 2006), malaria continues to be the pri-
mary cause of short-term morbidity in the country, with children
and adults experiencing up to five episodes of malaria per year
(NMCC, 2010; WHO, 2009). Since the planting season tends to over-
lap with the malaria season, health related absences from field work
are frequent, and are commonly cited by local farmers as primary
cause of lost field work and income.1
To evaluate the degree to which health affects agricultural pro-
ductivity, we conducted a cluster-randomized field experiment
with 516 farmers in Katete District, Zambia, from December 2009
to August 2010. As part of the experiment, farmers were ran-
domly selected for bed net programs, which allowed them to obtain
long-lasting insecticide treated nets (LLITNs) through agricultural
loan program schemes at differentially subsidized prices. The basic
intuition underlying the experiment is relatively straightforward:
as long as household labor and consumption decisions are non-
separable from household production decisions2 (Benjamin, 1992),

∗ Corresponding author at: Harvard School of Public Health, Boston, MA 02115,
USA. Tel.: +1 6174327389.
E-mail address: gfink@hsph.harvard.edu (G. Fink).
1
On average, farmers surveyed at baseline claimed that their harvest would
increase by 30% if field work was not interrupted by episodes of ill health.
2
If consumption and production decisions were perfectly separable, family labor
could be perfectly substituted for by hired labor.

http://dx.doi.org/10.1016/j.jhealeco.2015.04.004
0167-6296/© 2015 Elsevier B.V. All rights reserved.
152 G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164
decreased exposure to malaria should increase the time and energy
farmers can spend on their fields, and thus also increase the final
harvest amounts.
In a first paper based on this experiment, we analyzed the impact
of the additional LLITNs distributed on self-reported morbidity
(Fink and Masiye, 2012). In this paper, we analyze the impact of
the net programs on agricultural productivity, the main outcome
variable of the trial. In the first part of our analysis, we analyze the
impact of the interventions on net ownership and usage. Consistent
with recent work by Tarozzi et al. (2014), we find a substantial frac-
tion of farmers to be willing to purchase LLITNs at full or partially
subsidized prices when financing options are provided. On average,
farmers in the loan group acquired 0.9 nets, resulting in a 24% point
increase in the average fraction of sleeping spaces covered at the
household level.
In the second part of the paper, we estimate the impact of the
bed net programs on agricultural production. In order to facili-
tate a rapid distribution of bed nets, treatments were randomly
assigned at the cluster level prior to the collection of baseline
data in the experiment. The non-stratified cluster-level randomiza-
tion resulted in a rather unbalanced sample, with treated farmers
on average both larger and more productive than farmers in
the control group. To address these imbalances, we focus on
analyzing changes in production outcomes between the 2009 (pre-
intervention) and the 2010 (post-intervention) farming seasons.
The point estimates from our preferred specification suggest that
the returns to bed nets in the study sample were large: on aver-
age, we find that access to free bed nets (three nets for a typical
household) increased agricultural output by US$ 76, which corre-
sponds to 14.7% of the average annual harvest value. To address
omitted variable bias concerns, we include a large set of covariates
in our empirical models, and run an extensive series of robustness
and heterogeneity checks. Overall, treatment effects appear largest
among more educated farmers as well as farms with more diver-
sified portfolios, and larger for cotton (as the more labor intensive
crop) than for maize.
In the last part of our analysis, we explore potential mechanisms
underlying the productivity impacts observed. Unfortunately only
limited and self-reported data on malaria incidence (and no data
on parasitemia or asymptomatic malaria) was collected as part of
this project. However, the general patterns observed in the data
suggest that the programs likely induced substantial reductions in
the days of field work lost due to ill health. Given that full recov-
ery from acute malaria is often slow, reduced exposure to malaria
can increase the marginal product of labor (Nur, 1993), particu-
larly in cases where malaria induces anemia (Ehrhardt et al., 2006).
While there is theoretically also the possibility that the reduced
exposure to ill health may have been associated with a reduction
in direct medical expenditure, most malaria treatment in the area
appears to be provided for free, so that no evidence of lower health
expenditure was found.
Even though this paper is to our knowledge the first one using
experimental data to evaluate the productivity effects of malaria,
several studies have analyzed agricultural output in the context
of nutrition and other diseases. Following the initial work by
Strauss (1986) as well as Pitt and Rosenzweig (1986), Behrman
et al. (1997) document a rather robust association between
nutritional improvements and production in agricultural settings.
Loureiro (2009) and Ulimwengu (2009) find positive associations
between health and productivity using stochastic frontier regres-
sion techniques. Audibert and Etard (2003) examine the effect of
schistosomiasis among rice-growers, and find that exposure to
schistosomiasis reduces production by 26%. Fox et al. (2004) ana-
lyze the productivity declines associated with HIV positivity, and
find that HIV-positive workers earn on average 16–17% less over a
two year period. Similarly, Baranov et al. (2012) show that maize
production increases by up to 31% with HIV treatment, and attribute
this increase to increased overall labor supply and improved phys-
ical and mental health. Similar effects were, however, not found
for iron supplementation and deworming among tea pluckers in
Bangladesh (Gilgen et al., 2001). Most similar to the results pre-
sented in this paper are two cross-sectional studies using harvest
data to compute the agricultural output effect of malaria: Girardin
et al. (2004) analyze vegetable farming in Côte d’Ivoire, and find that
farmers who reported being sick more often had 47% lower yields.
Morel et al. (2008) use total farming output to quantify the agricul-
tural loss generated by work days lost due to malaria in Vietnam,
and find an average cost of US$ 11 per case of malaria, suggesting
returns to malaria prevention similar to the ones identified in this
paper. Conceptually, an overwhelming majority of this literature
suggests strong links between health and agricultural production in
low-income setting; this suggests that household production, labor
and consumption decisions are generally not separable (Benjamin,
1992), a finding which is also supported by recent evidence from
Zambia (Fink et al., 2014).
While this study primarily focuses on household-level out-
comes, the results presented here naturally also link to the broader
literature on the relation between health and income. Most of the
micro-level literature in this area has focused on the long term
benefits of improved childhood health in terms of education and
labor market outcomes (Bleakley, 2007; Bleakley and Lange, 2009;
Clarke et al., 2008; Kremer and Miguel, 2004). This paper highlights
a more immediate and direct effect of health on income similar
to the results shown in Thomas et al. (2010) for iron supplemen-
tation; this effect will clearly not apply in all low resource settings,
but may be of particular importance among rural and frequently
impoverished populations.
The rest of the paper is structured as follows: we provide a
detailed description of the study site and local agriculture prac-
tices in Section 2. In Section 3, we present the study design and
provide details on study implementation. In Section 4, we ana-
lyze the effects of the bed net programs on net ownership and net
usage. In Section 5, we estimate the impact of the net programs on
productivity. Section 6 shows some evidence on the mechanisms
underlying the main productivity results. We conclude with a short
summary and discussion in Section 7.
2. Study background
Fig. 1 shows the geographic location of the study site within
Zambia. Katete district is one of eight districts within Zambia’s East-
ern Province. Eastern Province is one of the least developed regions
of Zambia, with a majority of the population living below the one-
dollar-per-day poverty line, and an estimated under-5 mortality
rate of 151 per 1000 live births (Macro International, 2007). Katete
district is similar in its topography to the Western part of Malawi,
which is located about 100 km east of the district. The current
district population is estimated at 250,000, approximately half of
which live in the urban centers of Sinda and Katete (Zambia Central
Statistic Office, 2011a).
Malaria is endemic in most parts of Zambia, and the primary
cause of short term morbidity in the country (Zambia Ministry of
Health, 2012). The regional climate displays pronounced seasonal
fluctuations, with virtually no rainfall from May to November, fol-
lowed by a period of major rainfall from December to April. The
strong seasonal patterns are directly reflected in the seasonal fluc-
tuations of malaria. Malaria in the area is considered endemic and
seasonal, with a majority of the transmission occurring between
December and May, when continued rainfalls support the breed-
ing of the Anopheles mosquito larvae. According to the latest round
of the Malaria Indicator Survey, Eastern region is among the areas
 G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164
Fig. 1. Zambia (white) and Katete District (shaded red). (For interpretation of the references to color in this text, the reader is referred to the web version of the article.)
with the highest parasite prevalence in the country, with parasites
detected in 22 percent of children under the age of five in early
2010 (NMCC, 2011). Fig. 2 shows the clinical burden of Plasmodium
falciparum malaria in terms of the number of clinical cases per year
Fig. 2. Regional distribution of malaria incidence.
Source: Malaria Atlas Project.
153
and per 5 km2 as computed by the Malaria Atlas Project in 2007
(Hay and Snow, 2006). Katete district ranks among the most highly
exposed areas of the country, with an estimated annual burden of
approximately 500 cases per 5 km2 .
Since 2006, Zambia has made major efforts to reduce the burden
of malaria (Ashraf et al., 2010; Zambia Ministry of Health, 2006).
As part of the internationally supported Rollback Malaria Initia-
tive, four principal strategies have been employed by the country
through the National Malaria Control Centre: indoor-residual
spraying (for densely populated and primarily urban areas), mass
distribution of long-lasting insecticide treated nets (LLITNs), inter-
mittent preventive treatment of malaria in pregnancy (IPTp) and
case management through diagnostics and artemisinin-based com-
bination therapies (NMCC, 2007, 2009, 2010; Zambia Ministry of
Health, 2006). Between 2006 and 2008, 96,000 LLITNs were dis-
tributed in Katete district (NMCC, 2011). At the time of the 2008
Malaria Indicator Survey, the average number of LLITNs in Eastern
Province was 0.96 nets (NMCC, 2009), a level very similar to the
one observed in the study area at the beginning of the study in
December 2009.
The rural part of Katete targeted in this study is sparsely popu-
lated, with clusters of family-run farms grouped into small villages.
With an average size of approximately 4 ha (10 acres), the typi-
cal farm is small, and most planting and harvesting done without
machinery. Farm land is generally owned by communities, who
allocate the land to families via local headmen and chiefs. The
amount of farm land families can get access to is – at least the-
oretically – not limited; any individual can claim additional land
from the chief as long as they can show they have the manpower
and skills to use the land (Nolte, 2012).
3. Study design, randomization and descriptive statistics
3.1. Intervention background
As stated above, small-scale farming constitutes the primary
source of nutrition and income in the region. Cotton is the main
cash crops in the area, and sold to multinational cotton buyers (“gin-
ners”), who process the cotton and sell it on international markets.
154 G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164
Fig. 3. Cluster-level group assignment.
To enhance productivity and strengthen commercial links with
farmers, cotton ginners have set up a variety of agricultural loan
schemes, which allow farmers to receive cotton seeds and “chem-
icals” (fertilizer and pesticides) as well as agricultural machinery
throughout the planting and growing season on a loan basis. Upon
receipt of the cotton, ginners deduct the outstanding loan amount
from the final sales price, and pay out the remaining balance in cash.
Agricultural loans are generally provided free of interest, and are
given under the assumption that farmers will sell their harvest to
the cotton company offering them the agricultural loan. At the time
of the study, our partner organization (Dunavant Cotton) was pro-
viding loans to approximately 80,000 cotton farmers across Zambia.
According to the latest national estimates, 67% of the Zambia labor
force is employed in agriculture (Zambia Central Statistic Office,
2011b), which corresponds to approximately 1.3 million farming
households. The share of farmers working with Dunavant is rel-
atively small (approximately 6% of all farming households), both
because Dunavant is only one of about 10 cotton ginners in the
country and because only about 25% of small-scale farmers in the
country grow cotton (Goeb, 2011).
3.2. The intervention
In order to assess farm’s willingness to pay for nets within
existing agricultural loan schemes, two different net programs
were implemented as part of the experiment: a free net program
and a bed net loan program. Under the free net program, selected
farmers were allowed to obtain one free bed net for each uncov-
ered sleeping space in the household. Clusters in the loan arm
were assigned to one of two loan types: a “full price loan”, which
required farmers to repay the full price3 of the net (Zambian
Kwacha (ZK) 25,000, US$ 5) at the end of the harvesting season,
3
“full price” we charged reflects the current wholesale price, which is about 30%
below regional retail prices of about ZK 35,000 (US$ 7).
and a “subsidized price loan”, which required farmers to repay
ZK 12,500 (US$ 2.5) at the end of the harvesting season. Both
prices are substantially higher than what households reported
to have paid for bed nets in the past; 90% of bed nets found in
households at baseline were received through free governmental
or NGO programs (most likely received as part of the large national
programs); 7% had paid ZK 3000 (US$ 0.6) – the price public health
facilities charged prior to the national mass distribution, and the
remaining 3% reported to have paid a price between ZK 5000 and
ZK 10,000 acquiring nets through door-to-door sales.
All nets were distributed at the beginning of the weeding season
(late December) in order to provide protection for farmers during
the peak malaria season (December–May).
3.3. Sampling frame, enrollment and program assignment
The sampling frame for the study was provided by Dunavant
Zambia. Dunavant has nine regional offices, which distribute farm-
ing inputs and acquire cotton through local sheds and distributors.
Each distributor handles between 10 and 50 farmers in his commu-
nity. At the time the study was launched, Dunavant was working
with 96 distributors in the study area. In order to reduce the risk
of local spillovers, we restricted the sample to distributors operat-
ing in spatially separated areas, with a minimum distance of 3 km
between any two locations. This left us with a final sample of 49 dis-
tributors and their respective villages. On average, each of the 49
distributors was working with about 20 farmers, a listing of which
was provided to the study by Dunavant. We randomly selected 11
farmers from each distributor, and visited them for a baseline inter-
view in December 2009. Only the 11 farmers selected for the study
were allowed to receive the free or subsidized nets; with an average
village size of about 50 households, this means that the program
covered about 20% of the average village population. Out of 539
farmers invited to participate in the study, 516 farmers (95.7%)
were enrolled in the study, and completed the baseline interview
in December 2009.
 G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164 155

Table 1
Descriptive statistics.

Control (N = 153) Loans (N = 185) Free nets (N = 155) Equal means test (p-value)a

Mean St. dev. Mean St. dev. Mean St. dev. Control vs. Control vs. Loan vs. All Means
Loans Free nets Free nets equal

Farmer age 39.24 12.59 41.32 13.24 39.70 12.90 0.27 0.80 0.31 0.50
Farmer is married 0.83 0.38 0.83 0.37 0.88 0.32 0.96 0.21 0.15 0.30
Farmer years of education 4.34 3.41 4.14 3.66 4.21 3.88 0.63 0.81 0.90 0.89
Members under age 5 0.87 0.88 0.95 1.00 0.86 0.86 0.55 0.93 0.39 0.66
Members age 5–14 1.40 1.36 1.88 1.50 1.92 1.55 0.00 0.00 0.82 0.00
Members age 15–59 2.44 1.22 2.76 1.53 2.84 1.58 0.09 0.03 0.71 0.07
Members age 60+ 0.18 0.49 0.22 0.51 0.17 0.44 0.54 0.97 0.51 0.78
Chicken, geese and ducks 4.58 6.20 6.89 8.45 6.26 7.96 0.01 0.07 0.49 0.01
Goats, pigs and sheep 2.57 3.40 3.41 4.18 3.00 3.68 0.12 0.35 0.41 0.29
Cows 1.47 3.51 2.34 4.09 1.92 3.24 0.07 0.32 0.37 0.19
Bicycles 0.77 0.56 0.93 0.81 0.87 0.60 0.02 0.22 0.44 0.08
Mobiles 0.23 0.45 0.36 0.80 0.35 0.75 0.14 0.17 0.95 0.23
TVs 0.07 0.25 0.14 0.37 0.06 0.27 0.04 0.98 0.05 0.08
Mosquito nets 1.09 1.33 1.53 1.25 0.48 0.69 0.08 0.01 0.00 0.00
Cars, tractors and trucks 0.00 0.00 0.06 0.43 0.03 0.20 0.05 0.05 0.33 0.03
Maize planting area (ha) 1.78 1.17 2.16 3.16 1.86 1.13 0.19 0.71 0.27 0.38
Cotton planting area (ha) 1.36 1.09 1.38 1.25 1.17 0.73 0.94 0.21 0.32 0.33
Other crops (ha) 0.58 0.77 0.73 0.84 0.83 0.93 0.32 0.15 0.55 0.39
Cotton harvest 2009 (bales) 9.46 8.59 12.16 11.12 11.60 11.33 0.09 0.19 0.75 0.19
Maize harvest 2009 (bags) 20.13 15.62 29.83 35.96 28.16 25.98 0.01 0.00 0.68 0.00
Total harvest value 2009 (US$) 453.2 304.0 649.0 589.1 613.8 476.0 0.00 0.00 0.63 0.00

Notes: Based on 493 observations with complete information. All variables reflect baseline conditions as collected in December 2009.
a
p-Values based on cluster-bootstrapped standard errors. Each cluster corresponds to one distributor and 11 randomly selected farmers working with the distributor.

In order to accelerate the distribution of bed nets,4 bed net loan
programs were randomized prior to the collection of baseline data
at the distributor level. Randomization was done using a simple
random number draw generated by Stata. Out of the 49 eligible
distributors, 15 were assigned to the control group (30%), and 15
distributors (30%) were selected for the free net program. Since we
were particularly interested in the loan group and wanted to assess
differences in uptake with and without subsidy, a slightly larger
number of distributors (20% of distributors in each loan program)
were randomized into the net loan programs. The spatial distribu-
tion of treatment assignment is illustrated in Fig. 3. All farms in the
net program arms were informed about the programs at the end of
the baseline interview, and given 48 h to decide on the number of
nets they wanted to receive. Ordered nets were delivered within 10
days of the baseline survey, between December 20 and December
31, 2009.
A first follow-up or midline survey was conducted in April 2010,
during which information on recent illness episodes was collected.
The endline survey was conducted in July and August 2010 with a
primary focus on harvest outcomes and harvest sales.
Out of the 516 farmers initially enrolled in the study, 510 (98.8%)
were followed up successfully throughout the subsequent farm-
ing and harvesting season. One farmer passed away, three farmers
moved, and two farmers refused to participate in the follow-up
surveys. An additional two surveys were excluded from analysis
due to missing planting and harvesting information. Sixteen fur-
ther surveys have missing values on at least one of the extended
list of covariates used in some of the specifications, resulting in a
final analytical sample of 493 farmers.
3.4. Descriptive statistics
Table 1 shows descriptive statistics for the households enrolled
and followed up in the study by study arm. Average plot size was
4.15 ha (median 3.1) in 2009, and average harvest value in 2009
was US$ 577 (median US$ 463). With an average household size
of close to six members, this implies average per-capita resources
of approximately US$ 0.26 per day, placing the majority of these
households well below the international US$ 1.25 dollar per day
poverty threshold, even when input-related expenses (such as cot-
ton loans) are not accounted for. Cotton farmers are on average
slightly larger than other farms;5 at the national level, the average
plot size among small- and medium scale (non-commercial) farm-
ers is 3.1 ha (Jayne et al., 2008). On average, farms owned one bed
net at baseline; with a mean of three sleeping spaces per house-
hold, this implies that two thirds of household members were not
covered by nets at the beginning of the study.
While the randomized assignment of net programs across clus-
ters generated a fairly balanced sample with respect to household
head characteristics, the same was unfortunately not true for farm
size, with farms in the free net and net loan arms on average larger
and more productive than farms in the control group. As Table 1
shows, the largest and most productive farms were found in the
net loan group, followed by the free net group. Detailed data col-
lected on 2009 harvest outcomes suggests that the average value
of farm production (sum of all crops harvested multiplied with the
median sales prices of the respective crops in the area in 2009 – see
Table 1 for further details) was US$ 453 in the control group, US$
614 in the free net group, and US$ 649 in the net loan group. Similar
differences were found for household size: on average, households
in the loan and free net groups had 0.8 and 0.9 more household
members in the 5–59 age range than households in the control
group.
These differences between farmers in the control and farmers in
the two intervention groups are large and statistically significant,
and complicate statistical inference, since endline differences will
be at least be partially attributable to differences observed at base-
line. The imbalance does not appear to be driven by differences in
the spatial distribution or by spatial clustering of farms, but rather

4
Due to delays in funding and IRB approval, baseline surveys got pushed back
to December, so that we decided to do the distribution of nets immediately after
5
baseline to make sure farmers would benefit from them during the peak rainy Cotton ginners recommend to use at least one hectare of land for growing cotton,
season. which means that farms growing cotton rarely use less than two hectares of land.
156 G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164

Table 2
Program effect on ownership and coverage panel a: ownership and usage of nets.

Panel A: ownership and usage of nets

Number of nets Number of nets Number of nets Number of nets Number of nets Number of nets
received through owned at used at received through owned at used at
program endline endlinea programb endlineb endlineb

Loan program 0.811*** 0.992*** 0.823*** 0.676*** 0.710*** 0.555***

(0.147) (0.199) (0.189) (0.150) (0.168) (0.157)
Free nets 2.413*** 1.626*** 1.517*** 2.052*** 1.571*** 1.461***
(0.121) (0.190) (0.185) (0.113) (0.137) (0.141)
Control group 0.00 1.05 0.96 0.00 1.05 0.96
average
Controls No No No Yes Yes Yes
included
Observations 493 493 493 493 493 493
R-squared 0.46 0.23 0.20 0.64 0.45 0.46

Panel B: bed net coverage

Fraction of nets Fraction of Probability of Fraction of nets Fraction of Probability of

used at endlinec sleeping spaces not having any usedb sleeping spaces not having any
covered space covered coveredb space coveredb

Loan program −0.0479 0.243*** −0.279*** −0.0459 0.248*** −0.258***

(0.0318) (0.0687) (0.0700) (0.0364) (0.0667) (0.0688)
Free nets −0.0176 0.438*** −0.366*** −0.00924 0.472*** −0.383***
(0.0310) (0.0625) (0.0654) (0.0359) (0.0587) (0.0674)
Control group 0.93 0.41 0.39 0.93 0.41 0.39
average
Controls No No No Yes Yes Yes
included
Observations 419 492 493 419 492 493
R-squared 0.01 0.20 0.15 0.07 0.25 0.21
a
Usage is defined as the number of nets observed hanging by interviewers as part of the endline survey in July 2010.
b
Specifications control for plot sizes used for cotton, maize and other crops in 2009 as well as in 2010, family members under 5, family members 5–14, family members
60 and older, household wealth, mosquito nets at baseline, ownership of chickens, goats and sheep, farmer age, education and marital status, total maize harvest 2009, total
cotton harvest 2009 and total economic value of production 2009.
c
Fraction of nets used not defined for households without a bed net. Cluster-bootstrapped standard errors in parentheses.
p-Values based on cluster-bootstrapped standard errors. Each cluster corresponds to one distributor and 11 randomly selected farmers working with the distributor.
***
p < 0.01, ** p < 0.05, * p < 0.1.

seems to reflect presumably random variations in productivity level
across farmers.6 While a large number of meta-reviews in the medi-
cal literature suggest that results from imbalanced trials controlling
for baseline characteristics do on average not yield different results
from fully balanced trials (Berger, 2010; Knottnerus and Tugwell,
2012; Riley et al., 2013), there is clearly a concern that observable
differences may be correlated with other unobservable character-
istics such as malaria knowledge or farming skills. To deal with
these concerns, we follow the approach proposed by Glennerster
and Takavarasha (2013) as well as by Bennett et al. (2014) and
estimate both models where we control for lagged dependent vari-
ables and models where we use differences in outcome measures
between the 2009 (pre-intervention) and 2010 (intervention) sea-
sons as the dependent variable. These models exclusively identify
changes in the outcome measures over time, and thus directly
eliminate confounding or omitted variable bias concerns due to
time-invariant farm-specific differences prior to the intervention.
The resulting estimates will yield unbiased program impact assess-
ments as long as the random treatment assignment is not correlated
with changes in unobservable characteristics between baseline and
endline conditional on initial values, which seems reasonable given
the relatively short time period analyzed.
4. Program impact on bed net ownership, usage and
coverage
Table 2 shows the main results for bed net ownership and usage.
As described above, endline surveys were conducted in July and
August 2010. During both visits, net status was verified by inter-
viewers. We consider a bed net “used” if the net was observed
hanging by interviewers during the visit.7 In Panel A of Table 2,
we show the impact of the net programs on the number of nets
received (columns 1 and 4), the number of nets owned at end-
line (columns 2 and 5) and the number of nets in use at endline
(columns 3 and 6). In Panel B of Table 2 we show the impact of
the two treatments on bed net coverage in the household, i.e. the
number of nets hanging relative to the number of sleeping spaces
used by the household. In columns 1–3 of both panels we show
unadjusted differences between the three groups; in columns 4–6,
we show estimates with a full set of baseline covariates to con-
trol for pre-treatment differences in household size and bed net
ownership.
As documented in previous studies (Cohen and Dupas, 2010;
Dupas, 2014; Tarozzi et al., 2014), the demand for bed nets is highly
price elastic; on average, households in the loan group obtained
0.8 nets, compared to 2.4 nets in the free net group. It is worth

7
Nets are generally tied to a knot during the day to keep them clean, which means
6
The baseline table looks virtually the same when specific areas (Eastern, Western that observing a net as hanging does not necessarily mean it was used the previous
or central parts) or villages are excluded. night.
 G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164
highlighting that despite the high price elasticity, demand is strictly
positive in this setting with credit financing even when the full
price of the net is charged. This is rather different from the zero
demand found for nets with prices over US$ 1 in settings where net
acquisition requires an upfront cash payment (Cohen and Dupas,
2010; Dupas, 2014), but similar in magnitude to recent work by
Tarozzi et al. (2014) who find that 52% of household purchase bed
nets when financing options are provided.
Given that households in the control group had more nets at
baseline, the differences in ownership at endline are smaller than
the differences in the number of nets received; on average, house-
holds in the loan group owned 0.9 more nets than households in
the control group, while households in the free net group owned
1.6 nets more than households in the control group.
Consistent with Cohen and Dupas (2010) as well as Tarozzi et al.
(2014), we found no effect of net pricing on usage. On average,
90% of nets owned were actively used by the household during our
second follow-up, with no differences in utilization rates between
intervention and control groups. The remaining nets were generally
found stored for future use in the households, which is consis-
tent with the generally high levels of appreciation of nets in this
population.
In terms of sleeping space coverage, both treatments had a size-
able impact. As shown in Panel B of Table 2, 41% of sleeping spaces
were covered on average at endline in the control group; this frac-
tion increased to 65% in the loan group, and to 88% in the free net
group. Along the same lines, the likelihood of no sleeping space
being covered by a bed net decreased from 39% in the control group
to 13% in the loan group and to less than 1% in the free net group.
5. Impact on agricultural production
The main hypothesis investigated in this experiment is whether
short-term fluctuations in labor supply generated by ill health lead
to lower agricultural output. To measure production, we focus on
three different measures: maize harvest, cotton harvest, and total
production value. Maize is by far the most common crop in the
country, and grown on approximately 50% of all plots in the sample.
Maize is traded in standard bags of 50 kg, which makes measuring
total production of maize relatively easy. The second most common
crop in our sample is cotton. Cotton is used as a cash crop, and,
as described above, sold to cotton ginners at the end of the har-
vesting season. While ginners pay by kilogram (2009 prices were
US$ 0.30 per kg), farmers generally delivery cotton in large bags,
which are referred to as “bales”, and generally contain about 80 kg
each. Even though cotton and maize account for the large major-
ity of farming land and production in this sample, most farmers
use small plots to grow a variety of other plants such as sunflow-
ers, beans, groundnuts, and sweet potatoes. The diversity in crop
portfolios means that crop-specific quantities cannot easily be com-
pared across farmers or treatment groups. On the other hand, not
accounting for the resources generated by these crops would clearly
mean that the effects of additional labor inputs may not be fully
captured.
To measure total farm production, detailed harvest information
was obtained from all crops, and then converted into monetary val-
ues using the median 2009 market prices reported among farmers
who sold the respective crops. In theory, one may wish to use farm-
specific crop prices to account for differences in market access or
production quality; in practice, this is unfortunately not feasible
since a large number of farms do not sell specific crops at all, but
rather use it for their own consumption.8
8 10
The only cash crop in the sample is cotton; all other crops are mostly used for
own consumption, with some occasional sales to cover additional cash needs.
157
In order to have a comprehensive measure of farm productivity,
we defined the total economic value of production (TEVP) as

8
TEVP = qi P50,i (1)
i=1
where qi is the quantity of crop i and P50,i is the median price for
one unit of the respective crop.9 In the study sample, eight crops
were grown: cotton, maize, ground nuts, sweet potato, sunflower,
soy beans, tomato and cassava. In general, price variations across
farmers were low: for the two major crops (cotton and maize) prices
are negotiated and established at the national level by the govern-
ment. For some of the minor crops like sunflowers and soy beans,
prices are established at local markets. For all crops, individual
prices reported rarely deviated by more than 10% from the most
commonly reported market prices.
For our analysis, we focus on cotton and maize as the two most
commonly grown crops as well as the aggregate TEVP variable. Fig. 4
shows kernel density estimates of all three variables in 2010 on an
absolute (top panel) and logarithmic (bottom panel) scale.
The average cotton harvest in 2010 was 6 bales, which corre-
sponds to about 480 kg of cotton harvested with an average plot
of 1.3 ha. Average maize harvest was 28 bags, which implies an
average yield of about 1400 kg for an area of 2 ha. Both yields
are small when compared to industrial farmers, who frequently
achieve yields of over 10 tons of maize per hectare (13 times the
sample average) and over 2 tons of cotton per hectare (about 9 times
the sample average).10 The average total economic value of all crops
harvested in 2010 was 2.6 million Kwacha (US$ 517).
5.1. Empirical strategy
The basic empirical (intention-to-treat) model we estimate to
identify the treatment effects of interest is given by
yij = ˛ + Tj ˇ + Xij  + εij , (2)
where yij is the harvest outcome of interest observed for farm i
in cluster j in 2010, Tj is a vector of indicator variables capturing
the distributor-level treatment assignment, and Xij is a vector of
baseline covariates. Given the limited use of fertilizer in the region,
yields tend to display mean-reverting patterns over time, with good
yield years depleting the soil and being followed by less productive
years. Fig. 5 illustrates these patterns, showing the year-over-year
change in output as a function of total output in the 2009 (pre-
intervention) period.
To control for baseline differences as well as the observed mean-
reverting patterns, we first estimate models where we including
the lagged (2009) value of the outcome variable of interest. With
lagged outcome variables, the main estimated equation becomes
yijt = ˛ + yijt−1 + Tj ˇ + εij , (3)
where yi,t−1 is the lagged value of the dependent variable, i.e. the
harvest outcome in the 2009 season. Following the approach taken
in Bennett et al. (2014), we also test an alternative model where we
take changes in the outcome variables as dependent variable, and
control for an extensive set of baseline covariates, which is given
by
yij = ˛ + Tj ˇ + Xij  + εij . (4)
9
We assume that the quality of the produced crops are comparable across farm-
ers; this assumption is empirically always true for cotton and maize (were prices
are fixed), but may not necessarily reflect prices for local small-quantity trades.
See http://www.indexmundi.com/agriculture/ for a country ranking for crop
productivity.
158 G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164

Cotton (bales) Maize (bags) Total value (ZKR)

.0001 .0002 .0003 .0004

.03
.1
.02 .04 .06 .08

.02
Density

Density

Density
.01
0

0
0 20 40 60 0 100 200 300 0 5000 10000 15000 20000
kernel = epanechnikov, bandwidth = 0.9652 kernel = epanechnikov, bandwidth = 4.4401 kernel = epanechnikov, bandwidth = 355.2049

Ln(bales of cotton) Ln(Bags of maize) Ln(Total Value (ZKR))

.8

.6

.6
.6

.4

.4
Density

Density

Density
.4

.2

.2
.2
0

0
0 1 2 3 4 0 2 4 6 4 6 8 10
kernel = epanechnikov, bandwidth = 0.1381 kernel = epanechnikov, bandwidth = 0.1921 kernel = epanechnikov, bandwidth = 0.1668

Fig. 4. Agricultural outcomes: Kernel density estimates. Notes: total values are in rebased Zambian Kwachas (ZKR); one rebased Kwacha corresponds to 1000 “old” Zambian
Kwachas.

In this specification, yij is the change (difference) in the out- in total harvest value across the three groups are large, with farms
come between the 2009 (pre-intervention) and the 2010 harvests, in the loan and free net groups showing additional yields of US$
and Xij is a vector of baseline covariates. Table 3 shows the main 180 and US$ 156, respectively (Panel A, column 3). Given the large
treatment effect results. In Panel A of Table 3 we show uncondi- differences in total yields at baseline, a substantial fraction of this
tional differences across the three groups. The observed differences differential is clearly attributable differences in baseline covari-
ates. In Panel B of Table 3, we directly control for these baseline
differences in productivity by including lagged dependent vari-
ables in our model as outlined in Eq. (2). As expected, the lagged
300

dependent variables are highly significant in all models, and explain

a substantial fraction of the unadjusted differences observed in
Year-over-year change in output (%)

Panel A. Consistent with the convergence patterns seen in Fig. 5,

all coefficients on lagged variables are strictly positive and smaller
200

than one. The adjusted model suggests that the loan program
increased total harvest value by an average of US$ 69, while the
free net program increased yields by about US$ 65 – only the lat-
100

ter is marginally significant (Panel B, column 3). Once we estimate

Eq. (4) and control for baseline covariates in Panel C of Table 3,
we get smaller point estimates for maize, and larger point esti-
mates for cotton. Average cotton yields declined substantially from
0

2009 to 2010 (from 11.1 to 6.3 bales), which appears to be mostly

attributable to less favorable rain patterns. Declines were substan-
tially smaller in the intervention groups.
-100

In relative terms, this implies that the net programs increased

0 500 1000 1500
cotton yields by about 25% (1.3 additional bales relative to the
Total harvest value 2009 in USD
control group average of 5.2), while maize yields increased by 6%
(loans) and 12% (free nets) compared to the control group average
Fig. 5. Agricultural outcomes: year-over-year mean reversion. Notes: the percent-
of 22 bags. The relatively larger impact on cotton is consistent with
age (year-over-year) changes in output value between the 2009 and 2010 harvest
seasons as a function of the total 2009 harvest value in US$. the idea of cotton being the more labor-intensive crop; it could
 G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164 159

2000

1500
Production Value 2010 (US$)

Production Value 2010 (US$)

1500

1000
1000

500
500 0

0
0 500 1000 1500 2000 0 500 1000 1500 2000
Production value 2009 (US$) Production value 2009 (US$)

Control Loans Control Free nets

95% CI 95% CI

CONTROL VS. LOAN CONTROL VS. FREE NETS

Fig. 6. Fractional polynomial predictions: farm yields 2010 as function of 2009 farm yields.

also be interpreted as evidence of households prioritizing maize

Table 3
as the primary food crop, and therefore reducing (increasing) labor
Productivity impact. inputs on cotton rather than maize in cases where labor constraints
become (less) binding. Overall, the fully adjusted models shown in
Panel A: Unadjusted
Panel C suggest that the net loan program increased total harvest
Dependent Bags of maize Bales of cotton Total harvest value by (a not statistically significant) US$ 45, while the free net
value program increased harvest value by US$ 76 (Panel C, column 3).
(1) (2) (3) To provide a more direct sense of how overall production was
Loan program 10.25*** 1.917* 179.9*** affected by the program, we plot 2010 yields as fractional polyno-
(3.826) (1.101) (54.71)
mial functions of 2009 yields by treatment arm in Fig. 6. As the
Free nets 8.611*** 1.179 155.9***
(3.057) (0.751) (44.38) figure illustrates, farmers in the loan and free net groups did con-
Control group average 22.1 5.2 202.0 sistently better in 2010 than farmers in the control group. These
R-squared 0.022 0.014 0.032 differences appear to be larger for very small farmers (less than
Baseline covariates No No No US$ 300 production value) and appear to be particularly large for
farms at the upper end of the yield distribution with more than $
Panel B: Controlling for lagged dependent variables
1000 production value in 2009.
Dependent Bags of maize Bales of cotton Total harvest Given that the treatment assignment was made at the distribu-
value
tor level, one may find it more intuitive to analyze year-over-year
Loan program 4.107 1.197 69.25 changes at the distributor (cluster) level. In Fig. 7, we show the
(2.839) (0.799) (44.34) cluster-level distribution of year-over-year changes. Each obser-
Free nets 3.521 0.610 65.08*
(2.394) (0.693) (36.83)
vation in the plot corresponds to the average absolute change in
Lagged dependent 0.633*** 0.266*** 0.565*** production value in the cluster. The boxes at the center of the figure
(0.0916) (0.0899) (0.0731) show the median, 25th and 75th percentile for villages in each arm;
Control group average 22.1 5.2 202.0
R-squared 0.369 0.188 0.409
Baseline covariates No No No
300
YoY Change in Yields (Cluster Average)

Panel C: Differences in outcomes with full set of covariates

200

Dependent Bags of maize Bales of cotton Total harvest

value

Loan program 1.439 1.371* 45.82

100

(2.972) (0.737) (42.83)

Free nets 2.747 1.366* 75.60**
(2.701) (0.733) (35.85)
Control group average 1.99 −4.23 −51.19
0

R-squared 0.331 0.688 0.438

Baseline covariates Yes Yes Yes
-100

Notes: Specifications in Panel A do not include covariates. Specifications in Panel

B control for lagged dependent variables only. Specifications in Panel C control for
lagged dependent variables as well as for plot sizes used for cotton, maize and other
-200

crops in 2009, family members under 5, family members 5–14, family members
60 and older, household wealth, mosquito nets at baseline, ownership of chickens,
goats and sheep, farmer age, education and marital status. Cluster-bootstrapped
Control Loans Free nets
standard errors in parentheses. Each cluster corresponds to one distributor and
11 randomly selected farmers working with the distributor. *** p < 0.01, ** p < 0.05,
* p < 0.1 Fig. 7. Changes in farm yields at the cluster level.
160 G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164
the outside whiskers show the lower and upper adjacent values11
(Frigge et al., 1989) While the large majority of clusters in the con-
trol arm fared worse in 2010 than in 2009 (negative year-over-year
change), year-over-year changes were positive for a majority of vil-
lages in the loan group, and were positive for close to 75% of villages
in the free net group.
In order to make sure these results are not driven by individ-
ual farmers or clusters, we run a series of robustness checks in
Table 4, based on the fully adjusted specifications shown in Panel
C of Table 3. In column 1 of Table 4, we exclude households with
heads who did not receive any schooling. In column 2, we restrict
our analysis to farms with a medium degree of crop diversifica-
tion (3–5 crops) to ensure that the year-over-year changes are not
affected by differential exposure to price and crop risk. In order
to deal more directly with pre-existing differences in productivity,
we exclude the 20% least productive farmers in the baseline sea-
son from the regressions in column 3; in column (4) we exclude
the most productive farmers in 2009, and last, in column 5, we
exclude both the least and the most productive farmers from the
analysis. Overall, the results appear rather robust, with average pro-
gram estimates between US$ 44 and US$ 79 for net programs, and
estimates between US$ 54 and US 110 for free net programs, both
within the confidence intervals of our main impact estimate of US$
46 and US$ 76 (Table 3, Panel C). In terms of program impact, largest
effects are observed for more educated and diversified households;
in terms of baseline yields, effects seem to be larger at the very
bottom and the very top of the distribution, a pattern which is
consistent with the non-parametric estimates shown in Fig. 6.
6. Mechanisms and discussion
The primary mechanisms through which the programs were
intended to affect productivity was household health. In order
to monitor the health of households, two follow-up visits were
conducted, one in April 2010, and one in July/August 2010. Dur-
ing both follow-up visits, interviewers went over the household
roster collected during the baseline assessment, and asked respon-
dents to indicate for each household member whether they had
been sick since the last interview. In a previous paper (Fink and
Masiye, 2012), we used data on self-reported morbidity outcomes
from both follow-ups to identify the health impact of the additional
nets. Given the concerns with longer-term recall data highlighted
in the recent literature (Arnold et al., 2013; Das et al., 2012), we
restricted our analysis to morbidity episodes in the two weeks pre-
ceding the interview. Consistent with the estimates reported in the
bed net literature (Lengeler, 2004), we found large health benefits
attributable to the additional nets, with treated households repor-
ting reductions in the incidence of health problem of any kind of
about 40% and in the incidence of confirmed malaria of at least
50%.12
In order to directly assess the impact of ill health on farm labor
inputs, farmers were asked to report the number of days of field
work list for each illness episode. Specifically, farmers were asked
to report how many days of field work were lost because (i) the
sick person was not able to work and (ii) because somebody else
in the household had to take care of the sick person. Fig. 8 shows
the average number of days of field work lost per malaria episode13
11
Adjacent values are defined as the most extreme values within 1.5 interquartile
ranges of the lower and upper quartile, respectively. See Frigge et al. (1989) for
further details.
12
“Confirmed malaria” was defined as a fever episode where respondents indicate
that they were tested for malaria and that the test was positive.
13
Malaria episodes include all recent illnesses attributed to malaria by respon-
dents (based on the answer provided to the question “What health problem did the
person have”).
Average days of field work lost per bout of malaria
8
7
Days of caretaker field work lost
6 Days of patient field work lost
5
4
3
2
1
0
Under 5 Age 5-14 Age 15-59 Age 60+
Fig. 8. Work days lost by malaria episode.
by age group.14 The average number of days of field work lost by
malaria episode is 4.7; the burden is substantially lower for children
(slightly less than 4 days taken off for taking care of children on
average) than for adults, where the average number of days lost per
episode is about 7, with a majority of days lost directly attributable
to the patient’s absence from the field.
Given that the second follow-up was done after the harvest with
most farmers not working, we further restrict our analysis to illness
episodes reported in the first follow-up. In total, only 85 house-
holds reported illness episodes in the 2-week period preceding the
interview in April, which means that we have only very limited
power to detect effects, despite the substantial differences in the
likelihood of reporting an illness episodes: in the control group, 32
out of 153 (21%) of farms reported an illness episode in the two-
week period prior to the first follow-up. In the loan group, 33 out
of 185 (18%) reported a health problem, and last, in the free net
group, only 20 out of 155 (13%) of households reported an illness
episode; in relative terms, this means that the likelihood of expe-
riencing an acute morbidity episode was reduced by about 40% in
the free net group. In absolute terms, these reductions are of course
relatively small. Over a six-month period, the observed differences
would correspond to approximately one more illness episode per
household.
In Table 5, we analyze the program impact on field work labor,
as well as total health expenditure. In the first four columns, we
show unadjusted treatment effects. In columns 5–8, we show the
same estimates when a full set of household covariates is included;
different from the productivity impact estimates, the inclusion of
covariates does not make much of a difference in these specifica-
tions. On average, we find that net programs reduce the number
of days of field work lost by 0.3 days over a two-week period,
which corresponds to about 3.6 days over the agricultural season.
Given the small number of illness episodes the confidence inter-
vals around this estimate are wide however, and the estimated
coefficients not statistically significant from zero. In terms of the
estimated effects, about 60% of the field work losses appear to result
from the person sick not being able to complete field work, and
40% appear to result from field workers having to take care of other
household members. It is worth pointing out here that these meas-
ures only capture complete absence from field work; no data was
collected on worker strength or ability to work on the fields. In
14
Since no field work was reported during the second follow-up (scheduled in
the post-harvest season in July), all numbers reflect the morbidity reports provided
during the first follow-up round in April.
 G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164 161

Table 4
Robustness checks and heterogeneous treatment effects.

Change in production value 2010–2009 (US$)

Net loans 64.60 43.83 61.11 41.46 77.51

(47.27) (48.78) (53.41) (48.49) (59.06)
Free nets 110.4** 100.8** 67.75 60.23** 53.70
(45.21) (44.85) (44.82) (30.19) (39.66)
Full set of Yes Yes Yes Yes Yes
covariates

Sample Education of Medium Excluding bottom Excluding top 2009 Excluding clusters
head > 0 diversification: 2009 productivity productivity in top or bottom
3–5 crops quintile quintile quintile in 2009
Observations 341 395 386 400 293
R-squared 0.467 0.456 0.470 0.392 0.414

Notes: All specifications control for lagged dependent variables as well as for plot sizes used for cotton, maize and other crops in 2009, family members under 5, family
members 5–14, family members 60 and older, household wealth, mosquito nets at baseline, ownership of chickens, goats and sheep, farmer age, education and marital status.
Cluster-bootstrapped standard errors in parentheses. Each cluster corresponds to one distributor and 11 randomly selected farmers working with the distributor. *** p < 0.01,
** p < 0.05, * p < 0.1

Table 5
Labor supply and health expenditure effects.

Days of field Days of field Total days of Total health Days of field Days of field Total days of Total health
work lost due work lost due field work lost expenditure work lost due work lost due field work lost expenditure
to own sickness to other in last 2 weeks last two weeks to own sickness to other in last 2 weeks last two weeks
in last 2 weeks sickness in last in last 2 weeks sickness in last
2 weeks 2 weeks

Any programa) −0.174 −0.129 −0.303 0.0783 −0.170 −0.0818 −0.252 0.0740
(0.149) (0.129) (0.258) (0.0604) (0.154) (0.117) (0.234) (0.0573)

Loan program −0.206 −0.0764 −0.283 0.0842 −0.182 0.00153 −0.180 0.0972
(0.157) (0.136) (0.257) (0.0676) (0.145) (0.125) (0.239) (0.0668)

Free net −0.134 −0.192 −0.326 0.0713 −0.154 −0.196 −0.350 0.0423
program
(0.170) (0.138) (0.273) (0.0850) (0.192) (0.152) (0.292) (0.0905)

Control No No No No Yes Yes Yes Yes

included

Control group 0.391 0.360 0.752 0.072 0.391 0.360 0.752 0.072
mean

Observations 493 493 493 493 493 493 493 493

Notes: All specifications control for plot sizes used for cotton, maize and other crops in 2009 as well as in 2010, family members under 5, family members 5–14, family
members 60 and older, household wealth, mosquito nets at baseline, ownership of chickens, goats and sheep, farmer age, education and marital status, total maize harvest
2009, total cotton harvest 2009 and total economic value of production 2009. Cluster-bootstrapped standard errors in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1.

highly endemic areas like the one studied, asymptomatic malaria
infections are common, and likely to substantially reduce farm-
ers’ ability to complete field work tasks even when they do not
suffer from acute infections (Nur, 1993), As part of the follow-up
interviews, we also collected information on health expenditure.
Zambia’s health sector – in particular in rural areas – is dominated
by public health facilities which generally provide basic health ser-
vices (including malaria testing and drugs) for free. In our sample,
90% of respondents indicated zero out-of-pocket expenditure for
household members getting treated when sick. Given this, the very
small and not significant estimated program impact on total health
expenditure found in Table 5 is not surprising.
7. Discussion
The estimates presented in this paper suggest a rather large
positive impact of malaria prevention on agricultural productivity.
While the data collected as part of the project does unfortunately
not allow us to precisely identify the mechanism underlying this
impact, increased labor inputs appear to be the most plausible
causal pathway. Given that labor is in principal abundant in Zam-
bia, one might argue that short-term labor supply shocks should
not matter at all for farm production; in settings where labor can
be freely hired, morbidity should not affect farming output, since
any lacking labor could be hired in local markets. During the first
follow-up rounds, we directly questioned farmers about labor sub-
stitution. In total, 167 instances were reported where the head of
household was not able to work on the field because he or she was
sick.15 Out of these 167 episodes, substitute labor was hired only in
10 cases (6%). In 7 out of these 10 substitution cases, farmers found
somebody to work for free; only three farmers reported to pay for
labor, with wages ranging between 0.5 and 11 dollars per day. Anec-
dotally, local piece work (“ganyu”) labor is widely available (Fink
et al., 2014); in practice, most small-scale farmers do however not
appear to have the resources to hire such labor.
Even if one is willing to accept that hiring short-term labor
may be hard for farmers, the estimated impact numbers appear
large. As shown in Goldberg (2014), agricultural wages for day
labor are frequently less than US$ 1 in rural Malawi. In the Zambia
settings, wages appear slightly higher, with median daily wages
varying between KR 12.5 and KR 25 (US$ 2.5–5) reported for the
peak labor season in focus groups. Even at US$ 5, direct labor costs
are unlikely to account for the full differences in output observed.
15
Note that the total number of episodes here related to the four months period
between the baseline and the midline survey, and thus is different from the 85 illness
episodes reported for the two-week period preceding the interview analyzed in the
previous section of the paper.
162 G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164
Our estimates suggest that having all sleeping spaces covered with
bed nets will save the average farm approximately 3–5 working
days, which translates to about US$ 15–25, and therefore to 30%
of the estimated effects at most. Two factors may at least partially
explain the remaining gap: first, recovery from malaria is a slow
process, frequently taking more than two weeks, during which
farm workers are likely less productive (Nur, 1993) even if they
report back to work on the field and thus would not be counted
as “able to work” in our analysis. A second possibility is that farm-
ers in the study may have only reported major health events, so
that the reported numbers do not fully capture the true program
impact. In the 2007 Zambia Demographic and Health Survey (Macro
International, 2007), 20% of children under-5 were reported to have
suffered from fever or diarrhea over the 2-week period preceding
the survey. Even if adults are substantially less prone to be sick, the
reported morbidity prevalence seems very low: 32 episodes across
153 households in the control group implies about one episode for
every 25 individuals or a fever prevalence of about 4%, which is
about one fifth of the under-5 prevalence in the DHS.
One alternative explanation for the relatively large effects
observed are local spillover effects, which could potentially also
undermine the internal validity of the study: as demonstrated
by Apouey and Picone (2014), social interactions in the realm of
malaria and malaria prevention are likely, and may lead to stronger
associations between health behaviors and health outcomes at the
village or regional level. Even though it is quite likely that farmers
interact with other farmers outside the 3 km radius chosen for the
randomization, large spillover effects do not seem very likely in our
setting: first, by working only with small-scale farmers having an
outgrower contract with Dunavant, we covered only a small frac-
tion (<25%) of farmers in each village, which means that changes in
net coverage levels at the village level were relatively small. The-
oretically, our intervention could also have increased knowledge
or improved behavior. However, given that we did not provide any
information or encourage usage of nets at all but simply provided
access to subsidized nets large changes in knowledge or behavior
among untreated farms do not seem likely. In terms of the actual
nets distributed, we closely monitored their usage and ownership,
and found that virtually all nets (97%) remained in the house-
holds who originally acquired throughout the study period, which
makes us relatively confident that direct spillovers to control vil-
lages (nets reaching control villages) did not occur. One could also
have expected the intervention to increase the appreciation and
utilization of bed nets – we do not find any differences in utilization
across groups.
Even though we think that the most likely mechanism from
bed nets to agricultural productivity are the direct and indirect
costs associated with ill-health, one alternative interpretation of
the results presented is that subsidized or free nets constitute an
upfront financial transfer, which allows farmers to spend money
originally earmarked for bed nets on other farming related items.
One could argue that farmers could sell bed nets, and use the
resources for their own consumption or agriculture related expen-
diture. Given the data collected as part of this project, this appears
rather unlikely. As discussed in Section 3 of the paper, less than 10%
of nets available at baseline were actively purchased by farmers;
in general, bed net purchases appear to be more of an exception.
We also checked all households for the nets distributed: as stated
above, out of 547 nets distributed in December 2009, 528 (97%)
were located in the original households16 during the second follow-
up interview in July, so that frequent sales of nets can be excluded
as alternative pathway.
16
Respondents were asked about the source of each net in their household as part
of both follow-up surveys.
Table 6
Treatment and agricultural loan performance.
No payment Partial Cotton sales to
payment Dunavant (kgs)
(4) (5) (6)
Net loan program −0.003 −0.038 98.1
(0.06) (0.02) (92.6)
Free net program −0.001 −0.006 89.3
(0.06) (0.02) (108.6)
Observations 450 450 450
R-squared 0.045 0.064 0.119
Notes: All specifications control for plot sizes used for cotton, maize and other crops
in 2009 as well as in 2010, family members under 5, family members 5–14, family
members 60 and older, household wealth, mosquito nets at baseline, ownership
of chickens, goats and sheep, farmer age, education and marital status, total maize
harvest 2009, total cotton harvest 2009 and total economic value of production 2009.
Robust standard errors in parentheses are clustered at the cluster level.
*** p < 0.01, ** p < 0.05, * p < 0.1.
Given that the field work and net distribution were supported
by Dunavant and respondents may associate interviewers with
the company (even though no information provided by the farmer
was shared with the company), farmers in the loan groups might
selectively under-report their cotton production in order to not
be obliged to repay their full loan to Dunavant. To investigate this
hypothesis, we compared repayment and cotton sales (the amount
of cotton sold from farmers to Dunavant) across the three groups
in Dunavant records.
Out of the 493 farmers in our main analytical sample, admin-
istrative records could be found for 450 farmers (91.3%), with no
differences in tracking rates across arms. Table 6 compares three
outcomes from Dunavant perspective: the likelihood of farmers’
default, the likelihood of a farmer’s partial default, and the total
amount of cotton sold to Dunavant. While all the estimates suggest
that farmers in the two nets arms did slightly better than farmers
in the control arm (consistent with the documented productivity
effects), we do not find any evidence of farmer in the loan program
displaying higher default rates; if anything, the point estimates sug-
gest that farmers in the loan program perform best of all three
groups. It also seems important to highlight that general repay-
ment levels are high, with 84% of farmers fully repaying loans right
after the harvest, and an additional 3% making partial payments.
This suggests that farmers are fully cognizant of their credit com-
mitments, and unlikely to just have signed up for the loan programs
with the expectation of not repaying them.
In terms of cotton production, the overall sales estimates are not
statistically significant, but consistent with the numbers reported
in Table 3. The numbers reported in column 3 of Table 6 indicate
that farmers sold on average an additional 90 kg of cotton to Duna-
vant. Given that each bale corresponds to about 80 kg of cotton,
these numbers are very similar in magnitude to the 1.37 bales
reported in Panel C of Table 3.
8. Summary and conclusion
All results presented from our experiment suggest that preven-
tive health investment in the form of LLITNs can lead to substantial
improvements in agricultural output among small-scale farmers.
The estimates presented in this paper suggest that a typical house-
hold can increase harvest revenues by approximately US$ 76 if it
covers all sleeping places with bed nets. These numbers seem rather
large in absolute terms, but also seem consistent with farmer per-
ceptions: when prompted regarding the expected health benefits
of reducing the burden of malaria at the beginning of the harvesting
season, farmers indicated that their farms would on average have
30% higher yields if nobody in the household would fall sick; given
that bed nets reduce the incidence of fevers by 30–50% (Fink and
 G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164
Masiye, 2012; Lengeler, 2004), the 15% increase in yields observed
in this study appears to be within the range expected by farm-
ers. It is worth highlighting that the study site was intentionally
identified as an area with traditionally high malaria exposure: even
though the overall farm structure in the sample analyzed appears
fairly similar to other small-scale farms in rural Zambia and neigh-
boring countries in terms of their size and profitability, malaria
prevalence and incidence differs largely across regions, and will
almost certainly affect the overall impact of similar programs in
other areas.
Nevertheless, the high expected and realized returns to health
investment documented in this paper naturally raise the question
of why private investment in bed nets remains low. From a private
sector perspective, the provision of bed nets may seem an attractive
investment for larger companies interacting with farmers such as
cotton ginners, both to generate goodwill and higher yields. Follow-
ing the study presented in this paper, our study partner, Dunavant
Cotton, did indeed decide to distribute bed nets to close to half
of the farmers with the support of the World Bank. After an ini-
tial review, Dunavant opted against a continuation of the program
because the overall benefits in terms of contracts signed did not
appear large enough to support net programs on a continued basis
from the ginner’s perspective.17
While ginning companies may not be able to capture a large
enough fraction of additional farming outputs, the limited willing-
ness of farming households to invest in bed nets and their own
health appears a bit puzzling. Even though similarly low farmer
willingness to invest has been documented for other high-yield
investment options such as fertilizer or enhanced farming tech-
nologies (Duflo et al., 2008; Udry and Anagol, 2008), the low levels
of (unsubsidized) adoption appear particularly puzzling in the con-
text of nets. Nets are a technology well-known by farmers, and, if
farmers’ own statements are to be believed, a technology widely
recognized as effective. When prompted at baseline, farmers in the
study indicated that the burden of malaria in their households could
be reduced by approximately 50% if full bed net coverage was avail-
able, a number which largely coincides with the WHO’s current net
effectiveness estimates (Lengeler, 2004). The fact that farmers of
the study area believe in the effectiveness of nets in the area is
also underlined by the fact that more than 90% of nets were used
throughout the study – overall, lacking belief in net effectiveness
does not appear to be a big issue in the studied setting.
Two possible explanations for the lack of preventive investment
in the context of malaria are lacking capital markets or credit con-
straints more generally. The results presented in this study clearly
show some evidence in support of this hypothesis. Compared to the
very small fraction of women purchasing bed nets at prices above
US$ 1 when upfront payments are required (Cohen and Dupas,
2010) the demands for nets appears substantially increased when
financing options are made available as it was done in this study
and in Tarozzi et al. (2014). These increases in demand are both con-
sistent with lacking access to credit (Udry and Anagol, 2008) and
with models of hyperbolic discounting (Duflo et al., 2010); demand
for nets may also have been particularly strong because repayment
was due during the harvesting season, when farmers are relatively
well endowed with cash (Duflo et al., 2010).
However, even when full (zero-interest) financing was offered
to farmers in this study, uptake was only modest. One possible
explanation for this is that farmers may expect to receive bed nets
for free due to the large number of free or highly subsidized distri-
bution campaigns run by the Government of Zambia in the region
over the past years. Dupas (2014) investigates this hypothesis
17
See http://papers.ssrn.com/sol3/papers.cfm?abstract id=2358045 for a more
detailed report on this initiative.
163
empirically, but does not find any evidence for free distribution
campaigns affecting subsequent willingness to pay in the context of
bed nets. A similar argument would be that free nets do not change
willingness to pay directly, but may affect the perceived probability
of receiving free nets in the future. While we cannot fully rule
out this hypothesis, waiting for future distributions appears to be
rather risky as a strategy given that net benefits are immediate and
go substantially beyond the productivity improvements observed.
Under the assumption that farmers are risk-averse, a more plau-
sible explanation for the reluctant uptake of pre-financed nets
might simply be that nets constitute a relatively large investment,
which may not be attractive to farmers even if the mean return
to the investment is positive and large. With an average dispos-
able (cash) income of less than US$ 200 per year, committing to
an end-of-season payment of US$ 10–15 may appear difficult to
farmers given the already high uncertainty faced with respect to
final harvest outcomes. Recent evidence from Ghana suggests that
reducing risk exposure with insurance programs may be central to
increasing farm investment and productivity (Karlan et al., 2012);
further research in this area will be needed to better understand
the determinants of farm behavior and to design optimal policies
in this area.
Acknowledgements
The authors would like to thank the Milton Foundation for
funding this project, as well as Dunavant Cotton and in particu-
lar Rodrick Masaiti for the invaluable logistical support during all
stages of the field work. We would also like to thank Richard Sedl-
mayr and Felix Lam for their input into the study design and the
coordination of the field work, Peter Mulenga for the coordination
of data entry, and Jenny Aker, Nava Ashraf, David Atkin, Jessica
Cohen, Erica Field, Maggie McConnell, Kelsey Jack, Michael Kre-
mer, Zoe McLaren and John Strauss as well as the participants at
the NEUDC conference and the development seminar at Bocconi
University for their comments and suggestions.
References
Apouey, B., Picone, G., 2014. Social interactions and malaria preventive behaviors in
sub-Saharan Africa. Health Economics 23, 994–1012.
Arnold, B.F., Galiani, S., Ram, P.K., Hubbard, A.E., Briceno, B., Gertler, P.J., Colford Jr.,
J.M., 2013. Optimal recall period for caregiver-reported illness in risk factor and
intervention studies: a multicountry study. American Journal of Epidemiology
177, 361–370.
Ashraf, N., Fink, G., Weil, D.N., 2010. Evaluating the effects of large scale health inter-
ventions in developing countries. In: The Zambian Malaria Initiative (Ed.), NBER
Working Paper, vol. 16069. National Bureau of Economic Research, Cambridge,
MA.
Audibert, M., Etard, J-F., 2003. Productive benefits after investment in health in Mali.
Economic Development and Cultural Change 51, 769–782.
Baranov, V., Bennett, D., Kohler, H.-P., 2012. The Indirect Impact of Antiretroviral
Therapy. PSC Working Paper Series, 9-27-2012.
Behrman, J.R., Foster, A.D., Rosenzweig, M.R., 1997. The dynamics of agricultural
production and the calorie-income relationship. Journal of Econometrics 77,
187–208.
Benjamin, D., 1992. Household composition labor markets, and labor demand: test-
ing for separation in agricultural household models. Econometrica 60, 287–322.
Bennett, D., Naqviy, S.A.A., Schmidt, W-P., 2014. Learning, Hygiene and Traditional
Medicine. Working Paper.
Berger, V.W., 2010. Testing for baseline balance: can we finally get it right? Journal
of Clinical Epidemiology 63, 939–940, author reply 940-932.
Bleakley, H., 2007. Disease and development evidence from hookworm eradica-
tion in the American South, February 2007. Quarterly Journal of Economics 122,
73–117.
Bleakley, H., Lange, F., 2009. Chronic disease burden and the interaction of education,
fertility and growth. Review of Economics and Statistics 91, 52–65.
Clarke, S.E., Jukes, M.C.H., Njagi, J.K., Khasakhala, L., Cundill, B., Otido, J., Crudde, C.,
Estambale, B.B.A., Brooke, S., 2008. Effect of intermittent preventive treatment
of malaria on health and education in schoolchildren: a cluster-randomised,
double-blind, placebo-controlled trial. Lancet, 127–138.
Cohen, J., Dupas, P., 2010. Free distribution or cost-sharing. Evidence from a ran-
domized malaria prevention experiment. Quarterly Journal of Economics 125,
1–45.
164 G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164
Das, J., Hammer, J., Sánchez-Paramo, C., 2012. The impact of recall periods on
reported morbidity and health seeking behavior. Journal of Development Eco-
nomics 98, 76–88.
Duflo, E., Kremer, M., Robinson, J., 2008. How high are rates of return to fertilizer.
Evidence from field experiments in Kenya. American Economic Review Papers
and Proceedings 98, 482–488.
Duflo, E., Kremer, M., Robinson, J., 2010. Nudging farmers to use fertilizer: the-
ory and experimental evidence from Kenya. American Economic Review 101,
2350–2390.
Dupas, P., 2014. Short-run subsidies and long-run adoption of new health products:
evidence from a field experiment. Econometrica 82, 197–228.
Ehrhardt, S., Burchard, G., Mantel, C., Cramer, J., Kaiser, S., Kubo, M., Otchwemah, R.,
Bienzle, U., Mockenhaupt, F., 2006. Malaria, anemia, and malnutrition in African
children – defining intervention priorities. Journal of Infectious Diseases 194,
108–114.
Fink, G., Masiye, F., 2012. Assessing the impact of scaling-up bednet coverage
through agricultural loan programmes: evidence from a cluster randomised
controlled trial in Katete, Zambia. Transactions of the Royal Society of Tropical
Medicine and Hygiene 106, 660–667.
Fink, G., Jack, B.K., Masiye, F., 2014. Seasonal Credit Constraints and Agricultural
Labor Supply: Evidence from Zambia. NBER Working Paper., pp. 20218.
Fox, M.P., Rosen, S., MacLeod, W.B., Wasunna, M., Bii, M., Foglia, G., Simon, J.L., 2004.
The impact of HIV/AIDS on labour productivity in Kenya. Tropical Medicine &
International Health 9, 318–324.
Frigge, M., Hoaglin, D.C., Iglewicz, B., 1989. Some implementations of the box plot.
The American Statistician 43, 50–54.
Gilgen, D.D., Mascie-Taylor, C.G., Rosetta, L.L., 2001. Intestinal helminth infections,
anaemia and labour productivity of female tea pluckers in Bangladesh. Tropical
Medicine & International Health 6, 449–457.
Girardin, O., Daoa, D., Koudou, B.G., Essé, C., Cissé, G., Yao, T., N’Goran, E.K., Tschan-
nen, A.B., Bordmannd, G., Lehmannc, B., Nsabimana, C., Keiser, J., Killeen, G.F.,
Singer, B.H., Tanner, M., Utzinger, J., 2004. Opportunities and limiting factors of
intensive vegetable farming in malaria endemic Côte d’Ivoire. Acta Tropica 89,
109–123.
Glennerster, R., Takavarasha, K., 2013. Running Randomized Evaluations: A Practical
Guides. Princeton University Press, Princeton, NJ.
. Goeb, J.C. (Ed.), 2011. Impacts of Government Supports on Smallholder Cotton
Production in Zambia, vol. Master of Science. Michigan State University, Lansing.
Goldberg, J., 2014. Kwacha gonna do? Experimental evidence about labor supply in
rural Malawi. Working paper.
Hay, S.I., Snow, R.W., 2006. The malaria atlas project: developing global maps of
malaria risk. PLOS Medicine 3, 473.
Jayne, T.S., Zulu, B., Kajoba, G., Weber, M.T., 2008. Access to land and povery reduction
in rural Zambia: connecting the policy issues. In: Food Security Research Project
Working Paper, p. 34.
Karlan, D., Osei, R., Osei-Akoto, I., Agricultural, Udry C., 2012. Decisions after Relaxing
Credit and Risk Constraints. Mimeo.
Knottnerus, J.A., Tugwell, P., 2012. Good baseline balance – a prerequisite for valid
comparison. Journal of Clinical Epidemiology 65, 119–120.
Kremer, M., Miguel, E., 2004. Worms identifying impacts on education and health
in the presence of treatment externalities. Econometrica 72, 159–217.
Lengeler, C., 2004. Insecticide treated bednets and curtains for preventing malaria.
Cochrane Database of Systemic Reviews, CD000363.
Loureiro, M.L., 2009. Farmers’ health and agricultural productivity. Agricultural Eco-
nomics 40, 381–388.
Macro International, 2007. Zambia: DHS, 2007 – Final Report (English). Macro Inter-
national, Calverton, MD.
Morel, C.M., Thang, N., Xa, N., Hung, L.X., Thuan, L.K., Ky, P.V., Erhart, A., Mills, A.J.,
D’Alessandro, U., 2008. The economic burden of malaria on the household in
south-central Vietnam. Malaria Journal, 7.
NMCC, 2007. Zambia Malaria Indicator Survey 2006. Zambia Ministry of Health,
Lusaka.
NMCC (Ed.), 2009. Zambia Malaria Indicator Survey 2008. Zambia Ministry of Health,
Lusaka, Zambia.
NMCC, 2010. Zambia Malaria Indicator Survey 2010. Zambia Ministry of Health,
Lusaka.
NMCC (Ed.), 2011. ITN Distribution Data Base. NMCC, Lusaka.
Nolte, K., April 2012. Large scale agricultural investments under poor land gover-
nance systems: actors and institutions in the case of Zambia. In: World Bank
Conference on Land and Poverty Paper 2012.
Nur, E.T.M., 1993. The impact of malaria on labour use and efficiency in the Sudan.
Social Science & Medicine 37, 1115–1119.
Pitt, M.M., Rosenzweig, M.R., 1986. Agricultural prices, food consumption, and the
health and productivity of Indonesian farmers. In: Singh, I.J., Squire, L., Strauss,
J. (Eds.), Agricultural Household Models. Johns Hopkins University Press, Balti-
more.
Riley, R.D., Kauser, I., Bland, M., Thijs, L., Staessen, J.A., Wang, J., Gueyffier, F.,
Deeks, J.J., 2013. Meta-analysis of randomised trials with a continuous outcome
according to baseline imbalance and availability of individual participant data.
Statistics in Medicine 32, 2747–2766.
Strauss, J., 1986. Does better nutrition raise farm productivity. Journal of Political
Economy 94, 297–320.
Tarozzi, A., Mahajan, A., Blackburn, B., Kopf, D., Krishnan, L., Yoong, J., 2014.
Micro-loans insecticide-treated bednets malaria evidence from a randomized
controlled trial in Orissa (India). American Economic Review 104, 1909–1941.
Thomas, D., Frankenberg, E., Friedman, J., Habicht, J.-P., Ingwersen, N., McKelvey, C.,
Mohammed Hakimi, J., Pelto, G., Sikoki, B., Seeman, T., Smith, J.P., Sumantri,
C., Suriastini, W., Wilopo, S., 2010. Causal Effect of Health on Labor Market
Outcomes: Experimental Evidence.
Udry, C., Anagol, S., 2008. The Return to Capital in Ghana.
Ulimwengu, J., 2009. Farmers’ health and agricultural productivity in rural Ethiopia.
African Journal of Agricultural and Resource Economics 3, 83–100.
WHO (Ed.), 2009. World Malaria Report 2008. WHO, Geneva.
World Bank (Ed.), 2007. World Bank Development Indicators CD-ROM.
Zambia Central Statistic Office (Ed.), 2011a. 2010 Census of Population and Housing.
Zambia Central Statistic Office, Lusaka, Zambia.
Zambia Central Statistic Office (Ed.), 2011b. Living Conditions Monitoring Survey
Report 2006 and 2010. Zambia CSO, Lusaka, Zambia.
Zambia Ministry of Health, 2006. A 6-year Strategic Plan A Road Map for Impact on
Malaria in Zambia 2006–2011s. NMCC, Lusaka.
Zambia Ministry of Health (Ed.), 2012. Health Management Information System
(HMIS). Zambia Ministry of Health, Lusaka.
 Journal of Health Economics 42 (2015) 165–173

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Welfare implications of learning through solicitation versus

diversification in health care
Anirban Basu a,b,∗,1
a
Department of Pharmacy, Health Services and Economics, University of Washington, and The NBER, Cambridge, MA 1959 NE Pacific St, Box 357660,
Seattle, WA 98195-7660, United States
b
NBER, Cambridge, MA, United States

a r t i c l e i n f o a b s t r a c t

Article history: Using Roy’s model of sorting behavior, I study welfare implications of learning about medical care quality
Received 3 April 2014 through the current health care data production infrastructure that relies on solicitation of research
Received in revised form 9 April 2015 subjects. Due to severe adverse-selection issues, I show that such learning could be biased and welfare
Accepted 10 April 2015
decreasing. Direct diversification of treatment receipt may solve these issues but is infeasible. Unifying
Available online 20 April 2015
Manski’s work on diversified treatment choice under ambiguity and Heckman’s work on estimating
heterogeneous treatment effects, I propose a new infrastructure based on temporary diversification of
Keywords:
access that resolves the prior issues and can identify nuanced effect heterogeneity.
Learning
Diversification © 2015 Elsevier B.V. All rights reserved.
Comparative effectiveness research
Economic evaluation
Instrumental variables
Heterogeneity

JEL classification:
C1
C9
D6
I1

One of the fundamental challenges in health care markets is
lack of information about the quality of medical care and tech-
nology (Arrow, 1963). Information on medical product quality is
usually generated by employing an artificial form of ‘learning by
doing’ mechanism where a selected group of individuals (doers) is
allowed to consume alternative medical products (e.g. using stan-
dard statistical designs, such as randomized assignment of patients
to products). Wisdom from their experiences is disseminated to
other individuals, who will face the choice of using these medical
products in the near future, and to inform other decision makers,
who device social policies on access2 . Most public and private
stakeholders that are engaged in data production on medical qual-
ity signals have employed such mechanisms. Recently, substantial
public investments were made in the US, under the umbrella term
“comparative effectiveness research” (CER) and patient-centered
outcomes research (PCOR)3 , to facilitate production of such data
on alternative medical technologies that are currently being used
in clinical practice, albeit with incomplete knowledge about their
comparative qualities4 .

∗ Corresponding author at: Department of Pharmacy, Health Services and
Economics, University of Washington, Seattle, WA 98195-7660, United States.
Tel.: +1 206 616 2986; fax: +1 206 543 3964.
E-mail address: basua@uw.edu
1
I am grateful for comments from Karl Claxton, David Meltzer, Justin Robertson
and two anonymous reviewers and support from NIH research grants RC4CA155809
and R01CA155329. Opinions expressed are mine and do not reflect those of the
University of Washington or the NBER.
2
There are situations where learning from own’s doing is popular, aka the
repeated use of pharmaceutical products in chronic illnesses.
3
Patient Protection and Affordable Care Act of 2009, H.R. 3590, 111th Congress
§6301 (2010).
4
Throughout our paper, I assume the CER compares two medical technologies
that have been approved for use based on meeting the minimum safety thresh-
olds as those set by the Food and Drug Administration of the United States. Our
discussions do not encompass evaluation of experimental therapies. Such discuss-
ions are delegated to future work. Also see Philipson (1997) and Malani (2008) who
make distinct arguments about selection in trials of experimental therapies in the
presence of health insurance.

http://dx.doi.org/10.1016/j.jhealeco.2015.04.001
0167-6296/© 2015 Elsevier B.V. All rights reserved.
166 A. Basu / Journal of Health Economics 42 (2015) 165–173
In this paper, using a simple Roy’s model (Roy, 1951) of sorting
behavior, I prove that, when incremental treatment effects are het-
erogeneous across patients who have access to these treatments
under insurance, a data production infrastructure for comparative
medical quality that relies on soliciting voluntary participation of
subjects fails to identify any interpretable treatment effect param-
eter. Therefore, evidence generated through this process fails to
inform, objectively, either the individual patient on optimal med-
ical care use or a social insurer on optimal medical care insurance
coverage5 .
Unfortunately, such a data production infrastructure is and has
been the norm for CER randomized clinical trial (RCT) studies. There
are many examples of such failures in the literature. For example,
Ioannidis and Lau (1997) show that in human immunodeficiency
virus-related trials and trials of magnesium in acute myocardial
infarction, when the benefit or toxicity from a treatment varies
with the baseline risk of each patient, the treatment effect may
be markedly different in populations with a different representa-
tion of high- and low-risk patients. I show that such differential
representation of the population in trials may be driven more fun-
damentally by patient and physician behaviors and therefore the
problems of interpretation of trial results are systemic.
The implications of this finding are substantial. Incomplete com-
parative quality information generated by CER RCTs research has
the potential to misguide treatment choices since ex-ante percep-
tion of benefits do not coincide with the ex-post accrual of the same,
resulting in welfare losses (Basu, 2011). These inefficiencies in the
choice of medical products can also accentuate the inefficiencies
due to moral hazard stemming from health insurance (Arrow, 1963;
Pauly, 1968), translating to higher premiums and less protection
against risk, in both competitive and non-competitive insurance
markets (Basu, 2011). In this paper, I focus on understanding why
current data production infrastructure leads to incomplete infor-
mation.
I begin in the next section by laying out our ideal target param-
eters in CER evaluations; those that we would like to obtain
estimates for in order to guide treatment decisions at the individual
level and policy decisions on coverage at the social level. In Sec-
tions 2 and 3, I highlight the current data production infrastructure
and prove why it would produce incomplete information. I study
the implications for such incompleteness on decision-making and
welfare. In Section 4, I introduce a new framework for data produc-
tion that can efficiently resolve the biases inherent in the current
data production infrastructure by using diversification of access
to create a conduit for learning about meaningful and decision-
relevant effect parameters. This work unifies two broad themes
in the econometrics literature, one based on Manski’s work on
treatment choice under ambiguity (Manski, 2000, 2004, 2009) that
utilizes the concept of diversification of treatment and the other
based on Heckman, Vytlacil and others’ works on estimating het-
erogeneous treatment effects (Heckman, 1997, 2001; Heckman and
Vytlacil, 1999, 2001; Heckman et al., 2006). I show how this frame-
work can help overcome inefficiencies in health care markets that
stem from incomplete information.
1. Defining the true population average effect of a
treatment
Let us begin with a problem of evaluating the compara-
tive effectiveness of a new (approved) treatment compared to a
5 6
Note that my assertions about optimality are very general and does not depend
on specific welfare functions. What I prove is that the structural target parameters
on which information is required to maximize any welfare function is not informed
by current data production infrastructure.
control/standard treatment for a population of N patients indexed
by i. Standard treatment may also include the do-nothing option.
Let the individual-level true treatment effects represent the bene-
fits (net of harms) of the new treatment over the control and are
denoted by bi . Let p denote the price of the new treatment, which
is also the marginal cost for manufacturing the new treatment6 .
Patients are members of risk classes ˝, ˝ = 1,2,.k; k ≤ N, which
determine heterogeneity in treatment effects across individuals
through a production function:

bi = ˛k × I(˝i = k) (1)
k
where I() is an indicator function and ˛k is interpreted as the
true comparative effect of the new treatment over the standard
treatment in risk class k. Let’s assume that this comparative effect
is expressed in monetary terms. That is the effectiveness unit is
multiplied with the some predefined threshold representing the
monetary value of the marginal unit of benefit7 .
A population-level average effect parameter is given as

= Pr(˝ = k) × ˛k (2)
k
There are two types of decision makers, (1) the patient-
physician dyad, which I will refer to as the individual decision
maker, is assumed to always have knowledge about their risk class;
and (2) an insurer or social planner who decides the coinsurance
rate for providing health insurance coverage for the new treatment.
2. Data production and incompleteness in quality
information
A first-best scenario can be achieved under complete informa-
tion, where both the insurer and the individuals are aware of the
risk classes and the production function and are able to perfectly
predict bi . If individuals had full insurance they would choose treat-
ment if bi ≥ 0. Since the insurer can fully anticipate this individual
behavior, she can provide full coverage for treatment only for those
individuals who would experience benefits greater than cost and
not provide coverage for the rest. Thus, there is no efficiency loss
due to moral hazard.
Under the second-best scenario, there exist asymmetry of infor-
mation where, even though, individuals are assumed to be aware
of ˝k and b() and to be able to combine them to predict bi perfectly,
the insurer cannot as they have either no or only partial informa-
tion on ˝k (Arrow, 1963; Pauly and Blavin, 2008). Consequently,
the insurer cannot exclude patients from coverage who would get
treatment benefits lower than the cost of treatment (i.e. bi − p < 0).
This leads to moral hazard (Pauly, 2008). To counter this, the insurer
may offer coverage with a fractional coinsurance rate (r), which is
the fraction of price a patient must pay in order to receive treat-
ment. When r = 1, the new medical product is not covered through
insurance.
I assume individuals choose treatment by maximizing a generic
Net-Benefit criterion that is based on their perceived benefits from
treatment net of the demand price they face in acquiring the treat-
ment. I also assume that the social insurer’s goal is to maximize
consumer surplus as is realized ex post based on individual level
choices. Throughout this paper, I express the realized population
level benefits under different levels of coverage for the new treat-
ment as changes to the total outcomes had all patients taken the
Assume for now that the marginal cost is constant.
7
Under the welfare economic foundations, this threshold is the inverse marginal
utility of income (Weinstein and Zeckhauser, 1973; Garber and Phelps, 1997;
Meltzer, 1997).
 A. Basu / Journal of Health Economics 42 (2015) 165–173 167

standard treatment. Under any co-insurance rate r, r ∈ [0, 1], this have perfect information on either ˛k or ˛i . However, she may have
population level benefit, H0 , is given as information about the average effect, ˛.¯ The best a social insurer can
 do at this point is to calculate the average net monetary benefits of
H0 = I(˛k − r × p ≥ 0) × ˛k × I(˝i = k) (3) treatment,
i k
˛
¯ −p (5)
When individuals have complete information, they choose to
and recommend coverage if ˛ ¯ −p≥0 9.
receive the new treatment only if ˛k ≥ r × p. Individuals who would
expect to get harmed by treatment (i.e. ˛k < 0) would not select Without loss of generality,
treatment even if it were available to them for free, thereby self- Assumption 1. Let  > 0, the true population average treatment
limiting the magnitude of moral hazard. effect is positive, but ˛k ’s span the whole real line.
For a social insurer’s point of view, an optimal co-insurance rate
may be expressed as a solution to maximizing H0 net of costs and Assumption 2. Let ˛
¯ − p ≥ 0 and full coverage was recommended,
taking into account the social value of risk protection provided i.e. r* = 0.
by the insurance (Manning and Marquis, 1996). Consequently, the
Theorem 1. Under Assumptions 1 and 2, LPRE (0) > L2nd (r* ) for
moral hazard (welfare loss) under optimal coinsurance rate (r* ) in
∀r* ∈ [0, 1] if ˛i ≥ 0, ∀i. The welfare loss under pre-CER information
a second-best scenario is given as
with full insurance coverage is strictly larger than the welfare loss
 
   under any second-best scenario as long as all individuals perceive
∗ ∗
L2nd (r ) = I(r × p ≤ ˛k < p) × (p − ˛k ) × I ˝i = k a positive benefit from treatment.
i k
Proof. Under the Pre-CER scenario, two groups of individuals

p making inefficient choices drive the welfare loss. The first group
= Nk × (p − ˛k ) (4) consists of people who fail to receive treatment because their ˛i < 0
˛k =r ∗ ×p but they belong to risk groups where the treatment produces incre-
mental benefits that are more than the price of the treatment (i.e.
which constitutes the welfare loss due to the total number of indi- ˛k > p). The second group consists of individuals who would receive
viduals in each risk group (Nk ) who would choose treatment given treatment as they are led to believe that they would get a positive
the lower demand price (r × p) but ultimately obtain benefits lesser benefit (˛i ≥ 0) but realize a benefit less than the price ((i.e. ˛k < p).
than the price of treatment, i.e. r* × p ≤ ˛k < p Therefore total welfare loss is given by
Reality, however, deviates from both the first and second
best scenarios, because both individuals and the social decision  
maker face incomplete comparative information. To understand     
this incompleteness, one must study the data production mech- LPRE,CER = I ˛i < 0 × I (˛k − p > 0) × (˛k − p) × I ˝i = k
anisms in place. I consider and compare the circumstances before  i k

and after a CER study. I begin by understanding the consequences of     
incomplete information before a CER is conducted and why added + I ˛i ≥ 0 × I (˛k − p < 0) × (p − ˛k ) × I ˝i = k
investments for data productions, such as those provisions by the i k
∞  
 

latest legislations (2009 American Recovery and Reinvestment Act;  ˛

¯

p
˛
¯
= 1−˚ × Nk × (˛k − p) + ˚ × Nk × (p − ˛k )
Patient Protection and Affordable Care Act of 2009, H.R. 3590, 111th s s
Congress §630 (2010)), are called for. I then study how the current ˛k =p ˛k =−∞

mechanisms of CER may continue to propagate and even enhance (6)

the welfare losses due to incomplete information.
where ˚(.) is a cumulative normal distribution. Comparing (6) to
2.1. Pre-CER information and choices (4),
∞ 
  ˛¯

Before a CER RCT is conducted, it is safe to assume that ˛k is not

LPRE,CER (0) –L2nd (0) = 1−˚ × Nk × (˛k − p)
known with certainty at both the individual and the societal level. s
˛k =p
However, prior knowledge, obtained from evidence (of size n) gen- p  
 ˛
¯

erated during the process of approving the use of this new medical
+ ˚ −1 × Nk × (p − ˛k ) (7)
product would determine an individual patient’s anticipated belief s
˛k =0
about the incremental benefits of treatment given one’s own risk
   ˛¯

0

class. Let this evidence suggest that the average effectof treatment
2 /n 8 , where  + ˚ × Nk × (p − ˛k )
is ˛
¯ that is a random draw from Normal ,  s
is the average effect parameter defined in (2) and   represents ˛k =−∞

the heterogeneity and is the standard deviation of the effect in the The first and the third terms in (7) are the incremental losses
population. Let individual beliefs, ˛i , be given as a single draw from due to incomplete information pre-CER. The first term is the same
the distribution Normal (˛, ¯ s2 ) where s is the estimated standard as in (6) and comprises of individuals who fail to take treatment
deviation from prior evidence. It is assumed that s2 is a consis- but would have benefited more than its price. The loss represented
tent estimator of  2 . The schedule of ˛ across individual patients
i in the third term emanate from risk groups where ˛k < 0 and a frac-
determines the marginal benefits curve in the population in the tion of individuals in these risk groups take treatment based on
absence of a CER, which is based on aggregation of individual’s their perceived positive benefits, which was not the case under the
perceived marginal benefits. Moreover, the social insurer may not second-best scenario.

8
I take a conservative approach is assuming that ˛ ¯ and ˛i are consistent estima- 9
This is, in fact, the standard method used in most cost-effectiveness modeling
tors of . To the extent this is not true, the welfare losses described below may be studies that try to evaluate the cost-effectiveness of a new approved treatment for
higher. which there is no head-to-head comparison with its alternatives.
168 A. Basu / Journal of Health Economics 42 (2015) 165–173

The second term in (7) is a pervasive benefit of incomplete
information compared to the second-best scenario (expressed
as negative loss). The benefit emanate from risk groups where
0 < ˛k < p and a fraction of individuals in these risk groups forgo
treatment based on their perceived benefits (i.e. their ˛i < 0),
which was not the case under the second-best scenario, thereby
generating welfare gains.
Under Assumptions 1 and 2, if one assumes that ˛i ≥ 0, ∀i, that
is every individual perceives a positive benefit from treatment,
Theorem 1 is proved from (7) as the first two terms drop out
and LPRE,CER (0) > L2nd (0). Thus, naturally, LPRE,CER (0) > L2nd (r* ) for
∀r* ∈ [0, 1], since L2nd (0) > L2nd (r* ) for ∀r* ∈ [0, 1].
2.2. An ideal role for CER
Often an ideal CER is construed as one having larger sample size.
In fact, much of the value of information literature in medicine has
focused estimating the marginal value of a trial with additional
patients enrolled (see literature on the Expected Value of Sample
Information, EVSI). However, it is not clear whether, in the presence
of heterogeneity, such an approach to CER, is welfare enhancing.
For example, increase in sample size cannot overcome the issues
about selection into trial that we discuss later. Nevertheless, for
now, let’s assume that those concerns are not relevant. Then, as
n→ ∞, ˛−→¯ , s2 −→ 2 . Consequently, following Eq. (6) and sub-
p p
stituting the post-CER estimate of average effect in it, the welfare
loss post CER with an infinite sample will be:
∞ 
  
studies—Understanding selection into randomized trials
LPOST,CER(n→∞) = 1−˚ × Nk × (˛k − p)

˛k =p
   
for ˛K ≤ p as by construction ˛K < ˛ ¯ for all ˛K ≤ p. Therefore,
LPOST∗ (r ∗ ) < LPRE (r ∗ ) ∀r ∗ ∈ [0, 1].
This unambiguous dominance of an ideal CER over pre-
CER scenario arises because individuals are able to better
self-select their optimal treatment based on the risk group

specific knowledge  generated from an ideal CER.  In fact, as
 k → 0, ˚ ˛k /k → 1 for ˛k ≥ 0 and ˚ ˛k /k → 0 for ˛k < 0.
Consequently, LPOST * (r) → L2nd (r). Therefore, one can potentially
approach a second-best scenario under any level of insurance cov-
erage if new CER studies are able to generate information that can
enable individuals to better self-select treatments based on their
risk classes, even if the social insurer is unaware of these heteroge-
neous effects. The growing awareness of the potential value of such
has led to considerable federal investment in CER. New legislation
such as the Affordable Care Act of 2009 has also identified the need
to risk stratify comparative effectiveness.
However, the current data production infrastructure for CER
may not be aligned with the goals of such legislation. The gold
standard of data production in medical care involves controlled
experiments, where alternative treatments under investigation are
allocated to a selected group of patients by a chance mechanism.
I will refer to such a mechanism as a randomized clinical trial
(RCT) henceforth. I consider two issues within this data-generating
infrastructure that contributes towards the inability of current CER
infrastructure to resolve incompleteness in information: selection
in RCT enrollment and target parameters for RCTs.
2.3. Non-ideal design and implementation of CER
Unlike evaluation of experimental therapy where enrollments
may be more likely driven by altruistic motives, CER and economic

+ ˚ × Nk × (p − ˛k ) (8) evaluation is about approved and existing therapies available to

 patients. Therefore, there must be a strong implicit selection pro-
˛k =−∞
cess for patients who provide consent to enroll in a CER RCT, in
    which they have a non-trivial probabilistic expectation of receiving
Note that if ˚ / < ˚ ˛/s ¯ 2 , a CER of infinite size will
a treatment (most likely the new treatment) that they have some
decrease the magnitude of loss in the second term but increase the
difficulty in accessing outside of the RCT. Such difficulties must be
magnitude of loss in the first term as compared to LPRE and vice
because the cost of accessing the new treatments outside RCT must
versa. Therefore, the value of such a CER study, even with infinite
be high either due to differential insurance coverage of the treat-
sample, is indeterminate.
ments or due to strong physician preferences for one therapy over
An ideal role for CER would be when, even though the social
other. Consequently, this selection process into randomized trial
 indi-
insurer continue to believe in the average effect, it can enable
implies that patients who anticipate that their expected incremen-
vidual beliefs, ˛i to be a draw from the distribution Normal ˛k , k2
tal benefits from the new treatment, compared to the control, is
where k2 is variance for risk-group-specific effects10 . If coverage is more than the cost of acquiring the new treatment outside of RCT,
recommended, welfare losses under post ideal-CER scenario will would not enroll in the RCT and would avail the new treatment out-
be a modification of (6) to side RCT. Similarly, patients anticipating a very small incremental
∞   ˛
benefit from the new treatment, even compared to the subsidized
 K
LPOST∗ = 1−˚ × Nk × (˛k − p) cost of treatment within the RCT, would not enroll. That leaves
k patients who enroll in a CER RCT to be those who anticipate that
˛k =p
their incremental benefits must be less than the cost of acquiring

p   ˛
the new treatment outside of RCT but more that the subsidized cost
K
+ ˚ × Nk × (p − ˛k ) (9) of new treatment within an RCT11,12 .
k
˛k =−∞
  to question whether the anticipated benefits of
This brings
treatment ˛i are related in any form to the true benefits of
Note that both the terms in LPOST∗ in (9) are smaller treatment bi , even in the absence of formal CER. Obviously, if
than the corresponding terms in LPRE,CER in (6). This is ˛i = bi , then the value of any additional CER becomes zero as each

because, under 
Assumptions 1 and 2 and also assuming k = patient already know their true benefits. On the other hand, the
s, ∀k, ˚ ˛K /k > ˚(˛/s)
¯ for ˛K > p as by construction ˛K >
 
˛
¯ for all ˛K > p, since ˛
¯ − p ≥ 0. Similarly, ˚ ˛K /k > ˚(˛/s)
¯
11
Often enrollees are paid a monetary sum to compensate for their time spent
participating in the RCTs.
12
It is not necessary for only the patient to be choosing the treatment. Even in acute
10
In practice, even in the absence of CER such a situation may arise, when individ- settings, the physician choice may also be subject to the same selection principles
uals learn by repeated consumption of therapy (e.g. pharmaceuticals) or physicians where they have some anticipation of what works for individual patients. Therefore,
are able to anticipate effect heterogeneity based on baseline risks. physicians may often select to enroll patients to their trials.
 A. Basu / Journal of Health Economics 42 (2015) 165–173
value of CER is maximized when ˛i bi , where denotes statistical
independence. In practice, however, it is common to find some
dependency between ˛i and bi . Such dependencies may arise from
biological knowledge about the treatment’s mechanism of actions,
past experiences by physicians on using similar treatments on
certain patient risk-groups and by patient’s own learning by doing
mechanism in a chronic disease setting. Under such dependencies,
effect of selection into RCT becomes non-trivial. Specifically, I show
Theorem 2. A CER randomized trial produces an unbiased esti-
mate of the populationaverage treatment effect  (in Eq. (2)) if
and only if ˛i bi . If Corr ˛i , bi , > 0, RCTs will typically find small
positive benefits of treatment that are likely to be biased estimate
of true average treatment effect.
Proof. I formalize selection into a CER RCT following Roy’s model
(Roy, 1951) of self-selection using the following notation
 
Si = I Ui∗ ≥ 0
where S is an indicator for enrolling in an RCT that is driven by
the latent utility U* for enrolling. Again, without loss of generality,
Ui * is interpreted as the anticipated incremental net benefits (net
of costs) of enrolling versus not enrolling in an RCT for individual
i given that the individual anticipates a positive benefit from
treatment (i.e. ˛i > 0)13
    a consistent estimator for either population average 
Ui∗ = R × ˛i − CRCT − ˛i − COUT
= (COUT − CRCT ) − (1 − R ) × ˛i , if ˛i > 0
where CRCT and COUT are the costs of accessing the treatment
within and outside an RCT, respectively; R is the known random
probability of receiving the new treatment within the CER RCT.
In the presence of uncertainty, the population probability of an
RCT enrollment is given by
  
(COUT − CRCT )
 = Pr (Si ) = E Ui∗ ≥ 0 = Pr 0 < ˛i <
(1 − R )
Therefore, only patients who anticipate positive benefits but
whose magnitudes are less than the expected incremental cost of
accessing treatment outside RCT would enroll. Interestingly, when
COUT ≈ CRCT , enrollment in CER can be quite difficult. In contrast, if
COUT  CRCT , then the probability of enrollment will approach one
for everyone. Similarly, as R decreases, it reduces the cost differ-
ential between accessing the new treatment outside and within the
RCT, thereby lowering the probability of RCT enrollment. These fac-
tors severely limit the generalizability of results from CER RCTs. For
example, in one of the few surveys ever conducted to understand
the factors that determine RCT enrollment, it was found that only
2.7% of eligible patients enrolled in clinical oncology trials (Movsas
et al., 2007).
Target Parameters for RCT The goal of RCT is to estimate a
structurally meaningful population level parameter such as the
average treatment effect (ATE) of treatment compared to the con-
trol. Instead, the target population ends up being defined by the
RCT enrollees. Consequently, its target population defines the tar-
get parameter that an RCT tries to estimate. This parameter is
an average effect that is a weighted average of risk-class-specific
effects, where the weights are arbitrarily defined based on the risk
class-specific propensity to enroll in the RCT. Therefore, the target
parameter for RCT is given by

RCT = wk × ˛k
k
13
If individual anticipates a negative benefit from treatment he would not consider
enrolling in the first place.
169
      
where the weights wk = k × Pr ˝ = k / k × Pr ˝ = k
 k 
and k = Pr 0 < ˛i < (COUT − CRCT ) / (1 − R ) |˝i = k . The degree
of selection in the trial determines the target parameter of an RCT.
The touted internal validity of RCTs rests on obtaining a consistent

estimate for this target   
parameter.
When ˛i bi , F ˛i |˝i = k = F ˛i ∀k, where F() is the cumu-
function. This implies, k = , ∀k ⇒ wk =
 distribution
lative 
Pr ˝ = k , ∀k ⇒ RCT =  (according to Eqs (2) and (13)). On
the contrary, if , RCT =/  since the weights would
vary depending on which risk classes are more likely to enroll in
the RCT.

In fact, if perceived benefits are positively correlated with true
benefits, i.e. Corr ˛i , bi > 0, it implies Corr (wk , ˛k ) < 0 for ˛i > 0
and Corr (wk , ˛k ) > 0 for ˛i ≤ 0. Individuals who correctly antic-
ipate large positive or negative benefits from treatment are less
(10) likely to enroll in RCTs. In fact, Eqs (12) and (13) suggest that the
margin of individual who enroll in RCT anticipates a moderated
positive magnitude of benefits from treatment. This implies that
RCT results would typically find small positive benefits of a newer
treatment and the generalizability of these results to the whole
target population remains severely compromised14 .䊏
Consequently, in the presence of any anticipatory knowledge
about true treatment effects, the average effect from an RCT is not
effect or the
average effect of any segment of the population: E  ˆ RCT = / 
 
(11) and E  ˆ RCT = / ˛k × ∀k. Next, I study how such results can mis-
lead individual level decision-making and create inefficiencies both
through population-level coverage decisions and individual treat-
ment selections.
 
Assumption 3. In what follows, I will assume Corr ˛i , bi > 0
even in the absence of a formal CER.
(12) 2.4. Implications of incompleteness for decision-making
Theorem 3.
(a) Under Assumption 3, CER RCT may misguide a social planner
to provide coverage on treatments with negative average net
health benefits and to withhold coverage on treatments with
positive average net health benefits.
(b) Under Assumptions 3, LPOST (0) >=< LPRE (0). The welfare loss
under post-CER information can be larger than that under pre-
CER scenario with full insurance coverage.
Proof.
(a) Based on CER RCT results, the social planner updates her belief
over the average effect of the new treatment using a Bayesian
updating rule (Basu et al., 2011):
 
¯¯ =  × ˛
˛ ˆ RCT
¯ + 1− × (14)
where the weight  is determined by a weighted average
of prior uncertainty  2 and the sampling variance of  ˆ RCT ,
and calculates the average net monetary benefits of treat-
ment to be ˛¯¯ − p. Under Assumption 3, Theorem 2 proves
 
that E ˆ RCT > 0 but E  ˆ RCT >=< . Therefore, since E (˛)
¯ =
(13)
14
It is possible that under an ideal symmetric condition, the weights are such
that equivalent portions of the risk-groups with large positive effects and those
with large negative effects select out of enrolling and the average effect among the
enrollees still reflects the population average. However, such a scenario is highly
unlikely.
170 A. Basu / Journal of Health Economics 42 (2015) 165–173
   
, E ˛ ˆ RCT < a and E ˛
¯¯ − p < E (( − p)) if E  ¯¯ − p >
 
E (( − p)) if E ˆ RCT > .
This implies that CER RCT may misguide a social planner to
provide coverage on treatments with negative average net health
benefits or to withhold coverage on treatments with positive aver-
age net health benefits. This also highlights the fact that economic
evaluations based on CER RCT studies can be misleading. 䊏
(a) Individual beliefs, ˛i , about comparative effects will also evolve
following the CER RCT using a similar Bayesian updating rule
(Basu et al., 2011):
 
˛i = i × ˛i + 1 − i × 
ˆ RCT )
where the weights  i are determined by a weighted average
of prior uncertainty s2 and the sampling variance of  ˆ RCT . It
is important to note that even  though original beliefs
  may
have been consistent, i.e., E ˛i = bi , after CER, E ˛i = / bi .
Most importantly, ˛i > ˛i if ˛i < 0, under Assumption 3,
 
since E ˆ RCT > 0. That is, some patients who would have had
originally anticipated a negative effect from treatment, may
be rightfully so, are now led to believe in a larger, presum-
ably, positive effect from treatment. Similarly, patients who
would have, rightfully anticipated large benefits from treat-
ment, would have their updated anticipation moderated by
the small effect size estimated in the RCTs. Thus, the average
result from a CER study that is based on voluntary participa-
tion actually misleads individuals about their own comparative
effectiveness. Following (6), the welfare loss with the post CER
information is given by
 
k (˛i )/sk (˛i ) < ˛/s
¯ for ˛k < p, implying that more individuals
with ˛k < p are drawn to not use the treatment, the CER infra-
structure is welfare enhancing (i.e. LPOST CER < LPRE,CER ).䊏
3. Learning through diversification (LtD): A new
framework for data production
As I have shown in the previous sections, under some gen-
eral assumptions, the current CER-RCT framework that relies on
voluntary participation fails to consistently inform either the
population-level or individual-level comparative effect parameters
and cannot potentially lead us towards the second—best solutions
(in fact, it may lower welfare through evidence-based misguid-
ance).
Manski (2009) proposed that one way a social decision maker
can maximize welfare is through fractional allocations, where a
random fraction of the patient population received one treatment
while the other receives the alternative. Manski argued that, given
the ambiguity of evidence on counterfactual outcomes, such an
allocation would maximize a broad set of utilitarian welfare func-
tion for the social decision maker. Manski (2009) also pointed out
that such an allocation automatically created randomized experi-
ments, which were particularly important for learning treatment
responses. The current proposal builds on this idea of “diversi-
fied treatment” proposed by Manski (2009). However, our proposal
takes into account two realities in the context of health care.
First is that it is almost impossible, at least in the United States,
to completely restrict “receipt” of a treatment that has crossed the
regulatory and evidentiary hurdles and has been approved on the
basis of safety and efficacy. Therefore, diversification of treatment
    
allocation in terms of “receipt”, which is essential to answer CER

∞
k ˛i and PCOR type question, is usually not possible.
LPOST,CER = 1−˚    × Nk × (˛k − p)
sk ˛i Second, the social decision maker in the context of health care
˛k =p
is typically involved on deciding on insurance coverage of medi-
    
cal treatment, while individual subjects are typically left to decide

p
k ˛i
+ ˚    × Nk × (p − ˛k ) , (16) on the choice of treatment given insurance coverage. Therefore, a
sk ˛i social decision maker’s problem can be viewed to be a two-step
˛k =−∞
        process (Dehejia, 2005). Under any information set (i.e. alternative
where k ˛i = E ˛i |˝i = k and sk2 ˛i = Var ˛i |˝i = k . coverage decisions and CER information), first physician decides
Since k (˛i )/sk (˛i ) <=> ˛/s
¯ for any k, it proves that wel- whether to prescribe treatment for each individual. Second, given
fare loss under post-CER information can be larger than that these potential allocations, the social decision maker decides on
under pre-CER scenario with full insurance coverage. Only the level of coverage for treatment that would improve population
when k (˛i )/sk (˛i ) ≥ ˛/s
¯ for ˛k ≥ p, implying that more indi- health either by sustaining the individual choices or by incen-
viduals with ˛k ≥ p are drawn to take up the treatment, and tivizing to alter them if needed. That is, a social decision maker

Fraconal Coverage
via Techn
Technology
ology dras
Full Coverage Outcomes Evaluaon
sub-groups
for sub -groups

No Coverage
for sub-groups

Sufficiency of
Cross
Crossing
ing Evidenary
Thresholds

Fig. 1. The Learning through diversification (LtD) infrastructure.

 A. Basu / Journal of Health Economics 42 (2015) 165–173 171

studies these potential allocations of treatments under different be used to identify Marginal Treatment Effect (MTE) parameters
information sets and makes the optimal coverage policy that would ((Heckman and Vytlacil, 1999):
generate the highest population benefits driven by the individual  
allocation of treatments that would follow that policy. To the extent
∂Eϑ Y |Ð , ˝    
= E (Y1 − Y0 ) |˝, V = v = MTE ˝, v , (17)
that one can combine the ideas of diversified treatment for learning ∂p
to that of the two-step process of social decision making on opti- where Y = D × Y1 + (1 − D) × Y0 is the observed outcome, unobserved
mal coverage, one can improve the decision making for both the confounders are V ∼ Uniform[0,1] by construction and probability
individual patients and the social decision maker. Here I propose a of treatment choice given the lottery can be represented by p(Ð, ˝).
“Learning through Diversification” (LtD) infrastructure (Fig. 1) that Basu (2014) extends the LIV methods to identify Person-centered
can potentially mimic the ideal CER designs discussed in Section treatment (PeT) effects, which, for persons who choose treatment,
2.2. The three main features of a LtD infrastructure are: follow
 
(1) Fractional coverage can be achieved using a technology lottery, EV |˝,P(Ð ),D E Y1 − Y0 |˝, P (Ð ) , D = 1
Ð: For each new product, develop a random order based on,

P(Ð )
say, birth dates so that this new product with uncertain effec-   −1
  (18)
tiveness profile will be paid at varying levels by insurance (in = E Y1 − Y0 |˝, V < P (Ð ) = P(Ð ) MTE ˝, v dv
continuous fashion) in the first year. That is, such coverage cre- 0
ates a completely stochastic distribution of co-insurance rates
Similarly, conditional effect for a person who did not choose
in the population, F(bi |Ð) = F(bi ). Note that the lottery is done
treatment is obtained by integrating MTEs over values of V greater
anew for each new technology so that the probability that any
than p. Mean treatment effect parameters, ˛k (Eq. (1)) or  (Eq.
one person would be denied coverage for all new technologies
(2)) are readily obtained by average PeT effects over respective
will approach zero with increasing number of technologies.
subgroups (Basu, 2014).
(2) Outcomes evaluation: Using the randomization inherent in the
lottery, evaluating patient outcomes across different levels of
3.1.3. Welfare effects
coinsurance rates will directly answer the economic evalua-
Let’s take a two period model in which the first period is the
tion questions on expanding coverage for the target population.
Pre-CER period during which a CER study is being conducted. As
Additionally, the lottery would serve as a perfect instrumental
the end of the first period, the CER study results are disseminated
variable (IV) to study the comparative effectiveness of receiving
and therefore the second period represents the post-CER world.
the new product versus its competitor and the heterogeneity
Therefore, under Assumptions 1–3, total welfare loss over the two
in these effects in the population (Heckman, 1996, 1997, 2001;
periods in a CER-based data production world is given as:
Heckman and Vytlacil, 1999, 2005; Heckman et al., 2006; Basu,
2014) without the challenges of voluntary participation in CER LCER = LPRE,CER +LPOST,CER (19)
RCTs.
(3) Sequential decision making: Based on the outcome evaluation where LPRE,CER and LPOST,CER are given in Eqs (7) and (16), respec-
results, fractional allocation rules can be adapted over time tively.
for specific risk groups. Fractional allocation would continue Under the LtD framework of data production, welfare losses in
within risk groups where ambiguity persists. Optimal stopping both periods will be different. Let the total welfare loss over the
rules for fractional coverage can be developed using Bayesian two periods in a LtD-based data production world is given as:
methods. LLtD = LPRE,LtD + LPOST,LtD

3.1. Key Features of the Learning through Diversification Since the LtD infrastructure allows for consistent estimation of
Infrastructure the mean treatment effect parameters, in the second period, sub-
jective beliefs about the benefits of treatment  will align with the
3.1.1. Coinsurance (demand price) as an instrument true values for subjects in each risk groups, E ˛ ˆ k = ˛k . Moreover,
Traditional IV analyses focus around the debate on whether a  estimates
given that these   data
were generated using  at the popula-
chosen instrument is contaminated, given that the strength of the tion scale, ˚ ˛ ˆ k /ˆ k → 1 for ˛k ≥ 0 and ˚ ˛
ˆ k /ˆ k → 0 for ˛k < 0.
instrument is testable. In the LtD framework, the lottery, by design, Consequently,
is orthogonal to all confounders and therefore side steps the typical ∞   
 ˛ˆk
debates in this literature. The strength of the instrument is driven LPOST,LtD = 1−˚ × Nk × (˛k − p)
by variation in out-of-pocket payments by patients that in turn will ˆ k
˛k =p
depend on the market price of the new technology and the price
elasticity of demand. The LtD infrastructure would be most efficient 
p   
˛
ˆk
in data production for CER for new technologies that come at a high + ˚ × Nk × (p − ˛k ) → 0 (20)
ˆ k
price tag, which aligns with the notion, that most welfare can be ˛k =−∞
generated if we can properly identify people who would and would
Therefore, an LtD infrastructure will always be welfare enhancing
not benefit from the most expensive technologies.
compared to the CER infrastructure as long as
   
3.1.2. The target parameters in the LtD infrastructure LLtD − LCER = LPRE,LtD + LPOST,LtD − LPRE,CER + LPOST,CER
Meaningful and interpretable structural parameters for evalua-
tion can be recovered using data arising out of an LtD infrastucture < 0 → (LPRE,LtD ) < (LPRE,CER + LPOST,CER )
(Heckman and Vytlacil, 1999, 2001; Heckman et al., 2006, Basu,
2014)15 . For example, local instrumental variable approaches can

correlated with factors such as income (because of the price differentials that the
15
Note that it is important to pay close attention to dealing with essential het- lottery creates), which in turn may be correlated with gains and losses from the new
erogeneity within the LtD infrastructure. This is because treatment receipt will be treatment.
172 A. Basu / Journal of Health Economics 42 (2015) 165–173
That is fractional allocation should be designed in a way that
the loss during the initial data generation process is not greater
than the combined losses under the CER infrastructure both during
and after the data generation process. To obtain the most power
for analyses and consistency within the LtD infrastructure, it may
be useful to set the mean coinsurance rate to be 0.5. The welfare
losses, if any, during the data production period (that may be one
or few years) can be easily recuperated from the welfare gains from
LtD in the post data-production period that typically lasts for many
years.
4. Conclusions
Regulatory bodies often approves a new medical treatment
based on its potential safety profile and its incremental efficacy
compared to either placebo or a basic control treatment. Often
superiority of the new treatment is not established and its compar-
ative effectiveness compared to current clinical practice remains
ambiguous. Nevertheless, the treatment becomes available for
consumption at a substantial price in anticipation of a positive
effectiveness claims based on efficacy results. Variability in effec-
tiveness profile remains far from known. Under such ambiguity, a
social insurer faces the challenge of deciding whether to pay for
the treatment. In the US, public health insurance provider like the
Medicare usually extend full coverage of these new treatments as
long as there is positive efficacy signals. Other countries, like UK and
Canada, formally look at the budget impact of coverage by compar-
ing the costs of treatments (inclusive of its price) to the projected
effectiveness based on efficacy signals. When coverage is allowed,
a large welfare loss may ensue even when the new treatment can
genuinely produce higher effectiveness in a certain margin of the
population. This is due to the lack of evidence of how to match
patients to alternative treatments.
In this paper, I show that under the status quo policy of extend-
ing coverage to a new treatment in the absence of complete
information on its effectiveness profile, welfare loss can be substan-
tial. These losses can be minimized by investments in studies that
aims at generating such evidence. However, I also show, following
a Roy’s model of sorting behavior, that the current infrastructure
on data production for this purpose, suffer for severe self-selection
issues since the incentives to enroll in research studies is eroded
by the low demand prices of obtaining medical care outside these
studies. Consequently, the parameters identified from these stud-
ies do not inform any of the decision-relevant parameters, either
at the individual or the population level. I show that if one takes
the normative approach of a social insurer who is forward look-
ing and wants to maximize any given social welfare function
over a duration of time period (typically over the longevity of
the new technology being considered), then it makes sense for
the social insurer, irrespective of what coverage decision is made
today, to device ways to learn about variations in incremental
effectiveness of treatment in the population so that she can encour-
age/discourage appropriate subgroups to uptake/discard the new
treatment. In fact, generating such public evidence can directly
inform individuals within the population to use this new treat-
ment appropriately without additional effort by the social insurer,
thereby approaching the second-best solutions. Based on this
normative framework, I propose a positive Learning through Diver-
sification (LtD) infrastructure, through which a social insurer can
achieve her objectives.
The LtD infrastructure comprises of introducing the new
treatment with fractional coverage based random individual-level
co-insurance rates. One then uses these co-insurance rates as an
artificially created, but an almost perfect, instrumental variable
to study treatment effect heterogeneity based on a spectrum of
econometric tools available to researchers. Both clinical guidelines
and coverage decisions can then be sequentially revised to reflect
this evidence. I show that under non-stringent conditions, the LtD
infrastructure will be welfare enhancing compared to the current
data production infrastructure, such as CER.
One aspect of the LtD infrastructure that would appear to be
politically challenging is the notion of fractional coverage, albeit it
is for a short time during the introduction of the new treatment.
A full legal and ethical consideration of such random allocation is
beyond the scope of this paper. However, it is important to note
that unlike earlier discussions in this line of reasoning that revolved
around quasi-random treatment prescription (Manski, 2009), the
LtD infrastructure does not withhold treatment from anyone, but
rather changes the cost of accessing it in a random fashion. The
potential for patient welfare and the richness of scientific and policy
question that this infrastructure can answer should play a part in
deciding its ultimate feasibility.
References
Arrow, K.J., 1963. Uncertainty and the welfare economics of medical care. The Amer-
ican Economic Review 53 (5), 941–973.
Basu, A., 2011. Economics of individualization in comparative effectiveness research
and a basis for a patient-centered health care. Journal of Health Economics 30
(3), 549–559.
Basu, A., 2014. Person-centered treatment (PeT) effects using instrumental vari-
ables: An application to evaluating prostate cancer treatments. Journal of
Applied Econometrics 29 (4), 671–691.
Basu, A., Jena, A.B., Philipson, T.J., 2011. The impact of comparative effectiveness
research on health and health care spending. Journal of Health Economics 30
(4), 695–706.
Dehejia, R.H., 2005. Program evaluation as a decision problem. Journal of Economet-
rics 125, 141–173.
Garber, A., Phelps, C., 1997. Economic foundations of cost-effectiveness analysis.
Journal of Health Economics 16, 1–31.
Heckman, J.J., 1996. Randomization as an instrumental variable. The Review of Eco-
nomics and Statistics 78 (2), 336–341.
Heckman, J.J., 1997. Instrumental variables: a study of implicit behavioral assump-
tions used in making program evaluations. Journal of Human Resources 32 (3),
441–462.
Heckman, J.J., 2001. Accounting for heterogeneity, diversity and general equi-
librium in evaluating social programmes. The Economic Journal 111,
F654–F699.
Heckman, J.J., Urzua, S., Vytlacil, E., 2006. Understanding instrumental variables in
models with essential heterogeneity. Review of Economics and Statistics 88 (3),
389–432.
Heckman, J.J., Vytlacil, E., 1999. Local instrumental variables and latent variable mod-
els for identifying and bounding treatment effects. Proceedings of the National
Academy of Sciences 96 (8), 4730–4734.
Heckman, J.J., Vytlacil, E., 2001. Local instrumental variables. In: Hsiao, C., Morimue,
K., Powell, J.L. (Eds.), Nonlinear Statistical Modeling: Proceedings of the Thir-
teenth International Symposium in Economic Theory and Econometrics: Essays
in the Honor of Takeshi Amemiya. Cambridge University Press, New York, NY,
pp. 1–46.
Heckman, J.J., Vytlacil, E., 2005. Structural equations, treatment effects and econo-
metric policy evaluation. Econometrica 73 (3), 669–738.
Ioannidis, J.P., Lau, J., 1997. The impact of high-risk patients on the results of clinical
trials. Journal of Clinical Epidemiology 50 (10), 1089–1098.
Malani, A., 2008. Patient enrollment in medical trials: selection bias in a randomized
experiment. Journal of Econometrics 144, 341–351.
Manning, W.G., Marquis, M.S., 1996. Health insurance: the trade-off between
risk pooling and moral hazard. Journal of Health Economics 15 (5),
609–639.
Manski, C., 2000. Identification problems and decisions under ambiguity: empiri-
cal analysis of treatment response and normative analysis of treatment choice.
Journal of Econometrics 95, 415–442.
Manski, C., 2004. Statistical treatment rules for heterogeneous populations. Econo-
metrica 72, 1221–1246.
Manski, C., 2009. The 2009 Lawrence R. Klein Lecture: Diversified treatment under
ambiguity. International Economic Review 50 (4), 1013–1041.
Meltzer, D., 1997. Accounting for future costs in medical cost-effectiveness analysis.
Journal of Health Economics 16, 33–64.
Movsas, B., Moughan, J., Owen, J., Coia, L.R., Zelefsky, M.J., Hanks, G., Wilson, J.F.,
2007. Who enrolls onto clinical oncology trials? A radiation patterns of care
study analysis. International Journal of Radiation Oncology, Biology, Physics 68
(4), 1145–1150.
Pauly, M.V., 1968. The economics of moral hazard: comment. The American Eco-
nomic Review 58 (3), 531–537.
 A. Basu / Journal of Health Economics 42 (2015) 165–173 173

Pauly, M.V., 2008. Adverse selection and moral hazard: implications for health insur-
ance markets. In: Sloan, F, Kasper, H (Eds.), Incentives and Choice in Health and
Health Care. MIT Press, Cambridge, MA.
Pauly, M.V., Blavin, F.E., 2008. Moral hazard in insurance, value-based cost shar-
ing, and the benefits of blissful ignoring. Journal of Health Economics 27,
1407–1417.
Philipson, T.J., 1997. The evaluation of new health care technology: the labor eco-
nomics of statistics. Journal of Econometrics 76, 375–395.
Roy, A.D., 1951. Some thoughts on the distribution of earnings. Oxford Economic
Papers 3 (2), 135–146.
Weinstein, M., Zeckhauser, R., 1973. Critical ratios and efficient allocation. Journal
of Public Economics 2, 147–158.
 Journal of Health Economics 42 (2015) 174–185

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Short- and medium-term effects of informal care provision on female

caregivers’ health夽
Hendrik Schmitz a,∗ , Matthias Westphal b
a
University of Paderborn, RWI & CINCH, Germany
b
Ruhr Graduate School in Economics, University of Duisburg-Essen & CINCH, Germany

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, we present estimates of the effect of informal care provision on female caregivers’ health.
Received 24 October 2013 We use data from the German Socio-Economic Panel and assess effects up to seven years after care
Received in revised form 13 January 2015 provision. The results suggest that there is a considerable negative short-term effect of informal care
Accepted 4 March 2015
provision on mental health which fades out over time. Five years after care provision the effect is still
Available online 3 April 2015
negative but smaller and insignificant. Both short- and medium-term effects on physical health are virtu-
ally zero throughout. A simulation analysis is used to assess the sensitivity of the results with respect to
JEL classification:
potential deviations from the conditional independence assumption in the regression adjusted matching
I10
I18
approach.
C21 © 2015 Elsevier B.V. All rights reserved.
J14

Keywords:
Informal care
Regression adjusted matching
Propensity score matching
Mental health
Physical health

1. Introduction
Europe’s societies are getting older. Low birthrates and popu-
lation ageing due to technological progress in medicine shift the
age structure towards higher shares of elderly individuals. This has
strong implications for labour markets and social security systems
with the long-term care sector as one important part of those. The
World Alzheimer Report, for instance, expects, as a result of grow-
ing numbers of people in need of long-term care, publicly funded
夽 We thank Martin Fischer, Pilar García-Gómez, Audrey Laporte, Jan Marcus,
Jürgen Maurer, Alfredo Paloyo, Stefan Pichler, and Arndt Reichert for valuable sug-
gestions. We further thank two anonymous referees and two editors for very helpful
comments. Moreover, we are grateful for comments at the 22nd European Work-
shop on Econometrics and Health Economics (Rotterdam), the meeting of the health
economics section of the VfS (Hamburg), the annual meeting of the dggö (Essen),
the CINCH health economics seminar in Essen, the CINCH academy, the economics
of disease conference in Darmstadt, and seminars in Bayreuth and Paderborn. All
errors are our own. Financial support by the Fritz Thyssen Stiftung is gratefully
acknowledged.
∗ Corresponding author at: University of Paderborn, Warburger Strasse 100, 33098
Paderborn, Germany. Tel.: +49 5251 603213.
E-mail address: hendrik.schmitz@uni-paderborn.de (H. Schmitz).
costs of long-term care in the European Union (EU 27) to increase
from 1.2% of GDP in 2007 to 2.5% in 2060 (Alzheimer’s Disease
International, 2013).
Already today, costs are one reason why many governments pre-
fer informal care (care provision of close relatives and friends) over
professional formal care provision. In Germany, for instance, the
public long-term care insurance paid 700D per month in 2012 for
care recipients of the highest care level who are cared by family
members and 1550D per month to the same recipient cared by
professional caregivers. Germany is a country in which long-term
care is still predominantly regarded the task of the family (Schulz,
2010) and informal care is more common than in comparable states
like the Netherlands (Bakx et al., 2015). More than one million offi-
cial care recipients (about 46% of all) are exclusively cared by family
members rendering informal care the most important part of the
German long-term care system.
However, provision of informal care is both mentally and phys-
ically challenging. We, therefore, analyse the question of whether
there are some hidden costs – or costs often neglected in the pub-
lic debate – that make informal care provision not as economic as
often thought. This could be the case if informal care provision goes
along with health impairments of the caregivers. Other costs (not

http://dx.doi.org/10.1016/j.jhealeco.2015.03.002
0167-6296/© 2015 Elsevier B.V. All rights reserved.
 H. Schmitz, M. Westphal / Journal of Health Economics 42 (2015) 174–185
considered here but heavily analysed in the economic literature1 )
are forgone income for those who leave the labour force to provide
care.
The economic literature on health effects of caregiving is fairly
scarce.2 To the best of our knowledge, there are only three stud-
ies on the effect of care provision on health in a narrow sense. Coe
and van Houtven (2009) estimate health effects of informal care-
giving in the US using seven waves of the Health and Retirement
Survey (HRS). They use sibling characteristics and the death of the
mother as instrumental variables that control for selection into and
out of caregiving in order to identify causal effects. They find that
continued caregiving leads to a significant increase in depressive
symptoms for both sexes while physical health does not seem to be
affected. Do et al. (2015) use data from South Korea where informal
care is quite common among females caring for their parents-in-
law. The data allow identifying a health effect for daughters-in-law
where selection into care is taken into account by instrumenting
the informal care decision with parents-in-law’s health endow-
ment. Their findings suggest that there is an increased probability
of worse physical health by providing informal care. Di Novi et al.
(2013) use the first two waves of SHARE to estimate the effect of
caregiving on self-rated health and quality of life, measured by the
CASP-12. They find positive effects of care provision on self-rated
health (seen as a measure of physical health) and mixed evidence
regarding quality of life (seen as a measure of mental health).
Two further papers evaluate the relationship of caregiving and
caregiver drug utilisation. On the one hand, drug intake could be
seen as an objective measure of poor health. On the other hand, it
sheds light on direct costs of caregiving. Van Houtven et al. (2005)
assess the impact of caring on the intake of drugs using data on
caregivers for US veterans. One finding is that the intensive care
margin is an important factor for drug intake. Schmitz and Stroka
(2013) exploit data of a large German sickness fund that enables
to consider prescriptions of anti-depressants and drugs to restore
physical health. Their results support Van Houtven et al. (2005),
providing some evidence that caregiving increases the intake of
anti-depressants in particular if coupled with having a job. Other
studies look at broader welfare consequences of caring and use
life satisfaction as a proxy (Bobinac et al., 2010, Van den Berg and
Ferrer-i Carbonell, 2007, Leigh, 2010, van den Berg et al., 2014). One
issue with these studies is that they do not address reverse causality
and selection problems based on time-varying unobserved hetero-
geneity.
We use representative household data from the German Socio-
Economic Panel to estimate the effects of informal care provision
on female caregivers’ health. The outcome variables are mental
and physical summary scale measures (called MCS and PCS) for
the years 2002 to 2010 that capture the multidimensional nature
of health. Our contributions to the literature on health and infor-
mal care are twofold: First, we use a different approach to address
selection into and out of care provision. Except for Di Novi et al.
(2013), previous studies that deal with endogeneity problems all
use instrumental variables approaches. We try to identify the effect
of caring using different assumptions that can put the literature on
a broader basis and thereby complement it. Our approach is to fully
exploit the time dimension and richness of panel data in order to
1
E.g., Carmichael and Charles, 2003; Heitmueller, 2007; Heitmueller and Inglis,
2007; Bolin et al., 2008; Leigh, 2010; Van Houtven et al., 2013; Meng, 2013.
2
In the medical literature, there is a fair amount of studies on the relationship of
health and care provision. They mainly stem from the US (see e.g., Schulz et al., 1995;
Stephen et al., 2001; Gallicchio et al., 2002; Tennstedt et al., 1992; Beach et al., 2000;
Ho et al., 2009; Shaw et al., 1999; Lee et al., 2003; Dunkin and Anderson-Hanley,
1998; Colvez et al., 2002). In general, these studies use non-representative samples
and widely disregard endogeneity problems. Furthermore, they often concentrate
on more specific definitions of care, such as caring for people with dementia.
175
justify the conditional independence assumption that would allow
for a causal interpretation of the results. To be more precise, we
use a regression adjusted matching approach. Although we argue
below that, given our, data we can justify the conditional indepen-
dence assumption, we allow in a sensitivity analysis that follows
Ichino et al. (2008) for certain deviations from this assumption.
Second, to the best of our knowledge, this is the first study
that does not only look at contemporary, or short-term effects
of informal care provision on health, but also on medium-term
effects of up to seven years after care provision. By medium-term
effects we mean: if a women provides care in a certain year, what
is her expected change in health up to seven years afterwards.
This adds on work by Coe and van Houtven (2009) who also dis-
cuss persistence of health effects but need to stick to a two year
period. Medium-term consequences could be more severe than
instantaneous short-term health impacts restricted to the period
of providing care. Moreover, knowledge about the persistence of
health effects is arguably more important for policy makers than
about short-run effects only.
The results suggest that there is a considerable negative short-
term effect of informal care provision on mental health which,
however, fades out over time. Five years after care provision the
effect is still negative but smaller and insignificant. Both short- and
medium-term effects on physical health are virtually zero through-
out. The sensitivity analysis suggests that sensible deviations from
the conditional independence assumption do not change these
results.
The paper is organized as follows. Section 2 briefly outlines the
institutional setting of long-term care in Germany. Section 3 dis-
cusses the empirical approach, Section 4 presents the data. The
results are reported in Section 5 while Section 6 assesses the sensi-
tivity of the results. Section 7 concludes.
2. Institutional background
The German social long-term care insurance system was
introduced in 1995 as a pay-as-you-go system. It is financed by
a mandatory pay payroll tax deduction of currently 2.35% of gross
labour income (2.6% for employees without children). In order to
qualify for benefits, individuals need to be officially defined as care
recipients and be classified into one of three care levels. In care
level one individuals need support in physical activities for at least
90 min per day and household help for several times a week. Indi-
viduals in need of more care are classified into care levels two or
three, where the benefits increase in care levels.
Benefits also depend on the type of care, where monthly pay-
ments for informal care range from 235D (level one) to 700D (level
three), for professional ambulatory care from 450D to 1550D and
for professional nursing home care from 1023D to 1500D . The lat-
ter, in particular, does not fully cover the expenses for nursing home
visits and copayments of up to 50% are standard. Copayments for
professional ambulatory care are smaller and amount to an aver-
age of 247D or about 20% (Schmidt and Schneekloth, 2011). Social
welfare may step in if individuals are not able to bear the copay-
ment. Thus, the decision for formal or informal ambulatory care is
usually not driven by financial aspects as each care recipient who
is assigned a care level is entitled to benefits for all kinds of care.
The introduction of the insurance system in 1995 stressed the
family as the main provider of care, as it is thought to provide
care cheaper, more agreeable, and more efficiently. From the care
recipient’s perspective, the decision to receive informal care typi-
cally expresses a preference for being cared by familiar relatives
or friends. In some cases, informal care recipients are addition-
ally supported by professional carers. These are, on average older
recipients with a higher care level and, thus, a higher care burden
176 H. Schmitz, M. Westphal / Journal of Health Economics 42 (2015) 174–185
(Schulz, 2010). Apart from the care burden, a reason for profes-
sional care can be the absence of appropriate informal caregivers,
either because they chose to only participate in the labour market
or because their own physical or mental health conditions prohibits
the full amount of necessary care provision.
From the caregiver’s perspective, affection and sense of respon-
sibility towards a loved parent or spouse mainly drive the decision
to provide care. Although the insurance benefits for informal care
are often passed on to the care provider this comparably small
amount cannot be regarded a financial incentive to provide care,
as it is also needed to cover other expenses for care provision (see
Schmidt and Schneekloth, 2011 for all points). However, the insur-
ance funds do pay pension contributions for informal carers who
provide care at least 14 h a week (Schulz, 2010). In 2002, people
cared on average 14 h per week for care recipients whose assess-
ment of needs is at least classified as the lowest official category
(Schneekloth and Leven, 2003).
Between 2001 and 2011 there were only minor adjustments
to the German long-term care system. They were minor because
benefits were increased but only to keep pace with the inflation
(Rothgang, 2010) and, thus, did not change the incentives to provide
care. As of 2008, employed individuals are allowed to take a 10 day
(not repeatable) unpaid leave to organize or provide care in case of
an incidence of care dependency in the family. However, only very
few caregivers make use of this.3 Thus, the tasks of informal care-
givers, the composition of caregivers and care recipients as well as
financial incentives remained fairly similar over time.
3. Empirical strategy
We aim at estimating the effect of informal care provision on
health. Certainly, the decision to provide care is not random per se.
Given that someone close becomes care dependent, some individ-
uals choose to provide care while others do not. The willingness to
provide care depends on factors such as the financial and temporal
affordability, own health endowment as well as innate tendencies
such as personality traits.
To deal with this problem we apply the model of Rubin (1974).
Following his notation we observe Y = T · Y1 + (1 − T) · Y0 , where T
indicates whether an individual is assigned to treatment (2 h of
informal daily care provision, but we will also consider alterna-
tive definitions) or control group, Y is the outcome (health), and
the index {0, 1} indicates the potential health outcome of being a
caregiver or not. If we simply compare the realized outcomes, i.e.,
E(Y1 |T = 1) − E(Y0 |T = 0), selection bias will most likely arise due to
the non-randomness of care provision. However, the average treat-
ment effect on the treated (ATT) can be identified if the conditional
independence assumption holds and assignment to treatment is
random conditional on controls: Y1 , Y0 ⊥ T|X. That is, if all the deter-
minants that simultaneously influence the health outcome and the
selection into treatment are observed. Then, ATT = E(Y1 − Y0 |T = 1, X)
is the causal ceteris paribus impact of informal care provision on
health.
We use propensity score methods to estimate this effect and
combine matching with regression methods, thus employing the
so called regression adjusted matching approach (see, for exam-
ple, Rubin, 1979). The advantage to using either only matching or
linear regression is that it yields consistent estimates if either one
of each method fails to remove the selection bias. This is called the
double robustness property (Bang and Robins, 2005). Nevertheless,
this method rests on the conditional independence assumption and
3
Schmidt and Schneekloth (2011) report that only 9000 out of possibly 150,000
made use of this until 2011. The most frequent reason for not making use in their
survey was that individuals were not aware of the possibility.
Fig. 1. Basic time structure.
if it does not hold and both, regression model and propensity score
estimation are wrongly specified, the estimates are biased. The esti-
mation strategy is a two-step process, originally proposed by Bang
and Robins (2005). As a first step, the probability of being a care-
giver (the propensity score) conditional on relevant covariates is
estimated with a probit model. Subsequently, treatment and con-
trol group are matched. We use an Epanechnikov kernel with a
bandwidth of 0.03 in the basic specification. To further increase
the comparability, the sample is restricted to the common support
of the propensity scores of the treatment and control group.
As a second step, the health outcome is regressed on informal
care and, again, all control variables where the observations are
weighted by the kernel weights W estimated by the matching algo-
rithm: ˇ ˆ = (X  WX)−1 X  y. Standard errors are computed according
to the suggestion of Marcus (2014) who employs robust standard
errors of the regression above since they are slightly more con-
servative but easier to estimate than bootstrapped standard errors
that, in addition, are not formally justified.4 However, we cluster
standard errors on the individual level since individuals appear
several times in the data set.
We employ the time structure as presented in Fig. 1. Assign-
ment to treatment T occurs in t = 0. We condition on a large set
of covariates in t = −1, thus reducing the potential problem that
covariates are affected by the treatment status. We, then, compute
the treatment effect four times: 1 year after treatment, 3 years after
treatment, 5 years after treatment, and 7 years after treatment. Note
that conditioning variables and treatment group assignment are
always the same and determined in t = −1 and t = 0, respectively. As
explained in Section 4, the outcome variable is available biannually
between 2002 and 2010 in our data set. Since we condition on pre-
treatment outcome (see explanation below), the earliest possible
treatment year is 2003. We use the maximum available informa-
tion in the data and pool it to one sample. Then, individuals treated
in 2003 (call this wave 1) can be followed until t = 7 in 2010 whereas
individuals treated in 2009 (call this wave 4) can only be followed
until t = 1. Hence, the effect in t = 1 will be measured more precisely
than the one in t = 7.
Even though we condition on a large set of covariates that are
supposed to capture the process of the decision to provide care,
there are probably some threats to the conditional independence
assumption. First, there might be health driven selection into
treatment. Individuals who are confronted with the question to
provide care but are themselves in poor health might not be able
to do so. As informal care provision is both physically and mentally
challenging, this possible selection holds for both dimensions of
health. If this is indeed the case and informal care provision has
negative health effects, ignoring this reverse causality problem
would lead to an underestimation of the true effects (in absolute
values). We follow, e.g., Lechner (2009a) and García-Gómez (2011)
and match individuals on pre-treatment outcomes (here, health
status in t = −1), thus only comparing individuals of the same
4
We can confirm this finding in our data. Bootstrapped standard errors yield
slightly less conservative standard errors.
 H. Schmitz, M. Westphal / Journal of Health Economics 42 (2015) 174–185 177

Table 1 transition of care provision between t = 2 and t = 3. Hence, we follow,

Stratified sample.
e.g., Lechner et al. (2011) and use the standard static version. This
Stratum t = −1 t=0 enables us to estimate the average effect of care provision in t = 0 on
1 Care Care health in later years. This effect is generated by dynamics in care
No care provision which are not explicitly modelled but implicitly taken
2 No care Care into account. The descriptive statistics in the next section show that
No care the vast majority of care durations is between one and three years.
Hence, in reality, there is much less heterogeneity in care durations
than implied by all theoretically possible paths in Fig. 2(b).
baseline health status before treatment. This rules out that indi-
Above, we have set out selection issues and the responses to
viduals in the control group are in worse health due to a selection
those that are facilitated by our data. To assess the adequacy of
of healthy individuals into care provision.
our responses we report a sensitivity analysis in Section 6. We esti-
A second issue is unobserved heterogeneity, confounders that
mate a short- and a medium-term effect of care provision, where
both affect treatment and outcome, but are not observable for
by medium-term effect we mean the expected health effect of care
the researcher. As Lechner (2009a) suggests, this problem can be
provision in a certain year, five or seven years after. Given that an
mitigated by stratifying the sample according to care provision
individual cannot foresee her care provision path in the future, this
in t = −1. Comparing only individuals with the same care sta-
expected effect (though probably a composite of effects from dif-
tus in t = −1 accounts for a lot of unobserved heterogeneity that
ferent paths) is arguably the most relevant one from an individual
affects treatment participation. Hence, the conditional indepen-
perspective when deciding about providing care in t = 0 or not.
dence assumption is much more likely to hold within the strata of
previous care provision.5 Moreover, stratifying the sample at least
4. Data
mitigates the problem that control variables, though dated back
to t = −1, could be determined by care provision in t = 0 through
We use data from the German Socio-Economic Panel (SOEP)
confounders that both affect past control variables and current
which is a yearly repeated representative longitudinal survey of
treatment status.
households and persons living in Germany that started in 1984.
Hence, we generate two samples based on information in t = −1
The SOEP covers a wide range of questions on the socio-economic
and estimate the treatment effects independently for each sample
status like on work, education, health, and personal attitudes (see
as depicted in Table 1. Both estimated treatment effects and their
Wagner et al., 2007, for details). Currently, some 22,000 individ-
variances for each stratum are merged as weighted means.6
uals above the age of 18 from more than 10,000 households are
Note that treatment is only defined as care provision in t = 0
interviewed each year.
while we leave future care status unrestricted as exemplarily
We restrict the sample to women that have complete informa-
shown in Fig. 2(a) for care starters. The most important advan-
tion on treatment status in t = 0 and control variables in t = −1. Since
tage of this is that selection out of care provision due to bad health
caregiving among men is much less common and we observe con-
is no problem in identifying medium-term effects of care provi-
siderably fewer male caregivers we drop men, as it turned out to
sion because future health status – potentially affected by care and
be very difficult to properly model the treatment participation (the
potentially leading to selection out of care provision in later years –
propensity scores yielded only very low values). Moreover, we drop
does not affect the treatment group assignment at all. A drawback
female professional caregivers from the sample, as they might mix
might be that in this static model sequential paths of care provision
up professional and personal affairs. Beyond that, no further restric-
are not explicitly modelled. Fig. 2(b) shows some examples of paths
tions are imposed on the sample. Pooling all waves as shown in
after t = 0. Individuals who care in t = 0 might either stop in t = 1 or
Fig. 1, we end up with a sample of 31,177 person-year observations
go on and stop later, or even stop and take up care provision again.
in t = 0. The lowest line of Table 2 shows the number of observa-
The same holds for the control group that includes individuals who
tions in the sample. Of the 31,177 observations in t = 0, we observe
cared later on.
the health status of 28,622 in t = 1, of 20,288 in t = 3, and so on. This
Lechner (2009b) and Lechner and Miquel (2010) present a
number strongly drops over time, mainly because more and more
dynamic matching model that enables to compare the effect of, say,
episodes are right censored (again, see Fig. 1).
providing care in each period between t = 0 and t = 7 with, say, not
We identify caregivers depending on how individuals respond
caring in any of the periods.7 We do not make use of such a frame-
to the following question: “What is a typical day like for you? How
work. Most importantly, because the numbers of observations in
many hours do you spend on care and support for persons in need
the 256 (= 28 ) different paths become very small in our sample,
of care on a typical weekday?” which has been included into the
except for the path of never carers (0 − 0 − . . . −0). E.g., only 23
SOEP questionnaire since 2001.8 Answers to this question are also
observations in our data set provide care in each year between t = 0
shown in Table 2. 862 or 41% of all individuals who care a positive
and t = 7. Second, given that we condition on pre-treatment out-
number of hours per day, care for 1 h. 24% care 2 h, whereas 10%
come, we would need to condition on the health status at each node
care 3 h per day. Note that the numbers in t = 1 (and later) do not
in Fig. 2(b) in order to make the “dynamic conditional indepen-
refer to care provision in t = 1 but to the number of observations
dence assumption” (Lechner and Miquel, 2010) credible. However,
who care in t = 0 and are still observed in t = 1.
as the outcome variable is only available in every other year, we
If an individual states caring at least 2 h per day we consider
cannot, for instance, condition on health in t = 2 in modelling the
her a caregiver. That is, the treatment indicator is the binary vari-
able of caring for at least 2 h per day. This comes closest to other
definitions in the literature, e.g., Leigh (2010) who defines care pro-
5
However, for stratum 1 (individuals who already provided care in t = −1) there vision as caring at least for 10 h per week. Below we show that the
is presumably more unobserved heterogeneity left, since here, all individuals that
results are robust to higher or lower thresholds. The question does
have been caring, potentially for several years, are pooled. Thus, we identify only an
average effect over all conceivable care spells. This, however, holds for all studies
that cross-sectional data or panel data and do not explicitly model the dynamics of
care.
 1 8
This question does not refer to child care which is a separate category in the
6 
ATT = 1
n i∈1,2
ni · 
ATT i , se =
n2 i∈1,2
n2i · se2i . time use questionnaire. The Supplementary Material includes a paragraph on the
7
See also Augurzky et al. (2012) for an application. justification for the validity of self-reported answers to these kinds of questions.
178 H. Schmitz, M. Westphal / Journal of Health Economics 42 (2015) 174–185
Fig. 2. Group assignment rules. Note: 1 = providing care; 0 = not providing care; X = care status not specified (= either 1 or 0). Right panel does not include all possible paths
but only a small excerpt.
Table 2
Sample size.
t=0 t=1
Hours of care = 0 29,080 26,667
Hours of care = 1 862 (= 41%) 800
Hours of care = 2 507 (= 24%) 479
Hours of care = 3 203 (= 10%) 193
Hours of care = 4 167 (= 8%) 152
Hours of care > 4 358 (= 17%) 331
All observations 31,177 28,622
Source: SOEP, own calculations. Number in parentheses is the share among all individuals with positive hours of care. Hours of care are measured in t = 0 only.
not allow for a link between caregiver and care recipient. Hence,
we have no information on the care recipient and we are not able
to stratify our analysis with respect to her (e.g., in order to eval-
uate differences between caring for spouses or parents). This is a
common shortcoming in this literature.
Table 3 gives a notion of the duration of care episodes. It counts
the consecutive years individuals provide care of at least 2 h per
day. In presenting the numbers we distinguish between uncen-
sored spells (of individuals that are observed to provide no care
before and after a care episode) and censored spells (individuals
that either enter the sample as caregivers or are caregivers at the
end of the observation period). Due to the sample construction,
there are many right censored individuals which complicates the
interpretation of the table somewhat. What should be taken away
from it is that the vast majority has care spells of about one to three
years. Therefore, the effects after seven years are mainly driven by
individuals who had shorter caregiving episodes. Individuals who
constantly care over many years hardly add to the results.9
The two outcome measures are a mental and a physical health
score that are based on information from the SF-12v2 question-
naire, a component of the SOEP, which includes twelve questions
on mental and physical health. All items capture the general cur-
rent mental and physical health status since all questions relate
to the past four weeks, see the questionnaire in Table A2 in the
Appendix. Answers to these questions are collapsed into the Men-
tal Component Summary Scale (MCS) and the Physical Component
Summary Scale (PCS) by explorative factor analysis (see, Andersen
9 10
This is due to the very low number of observations. Moreover, these 19 indi-
viduals caring throughout in our sample exhibit a mean MCS of 45.81 (compared
to 49.38 overall). Thus, they do not affect the results in a quantitatively important
way.
t=3 t=5 t=7
18,956 11,455 5,194
564 357 160
317 197 85
140 81 36
100 53 24
211 111 53
20,288 12,254 5,552
et al., 2007). Thus, both variables capture the multidimensional
aspect of health. The scales range from 0 to 100, normalised to
mean values of 50 and standard deviations of 10 in the 2004 refer-
ence sample. Higher values mean a better health status. MCS loads
information on perceived melancholy, time pressure, mental bal-
ance and emotional problems into one summary scale.10 The SF-12
is commonly used to measure general health and functioning in
epidemiological research (Ware et al., 1996). It includes informa-
tion on subjective health but the component summary scales are
correlated actual with health diagnoses. For example, Gill et al.
(2007) find that MCS”is a useful screening instrument for depres-
sion and anxiety disorders in the general community, and thus, a
valid measure of mental health”. This view is supported by Vilagut
et al. (2013) who find”acceptable results for detecting both active
and recent depressive disorders in general population samples”.
This property could build the bridge between the short-term symp-
toms that are measured to longer-lasting health consequences that
are thus also captured by this summary scale. Salyers et al. (2000)
regard it as a valid and reliable instrument to measure health-
related quality of life. Recently, MCS has also been used in the
economic literature where it was shown to be correlated with, e.g.,
unemployment (Schmitz, 2011; Reichert and Tauchmann, 2011),
and unemployment of spouses (Marcus, 2013). MCS and PCS were
first introduced in the SOEP in 2002 and subsequently sampled
every other year. This is why we restrict our observation period
to the years 2002–2010.
The physical component comprises: Physical fitness (2 Questions), general
health, bodily pain, role physical (2). The mental component comprises: Mental
health (2), role emotional (2), social functioning, vitality. See the questionnaire in
Table A2 in the Appendix.
 H. Schmitz, M. Westphal / Journal of Health Economics 42 (2015) 174–185 179

Table 3
Care duration.

Years of consecutive care as of t = 0 1 2 3 4 5 6 7 8 Total

Uncensored Observations 348 107 35 29 8 7 6 - 542

Share 65% 19% 7% 6% 1% 1% 1% - 100%

Censored Observations 238 183 77 90 37 39 12 19 693

Share 35% 26% 12% 13% 5% 5% 1% 3% 100%

thereof:
Left censored Observations 80 27 16 11 10 6 4 19 173
Share 46% 16% 9% 7% 6% 3% 2% 1% 100%

Total Observations 586 290 112 119 45 46 18 19 1235

Share 47% 23% 9% 10% 4% 4% 1% 2% 100%

Source: SOEP, own calculations. Uncensored individuals did not provide care in t = −1 and stopped caregiving some time before t = 7. Therefore, the maximum observable
care duration is 7 years. In contrast to the empirical analysis in the rest of the paper, this table uses information up to the wave of 2011 or t = 8 in order to be able to calculate
the number of individuals who exactly care for 7 years.

We now turn to the selection of the control variables. Taking on the SOEP and in years after the treatment assignment,13 they are
the burden of care could theoretically be modeled as a three-stage useful controls because these measures are supposed to be stable
process. Women provide care if (i) they need to. Given that they over a shorter period of time. The individual average of each mea-
need to provide care, they (ii) must be willing to do so. Finally, (iii), sure is taken over all years as a proxy for time invariant personality.
they need to be able to provide care.11 Finally, on the third stage, the own health status determines the
At the first stage, the event that someone close becomes care ability to provide care. As discussed in Section ‘2, we control for pre-
dependent is a prerequisite of the need to provide informal care. treatment health (MCS and PCS). Moreover, we control for health
This first stage in general depends on the age and the intra-familial satisfaction and life satisfaction. All control variables are listed in
social environment. We model the social environment by using Table 4. Variables that might theoretically belong into the model
indicators whether parents are alive, their age as well as the num- but were not significant in the propensity score regression are left
ber of siblings.12 The latter can reduce the need to provide care for out. This holds, for instance, for income, the age of the father, the
frail parents as siblings could step in. Variables on this stage are number of brothers, or calendar year dummies.
sometimes employed as instruments for care provision in other
studies.
5. Results
At the second stage, given that someone close is in need of
care, the willingness to provide care can be modeled as a function
5.1. Matching quality
of socio-economic characteristics and personality traits. Socio-
economic characteristics grouped in here are, e.g., own age, marital
Table 4 reports descriptive statistics of all covariates for different
status, employment status, and level of education. Note, however,
subgroups. It reveals that the mean as well as the standard deviation
that family background variables might also belong to the first
of the covariates are significantly different in the unweighted base-
stage. For instance, singles do not need to care for a spouse or
line sample. Column 4 gives the standardized difference between
parents-in-law. Furthermore, we use character traits measured
both means. Without matching almost all confounders are differ-
in the Big Five Inventory (Big5), well-known in psychology for
ent at the 5% significance level between the carer and non-carer
being a proxy of human personality (see McRae and John, 1992
sample. In particular age, the age of the mother, and marital sta-
or Dehne and Schupp, 2007) as well as positive and negative reci-
tus exhibit large differences but also personality traits seem to be
procity. Although the SOEP captures each item of the Big5 with
quite strong predictors of care provision. The kernel matching algo-
relatively few questions in the 2005 and 2009 questionnaires, sur-
rithm equalizes both samples by assigning different weights to each
veys revealed sufficient validity and reliability (see Dehne and
member of the control group. In order to compute these weights, we
Schupp, 2007). The items of the Big5 are: neuroticism, the ten-
employ an Epanechnikov kernel with a bandwidth of 0.03. Whereas
dency of experience negative emotions; extraversion, the tendency
a bandwidth of 0.06 does not accomplish to equalize all covariates,
to be sociable; openness, the tendency of being imaginable and
a bandwidth of half the size balances every control variable to a
creative; agreeableness, the dimension of interpersonal relations
standardized bias around 5 or less.
and conscientiousness the dimension of being moral and orga-
As regards the propensity score, the regions of common sup-
nized (see Budria and Ferrer-i Carbonell, 2012). There are three
port are roughly [0.04, 0.14] for the stratum of women who did not
questions for each of these items which are gathered on a 7-item
provide care in t = −1 and [0.23, 0.87] for those who did provide
scale. Furthermore, there is positive reciprocity, the tendency of
care. The overlap within each stratum is good as we do not lose
being cooperative and negative reciprocity, the tendency of being
treatment observations by restricting the sample to the common
retaliatory. For each personality measure, the score is generated
support.14 The low probabilities in the first stratum are simply due
by averaging over the outcome of the corresponding questions per
to the small amount of caregivers. This indicates that there is a
individual. Although these questions are only prompted twice in
large unobserved component determining caregiving. But we argue
that this unobserved heterogeneity is not a big concern given the
estimation strategy outlined in Section 3. Yet, there is one advan-
tage of this fuzziness: It brings about a sufficiently large amount
11
Note that we do not explicitly model this three-stage process but that we just
have it in mind. Which variable belongs to which stage is then just a matter of
interpretation.
12 13
However, the number of brothers does not seem to play a role statistically. Thus, The Big5 are included in the surveys in 2005 and 2009, whereas questions on
in the empirical model we only focus on the number of sisters. An alternative spec- negative and positive reciprocity are asked in 2005 and 2010.
14
ification using that – among others – also uses the number of brothers can be found Of course, this also means that the required overlap condition stating that some
in the Supplementary Material. randomness is needed is ensured in our model (see Heckman et al., 1998).
180 H. Schmitz, M. Westphal / Journal of Health Economics 42 (2015) 174–185

Table 4
Descriptive statistics according to treatment and matching status.

Treated Controls Matched controls Standardized bias

mean SD mean SD mean SD unmatched sample matched sample

(0.06) (0.03)

Stage (I): care obligations

Age of mother
∈[30, 39] 0.01 0.09 0.02 0.16 0.01 0.10 -13.31 -5.15 -2.35
∈[40, 49] 0.03 0.18 0.10 0.30 0.04 0.20 -27.00 -9.33 -3.24
∈[50, 59] 0.08 0.27 0.13 0.34 0.08 0.28 -18.15 -6.47 -2.22
∈[60, 69] 0.12 0.32 0.12 0.32 0.11 0.32 0.25 0.96 1.12
∈[70, 79] 0.09 0.28 0.06 0.24 0.08 0.28 9.16 4.00 1.71
∈[80, 89] 0.09 0.28 0.02 0.14 0.08 0.28 30.58 6.81 1.15
∈[90, 99] 0.01 0.10 0.00 0.03 0.01 0.09 12.04 3.36 2.10
Mother alive 0.46 0.50 0.48 0.50 0.46 0.50 −5.72 −3.05 −0.97
Age of father
∈[30, 39] 0.00 0.07 0.01 0.09 0.01 0.07 −5.14 −2.17 −1.03
∈[40, 49] 0.02 0.12 0.08 0.27 0.03 0.16 −30.29 −10.98 −4.58
∈[50, 59] 0.04 0.20 0.10 0.30 0.05 0.22 −20.75 −7.05 −2.12
∈[60, 69] 0.07 0.25 0.08 0.27 0.07 0.25 −6.43 −2.13 −0.39
∈[70, 79] 0.04 0.19 0.04 0.19 0.04 0.19 −0.40 −0.84 −0.87
∈[80, 89] 0.01 0.11 0.01 0.10 0.01 0.11 2.34 0.44 −0.17
∈[90, 99] 0.00 0.05 0.00 0.01 0.00 0.03 7.10 4.80 4.80
Father alive 0.19 0.40 0.34 0.47 0.21 0.41 −32.50 −11.54 −4.01
Number of sisters 1.08 1.21 1.09 1.21 1.09 1.23 −1.12 −1.28 −1.20
Partner existent 0.81 0.39 0.68 0.47 0.80 0.40 29.91 9.71 2.28
Age of partner 47.73 26.02 35.53 27.08 46.56 25.96 45.93 15.60 4.40

Stage (II): willingness to provide care

NEURO 4.53 0.67 4.37 0.72 4.51 0.71 22.53 7.58 2.05
CONSC 6.04 0.74 5.97 0.79 6.04 0.77 9.76 2.74 0.29
AGREE 5.61 0.83 5.58 0.84 5.60 0.84 3.66 1.40 0.90
OPENN 4.37 1.15 4.51 1.12 4.40 1.13 −11.92 −4.86 −2.20
EXTRA 5.02 0.91 5.04 0.95 5.02 0.94 −1.97 −1.07 −0.49
Positive reciprocity 5.66 0.95 5.55 0.99 5.67 0.96 11.40 2.92 −0.39
Negative reciprocity 2.71 1.19 2.87 1.24 2.73 1.22 −12.81 −4.20 −1.34
Acceptance of private funding 3.31 0.81 3.29 0.8 3.31 0.82 2.68 1.12 0.44
Age 56.28 12.85 49.57 16.34 55.31 13.56 45.68 16.86 6.65
Age squared 3333.01 1419.70 2724.16 1691.40 3242.65 1471.10 38.99 14.54 5.79
Married 0.80 0.40 0.63 0.48 0.79 0.41 39.14 13.06 3.67
Divorced 0.07 0.25 0.09 0.28 0.07 0.25 −7.60 −2.67 −0.77
Single 0.07 0.25 0.17 0.38 0.08 0.27 −32.80 −11.03 −3.57
Children hh 0.18 0.38 0.30 0.46 0.19 0.40 −29.50 −11.17 −4.27
Educ general 0.17 0.37 0.17 0.38 0.17 0.37 −2.31 −0.99 −0.73
Educ middle 0.55 0.50 0.49 0.50 0.54 0.50 11.66 4.11 1.45
Foreign 0.04 0.20 0.06 0.24 0.05 0.21 −9.91 −4.15 −2.07
West 0.69 0.46 0.75 0.43 0.70 0.46 −13.54 −5.43 −2.10
Full time 0.13 0.34 0.26 0.44 0.15 0.35 −33.85 −11.66 −3.85

Stage (III): ability to provide care

MCS 47.38 10.52 49.47 10.12 47.59 10.91 −20.23 −6.82 −2.01
PCS 46.44 10.01 49.02 10.14 46.79 10.47 −25.57 −9.63 −3.48
Satisfaction health 6.19 2.21 6.58 2.17 6.25 2.24 −17.62 −6.56 −2.42
Satisfaction life 6.60 1.85 6.97 1.76 6.62 1.95 −20.64 −6.56 −1.07

N 1,235 29,942 29,942

The standardized difference is calculated according to: Diff = 100 ·  x̄1 −x̄0
where 0.06 and 0.03 refer to the employed Kernel bandwidth. While the bandwidth of 0.06
1 ( 2 + 2 )
2 1 0
is only shown for sake of illustration, 0.03 is used in the estimations.

of observations in the control group having a similar value of the
estimated propensity score. This provides a hint that the results are
not sensitive to a different choice of the matching methods.
5.2. Estimation results
The baseline estimation results are reported in Fig. 3 for both
outcome variables MCS (3(a)) and PCS (3(b)). For convenience,
we restrict this section to a graphical presentation of the results.
Table A1 in the Appendix gives an overview of all results shown
in this section. The dotted lines denote 95% confidence bands for
the corresponding effect. Fig. 3 reports the results for both pre-
treatment strata separately. Care starters (black points) are those
who did not care in t = −1 and care continuers (light grey points)
those who did care in t = −1. The confidence bands are wider for
care continuers, since this is a much smaller group. The weighted
average over both effects has confidence bands comparable to the
black ones in Fig. 3.
The effects are remarkably similar for both groups. If a woman
cares at least 2 h per day, her mental health score decreases by
2.00 units (or 20% of a standard deviation, SD)15 in the first year,
all other things equal. Three years after treatment assignment, this
effect reduces to 16% of a SD before settling at below 12% five and
seven years afterwards. That is, women who provide care in t = 0 can
expect to have a reduced mental health score by 12% of a SD seven
15
For convenience we already report the average effect over both groups here.
 H. Schmitz, M. Westphal / Journal of Health Economics 42 (2015) 174–185
(a) MCS
Fig. 3. Baseline results MCS and PCS. Source: SOEP. Own calculations. Note: The dotted lines indicate 95% confidence bands.
years after. The confidence bands indicate significant results at the
5% level one and three years after assignment to treatment. The
effects five and seven years after are insignificant because the point
estimates attenuate but in the first place because the numbers of
observations strongly drop. The magnitude of the effect after seven
years, however, is still 60% the amount of the baseline effect and
thus, not negligible. All in all it is fair to note that, independent of
the previous care status, there is a considerable short-term effect of
care provision on mental health (in line with findings from previous
studies, e.g., Coe and van Houtven, 2009) which decreases over time
without being completely irrelevant in its extent to those who care.
In contrast, for PCS (right panel), there is basically a zero effect
throughout all periods and for both strata, providing evidence for
negligible effects of informal care provision on physical health.
Given the absence of physical health effects, we restrict our analysis
to mental health in the following. Moreover, we only report aver-
aged effects of both strata of care provision in t = −1. Fig. 4 presents
the results for alternative daily care intensities and different def-
initions of the control group.16 In Fig. 4(a) we compare the effect
when care provision of at least 2 h per day are used to define the
treatment indicator (light grey-dashed line, the baseline specifica-
tion) with 1 h per day (black line) and 3 h per day (dark grey-dashed
line). There are basically no differences in the effect between 1 and
2 h of care as a definition. As regards 3 h of care we find a consid-
erably stronger short-term effect with a reduction of MCS by 31%
of a sd. This probably reflects a higher burden of higher care inten-
sities. Subsequently, however, the effect does not remain on this
high level. It immediately drops back to regions similar to those
for 1 and 2 h. Most notably, the qualitative result of a considerable
short-term effect and a much smaller medium-term effect remains
unchanged regardless of the care intensity.17
The definition of treatment and control group only in t = 0 allows
for cases where individuals in the control group start to provide care
in later years. This is in fact the case for some 15% of all observations
in the control group. It might be suspected that these individuals
suffer from a short-term mental health drop later which, compared
with the effects in the treatment group, lead to the observed rel-
ative decline in the mental health drop of the treatment group.
16
In the Supplementary Material, we also report the results for females caring 4 h
and more. The results are comparable.
17
Although not shown here, also the PCS results are robust to these different
definitions.
181
(b) PCS
In order to test if this drives the results, we exclude all individ-
uals from the control group that provided care in any year between
t = 1 and t = 7. That is, we only use individuals in the lowest path in
Fig. 2(b). In principle, this is not a desirable specification as it bases
the control group definition on later outcomes. Thus, it should only
be regarded as a brief check whether these individuals drive the
results observed above. Fig. 4(b) shows that this is not the case. The
results are largely the same.
The results suggest a significant short-term effect of informal
care-provision on mental health while there is a smaller and not
significant medium-term effect. Given that the vast majority of
individuals provide care for about one to three years, the main
pathway of these effects is probably the following one. Contem-
poraneously, care provision is a mentally burdensome task. The
short-term effects are mostly generated by individuals who just
stopped to provide care or who are still providing care in t = 1. As
to be expected, this effect increases in care intensity. Yet, after the
care episode ceased, individuals recover and their mental health
status approaches former levels.
The short-term effect is not necessarily entirely due to care pro-
vision. It might be a joint effect of care and the observation of the
decline of a beloved person. As most of the previous literature,
we cannot disentangle the family effect from the active caregiving
effect. As results of Bobinac et al. (2010) suggest, the overall effect
is a mixture of both but a caregiving effect remains after control-
ling for the family effect. Yet, this does not affect the interpretation
of the medium-run effect of almost no mental health impairment
a couple of years after care provision. Given that the effect in t = 7
is very small, it can be concluded that there is less evidence for
a scarring effect of care provision. Moreover, since only handful
individuals in the sample care throughout the entire observation
period, this result can apparently not be explained by an adaptation
effect of care providers to their new situation.
In Section 4 we mentioned that we cannot stratify the analysis
with respect to the care recipient as we do not have information
on who is being cared. We can, however, approach such an anal-
ysis by splitting the sample into caregivers below and above the
age of 60. The former group has a higher likelihood to care for a
parent while the latter should be more likely to care for a spouse.
Note that stronger restrictions such an age cutoff at 70 or groups
such as unmarried women with at least one parent alive are hardly
feasible due to strongly reduced numbers of observations. Fig. 5
shows the effect over time for both subgroups. Initially, they coin-
cide nearly perfectly. Five years after care is observed, they deviate
182 H. Schmitz, M. Westphal / Journal of Health Economics 42 (2015) 174–185
Fig. 4. Alternative definitions of treatment and control groups (MCS only). Source: SOEP. Own calculations. Note: The dotted lines indicate 95% confidence bands.
from each other. Whereas younger carers drop back almost to the
initial level, for older carers the impact on their mental score is
even stronger. The results could be interpreted such that the active
caregiving effect does not depend on the care recipient. However,
a likely family effect might arguably be stronger in case of carepro-
vision for a spouse than for oldest old parents. However, the effects
come closer after seven years and due to large confidence bands
one should interpret these results cautiously.
Altogether, the results from this section could be interpreted as
good news. While there is a considerable negative short-term effect
of contemporaneous caregiving, the scarring effect is less likely to
be prevalent. One negative interpretation for these results could,
however, be an increased consumption of antidepressants as found
by Van Houtven et al. (2005) and Schmitz and Stroka (2013) for the
short run. If this would hold for the long run, the mental health
score might increase over time due to drug consumption and not
due to improved health. Whether this is the case or not requires
long-term data on care and drug consumption and is left for future
research.
In the Supplementary Material we report results from alterna-
tive specifications of the propensity score, the treatment indicator
and a subgroup analysis for unmarried women with at least one
Fig. 5. Alternative definitions of treatment and control groups (MCS only). Source:
SOEP. Own calculations. Note: The dotted lines indicate 95% confidence bands.
parent alive who arguably can be identified as caring for their par-
ents.
6. Sensitivity analysis
Thus far, we have argued that our estimation strategy allows us
to interpret the results in a causal manner since, by fully exploiting
the panel information in the SOEP, the conditional independence
assumption is likely to hold. However, this inherently untestable
assumption might nevertheless fail. For example, in the context of
care, it might be particularly challenging to properly control for
intrinsic willingness to provide care. Yet, the conditional indepen-
dence assumption is not necessarily an “all or nothing” assumption
and there might be different degrees of its violation. To examine
to what extent the magnitude and the significance of our results
depend on the potential exclusion of a relevant variable, we follow
an approach by Ichino et al. (2008) who refined the suggestions for
sensitivity analyses by Rosenbaum and Rubin (1983) and Imbens
(2003). This analysis is also in the spirit of the one suggested by
Altonji et al. (2005) without the need to make strong parametric
assumptions.
Assume that the conditional independence assumption does not
hold but that the failure is due to an unobserved variable U. If
we could condition on it, we would be able to restore conditional
independence:
Y0 ⊥⊥ T |(X, U).
Hence, all the unobserved heterogeneity that results in bias is cap-
tured by U. For simplicity, Ichino et al. (2008) follow Rosenbaum
and Rubin (1983), who proposed a binary U.
We simulate U by drawing 200 times from the Bernoulli dis-
tribution for each individual and estimate the ATT 200 times,
conditioning on X as before, but also on U.18 In simulating U, we
make sure that it is both correlated with T and Y such that leaving
it out would result in a violation of the conditional independence
assumption. Taking the average over all effects provides us with
18
This section contains a non-technical and intuitive discussion of the analysis. A
more detailed account is provided in the Supplementary Material published online.
For an extensive treatment, refer to Ichino et al. (2008).
 H. Schmitz, M. Westphal / Journal of Health Economics 42 (2015) 174–185
robust point estimates as well as standard errors of the average
treatment effect on the treated.19
The major question is how strong and in what direction the cor-
relation between U and Y resp. T should be defined. We follow one
of the two approaches suggested by Ichino et al. (2008) and set it
such that we control the “outcome effect” (own effect of U on Y)
and the “selection effect” (effect of U on T). As an illustration, think
of U as general intrinsic willingness to provide care: U = 1 indicates
generally willing, U = 0 means not willing. This unobserved variable
certainly has a positive selection effect such that willing people are
more likely to provide care. It may also have a positive outcome
effect if the general willingness is positively correlated with health
endowment independent of treatment.
The magnitudes of outcome and selection effects could be arbi-
trarily chosen. One way to find reasonable values is to use observed
binary variables in the data set and calculate the observed selection
and the outcome effects of these variables. This gives an indica-
tion of the distribution of selection and outcome effects in the data.
To bound these effects, one could argue that the unobserved vari-
able U should not have much larger selection and outcome effects
than important observed variables, for which we have a long vector,
including age, education, and initial health.
We compute these effects for all variables in the sample. Results
are reported in the Supplementary Material. We then choose the
parameters to simulate U such that is has an effect on treatment and
outcome in the same magnitude as the control variable with the
highest effect (which is the potential caregiver’s age). With these
calibrations, no other confounder in the sample (except for age)
features such a high effect on mental health and no other makes
people select into treatment like the simulated binary confounder
U. The first assumed selection effect reflects a positive selection
into treatment, i.e., more people with high values of U will take
the treatment. Together with a positive outcome effect, we should
underestimate effects of care provision on health. The second pair
of selection and outcome effects reflects a negative selection into
treatment leading to overestimation if this was neglected by the
analysis so far.
Fig. 6 presents the results of both specifications and the baseline
specification for MCS. Including a confounder U with characteristics
that lead to a positive selection into treatment (the dark grey-
dashed line) leads to larger effects of care provision than in the
baseline case while we find weaker effects when including a con-
founder that induces as negative selection (the light grey-dashed
line). The lines are parallel-shifted by the confounder. However, in
all three cases, we find a significant (both statistically and econom-
ically) short-term effect of care provision on mental health which
reduces over time. After seven years, the effects are insignificant for
most specifications, marginally significant though for the one with
a confounder inducing a positive selection into treatment. Thus, if
there are further confounders that point in the same direction as
most of the variables in our sample, our result will define a lower
bound. Furthermore, this would raise the likelihood of significant
medium-term effects.
Thus, as long as unobserved effects that are necessarily left out in
our analysis do not have a drastically higher impact than observed
control variables, we find that the average treatment effects we
received in the main analysis are robust. Given that we control for
a large set of important determinants of care provision, it seems
unlikely that there are actually unobserved variables that have
such a drastic effect or an effect much stronger than observable
covariates.
19
We use a modified version of the user-written Stata command sensatt
(Nannicini, 2007).
183
Fig. 6. Results of the sensitivity analysis (MCS). Source: SOEP. Own calculations.
Note: The dotted lines indicate 95% confidence bands. Strong positive selection
assumes a selection effect of s = 0.25 and an outcome effect of d = 0.25. See the Supple-
mentary Material for exact definitions of s and d and justifications for these values.
Strong negative selection assumes a selection effect of s = −0.25 and an outcome
effect of d = 0.25.
7. Conclusion
This paper examines whether informal care provision affects the
report of measures that indicate mental or physical strain among
women. We use the German Socio-Economic Panel that identifies
informal caregivers by the daily time spent caring. We define care-
givers as women who care at least 2 h per weekday (but other
definitions lead to a similar picture). We evaluate the impact of
caregiving on health by help of a regression adjusted matching
technique. The problems of unobserved heterogeneity and reverse
causality are tackled by exploiting the panel structure of the data
set and controlling for pre-treatment outcome as well as stratifying
by pre-treatment care status.
While we do not find effects of informal care on physical health
in the short- and in the medium-run, our results suggest that there
are considerable short-term effects of informal care provision on
mental health which, however, attenuate over time. Five years after
care provision the effect is still negative but smaller and insignifi-
cant. It seems that, contemporaneously, care provision is a mental
burden but there is not a large scarring effect. The sensitivity anal-
ysis according to Ichino et al. (2008) suggests that these results
are stable even for considerable deviations from the conditional
independence assumption: the effects are still similar in magni-
tude even if we falsely have not incorporated a confounder that is
stronger than any other one that we have controlled for before.
We contribute to the current debate on how to realign the
care system in Germany and countries with similar demographic
developments. Our results suggest that there are considerable
short-term health effects and although it seems to be good news
that the effects are abating, it should not be concluded that there is
no need to improve the system as apart from health there are other
additional effects of care provision (e.g., labour force participation
and wages) that are not analysed here. The current German govern-
ment put the enhancement of informal care high on the agenda.20
In particular, the supply of low-threshold services is planned to
be expanded and increases in benefits from the long-term care
insurance are meant to be spent on those. These services are, e.g.,
additional help in the household, contact persons in case of any
20
http://www.bmg.bund.de/ministerium/presse/english-version.html
184 H. Schmitz, M. Westphal / Journal of Health Economics 42 (2015) 174–185

problems, or professional short-term care (also overnight) in case health outcome in seven years (irrespective of future events that I
of short-term absence of informal care providers due to sickness, cannot control today).
obligations in the job, or holidays. Thus, while family members
will certainly continue to play an important role in care provision, Appendix A.
these measures are thought to assist them and to reduce the most
stressful aspects of care. Tables A1 and A2
The measured effect in this study is an average effect over dif-
ferent groups of care providers. Schmitz and Stroka (2013), for
instance, focus on individuals who not only provide informal care Table A1
but also work full-time. This double burden might well also have Table of results.
health effects in the longer run. This question is left for future t=1 t=3 t=5 t=7
research. The main limitations in this study arise from the imperfect
Care 2 h per day (baseline) −2.00*** −1.64*** −1.01 −1.19
data set. Both measures of care provision as well as health indica- (0.39) (0.47) (0.62) (0.86)
tors are self-reported and potentially measured with error. We do ...care starters −2.03*** −1.67*** −1.02 −1.21
not observe any characteristics of the care recipient. Hence, we can- (0.40) (0.49) (0.64) (0.88)
not distinguish between the family effect that occurs just because a ...continued care −1.42** −0.93* −0.70 −0.94
(0.61) (0.74) (0.89) (1.51)
close relative is in need of care and the caregiving effect. However, Care 3 h per day −3.02*** −1.44** −1.00 −1.64
this should not qualitatively affect the interpretation of the already (0.53) (0.68) (0.78) (1.12)
small medium-term effect. Likewise, as it is not observed whether Care 1 hr per day −1.90*** −1.59*** −0.48 −0.97
care recipients receive additional professional care or only informal (0.31) (0.39) (0.48) (0.67)
Observations 28,622 20,288 12,254 5,552
care, we cannot discriminate between cases in which the caregiver
assists professional care and in which she is the only care provider. Only never carers in control group
Moreover, due to data restrictions we are not able to identify the Care 2 h per day −2.08*** −1.56*** −1.25** −0.787
(0.27) (0.34) (0.46) (0.69)
cumulative effect of care provision for many consecutive years. This
Observations:
might go along with even long-run health impairments. However, 25,914 18,464 11,301 5,166
our representative data suggest that only a very small group of
PCS as outcome:
women is faced with the need (and willingness) to provide care Care 2 h per day 0.14 0.14 0.08 −1.01
for many consecutive years. Moreover, we argue that our approach (0.33) (0.40) (0.49) (0.84)
allows us to answer a question that is more relevant from an indi- Source: SOEP, own calculations. Note: * p < 0.1; ** p < 0.05; *** p < 0.01 indicate the cor-
vidual perspective: if I provide care today, what is my expected responding significance level. Standard errors are in parantheses.

Table A2
SF-12v2 questionnaire in the SOEP.

Very good Good Satisfactory Poor Bad

How would you describe your current health?

Greatly Slightly Not at all – –

When you ascend stairs, i.e. go up several floors on foot: Does your
state of health affect you greatly, slightly or not at all?
And what about having to cope with other tiring everyday tasks,
i.e. where one has to lift something heavy or where one requires
agility: Does your state of health affect you greatly, slightly or
not at all?

Please think about the last four weeks. How often did it occur Always Often Sometimes Almost never Never
within this period of time, . . .

♦that you felt rushed or pressed for time?

♦that you felt run-down and melancholy?
♦that you felt relaxed and well-balanced?
♦that you used up a lot of energy?
♦that you had strong physical pains?
♦that due to physical health problems
. . .you achieved less than you wanted to
at work or in everyday tasks?
. . .you were limited in some form
at work or in everyday tasks?
♦that due to mental health or emotional problems
. . .you achieved less than you wanted to
at work or in everyday tasks?
. . .you carried out your work or everyday tasks
less thoroughly than usual?
♦that due to physical or mental health problems you were limited
socially, i.e., in contact with friends, acquaintances or relatives?

Note. Source: SOEP Individual question form. Available at: http://panel.gsoep.de/soepinfo2008/.

 H. Schmitz, M. Westphal / Journal of Health Economics 42 (2015) 174–185
Appendix B. Supplementary Data
Supplementary data associated with this article can be found, in
the online version, at http://dx.doi.org/10.1016/j.jhealeco.2015.03.
002
References
Altonji, J.G., Elder, T.E., Taber, C.R., 2005. Selection on observed and unobserved
variables: assessing the effectiveness of catholic schools. Journal of Political
Economy 113 (1), 151–184.
Alzheimer’s Disease International, 2013. World Alzheimer Report 2013. Journey
of Caring – An Analysis of Long-Term Care for Dementia. Technical Report.
Alzheimer’s Disease International (ADI), London.
Andersen, H.H., Mühlbacher, A., Nübling, M., Schupp, J., Wagner, G.G., 2007. Compu-
tation of standard values for physical and mental health scale scores using the
SOEP version of SF-12v2. Schmollers Jahrbuch 127, 171–182.
Augurzky, B., Reichert, A., Schmidt, C.M., 2012. The effect of a bonus program for
preventive health behavior on health expenditures. Ruhr Economic Papers 373,
Essen.
Bakx, P., de Meijer, C., Schut, F., van Doorslaer, E., 2015. Going formal or informal,
who cares? The influence of public long-term care insurance. Health Economics
24 (6), 631–643.
Bang, H., Robins, J.M., 2005. Doubly robust estimation in missing data and causal
inference models. Biometrics 61 (4), 962–973.
Beach, S.R., Schulz, R., Yee, J.L., Jackson, S., 2000. Negative and positive health effects
of caring for a disabled spouse: longitudinal findings from the caregiver health
effects study. Psychology and Aging 15 (2), 259–271.
Bobinac, A., van Exel, N.J.A., Rutten, F.F., Brouwer, W.B., 2010. Caring for and caring
about: disentangling the caregiver effect and the family effect. Journal of Health
Economics 29 (4), 549–556.
Bolin, K., Lindgren, B., Lundborg, P., 2008. Your next of kin or your own career?
Caring and working among the 50+ of europe. Journal of Health Economics 27
(3), 718–738.
Budria, S., Ferrer-i Carbonell, A., 2012. Income comparisons and non-cognitive skills.
In: SOEPpapers No. 441., pp. 1–29.
Carmichael, F., Charles, S., 2003. The opportunity costs of informal care: does gender
matter? Journal of Health Economics 22 (5), 781–803.
Coe, N.B., van Houtven, C.H., 2009. Caring for mom and neglecting yourself? The
health effects of caring for an elderly parent. Health Economics 18 (9), 991–1010.
Colvez, A., Joel, M.-E., Ponton-Sanchez, A., Royer, A.-C., 2002. Health status and work
burden of Alzheimer patients’ informal caregivers: comparisons of five different
care programs in the European Union. Health Policy 60 (3), 219–233.
Dehne, M., Schupp, J., 2007. Persoenlichkeitsmerkmale im Sozio-ökonomischen
Panel (SOEP) – Konzept, Umsetzung und empirische Eigenschaften. Technical
report.
Di Novi, C., Jacobs, R., Migheli, M., 2013. The Quality of Life of Female Informal Care-
givers: From Scandinavia to the Mediterranean Sea. CHE Research Paper 84.
Centre for Health Economics, University of York.
Do, Y.K., Norton, E.C., Stearns, S., Houtven, C.H.V., 2015. Informal care and caregiver’s
health. Health Economics 24 (2), 224–237.
Dunkin, J.J., Anderson-Hanley, C., 1998. Dementia caregiver burden – a review of
the literature and guidelines for assessment and intervention. Neurology 51 (1),
53–60.
Gallicchio, L., Siddiqi, N., Langenberg, P., Baumgarten, M., 2002. Gender differences
in burden and depression among informal caregivers of demented elders in the
community. International Journal of Geriatric Psychiatry 17 (2), 154–163.
García-Gómez, P., 2011. Institutions, health shocks and labour market outcomes
across Europe. Journal of Health Economics 30 (1), 200–213.
Gill, S.C., Butterworth, P., Rodgers, B., Mackinnon, A., 2007. Validity of the men-
tal health component scale of the 12-item short-form health survey (mcs-12)
as measure of common mental disorders in the general population. Psychiatry
Research 152 (1), 63–71.
Heckman, J., Ichimura, H., Smith, J., Todd, P., 1998. Characterizing selection bias using
experimental data. Econometrica 66 (5), 1017–1098.
Heitmueller, A., 2007. The chicken or the egg? Endogeneity in labour market par-
ticipation of informal carers in England. Journal of Health Economics 26 (3),
536–559.
Heitmueller, A., Inglis, K., 2007. The earnings of informal carers: wage differentials
and opportunity costs. Journal of Health Economics 26 (4), 821–841.
Ho, S.C., Chan, A., Woo, J., Chong, P., Sham, A., 2009. Impact of caregiving on health and
quality of life: a comparative population-based study of caregivers for elderly
persons and noncaregivers. The Journals of Gerontology, Series A: Biological
Sciences and Medical Sciences 64 (8), 873–879.
Ichino, A., Mealli, F., Nannicini, T., 2008. From temporary help jobs to permanent
employment: what can we learn from matching estimators and their sensitivity.
Journal of Applied Econometrics 23, 305–327.
Imbens, G.W., 2003. Sensitivity to exogeneity assumption in program evaluation.
American Economic Review 93 (2), 126–132.
Lechner, M., 2009a. Long-run labour market and health effects of individual sport
activities. Journal of Health Economics 28, 839–854.
185
Lechner, M., 2009b. Sequential causal models for the evaluation of labor market
programs. Journal of Business and Economic Statistics 27, 71–83.
Lechner, M., Miquel, R., 2010. Identification of the effects of dynamic treatments by
sequential conditional independence assumptions. Empirical Economics 39 (1),
111–137.
Lechner, M., Miquel, R., Wunsch, C., 2011. Long-run effects of public sector sponsored
training in West Germany. Journal of the European Economic Association 9 (4),
742–784.
Lee, S., Colditz, G.A., Berkman, L.F., Kawachi, I., 2003. Caregiving and risk of coronary
heart disease in U.S. women: a prospective study. American Journal of Preventive
Medicine 24 (2), 113–119.
Leigh, A., 2010. Informal care and labor market participation. Labour Economics 17
(1), 140–149.
Marcus, J., 2013. The effect of unemployment on the mental health of spouses
– evidence from plant closures in Germany. Journal of Health Economics 32,
546–558.
Marcus, J., 2014. Does job loss make you smoke and gain weight? Economica 324
(81), 626–648.
McRae, R.R., John, O.P., 1992. An introduction to the five factor model and its appli-
cations. Journal of Personality and Social Psychology 60 (2), 175–215.
Meng, A., 2013. Informal home care and labor-force participation of household
members. Empirical Economics 44 (2), 959–979.
Nannicini, T., 2007. Simulation-based sensitivity analysis for matching estimators.
Stata Journal 7 (3), 334–350.
Reichert, A., Tauchmann, H., 2011. The Causal Impact of Fear of Unemployment on
Psychological Health. Ruhr Economic Papers 266.
Rosenbaum, P.R., Rubin, D.B., 1983. Assessing sensitivity to an unobserved binary
covariate in an observational study with binary outcome. Journal of the Royal
Statistical Society, Series B: Methodological 45 (2), 212–218.
Rothgang, H., 2010. Social insurance for long-term care: an evaluation of the German
model. Social Policy and Administration 44 (4), 436–460.
Rubin, D.B., 1974. Estimating causal effects of treatments in randomized and non-
randomized studies. Journal of Educational Psychology 56 (5), 688–701.
Rubin, D.B., 1979. Using multivariate matched sampling and regression adjustment
to control bias in observational studies. Journal of the American Statistical Asso-
ciation 74 (366), 318–328.
Salyers, M.P., Bosworth, H.B., Swanson, J.W., Lamb-Pagone, J., Osher, F.C., 2000. Reli-
ability and validity of the sf-12 health survey among people with severe mental
illness. Medical Care 38 (11), 1141–1150.
Schmidt, M., Schneekloth, U., 2011. Abschlussbericht zur Studie “Wirkungen des
Pflege-Weiterentwicklungsgesetzes”.
Schmitz, H., 2011. Why are the unemployed in worse health? The causal effect of
unemployment on health. Labour Economics 18 (1), 71–78.
Schmitz, H., Stroka, M.A., 2013. Health and the double burden of full-time work and
informal care provision: evidence from administrative data. Labour Economics
24, 305–322.
Schneekloth, U., Leven, I., 2003. Hilfe und Pflegebedürftige in Privathaushalten in
Deutschland 2002.
Schulz, E., 2010. The Long-Term Care System for the Elderly in Germany. DIW Dis-
cussion Paper 1039, DIW Berlin.
Schulz, R., O’Brien, A., Bookwala, J., Fleissner, K., 1995. Psychiatric and physical
morbidity effects of dementia caregiving: prevalence, correlates, and causes.
Gerontologist 35 (6), 771–791.
Shaw, W.S., Patterson, T.L., Ziegler, M.G., Dimsdale, J.E., Semple, S.J., Grant, I., 1999.
Accelerated risk of hypertensive blood pressure recordings among Alzheimer
caregivers. Journal of Psychosomatic Research 46 (3), 215–227.
Stephen, M.A., Townsend, A.L., Martire, L.M., Druley, J.A., 2001. Balancing parent
care with other roles: interrole conflict of adult daughter caregivers. The Jour-
nals of Gerontology, Series B: Psychological Sciences and Social Sciences 56 (1),
P24–P31.
Tennstedt, S., Cafferata, G.L., Sullivan, L., 1992. Depression among caregivers of
impaired elders. Journal of Ageing and Health 4 (1), 58–76.
Van den Berg, B., Ferrer-i Carbonell, A., 2007. Monetary valuation of informal care:
the well-being valuation method. Health Economics 16 (11), 1227–1244.
van den Berg, B., Fiebig, D.G., Hall, J., 2014. Well-being losses due to care-giving.
Journal of Health Economics 35 (0), 123–131.
Van Houtven, C., Wilson, M., Clipp, E., 2005. Informal care intensity and caregiver
drug utilization. Review of Economics of the Household 3 (4), 415–433.
Van Houtven, C.H., Coe, N.B., Skira, M.M., 2013. The effect of informal care on work
and wages. Journal of Health Economics 32 (1), 240–252.
Vilagut, G., Forero, C.G., Pinto-Meza, A., Haro, J.M., de Graaf, R., Bruffaerts, R., Kovess,
V., de Girolamo, G., Matschinger, H., Ferrer, M., Alonso, J., 2013. The mental
component of the short-form 12 health survey (sf-12) as a measure of depres-
sive disorders in the general population: results with three alternative scoring
methods. Value in Health 16 (4), 564–573.
Wagner, G.G., Frick, J.R., Schupp, J., 2007. The German Socio-Economic Panel
Study (SOEP), scope, evolution, and enhancements. Journal of Applied Social
Science Studies (Schmollers Jahrbuch: Zeitschrift für Wirtschafts- und Sozial-
wissenschaften) 127 (1), 139–169.
Ware, J.E., Kosinski, M., Keller, S.D., 1996. A 12-item short-form health survey: con-
struction of scales and preliminary tests of reliability and validity. Medical Care
34 (3), 220–233.
 Journal of Health Economics 42 (2015) 186–196

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Tobacco control campaign in Uruguay: Impact on smoking cessation

during pregnancy and birth weight
Jeffrey E. Harris a,∗ , Ana Inés Balsa b , Patricia Triunfo c
a
Department of Economics, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
b
Departamento de Economía, Facultad de Ciencias Empresariales y Economía, Universidad de Montevideo, Montevideo 11500, Uruguay
c
Departamento de Economía, Facultad de Ciencias Sociales, Universidad de la República, Montevideo 11200, Uruguay

a r t i c l e i n f o a b s t r a c t

Article history: We analyzed a nationwide registry of all pregnancies in Uruguay during 2007–2013 to assess the impact of
Received 23 April 2014 three types of tobacco control policies: (1) provider-level interventions aimed at the treatment of nicotine
Received in revised form 13 April 2015 dependence, (2) national-level increases in cigarette taxes, and (3) national-level non-price regulation
Accepted 20 April 2015
of cigarette packaging and marketing. We estimated models of smoking cessation during pregnancy at
Available online 29 April 2015
the individual, provider and national levels. The rate of smoking cessation during pregnancy increased
from 15.4% in 2007 to 42.7% in 2013. National-level non-price policies had the largest estimated impact
JEL classification:
on cessation. The price response of the tobacco industry attenuated the effects of tax increases. While
I18
I12
provider-level interventions had a significant effect, they were adopted by relatively few health centers.
D12 Quitting during pregnancy increased birth weight by an estimated 188 g. Tobacco control measures had
no effect on the birth weight of newborns of non-smoking women.
Keywords: © 2015 Elsevier B.V. All rights reserved.
Economic evaluation
Cigarette taxes
Package warnings
Advertising bans
Tobacco control

1. Introduction all advertising and promotion of tobacco products, mandated that

The tobacco epidemic continues to represent a serious public
health threat throughout the world. By one recent estimate, the
worldwide annual mortality burden has already reached 5 million
deaths from direct tobacco smoking and another 600,000 deaths
attributable to the effects of environmental smoke (World Health
Organization, 2012). Within the next 20 years, annual deaths from
tobacco are projected to continue to rise to 8 million, of which more
than 80% will occur in low- and middle-income countries (Mathers
et al., 2008).
Beginning in 2005, Uruguay instituted a series of aggressive anti-
smoking measures that placed this small South American country
of 3.3 million inhabitants in the forefront of tobacco control policy
worldwide. By 2012, the Uruguayan government had prohibited
smoking in enclosed public spaces and workspaces, banned nearly
∗ Corresponding author. Tel.: +1 617 253 2677; fax: +1 617 253 6915.
E-mail addresses: jeffrey@mit.edu (J.E. Harris), abalsa@um.edu.uy (A.I. Balsa),
patricia.triunfo@cienciassociales.edu.uy (P. Triunfo).
pictograms with warnings cover 80% of the front and back of
every pack, banned misleading marketing terms such as “light” and
“mild,” and outlawed multiple versions of the same brand such as
Silver or Blue. Tobacco taxes were increased, and all healthcare
providers were required to offer treatment for nicotine depend-
ence.
In a previous report, two of us (JH and PT) found that Uruguay’s
comprehensive nationwide antismoking campaign was associated
with a substantial, unprecedented decrease in tobacco use (Abascal
et al., 2012). During 2005–2011, per capita cigarette consumption
decreased by 4.3% per year, while the 30-day prevalence of cigarette
use among students aged 13–17 years and the overall population
prevalence of current tobacco use declined at annual rates of 8.0%
and 3.3%, respectively. The observed declines in each of these three
indicators of tobacco use were significantly larger than those seen
in neighboring Argentina, a culturally similar country that had not
conducted a comprehensive antismoking campaign and served as
a control.
While our previous study contributed to the evaluation of
the overall impact of Uruguay’s tobacco control campaign, it did

http://dx.doi.org/10.1016/j.jhealeco.2015.04.002
0167-6296/© 2015 Elsevier B.V. All rights reserved.
 J.E. Harris et al. / Journal of Health Economics 42 (2015) 186–196
not address the quantitative contributions of individual campaign
components. Pursuing that objective here, we classify the inter-
ventions implemented in Uruguay during 2007–2013 into three
categories: (1) provider-level interventions aimed at the treatment
of nicotine dependence, (2) national-level increases in cigarette
taxes, and (3) national-level non-price regulation of cigarette pack-
aging and marketing. We study the effects of these individual
campaign components on a critical target population – pregnant
women.
Studying the population of pregnant women is important not
only for the well-recognized adverse health consequences of
smoking during pregnancy (Permutt and Hebel, 1989; da Veiga
and Wilder, 2008; McCowan et al., 2009), but also for the nar-
row nine-month window during which pregnant women have
heightened susceptibility to health-related interventions. We take
advantage of a continuous nationwide registry of all live preg-
nancies from 2007 to 2013 to study the effects of the campaign
on two main outcomes: the probability that a pregnant smoker
will quit smoking by her third trimester and her infant’s birth
weight.
To identify the effect of the provider-level interventions, we
use a difference-in-differences (DID) approach, exploiting the fact
that these policies were implemented at different health centers
at different times. To assess the effect of taxes, we rely upon a
series of discrete tax increases during our study period. Finally, to
assess the effects of non-price regulation of packaging and market-
ing, we take advantage of the fact that these nationwide measures
went into effect at different times. As an additional control, we
compare the effect of these interventions on the birth weight of
children whose mothers smoked during pregnancy with the cor-
responding effect, if any, on the offspring of mothers who did not
smoke.
Our study contributes to an extensive literature evaluating
the impact of such tobacco control policies as tax increases,
control of environmental tobacco smoke, cigarette pack war-
nings, restrictions on cigarette marketing, regulation of tobacco
constituents, mass media anti-smoking campaigns, and the treat-
ment of addiction (Saffer and Chaloupka, 2000; Wakefield and
Chaloupka, 2000; Powell et al., 2005; Blecher, 2008; Carpenter and
Cook, 2008; DeCicca et al., 2008; Anger et al., 2011; Hammond,
2011; Hoek et al., 2011; Chaloupka et al., 2012; Emery et al.,
2012; Mons et al., 2013). Our work is distinguishable in that
we exploit an extensive micro database to evaluate the relative
impacts of multiple types of interventions in the context of a
nationwide tobacco control campaign conducted in a developing
country.
We find persuasive evidence on the impact of each of the three
policy categories analyzed – provider-level interventions, taxes,
and non-price policies – on the likelihood of quitting smoking
during pregnancy and on birth weight. In terms of the relative
contributions of each of these policies to the observed increase
in quit rates, the regulation of marketing and packaging had the
strongest effect, accounting for 71% of the total observed varia-
tion in quit rates during 2007–2013. While interventions to treat
nicotine dependence had a strong effect at the level of the indi-
vidual provider, relatively few prenatal care centers adopted these
interventions during the study period, thus contributing little to
the overall increase in the quit rate. Tax increases, on the other
hand, explained an estimated 25% of the variation in quit rates dur-
ing 2007–2013. While real taxes increased 122% during that time,
the tobacco-industry passed on only a fraction of the tax increases
to consumers, so that real cigarette price increased by only 17%.
Finally, we find that smoking cessation was associated with a sig-
nificant increase in birth weight. By contrast, the tobacco control
policies under study had no effect on the birth weight of offspring
of mothers who did not smoke.
187
2. Background and data
2.1. Nationwide anti-smoking policies
In 2005, one year after the legislature had ratified the Frame-
work Convention on Tobacco Control, Uruguay’s newly elected
administration launched a National Program for Tobacco Control
that formed the basis for a succession of progressively more strin-
gent tobacco control policies (Abascal et al., 2012). In March 2006,
all enclosed public spaces and all public and private workspaces
were declared 100% smoke-free. In June 2008, the scope of tobacco-
free spaces was extended to taxis, buses, airplanes and other public
transport.
These curbs on environmental tobacco smoke were paralleled
by a series of advertising restrictions on tobacco products. In May
2005 the government banned cigarette advertising on television
during children’s viewing hours (before 9:30 pm) and prohibited
advertising, promotion or sponsorship by tobacco companies of all
sporting events. These restrictions were subsequently codified in
March 2008, when comprehensive tobacco control legislation (Law
18.256) prohibited all advertising and promotion of tobacco prod-
ucts except at point of sale. In October 2008, logos, trademarks and
other tobacco-related symbols were banned on non-tobacco prod-
ucts. In May 2014, all advertising was prohibited, even at the point
of sale.
In addition, the Uruguayan government promulgated warning
requirements on cigarette packages and imposed restrictions on
manufacturers’ branding practices. A May 2005 ministerial decree
banned all references to “light,” “ultra light,” “mild,” “low tar” and
other descriptors that might misleadingly imply reduced harm. The
decree also mandated a series of rotating warnings with images
covering 50% of the front and back of each cigarette pack. The
deadline for compliance with the first round of these rotating war-
nings was April 2006. Subsequent rounds had respective deadlines
of December 2007, February 2009, February 2010, January 2012,
and April 2013. A “single presentation rule,” issued as a minis-
terial decree along with the third round of warnings, barred the
marketing of multiple versions of the same brand, such as Silver
or Blue. Finally, a 2009 decree mandated that the size of the war-
nings be increased to 80% of the front and back of each pack. This
requirement was implemented with the fourth round of warnings
and became effective by February 2010.1
Fig. 1 shows a timeline summarizing the major nationwide
non-price regulatory measures from 2005 to 2013. The blue text
describes each of the six rounds of package warnings, while the
boldface red text describes regulatory measures other than the
mandated warnings. The black lines point to the compliance dead-
lines for each regulatory measure.2
Fig. 2 further describes the six rounds of rotating package war-
nings. In each round, we show only one of several mandated images.
The relative sizes of the images in the figure correspond to their rel-
ative sizes on each pack, with the last three rounds reflecting the
required increase from 50% to 80% of the front and back surfaces.
2.2. Smoking cessation programs directed at healthcare providers
In 2008, the comprehensive tobacco control law mandated that
every primary care provider, whether public or private, incorporate
1
This “80% rule” was promulgated 3 months before the issuance of the fourth
round of images. However, we have no evidence of significant compliance with the
80% rule before the deadline for compliance with the fourth round of images.
2
With the exception of the comprehensive tobacco control law, all measures
provided for a 180-day compliance period. By specifying the end of the compliance
period as the effective date of each measure, we assumed that tobacco manufactur-
ers waited until each deadline to comply.
188 J.E. Harris et al. / Journal of Health Economics 42 (2015) 186–196

3rd round of
package warnings.
Comprehensive tobacco
Brands restricted to
control legislation.
a single presentation.
All advertising except
point-of-sale banned. 4th round of
package warnings.
Warnings must cover
2nd round of 80% of both front & back.
package warnings.

1st round of 5th round of

package warnings. package warnings.
Warnings must cover
50% of both front & back.
6th round of
Smoking prohibited package warnings.
in all enclosed public
spaces and all public
and private work spaces.

2005 2006 2007 2008 2009 2010 2011 2012 2013 2014

Fig. 1. Timeline of nationwide non-price tobacco control measures. The blue text refers to the deadlines for each of the six rounds of rotating package warnings, while the
boldface red text refers to other tobacco control measures.

the diagnosis and treatment of tobacco dependence into its menu
of basic services. Pursuant to this legislation, in 2009 the Ministry
of Public Health and the National Resource Fund (“Fondo Nacional
de Recursos” or FNR), the governmental agency responsible for
financing resource-intensive medical technologies, established
national guidelines for primary care providers on the diagnosis
and treatment of nicotine dependence. Through a set of agree-
ments (“convenios”) with the FNR, healthcare institutions were
eligible to receive training and free nicotine patches and bupro-
pion in return for setting up a smoking cessation program with
little or no patient copayments (Esteves et al., 2011). Health
centers that had no agreements with the FNR were required
to provide smoking cessation services in accordance with the
guidelines, but they were permitted to charge nontrivial copay-
ments to patients. In what follows, we refer to these agreements
between health centers and the FNR as “provider-level agree-
ments.”
Among all sites providing prenatal care to pregnant women, the
proportion with FNR agreements increased from 7 to 12% during
2005–2013. Concurrently, the proportion of all pregnant women
receiving prenatal care at sites with FNR agreements increased from
13% in 2005 to 36% in 2007, but then declined to 31% by 2013.

Fig. 2. Timeline of six rounds of rotating package warnings. Each round displays only one of several mandated images. The relative sizes of the images correspond to their
relative sizes on each pack, with the last three rounds reflecting the required increase from 50% to 80% of the front and back surfaces.
 J.E. Harris et al. / Journal of Health Economics 42 (2015) 186–196 189

80
Uruguayan Pesos per Pack of 20 Cigarettes

70
60
(Base = December 2010)

50
Real Price

40
Real Taxes
30
20 (Excise + VAT)

A B C D E F G H
10
0

2001 2003 2005 2007 2009 2011 2013

Year

Fig. 3. Real price and real taxes per pack of cigarettes, 2001–2013. The vertical lines show the timing of non-price nationwide policy measures, as described in Fig. 1.
(A) Smoking prohibited in all enclosed public spaces and all public and private workspaces. (B) 1st round of package warnings. (C) 2nd round of package warnings. (D)
Comprehensive tobacco control legislation. (E) 3rd round of package warnings; brands restricted to a single presentation. (F) 4th round of package warnings; warnings must
cover 80% of front and back. (G) 5th round of package warnings. (H) 6th round of package warnings.

2.3. Cigarette tax increases information at the level of the individual pregnancy on maternal
characteristics, self-reported smoking behavior, current and past
In addition to the foregoing policy interventions, the Uruguayan obstetric history, the timing of prenatal care, the sites of prenatal
government increased its indirect taxes on tobacco products. care and delivery, and birth outcomes including birth weight (CLAP,
Imposed solely at the national level, these taxes consist of an 2001). In 2012, the SIP covered an estimated 94% of all live births
excise tax (“impuesto específico interno” or IMESI) and a value in Uruguay.
added tax (“impuesto al valor agregado” or IVA). The IMESI, which Our analyses relied upon the following individual-level mater-
was first applied to cigarettes in 1993, underwent a series of dis- nal characteristics, derived from the SIP registry: the timing of
crete increases in June 2002, May 2003, July 2007, June 2009, and the first prenatal visit (first-trimester prenatal care); the mother’s
February 2010. The IVA, by contrast, was first applied to cigarettes age (<16, 17–19, 20–34, 35–39, and 40+ years), marital status
in July 2007 and since then has constituted 22% of the pre-tax (single, married, cohabiting, other), and educational attainment
price including the IMESI or, equivalently, 18% of the retail price. (primary, secondary, university); the number of prior deliveries
Fig. 3 shows the estimated real price and real tax on a pack of (0, 1, 2, 3, 4+); number of prior abortions; a history of diabetes
cigarettes during 2001–2013. Only 45% of the abrupt increase in or hypertension; whether any complications of pregnancy were
cigarette taxes in July 2007 was passed on to consumers in the observed, in particular, the presence of preeclampsia or eclamp-
form of higher retail prices. The construction of the real price sia; the mother’s body mass index based on her self-reported
and tax series is described in our working paper (Harris et al., height and weight prior to the pregnancy (underweight, normal
2014). weight, overweight, obese); the mother’s use of alcohol or illicit
In Uruguay, an estimated 99% of tobacco users smoke manu- drugs; the sites of prenatal care; and the newborn’s sex and birth
factured cigarettes, hand-rolled cigarettes, or both (Abascal et al., weight.3
2012). Manufactured cigarettes, in particular, make up more Prior to 2007, each individual record in the SIP database con-
than 85% of taxable cigarette consumption (Dirección General tained the pregnant woman’s smoking status only at the time of
Impositiva, 2012). During 2004–2012, by one estimate, contraband initiation of prenatal care. It did not show changes in smoking, if
cigarette sales constituted approximately 12% of total cigarette any, during the course of her pregnancy. Under a new data entry
consumption on average (Curti, 2013). With the possible excep- system beginning in 2007, the prenatal record noted the woman’s
tion of less densely populated provinces (“departamentos”) along smoking status separately in each trimester of her pregnancy. For
Uruguay’s borders with Brazil and Argentina, where contraband example, if a woman initiated prenatal care in her second trimester,
tobacco use appears more prevalent, there has been little effective the healthcare provider recorded her smoking status in the first
geographical variation in retail price. trimester, based on her recall, as well as in the current trimester.
Her smoking status would subsequently be recorded in a follow-up
2.4. Perinatal information system (SIP) prenatal visit during her third trimester. The perinatal data derived
from this new system, which we refer to as the “new SIP,” were the
Our source of micro data on the smoking practices of pregnant focus of our analysis.
women was the Perinatal Information System (“Sistema Infor-
mático Perinatal” or SIP), a mandatory nationwide electronic
registry operating in all prenatal care clinics in Uruguay since 1990. 3
To avoid loss of observations, we included dummy variables equal to 1 when
Developed and overseen by the Latin American Center for Peri- some maternal characteristics were missing. For further details on maternal and
natology (“Centro Latinoamericano de Perinatología” or CLAP) of pregnancy characteristics, including descriptive statistics, see our working paper
the Pan American Health Organization, the database contained (Harris et al., 2014).
190 J.E. Harris et al. / Journal of Health Economics 42 (2015) 186–196
3. Principal endpoints: smoking cessation and birth weight
To assess the impact of Uruguay’s tobacco control campaign, we
focused our attention primarily on pregnant women who smoked
cigarettes at any time from the first prenatal visit onward. Within
this target group of pregnant smokers, our principal endpoint
was smoking cessation by the third trimester. Our analysis of
this endpoint was confined to the interval from 2007 to 2013,
when data on smoking habits during each trimester were avail-
able through the new SIP system. Fig. 4 shows the progressive
increase in the annual mean quit rate among pregnant smokers
from 15.4% in 2007 to 42.7% in 2013. Our main research objec-
tive was to determine what proportion of this substantial rise in
quit rates could be attributed to Uruguay’s multi-component cam-
paign.
Why did not we focus on the prevalence of smoking at the onset
of pregnancy? Unfortunately, within the limitations of Uruguay’s
SIP system, this alternative endpoint was subject to a critical source
of potential measurement bias. Starting in July 2008, the Uruguayan
Ministry of Health established a new system of financial incentives
for providers to increase the completeness of data reporting on SIP
records. As Fig. 5 shows, the prevalence of smoking at the first
prenatal visit remained stable at about 25% from 2000 to 2005.
Thereafter, with the onset of the tobacco control campaign, the
prevalence had declined toward 15% by the early months of 2009.
In April 2009, however, nine months after the institution of the new
system of financial incentives, the prevalence abruptly increased.
At the same time, as shown in Fig. 5, the proportion of records with
missing data declined dramatically from about 20% in 2009 to 1–2%
by 2013.
The best explanation for the abrupt break in prevalence is
that women with missing data on smoking at the onset of
pregnancy were more likely to be smokers. Moreover, as the
Ministry of Health imposed increasingly strict goals for data com-
pleteness, the effect of including these previously unreported
smokers in the prevalence calculation is likely to have grown
ever larger. While we did have some data on the Ministry’s
data-completeness goals, we concluded that the task of correct-
ing for this missing data bias and at the same time identifying
the effects of post-2009 tobacco control measures would prove
intractable.
On the other hand, we did not detect any comparable break in
the monthly time series of smoking cessation rates. We thus con-
sidered the quitting to be a more reliable and sensitive endpoint
than smoking prevalence. Still, to the extent that the new system
of data-completeness goals recorded an increasing number of hard-
core smokers, our estimates in Fig. 5 of quit rate during 2009–2013
would be biased downward. We stress that our principal endpoint
represents the rate of cessation conditional upon smoking on or
after the first prenatal visit.
Smoking cessation during pregnancy is known to reduce its
adverse health effects. Accordingly, as an additional endpoint, we
studied the impact of Uruguay’s tobacco control campaign on
birth weight. As shown in Fig. 6, the difference in birth weight
between non-smokers and continuing smokers was on the order of
210 g, while the corresponding difference between non-smokers
and quitters was only about 25 g. Although the smaller number
of quitters in the SIP database during 2007–2008 decreased the
precision of the mean birth weight estimates, Fig. 6 still shows a
background increase in birth weight for all three groups. The anal-
ysis of birth weight also helped us address the concern that the SIP
registry contained self-reported data on smoking habits. If women
had falsely reported having quit smoking with increasing frequency
as the campaign progressed, we would have expected to see a grow-
ing reduction in the apparent favorable effects of cessation on birth
weight.
3.1. Impact on quitting during pregnancy
3.1.1. Individual-level analysis
We first investigate the effect of the provider-level agreements
and national-level tobacco control policies on the quit rate during
pregnancy. To that end, we begin with a linear probability model
based upon observations at the level of the individual pregnant
woman:
yist = ci ˛0 + xst

˛1 + zt ˛2 + Ds + eist (1)
where the subscript i indexes each woman, the subscript s refers
to the health center where she received prenatal care, and the sub-
script t refers to the calendar date corresponding to the midpoint
of her third trimester.4 The data yist are binary variables repre-
senting smoking cessation, where yist = 1 if the woman quit and
yist = 0 if she continued to smoke through her third trimester. The
vector of exogenous variables ci represents individual-level mater-
nal characteristics, while xst represents the presence or absence of
a provider-level agreement at health center s on calendar date t,
and the vector zt represents those national-level policy variables
including cigarettes taxes that were in effect at calendar date t. The
parameters Ds represent health center-specific fixed effects. We
assume that the unobserved error terms eist are uncorrelated with
the observed explanatory variables and have zero means.
Since the presence or absence of a provider-level agreement
xst varied by health center and calendar date, the impact of these
programs was identifiable via a DID model. With respect to the
national-level policies zt , however, we needed to be careful about
what policy impacts could and could not be identified from the
data. Successive increases in cigarette taxes during the 2007–2013
observation period, beginning with the inclusion of tobacco in the
value-added tax on July 1, 2007 (Fig. 3), permitted us to identify
the impact of this policy. On the other hand, we could not iden-
tify the impacts of the two non-price policies that went into effect
before the start of our observation period: the prohibition of smok-
ing in public places and enclosed workspaces (March 1, 2006) and
the requirement that all packs contain rotating images with war-
nings covering 50% of the front and back of each pack (April 18,
2006). Moreover, the effective dates of the second, third and fourth
rounds of warnings were either close to or coincident with other
policy measures, and thus their impacts could not be separately
identified without additional strong assumptions concerning their
effects over time (Fig. 1).
That left us with five non-price, national-level policies: the com-
prehensive tobacco legislation banning nearly all advertising (in
effect from March 6, 2008 onward); the single presentation rule (in
effect from February 14, 2009 onward); the increase in the warning
size from 50% to 80% of the front and back of each pack (in effect
from February 28, 2010 onward); the fifth round of warnings (in
effect from January 7, 2012 through April 7, 2013); and the sixth
round of warnings (in effect from April 8, 2013 onward).
Column (A) of Table 1 shows our results. We estimated the
parameters of Eq. (1) by ordinary least squares (OLS) with Huber-
White robust standard errors implemented at the individual level.
The presence of an agreement between the woman’s health center
and the FNR increased the probability of quitting by an esti-
mated 4.8 percentage points (p = 0.024). The coefficient of the log
tax per pack was 0.079 (p = 0.031). At the sample mean value
of the dependent variable equal to 0.377, the estimated elastic-
ity of the smoking cessation with respect to cigarette taxes was
4
For the individual-level model of Eq. (1), we assigned a woman to calendar date
t based on the midpoint of her prenatal care visits during her third trimester of
pregnancy. If she had no prenatal visits, we assigned her a date t equal to 30 days
prior to delivery.
 J.E. Harris et al. / Journal of Health Economics 42 (2015) 186–196 191

50
6623 6581

Proportion of Pregnant Smokers

Who Quit by Third Trimester (%)
6845

40
6583

5888

30
2582

20
638
10
0

2007 2008 2009 2010 2011 2012 2013

Year of Midpoint of Third Trimester

Fig. 4. Increase in mean quit rate among pregnant smokers, 2007–2013. Vertical bars represent 95% confidence intervals. Adjacent to each point is the number of pregnant
smokers in the SIP database with data on smoking status in the third trimester. Each smoker was assigned to the calendar year of the mean date of all her third-trimester
prenatal visits.

0.079/0.377 = 0.21. All five of the non-price nationwide policies then all indicators of non-price policies zt equal to 0. Based on this
had significant effects, with the comprehensive tobacco control decomposition procedure, we estimated the following attributable
law assuming the dominant role. All five non-price policies com- proportions: provider-level agreements, 4.3%; tax increases, 25.2%;
bined increased the probability of quitting by an estimated 26.9 and regulations of packaging and marketing, 70.5%.
percentage points (p < 0.001).
We used the results of our individual-level analysis to com- 3.2. Health center level analysis
pute the relative contributions of each of the three categories of
tobacco control policy to the overall change in smoking cessa- It is arguable that an individual-level analysis of Eq. (1) over-
tion rates observed during the 2007–2013 study period. To that states the precision of estimated policy impacts. The central thrust
end, we first computed the predicted values ŷist derived
 from Eq. of this criticism is that the agreements for smoking cessation ser-
(1) and then calculated the corresponding sum Ŷ = ŷ over
i,s,t ist vices were made with health centers rather than with individual
all observations. We then recomputed the predicted values ŷist patients. Moreover, the tax increases and non-price policies were
and corresponding sums Ŷ , successively setting all values of the carried out at the national rather than the individual level. To
provider-level agreement variable xst equal to 0, then all values of address these concerns, we performed a series of aggregate anal-
the log real tax rate equal to the initial level in January 2007, and yses at both the health center level and national level. Following
30
25
at First Prenatal Visit (%)
Prevalence of Smoking

20

% Prevalence
15

25
Proportion of Records
with Missing Data (%)
10

20
15
5

10
0

% Missing Data
0

2000 2002 2004 2006 2008 2010 2012 2014

Month of Birth

Fig. 5. Monthly prevalence of smoking at the first prenatal visit. Monthly proportion of records with missing data. Smoking prevalence is measured on the left axis, while
the proportion with missing data is measured on the right. The diameter of each data point is proportional to the number of observations.
192 J.E. Harris et al. / Journal of Health Economics 42 (2015) 186–196

3400
95% CI

Non-Smokers

3300
Birth Weight (gm)

3200 Smokers Who Quit

Continuing Smokers
3100
3000

2007 2008 2009 2010 2011 2012 2013

Year of Birth

Fig. 6. Annual mean birth weight among non-smoking pregnant women, continuing smokers, and smokers who quit, 2007–2013. Non-smokers were pregnant women who
did not report smoking at any prenatal visit. Continuing smokers were pregnant women who reported smoking at one or more prenatal visits, but had not quit smoking by
the third trimester. Smokers who quit likewise reported smoking at one or more prenatal visits, but had quit by the third trimester.

Table 1
Estimated effects on probability of quitting by the third trimester of pregnancy.

Independent variable (A) (B) (C) (D) (E)

Provider-level agreements 0.048** (0.021) 0.056* (0.034) 0.056** (0.025) 0.056** (0.025)
Log real tax per pack 0.079** (0.036) 0.106** (0.053) 0.105* (0.054) 0.136** (0.055)
Tobacco control law 0.094*** (0.017) 0.138*** (0.038) 0.138*** (0.026) 0.139*** (0.023)
Single presentation rule 0.075*** (0.014) 0.053** (0.023) 0.053*** (0.019) 0.039** (0.015)
80% rule 0.028*** (0.011) 0.029* (0. 015) 0.030* (0.015) 0.032* (0.016)
Fifth round of warnings 0.037*** (0.008) 0.030*** (0.011) 0.030*** (0.011) 0.036*** (0.007)
Sixth round of warnings 0.035*** (0.011) 0.0355** (0.0147) 0.035** (0.015) 0.038*** (0.010)
Five non-price measures combined 0.269*** (0.020) 0.286*** (0.044) 0.285*** (0.027) 0.283*** (0.022)
No. observations 31,230 1422 1400 1400 28
*
Significant at p < 0.10.
**
Significant at p < 0.05.
***
Significant at p < 0.01.
A. OLS estimation on individual maternal-level observations, based on Eq. (1). White-Huber robust standard errors. Coefficients of individual maternal characteristics and
health center fixed effects not shown.
B. OLS estimation on observations grouped by health center and calendar quarter, based on Eq. (3). Standard errors adjusted for clustering at level of health center. Observations
weighted by inverse of standard errors of fixed effects derived from Eq. (2).
C. FGLS estimation on observations grouped by health center and calendar quarter, based on Eq. (3). Standard errors adjusted for clustering at level of health center. Uniform
first-order serial correlation estimated to equal 0.0210. Observations weighted by inverse of standard errors of fixed effects derived from Eq. (2).
D. FGLS estimation on observations grouped by health center and calendar quarter, based on Eq. (3). Standard errors adjusted for clustering at level of health center. Uniform
first-order serial correlation estimated to equal 0.0237. Observations weighted by inverse of standard errors of fixed effects derived from Eq. (2).
E. OLS estimation on observations grouped by calendar quarter, based upon Eq. (4). Newey-West standard errors adjusted for heteroskedasticity and serial correlation.
Observations weighted by inverse of standard errors of fixed effects derived from Eq. (2).

(Amemiya, 1978; Hansen, 2007; Imbens and Woolridge, 2014) and
others, we specified the following two-stage model.
yist = ci ˇ0 + Fst + uist
 variables and with each other, and had zero means.
Fst = xst ˇ1 + zt ˇ2 + Gs + vst
Here, the subscript s = 1, . . ., S continues to index health centers,
while the subscript t = 1, . . ., T now indexes calendar quarters. The
subscript i = 1, . . ., Nst now indexes women who received prena-
tal care at health center s and whose third trimester occurred in
calendar quarter t.
In the first stage (Eq. (2)), the data yist are binary variables repre-
senting smoking cessation and the exogenous variables ci represent
individual-level maternal characteristics, while Fst are fixed effects
for each of the ST combinations of health center and calendar quar-
ter. In the second stage (Eq. (3)), the data xst represent the presence
or absence of a provider-level agreement at health center s during
calendar quarter t, while the data zt denote the national-level poli-
cies in effect in calendar quarter t. The parameters Gs are S health
center-specific fixed effects. We assumed that the unobserved error
(2) terms uist and vist were uncorrelated with the observed explanatory
(3)
To estimate the model parameters, we first ran OLS on Eq.
(2), thus obtaining estimates F̂st of the parameters Fst . In effect,
these OLS estimates represented the predicted quit rate in each
health center s and calendar quarter t of a pregnant smoker in
the reference category of maternal characteristics.5 We then esti-
mated the parameters of the DID model (3), where the dependent
5
A woman in the reference category was married, aged 20–34 years, had less
than a high school education, did not seek prenatal care in her first trimester, had
no prior abortions or deliveries, a pre-pregnancy body mass index 18.5–24.9 kg/m2 ,
no history of diabetes, hypertension, eclampsia or pre-eclampsia, and gave birth to
 J.E. Harris et al. / Journal of Health Economics 42 (2015) 186–196
variable Fst was replaced by its estimated value F̂st . We weighted the
observations by the inverse of the standard errors of the estimates
F̂st .
Within the context of this aggregate model, researchers have
expressed concerns that traditional OLS estimation of DID models
ignores serial correlation of errors and thus overstates the precision
of the coefficient estimates (Bertrand et al., 2004; Cameron et al.,
2008). In our context, serial correlation would arise when health
centers were subject to common unobserved shocks that per-
sisted for more than a calendar quarter. To address these concerns,
we estimated the parameters of Eq. (3) by feasible generalized
least squares (FGLS) under varying specifications of the covariance
matrix V = E[vst vs t  ].
Columns (B) and (C) in Table 1 show our regression estimates of
the parameters ˇ1 and ˇ2 of Eq. (3) under two different assump-
tions concerning the covariance matrix. In column (B), we assumed
clustering of errors within each health center. In column (C), we
assumed first-order temporally correlated errors with a uniform
correlation coefficient among health centers. We found quite simi-
lar results when we estimated Eq. (3) under the assumption that the
coefficient of serial correlation varied by health center (not shown
in Table 1). In comparison with the maternal-level estimates (col-
umn A), we observed larger coefficients for two policies: the log real
cigarette tax and the comprehensive tobacco control law. However,
their corresponding standard errors were also increased, so that we
could not reject the hypothesis that their impacts equalled those
estimated at the individual-level data. Finally, in both columns (B)
and (C), the combined effect of the five non-price policies was indis-
tinguishable from the effect estimated from individual-level data.
3.3. National level analysis
It is arguable that the 2-stage model of Eqs. (2) and (3) still
overstates the precision of the estimated impacts of cigarette taxes
and non-price regulations of cigarette packaging and marketing,
as these policies were carried out at the national level. To address
this criticism, we extended our two-stage model of Eqs. (2) and
(3) to three stages. In particular, we continued to specify the first-
stage model of Eq. (2), but replaced Eq. (3) with the following two
equations:

Fst = xst 1 + Hs + Jt + st
Jt = zt 2 + t
As before, we assumed that the unobserved error terms in (2),
(4) and (5) were uncorrelated with their respective explanatory
variables.
The second-stage Eq. (4) differs from (3) in that Fst now depends
on the presence or absence of a provider-level agreement xst at
health center s during calendar quarter t, as well as health center
fixed effects Hs , calendar quarter fixed effects Jt , and an unobserved
random error  st with mean zero. To estimate the policy impact
parameters  1 in (4), we replaced the fixed effects Fst with their
estimates F̂st from Eq. (2). As in the two-stage model, Eq. (4) was
estimated by feasible GLS under various assumptions concerning
the covariance matrix of the errors, where the observations were
weighted by the inverse of the standard errors of the estimates F̂st .
In third-stage Eq. (5), the calendar quarter fixed effects Jt in turn
depend on national level policies as well as an unobserved random
error t with mean zero. To estimate the policy impact parameters
 2 in (5), we replaced the fixed effects Jt with their estimates Ĵt
from the second stage Eq. (4). We then estimated the linear model
a singleton female. For further details on maternal and pregnancy characteristics,
including descriptive statistics, see our working paper (Harris et al., 2014).
193
(5) by OLS with Newey–West standard errors with a maximum lag
of 4 calendar quarters to take account of possible heteroskedas-
ticity and serial correlation of the error terms (Newey and West,
1987). Similarly, we weighted the observations by the inverse of
the standard errors of the estimates Ĵt .
Fig. 7 motivates the logic underlying the third stage of our anal-
ysis. The vertical axis measures the estimated fixed effects Ĵt , which
represent calendar quarter-specific quit rates adjusted for individ-
ual maternal characteristics, health center fixed effects, and the
presence or absence of a provider-level agreement. The horizon-
tal axis measures the corresponding calendar quarter t. Highlighted
are the dates on which the value added tax was imposed on tobacco,
comprehensive tobacco legislation was passed, only brands with a
single presentation were permitted, images with warnings were
mandated to cover 80% of the front and back of each pack, and the
fifth and sixth rounds of images went into effect.
Columns (D) and (E) show the results of our three-stage proce-
dure. Column (D) shows the estimate of  1 in Eq. (4) in the case
of FGLS with serially correlated errors. As in the previous models,
the presence of provider-level agreement increased the probability
of smoking cessation by 5.6 percentage points (p = 0.027). Column
(E) shows the estimates of the policy impact parameters  2 in Eq.
(5). The estimated policy impacts are comparable to those derived
from the 2-stage model (columns B and C). The point estimate of
the impact of cigarette taxes is larger and statistically significant
(p = 0.045), implying an estimated elasticity of quitting equal to
0.36. Still, we could not reject the hypothesis that the estimated tax
impact equalled that estimated from maternal-level data. Finally,
the estimated effect of the five non-price policies combined was
indistinguishable from the estimates based on the individual-level
and two-stage models.
4. Impact on birth weight
We relied upon maternal-level data to assess the effect of
quitting smoking during pregnancy and birth weight. Following
(Permutt and Hebel, 1989), we adopted a simultaneous equa-
tion framework in which Uruguay’s anti-smoking policies served
as instruments for the endogenous variable of smoking cessa-
tion. In addition, we took advantage of the birth-weight endpoint
(4) to strengthen our identification of the causal effect of Uruguay’s
tobacco control campaign. Employing the population of non-
(5)
smoking pregnant women as a control group, we performed a
falsification test to determine whether the same policy instru-
ments had any effect on the birth weight of infants delivered by
non-smokers.
We estimated the following equation on our sample of smoking
pregnant women:
wist = ı1 yist + ci ı0 + Ks + ς ist (6)
where the variable wist denotes the birth weight of the infant deliv-
ered by mother i in health center s and the calendar date t refers to
the midpoint of her third trimester of pregnancy. In Eq. (6), birth
weight depends on smoking cessation (yist ), individual maternal
characteristics (ci ), a health center fixed effect (Ks ), as well as a
random error term (ς ist ) with zero mean.
If the random error ς ist is correlated with yist , then the coeffi-
cient ı1 will be biased when Eq. (6) is estimated by OLS. There is,
in fact, good reason to believe that this is the case. For example, a
woman with a propensity to engage in risky behaviors will tend not
to quit smoking and deliver a low-weight baby. In that case, failure
to account for the unobserved heterogeneity will overestimate the
parameter ı1 . Alternatively, a woman who runs into complications
during her pregnancy will be under pressure to quit smoking and
194
.5
.4
Quit Rate Fixed Effects
.3
.2
.1
0
2007
Fig. 7. Trend in quit rate estimated from national-level model. The vertical axis measures the fixed effects Ĵt estimated from Eq. (4), which represent calendar quarter-specific
quit rates adjusted for individual maternal characteristics, health center fixed effects, and the presence or absence of a provider agreement. The horizontal axis measures the
corresponding calendar quarter t. Highlighted are the dates on which various national-level tobacco control measures went into effect.
tend to deliver a low-weight baby. In that case, the parameter ı1
will be underestimated.
To address the potential endogeneity of the quit variable yist ,
we therefore estimated Eq. (6) by two-stage least squares (2SLS),
where the first stage is defined by Eq. (1) and the instruments for
smoking cessation are the policy variables xst and zt . As shown in
Table 2, estimation of Eq. (6) by OLS, with clustering of standard
errors by health center, gave an estimate of ı̂1 = 123.2 g (p < 0.001).
By contrast, estimation of (6) by 2SLS, similarly with standard
errors clustered by health center, gave an estimate of ı̂1 = 187.9 g
(p = 0.028). While 2SLS increased the estimated standard error of
ı̂1 , all tests rejected the hypotheses of weak instruments or over-
identification. Although the 95% confidence intervals of the two
estimates of ı̂1 overlapped substantially, the results still suggest
that the OLS estimate may be biased downward.
For our falsification test, we used the parameters estimated
from the first-stage Eq. (1) to predict the values of yist for those
women who did not report smoking during pregnancy. We then
estimated the model of Eq. (6) on all women in this comparison
group, replacing yist with its predicted value. If the tobacco con-
trol policies captured in the variables xst and zt in fact improved
birth weight by increasing the rate of quitting, then the estimate
of ı̂1 in (6) should be indistinguishable from zero in the compari-
son group of non-smoking pregnant women. On the other hand, if
these policies are simply correlated with other unobserved factors
that improved birth weight, then the estimate of ı̂1 in (6) should be
significantly positive for non-smokers as well. In fact, as shown in
Table 2, the estimate of ı̂1 was indistinguishable from zero among
non-smokers (−27.8 g, p = 0.527).
Table 2
Estimated effect of smoking cessation on birth weight.a,b
Population OLS
Smokers 123.2*** (10.8)
Non-smokers
No. observations 31,186
**
Significant at p < 0.05.
***
Significant at p < 0.01.
a
All estimates correspond to the parameter ı1 in Eq. (6).
b
All standard errors adjusted for clustering by health center.
J.E. Harris et al. / Journal of Health Economics 42 (2015) 186–196
6th Round
Warnings
5th Round
80% Rule Warnings
Single
Presentation
Rule
Advertising Ban
Anti-Smoking Law
Value Added Tax
Imposed on Cigarettes
2008 2009 2010 2011 2012 2013 2014
Calendar Quarter
5. Discussion and conclusions
To assess the impact of Uruguay’s nationwide tobacco control
campaign, we analyzed a comprehensive nationwide registry of
pregnancies ending in a live birth during 2007–2013. We focused
sharply on smoking cessation among those women who reported
smoking at any time during pregnancy, as well as the conse-
quences of smoking cessation for birth weight. We observed a
striking increase in the proportion of pregnant smokers who had
quit by their third trimester, from 15.4% in 2007 to 42.7% in
2013.
We employed a difference-in-differences approach to evalu-
ate the effects of provider-level agreements on quit rates during
pregnancy. We found that smoking cessation programs estab-
lished under agreements between health centers and the National
Resource Fund increased quit rates by between 4.8 and 5.6 per-
centage points. Unfortunately, no more than one-third of pregnant
smokers received care at health centers with contractual agree-
ments during the period of analysis, and by 2013, the proportion of
women exposed to such treatments had declined. As a result, these
provider-level agreements contributed little to the overall increase
in smoking cessation observed during 2007–2013.
Although real taxes per pack increased by 122% during our study
period (Fig. 3), tax increases alone explained only about 20% of the
overall rise in smoking cessation during pregnancy. In addition,
our estimates of the tax elasticity of quitting, ranging from 0.21 to
0.36, were lower than those reported in the U.S. (Ringel and Evans,
2001; Colman et al., 2003). The principal explanation for this limited
influence is that manufacturers moderated their retail prices in
response to the application of the value added tax to cigarettes
in July 2007 and other non-price regulatory policies enacted dur-
ing 2007–2009 (Fig. 3) (Harris et al., 2014). As a result, the real
retail price of cigarettes increased by only 17% during 2007–2013.
2SLS Falsification test
Endogenous responses of the tobacco industry to state and nation-
187.9** (85.5)
wide tax increases have been previously documented in the U.S.
−27.8 (43.8)
31,186 126,504
(Harris, 1987; Harris et al., 1996; Chaloupka et al., 2010; Miura,
2010).
To the contrary, most of the observed increase in quit rates was
attributable to non-price regulation of the marketing and packaging
of cigarettes. While the combined effect of all five non-price policies
 J.E. Harris et al. / Journal of Health Economics 42 (2015) 186–196
was indistinguishable across the models that we tested (Table 1),
it was nonetheless difficult to identify with precision the quanti-
tative impacts of the individual component policies. We could not
distinguish immediate versus long-run effects of policies, nor could
we identify synergies between policies. Our analysis at best identi-
fied the average impacts of the tobacco control measures over time,
conditional upon a specific temporal sequence of policy interven-
tions. Thus, the increase in the size of warnings from 50 to 80% of
the front and back of each pack of cigarettes, in effect since February
28, 2010, was associated with a 3 percentage point increase in the
cessation rate (Table 1). However, this estimated impact was in the
context of the coincident fourth round of warnings, as well as a
ban on smoking in public places and private workspaces (March 1,
2006), the comprehensive tobacco control law (March 6, 2008), the
single presentation rule (February 14, 2009) and other policies in
effect at the same time (Figs. 1 and 2).
While we had micro data on smoking cessation at the individual
level, the tobacco control policies that we evaluated were carried
out at either the health center level or the national level. Accord-
ingly, if there were significant inter-correlation of quitting behavior
among women who attended the same health center or who were
pregnant in the same calendar quarter, then the effective number of
observations could be far less than the approximate 31,000 shown
in column A of Table 1. However, our estimates on data aggregated
at the health center level (columns B and C) and at the national level
(columns D and E) generally confirmed our individual-level results.
Some economists have suggested that the impact of smoking
cessation during pregnancy on birth weight, as estimated from
cross-sectional databases, may be exaggerated by the presence of
unobserved heterogeneity (Lien and Evans, 2005; Abrevaya, 2006;
Abrevaya and Dahl, 2008; Walker et al., 2009; Juarez and Merlo,
2013). That is, a woman who tends to engage in risky behaviors
will continue to smoke during pregnancy and have lower weight
babies. However, the presence of unobserved heterogeneity can
also result in an underestimate of the impact of smoking cessa-
tion. Thus, a woman who encounters complications in the third
trimester, such as intrauterine growth retardation, will quit smok-
ing and have a lower weight baby. Our OLS estimate of the effect
of quitting on birth weight was 123 g (95% CI, 102–145). When we
used Uruguay’s tobacco control policies as instruments for quit-
ting smoking, our 2SLS estimate was 188 g (95% CI, 20–356). While
our 2SLS estimate reinforces the conclusions that smoking cessa-
tion during pregnancy, even in the third trimester, has a significant
positive effect on birth weight, the confidence interval surrounding
that estimate was too wide to draw definitive conclusions about the
direction of bias, if any, in the OLS estimate. Our results confirm that
even delayed cessation of smoking can reduce the adverse effects
of smoking during pregnancy (Lieberman et al., 1994; Raatikainen
et al., 2007; Batech et al., 2013; Yan and Groothuis, 2013).
As we have already noted, Uruguay’s tobacco control campaign
was associated with declines in per capita cigarette consumption,
adult smoking prevalence and adolescent cigarette use that were
significantly greater than those observed in Argentina, a country
with a common international border, language and culture that
served as a control (Abascal et al., 2012). Unfortunately, we were
unable to locate reliable nationwide data from Argentina on the
smoking practices of pregnant women during 2007–2013, and thus
could not construct a comparable external control group for this
study. Still, our falsification test took advantage of an internal
control group, namely, pregnant Uruguayan women who did not
smoke during pregnancy. We found, in particular, that the tobacco
control policies implemented in Uruguay during 2007–2013 had no
effect on the birth weight of mothers in this non-smoking control
group.
Our data on smoking from the SIP registry were self-reported. It
is conceivable that Uruguay’s tobacco control campaign increased
195
the social pressure on a pregnant woman to deny smoking when
queried by her obstetrician. While we cannot completely rule out
misreporting bias, we note that if women had falsely reported
having quit smoking with increasing frequency as the campaign
progressed, we would have expected to see a growing reduction
in the apparent favorable effect of quitting on birth weight. That is
not what we observed in Fig. 6. Moreover, if the apparent increase
in smoking cessation during 2007–2013 were solely the result of
increasing misreporting, then we would have observed no relation
between quitting and birth weight. That is not what we observed
in Table 2. In the context of smoking during pregnancy, a number
of authors have found a strong correlation between self-reported
cigarette consumption and objectively measured levels of nicotine
metabolites (Castellanos et al., 2000; Althabe et al., 2008; Himes
et al., 2013).
During 2005–2009, the prevalence of smoking during pregnancy
declined from approximately 25% to 15% (Fig. 5). Unfortunately,
missing data biases seriously complicated the interpretation of
prevalence trends thereafter. Moreover, there is evidence that some
women who quit during pregnancy later resumed smoking and
then quit once again during the next pregnancy (Harris et al.,
2014). Still, if the trend observed during 2005–2009 is an accurate
indicator of an overall decline in prevalence, then our results on
smoking cessation may substantially understate the overall impact
of Uruguay’s tobacco campaign.
Our results have important implications for future research
and for the future design of tobacco control policies. Our find-
ings suggest that enhanced targeting of healthcare providers in the
implementation of tobacco cessation programs, as well as increased
recruitment of patients, could have a high payoff. At the same time,
our findings strongly suggest that non-price policies, in particu-
lar regulation of marketing and packaging, can have an important
impact in reducing tobacco use.
Financial support
We gratefully acknowledge the financial support of the
Bloomberg Foundation through an unrestricted grant to the Min-
istry of Public Health (Ministerio de Salud Pública), Uruguay.
Neither the Bloomberg Foundation nor the Ministry of Public Health
exerted any influence on the conduct of this study or the drafting
of this manuscript.
Conflicts of interest
We have no conflicts of interest to declare.
Authors’ contributions
All three coauthors contributed to the conceptualization and
design of this study, the analysis of the data, and the writing of
this report.
Acknowledgments
We thank the Area Sistema Informático Perinatal from the
Epidemiology Division of the Ministry of Public Health (UINS)
for providing us with the perinatal data. We acknowledge
valuable inputs from Winston Abascal, Rafael Aguirre, Wanda
Cabella, Fernando Esponda, Elba Estévez, Marinés Figueroa, Ana
Lorenzo, Luis Mainero, Anna Mikusheva, and Giselle Tomasso.
The opinions expressed in this paper are ours and ours
alone.
196 J.E. Harris et al. / Journal of Health Economics 42 (2015) 186–196
References
Abascal, W., Esteves, E., Goja, B., Gonzalez Mora, F., Lorenzo, A., Sica, A., Triunfo,
P., Harris, J.E., 2012. Tobacco control campaign in Uruguay: a population-based
trend analysis. Lancet 380 (9853), 1575–1582.
Abrevaya, J., 2006. Estimating the effect of smoking on birth outcomes using a
matched panel data approach. Journal of Applied Econometrics 21 (4), 489–519.
Abrevaya, J., Dahl, C.M., 2008. The effects of birth inputs on birthweight: evidence
from quantile estimation on panel data. Journal of Business and Economic Statis-
tics 26 (4), 379–397.
Althabe, F., Colomar, M., Gibbons, L., Belzan, J.M., Buekens, P., 2008. Tabaquismo
durante el embarazo en Argentina y Uruguay (Smoking during pregnancy in
Argentina and Uruguay]). Medicina (Buenos Aires) 68, 48–54.
Amemiya, T., 1978. A note on a random coefficient model. International Economic
Review 19 (3), 793–796.
Anger, S., Kvasnicka, M., Siedler, T., 2011. One last puff? Public smoking bans and
smoking behavior. Journal of Health Economics 30 (3), 591–601.
Batech, M., Tonstad, S., Job, J.S., Chinnock, R., Oshiro, B., Allen Merritt, T., Page, G.,
Singh, P.N., 2013. Estimating the impact of smoking cessation during pregnancy:
the San Bernardino County experience. Journal of Community Health 38 (5),
838–846.
Bertrand, M., Duflo, E., Mullainathan, S., 2004. How much should we trust
differences-in-differences estimates? Quarterly Journal of Economics 119 (1),
249–275.
Blecher, E., 2008. The impact of tobacco advertising bans on consumption in devel-
oping countries. Journal of Health Economics 27 (4), 930–942.
Cameron, A.C., Gelbach, J.B., Miller, D.R., 2008. Bootstrap-based improvements for
inference with clustered errors. Review of Economics and Statistics 90 (3),
414–427.
Carpenter, C., Cook, P.J., 2008. Cigarette taxes and youth smoking: new evidence
from national, state, and local Youth Risk Behavior Surveys. Journal of Health
Economics 27 (2), 287–299.
Castellanos, M.E., Munoz, M.I., Nebot, M., Paya, A., Rovira, M.T., Planasa, S., Sanroma,
M., Carreras, R., 2000. Validez del consumo declarado de tabaco en el embarazo
(Validity of the declared tobacco consumption in pregnancy]). Aten Primaria 26
(9), 629–632.
Chaloupka, F.J., Peck, R., Tauras, J.A., Xu, X., Yurekli, A., 2010. Cigarette Excise Tax-
ation: the Impact of Tax Structure on Prices, Revenues, and Cigarette Smoking.
Cambridge, Massachusetts, National Bureau of Economic Research, Working
Paper 16287, August.
Chaloupka, F.J., Yurekli, A., Fong, G.T., 2012. Tobacco taxes as a tobacco control
strategy. Tobacco Control 21 (2), 172–180.
CLAP, 2001. Sistema Informático Perinatal en el Uruguay 15 Años de Datos
1985–1999. Centro Latinoamericano de Perinatologia y Desarrollo Humano,
Publicación Científica del CLA, Montevideo, Uruguay, pp. P1485.
Colman, G., Grossman, M., Joyce, T., 2003. The effect of cigarette excise taxes on
smoking before, during and after pregnancy. Journal of Health Economics 22
(6), 1053–1072.
Curti, D., 2013. El comercio iliı́cito en Uruguay y su relacioı́n con los impuestos:
resultados de investigacioı́n (Illicit trade in Uruguay and its relation to taxes:
research results]). Centro de Investigación para la Epidemia de Tabaquismo
(CIET), May 9, Montevideo.
da Veiga, P.V., Wilder, R.P., 2008. Maternal smoking during pregnancy and birth-
weight: a propensity score matching approach. Maternal Child Health Journal
12 (2), 194–203.
DeCicca, P., Kenkel, D., Mathios, A., 2008. Cigarette taxes and the transition from
youth to adult smoking: smoking initiation, cessation, and participation. Journal
of Health Economics 27 (4), 904–917.
Dirección General Impositiva, 2012. Volúmenes físicos de bienes gravados por el
IMESI – Series anuales (archivo xls). Montevideo Dirección General Impositiva,
República Oriental del Uruguay.
Emery, S., Kim, Y., Choi, Y.K., Szczypka, G., Wakefield, M., Chaloupka, F.J., 2012. The
effects of smoking-related television advertising on smoking and intentions to
quit among adults in the United States: 1999–2007. American Journal of Public
Health 102 (4), 751–757.
Esteves, E., Gambogi, R., Saona, G., Cenández, A., Palacio, T., 2011. Tratamiento de la
dependencia al tabaco: experiencia del Fondo Nacional de Recursos (Treatment
of tobacco dependence: experience of the National Resource Fund]). Revista
Uruguaya de Cardiología 26 (3), 78–83.
Hammond, D., 2011. Health warning messages on tobacco products: a review.
Tobacco Control 20 (5), 327–337.
Hansen, C.B., 2007. Generalized least squares inference in panel and multilevel
models with serial correlation and fixed effects. Journal of Econometrics 140,
670–694.
Harris, J.E., 1987. The 1983 increase in the federal excise tax on cigarettes. In: Sum-
mers, L.H. (Ed.), Tax Policy and the Economy, vol. 1. MIT Press, Cambridge, MA,
pp. 87–111.
Harris, J.E., Balsa, A.I., Triunfo, P., January 2014. Tobacco Control Campaign in
Uruguay: Impact on Smoking Cessation During Pregnancy and Birth Weight.
National Bureau of Economic Research, Working Paper No. 19878, Cambridge,
MA.
Harris, J.E., Connolly, G.N., Brooks, D., Davis, B., 1996. Cigarette smoking before
and after an excise tax increase and an antismoking campaign – Mas-
sachusetts, 1990–1996. MMWR Morbidity and Mortality Weekly Report 45 (44),
966–970.
Himes, S.K., Stroud, L.R., Scheidweiler, K.B., Niaura, R.S., Huestis, M.A., 2013. Prenatal
tobacco exposure, biomarkers for tobacco in meconium, and neonatal growth
outcomes. Journal of Pediatrics 162 (5), 970–975.
Hoek, J., Wong, C., Gendall, P., Louviere, J., Cong, K., 2011. Effects of dis-
suasive packaging on young adult smokers. Tobacco Control 20 (3), 183–
188.
Imbens, G., Woolridge, J.M., 2014. New Developments in Econometrics. Economet-
rics of Cross Section and Panel Data. Lecture 7. Cluster Sampling. Centre for
Microdata Methods and Practice (CEMMAP), London.
Juarez, S.P., Merlo, J., 2013. Revisiting the effect of maternal smoking during
pregnancy on offspring birthweight: a quasi-experimental sibling analysis in
Sweden. PLOS ONE 8 (4), e61734.
Lieberman, E., Gremy, I., Lang, J.M., Cohen, A.P., 1994. Low birthweight at term and
the timing of fetal exposure to maternal smoking. American Journal of Public
Health 84 (7), 1127–1131.
Lien, D.S., Evans, W.N., 2005. Estimating the impact of large cigarette tax hikes: the
case of maternal smoking and infant birth weight. Journal of Human Resources
40 (2), 373–392.
Mathers, C.D., Boerma, T., Ma Fat, D., 2008. The Global Burden of Disease: 2004
Update. World Health Organization, Geneva.
McCowan, L.M., Dekker, G.A., Chan, E., Stewart, A., Chappell, L.C., Hunter, M., Moss-
Morris, R., North, R.A., 2009. Spontaneous preterm birth and small for gestational
age infants in women who stop smoking early in pregnancy: prospective cohort
study. British Medical Journal 338, b1081.
Miura, M., 2010. Regulating Tobacco Product Pricing: Guidelines for State and Local
Governments. Tobacco Control Legal Consortium, Saint Paul, MN.
Mons, U., Nagelhout, G.E., Allwright, S., Guignard, R., van den Putte, B., Willemsen,
M.C., Fong, G.T., Brenner, H., Potschke-Langer, M., Breitling, L.P., 2013. Impact
of national smoke-free legislation on home smoking bans: findings from the
International Tobacco Control Policy Evaluation Project Europe Surveys. Tobacco
Control 22 (e1), e2–e9.
Newey, W.K., West, K.D., 1987. A simple, positive semi-definite, heteroskedas-
ticity and autocorrelation consistent covariance matrix. Econometrica 55 (3),
703–708.
Permutt, T., Hebel, J.R., 1989. Simultaneous-equation estimation in a clinical trial of
the effect of smoking on birth weight. Biometrics 45 (2), 619–622.
Powell, L.M., Tauras, J.A., Ross, H., 2005. The importance of peer effects, cigarette
prices and tobacco control policies for youth smoking behavior. Journal of Health
Economics 24 (5), 950–968.
Raatikainen, K., Huurinainen, P., Heinonen, S., 2007. Smoking in early gestation
or through pregnancy: a decision crucial to pregnancy outcome. Preventive
Medicine 44 (1), 59–63.
Ringel, J.S., Evans, W.N., 2001. Cigarette taxes and smoking during pregnancy. Amer-
ican Journal of Public Health 91 (11), 1851–1856.
Saffer, H., Chaloupka, F., 2000. The effect of tobacco advertising bans on tobacco
consumption. Journal of Health Economics 19 (6), 1117–1137.
Wakefield, M., Chaloupka, F., 2000. Effectiveness of comprehensive tobacco control
programmes in reducing teenage smoking in the USA. Tobacco Control 9 (2),
177–186.
Walker, M.B., Tekin, E., Wallace, S., 2009. Teen Smoking and Birth Outcomes. South-
ern Economic Journal 75 (3), 892–907.
World Health Organization, 2012. WHO Global Report: Mortality Attributable to
Tobacco. World Health Organization, Geneva.
Yan, J., Groothuis, P.A., August 2013. Timing of Prenatal Smoking Cessation or Reduc-
tion and Infant Birth Weight: Evidence from the United Kingdom Millennium
Cohort Study. Appalachian State University, Department of Economics Working
Paper, Number 13–16, Boone, NC.
 Journal of Health Economics 42 (2015) 197–208

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Financing and funding health care: Optimal policy and political

implementability夽
Robert Nuscheler 1 , Kerstin Roeder ∗
University of Augsburg, Department of Economics, Universitätsstraße 16, 86159 Augsburg, Germany

a r t i c l e i n f o a b s t r a c t

Article history: Health care financing and funding are usually analyzed in isolation. This paper combines the corre-
Received 2 July 2014 sponding strands of the literature and thereby advances our understanding of the important interaction
Received in revised form 10 March 2015 between them. We investigate the impact of three modes of health care financing, namely, optimal income
Accepted 20 April 2015
taxation, proportional income taxation, and insurance premiums, on optimal provider payment and on
Available online 15 May 2015
the political implementability of optimal policies under majority voting. Considering a standard multi-
task agency framework we show that optimal health care policies will generally differ across financing
JEL classification:
regimes when the health authority has redistributive concerns. We show that health care financing also
H24
I14
has a bearing on the political implementability of optimal health care policies. Our results demonstrate
I18 that an isolated analysis of (optimal) provider payment rests on very strong assumptions regarding both
the financing of health care and the redistributive preferences of the health authority.
Keywords: © 2015 Elsevier B.V. All rights reserved.
Health care financing
Provider payment
Service quality
Cost containment
Political economy

1. Introduction Positive financing frameworks investigate how the politico econ-

Health care funding traditionally receives great attention by
health economists. The primary topic of interest is optimal provider
payment and how the environment, e.g., competition and infor-
mation, shapes the optimal reimbursement system (see Chalkley
and Malcomson, 2000, for an overview). The question of how the
required revenue to reimburse providers should be raised, that is,
how health care financing should be organized, has received less
attention. The normative literature typically asks whether a social
health care system is suited to improve social welfare and if so
what size of the system is optimal (see, e.g., Blomqvist and Horn,
1984; Cremer and Pestieau, 1996, and Breyer and Haufler, 2000).
夽 The paper gained from discussions with participants of the following work-
shops, seminars and conferences: EHEW in Sevilla, ECHE in Zurich, Conference of the
German Economic Association in Göttingen and seminars at the HCHE and the Uni-
versity of Siegen. We are particularly grateful to Mathias Kifmann, Friedrich Breyer,
and Hartmut Kliemt for their comments and suggestions.
∗ Corresponding author. Tel.: +49 821 598 4477; fax: +49 821 598 4175.
E-mail addresses: robert.nuscheler@wiwi.uni-augsburg.de (R. Nuscheler),
kerstin.roeder@wiwi.uni-augsburg.de (K. Roeder).
1
Tel.: +49 821 598 4202; fax: +49 821 598 4232.
omy equilibrium depends on voter heterogeneity (see, e.g., Epple
and Romano, 1996a,b, and Gouveia, 1997). Surprisingly, health care
financing and funding are mostly analyzed in isolation: research
on provider payment ignores how health care is being financed
and the financing literature neglects how the funds are being used.
By simultaneously analyzing health care financing and funding our
study fills this gap and thereby advances the understanding of the
important health care financing and funding interaction. Addition-
ally, the current article analyzes both, the optimal and the political
allocation. This allows us to identify inefficiencies that are rooted in
the political decision making process and to assess whether optimal
policies are politically feasible.
We investigate the impact of three modes of health care
financing, namely, optimal income taxation, proportional income
taxation, and insurance premiums, on optimal provider payment
and on the political implementability of optimal policies under
majority voting. We consider a standard multi-task agency frame-
work where the health care provider chooses the quality of care
and cost reducing effort. More quality in health care is beneficial to
patients and to the provider who is considered (partially) altruis-
tic. More quality increases treatment costs and cost reduction effort
lowers them. Individuals differ along two dimensions, i.e., risk and

http://dx.doi.org/10.1016/j.jhealeco.2015.04.003
0167-6296/© 2015 Elsevier B.V. All rights reserved.
198 R. Nuscheler, K. Roeder / Journal of Health Economics 42 (2015) 197–208
income. Given this heterogeneity an allocation is assessed along
three dimensions: quality, effort, and the distribution of income (or,
equivalently, the numéraire commodity). If optimal income taxa-
tion is feasible — or in the absence of redistributive concerns — and
if quality and effort are contractible, the first-best allocation can
be implemented. When quality and effort are non-contractible, the
health authority uses a linear cost-sharing arrangement to steer
the provider’s incentives to invest in quality and to exert effort.
As the health authority has two margins but only one instrument
the first-best allocation can no longer be implemented. The health
authority then uses the cost-sharing parameter to optimally trade
off the inefficiencies in quality and effort. If health care financing
is through optimal income taxes, this tradeoff is not blurred by
any redistributive consequences which the financing of health care
provision might have.
The second-best allocation is then contrasted with allocations
under alternative financing regimes, namely, proportional income
taxes and insurance premiums. We call the resulting allocations
third-best. With proportional income taxation, income is redis-
tributed from high-income agents to low-income ones and from
low-risk agents to high-risk ones. Depending on the distributional
characteristics of risk and income the third-best policy may imply
more cost-sharing than the second-best policy and with it higher
quality and less effort, causing health care expenses to be higher.
When health care financing is through insurance premiums the
second-best quality-effort tradeoff is affected if and only if insur-
ance premiums involve some pooling. Then, premiums redistribute
income from low-risk agents to high-risk ones with the extent
being governed by the degree of pooling. The comparison between
the second-best and third-best allocations hinges on the distribu-
tional characteristic of risk and on the extent of pooling.
To complete the picture, we derive the allocations under major-
ity voting and contrast them with the optimal policies for both
proportional income taxes and insurance premiums. While the
redistributive preferences of the health authority are governed by
the distributional characteristics, the preferences of the median
voter depend on individual heterogeneity. This implies that, only
in knife-edge cases, can the optimal (third-best) policies be imple-
mented as political equilibria. For the case of proportional income
taxes the comparison of the two allocations depends on how the
relative inequity between risk and income compares to the relative
distributional characteristics between these two dimensions. For
insurance premiums it is only the inequity in risk together with the
extent of pooling and its relation to the distributional characteristic
of risk that matters.
Finally, it should be noted that, rather remarkably, risk-rated
premiums imply second-best optimal health care provision for both
the third-best allocation and the political outcome. The reason
being that risk-rated premiums preclude any form of redistribution.
There is then no conflict in the electorate about how to shape the
health care system and second-best health care provision results.
From the normative end it does not pay off to distort the opti-
mal policy away from the second-best as the associated efficiency
losses are not compensated by redistributive gains. As the resulting
income distribution may not be optimal the equilibrium alloca-
tion may not be second-best efficient. Our results demonstrate that
studies on optimal provider payment that neglect health care finan-
cing rest on very strong assumptions regarding the redistributive
motives of the health authority, or on health care financing.
This article relates to two strands of the health economics
literature. First, provider payment. The papers of Chalkley and
Malcomson (1998a,b), Ma (1994), and, more recently, Eggleston
(2005), and Kaarbøe and Siciliani (2011), show that mixed payment
systems, i.e., a combination of capitation payments and cost-
sharing, will generally be optimal. Whether quality incentives
are high powered or low powered depends on the respective
environment. It may relate to the quality elasticity of demand as
in the first three papers, or to the complementarity between the
different dimensions under consideration (the latter two papers).
We use a simplified version of their models enabling us to integrate
health care financing. There are only two articles we are aware
of that consider a median voter approach to provider payment,
namely, Gravelle (1999) and Nuscheler (2003). These papers look at
how optimal capitation payments for physicians relate to the ones
that would be implemented by majority voting. Both papers do not
consider a multi-task agency framework and remain silent about
health care financing.
Second, the health care financing literature. The normative liter-
ature typically takes an optimal income taxation approach and asks
whether there is a case for redistributive social health care finan-
cing in the presence of progressive income taxation. Blomqvist and
Horn (1984) and Cremer and Pestieau (1996), for instance, show
that the desirability of social health insurance in parallel to an opti-
mal income taxation scheme crucially depends on the correlation
between income and health risk. For the empirically relevant case
of a negative correlation, a redistributive public health care sys-
tem can improve on a purely private health care market.2 Breyer
and Haufler (2000) advocate for a strict separation of income redis-
tribution and health care financing as this would allow for better
health insurance contracts (in terms of ex post moral hazard) and
more efficient public financing in general (lower shadow costs of
public funds). Political feasibility of optimal policies and provider
reimbursement are ignored. The positive literature on health care
financing aims at explaining the existence of public health care,
its size and its form of financing. Epple and Romano (1996a) and
Gouveia (1997) were the first to address these issues.3 The former
paper considers agent heterogeneity in income and shows that
there is an ‘ends against the middle’ equilibrium when public health
care can be topped up by actuarially fair private health insurance.
Gouveia (1997) shows that this result continues to hold when het-
erogeneity in risk is added to the framework. Both papers derive
conditions under which a mixed health care system with pub-
lic and private health care financing arises. The mode of public
financing, however, is taken as given. We explicitly analyze the con-
sequences of alternate financing regimes on economic allocations.
Rather than taking a reduced form approach where a health good
is uniformly distributed to those who need it, we add a multi-task
provider payment setting to the model. Finally, Epple and Romano
(1996a) and Gouveia (1997) offer no normative analysis. By con-
trast, the current article studies normative and positive allocations
and demonstrates how they compare to one another. Kifmann
(2005) extends Gouveia’s analysis by introducing a constitutional
stage where voters have a say on the mode of health care financing.
But, again, a normative analysis is missing as well as the integration
of provider payment.
Finally, our paper relates to the normative literature that ana-
lyzes both, health care financing and funding. Zeckhauser (1970)
was the first to simultaneously analyze provider payment and
health care financing. Ma and McGuire (1997) generalized this
framework. These papers analyze optimal health insurance in an
ex post moral hazard setting. We consider a multi-task agency set-
up instead and investigate a much richer set of financing regimes.
Moreover, their frameworks are normative in nature. An analysis
2
Kifmann and Roeder (2011) extend the analysis to premium subsidies and exam-
ine whether this approach is superior to social health insurance from a welfare
perspective. For a negative correlation they find that combining premium subsidies
with social health insurance is the optimal policy.
3
Epple and Romano (1996b) is another example. In this article the authors inves-
tigate a framework where individuals can opt out the public plan and buy private
health insurance. As a result, preferences are no longer single-peaked.
 R. Nuscheler, K. Roeder / Journal of Health Economics 42 (2015) 197–208
Table 1
Agent heterogeneity.
Income
p r  xij
Health l  pl  rl  l >0.5
risk h  ph  rh  h < 0.5
  p > 0.5  r < 0.5 1 respectively.
of how their outcomes relate to those that would be implemented
in a political process is missing. Our model is considerably richer
in this respect allowing us to analyze the important interaction
between health care financing and funding in normative and pos-
itive settings in great detail and thereby to contribute to a rather
slim financing and funding literature.
The remainder of the paper is organized as follows. Section 2
introduces the basic framework, followed by a normative analysis
in Section 3. Political outcomes are derived and compared to the
respective normative allocations in Section 4. We discuss model
extensions in Section 5. Section 6 concludes.
2. The model
2.1. Individuals
We consider a continuum of individuals of size one who differ
along the two most important dimensions when it comes to health
care financing and health care provision, namely, income and risk.
There are two income types i = r, p (rich and poor) with income lev-
els yr > yp > 0. The probability of falling ill is denoted j and assumes
the value l ∈ (0, 1) for low-risk individuals and h ∈ (l , 1) for high-
risk individuals. Both income and risk are exogenously given. The
two-dimensional heterogeneity gives rise to ij-types and we denote
their share in the population ij ∈ [0, 12 ), where the upper bound is
introduced for the sake of interest.4 To ease notation we define
 i ≡  il +  ih and  j ≡  pj +  rj . In the following, we assume  p > 0.5,
that is, median income, yp , is below average income, y = p yp + r yr .
In addition, the majority of individuals is exposed to a low health
risk,  l > 0.5. This implies that median risk, l , is smaller than aver-
age risk,  = l l + h h . Table 1 summarizes.
Sickness inflicts a disutility L > 0 on individuals. Through the
receipt of care these costs can be mitigated but not eliminated. We
refer to this reduction as the benefit of treatment and denote it
b ∈ [0, L). There is one health care provider (HCP), e.g., a hospital,
specialist, or general practitioner, who treats all patients in need of
care. We summarize all provider activities that aim at increasing
b in a single, one-dimensional quality index q.5 We let bq > 0 and
bqq ≤ 0, that is, an improvement in health care quality increases the
benefits from treatment but at a decreasing rate.
Health care is publicly financed. We distinguish between three
financing regimes indexed n ∈ {T, t, p}. The government either
uses individualized lump-sum transfers Tij (optimal income tax-
ation, n = T), proportional income taxes with tax rate t ≥ 0 (n = t),
or insurance premiums (n = p) to generate the revenue required to
reimburse the HCP. Insurance premiums can be either risk-based,
pooled, or be a mixture of both. More precisely, the insurance pre-
mium is given by [(1 − ϕ)j + ϕ]p, where p is the price of health
care and ϕ ∈ [0, 1] the extent of pooling.
In addition to the (expected) benefits from health care, agents
derive utility from consumption of a numéraire commodity
4
In our political game an ij-type with ij ≥ 12 could dictate the allocation.
5
This assumption rules out multi-task quality issues when it comes to optimal
provider payment (Chalkley and Malcomson, 1998a,b and (Kaarbøe and Siciliani,
2011)).
199
denoted x. Assuming quasi-linear preferences, the optimization
problem of individual ij can be written as6
max Uij = xij + j [b(q) − L], (1)
where the budget constraint amounts to xij = yi − Tij , xij = (1 − t)yi ,
and xij = yi − [(1 − ϕ)j + ϕ]p for financing regimes T, t, and p,
2.2. The health care provider
When treating a patient the provider incurs treatment costs K.
As health care always includes a random element, these costs are
uncertain. From an ex ante perspective only expected treatment
costs, c = E(K), matter. Despite this uncertainty the HCP has an influ-
ence on costs and we suggest that c is a function of the quality of
care, q, and the provider’s effort to keep treatment costs down, e.7
We let the expected treatment cost function c(q, e) satisfy cq > 0,
cqq > 0, ce < 0, cee ≥ 0 and ceq = 0. Expected treatment costs, thus,
increase with quality and do so at an increasing rate. More effort
implies lower treatment costs in expectation but at a decreasing
rate.8 The provider’s effort to contain costs involves a disutility v(e)
per patient, where ve > 0 and vee > 0. Finally, the HCP incurs a fixed
cost F.
Total provider reimbursement is denoted P and the actual
payment is determined by a simple linear cost-sharing arrange-
ment, P = K + . The parameter  ∈ [0, 1] is the extent to which the
health authority (HA) is willing to share treatment costs. As an addi-
tional compensation the HCP may receive a lump-sum payment
 ∈ R per patient treated. The expected reimbursement is given by
E(P) = c(q, e) + . (2)
In addition to reimbursement the HCP derives utility from the
patients’ benefits generated through treatment. This gives rise to
the HCP’s expected payoff:
H(q, e) =  [˛b(q) − c(q, e) − v(e) + E(P)] − F, (3)
where  is the share of agents treated, that is, all individuals who
are sick. The parameter ˛ captures the HCP’s degree of altruism
towards patients’ health benefits (see, e.g., Ellis and McGuire, 1990;
Eggleston, 2005; Jack, 2005). For ˛ = 1, the HCP is a ‘perfect agent’
and for ˛ < 1 an imperfect one.9
Substituting E(P) in Eq. (3) by Eq. (2) we arrive at the HCP’s
optimization program
maxH(q, e) =  [˛b(q) − (1 − )c(q, e) − v(e) + ] − F. (4)
q,e
Our assumptions about the benefit function b(q) and the cost func-
tions c(q, e) and v(e) ensure that the HCP’s problem is concave. The
first-order conditions with respect to q and e amount to
∂H(q, e)
= 0 ⇔ ˛bq − (1 − )cq = 0, (5)
∂q
6
We assume that patients passively accept the quality of treatment the HCP is
willing to provide.
7
Although we are not concerned with multiple quality dimensions we have a
multi-task agency problem: the HCP chooses quality and effort.
8
The assumption that the cross derivative vanishes is made for analytical conve-
nience.
9
Note that even in the case of perfect agency, the individuals’ well-being does not
fully enter the HCP’s utility function. The reason is that even though the HCP con-
siders the patients’ utility from treatment, it does not take into account the financial
costs of service delivery, that is, the taxes or premiums that are borne by the patients
(see also Jack, 2005). So, the HCP’s and the patient’s valuation of health care services
may differ even if ˛ = 1.
200 R. Nuscheler, K. Roeder / Journal of Health Economics 42 (2015) 197–208

∂H(q, e) for regimes T, t, and p, respectively. In equilibrium the participation

= 0 ⇔ −(1 − )ce − ve = 0. (6)
∂e constraint will be binding so that we can solve Eq. (9) for . Inserting
 into the budget constraints (10)–(12) we obtain one condition for
The above system of equations yields the optimal quality of health
each regime that guarantees both, a balanced public budget and
services q( ; ˛) and the optimal level of cost reducing effort e( ; ˛)
participation of the HCP:
as functions of the cost-sharing parameter  and the degree of altru-  
ism ˛. The optimal quality of health care services is determined such T () =  c(q, e) + v(e) − ˛ˇb(q) + F, (13)
that the internalized marginal benefit of treatment (the first term

 F

of Eq. (5)) is equal to the marginal cost of providing quality (the t() = c(q, e) + v(e) − ˛ˇb(q) + , (14)
second term). The optimal cost reducing effort is chosen to equal- y 
ize the marginal benefit of lower treatment costs (the first term of F
Eq. (6)) with the marginal disutility of effort (the second term). p() = [c(q, e) + v(e) − ˛ˇb(q)] + . (15)

The actual quality and effort levels the HCP is willing to provide
Since e = e() and q = q(), the budget balancing revenue
depend on the contract he has. Specifically, it depends on the cost-
depends on the cost-sharing parameter  only indirectly, that is,
sharing component . With the help of Cramer’s rule (see the
via the influence cost-sharing has on the HCP’s optimally chosen
Appendix A.2), we get
quality and effort. With the help of the implicit function theorem
dq(; ˛) −cq and observing the first order conditions of the HCP, Eqs. (5) and (6),
q ≡ = > 0, (7)
d ˛bqq − (1 − )cqq we have

de(; ˛) ce
e ≡ = < 0. (8) dT () dt() dp()
d (1 − )cee + vee =y = = [(cq − ˛ˇbq )q + (ce + ve )e ] > 0.
d d d
The more cost-based the HCP’s payment the larger the quality he is (16)
willing to supply and the smaller the effort he exerts to reduce cost.
The first effect is due to the lower price the HCP has to pay for quality A higher cost-sharing parameter requires higher public revenues.
improvements when the cost-sharing parameter  increases. The This is a direct consequence of the HCP’s response to an increase
less expensive the provision of quality the more the HCP is willing in cost-sharing: more cost-sharing implies higher quality (7) and
to offer. The second effect is a moral hazard effect. An increase in the less cost reducing effort (8) both increasing expected health care
share of reimbursed treatment costs undermines the HCP’s incen- spending and with it the revenues needed to balance the public
tives to contain costs. Additionally, q( ; ˛) positively depends on budget.
˛: the more the HCP cares about the patients’ health care benefits,
the higher the quality of health care services he delivers. 2.4. The economic equilibrium

2.3. The public health care scheme The following definition introduces the notion of economic
equilibrium into our model economy.
The government or the health authority (who is the purchaser
of health care services) faces two constraints, a participation con- 
Definition 1. (Economic equilibrium)
An allocation xpl , xph , xrl , xrh , q, e with policy instruments
straint and a budget constraint. We assume throughout that the Tij /t/p,  and  constitutes an equilibrium of the economy if the
benefit from treatment is sufficiently large so that the HA always following conditions hold:
wants to contract with the HCP. As the HCP cannot be forced to
provide health care services, the lump-sum transfer  per patient
(i) the utility of all agents is maximized, i.e., the program given in
and the cost-sharing parameter  must be chosen such that the HCP
Eq. (1) is solved,
is willing to accept the contract. With a reservation utility of zero
(ii) the health care provider’s participation constraint (9) is satis-
the participation constraint reads as10
fied and its payoff is maximized, i.e., the program given in Eq.
 
 ˛ˇb(q) − (1 − )c(q, e) − v(e) +  − F ≥ 0. (9) (4) is solved, and
(iii) the government’s budget constraint is balanced, i.e., for opti-
The parameter ˇ ∈ {0, 1} allows us to distinguish between two sce- mal income taxation Eq. (10) holds, for proportional income
narios. For ˇ = 0 the monetary part of the HCP’s utility needs to be taxation Eq. (11) holds, and for insurance premiums Eq. (12)
non-negative. In contrast, for ˇ = 1 the participation constraint is holds.
less demanding as the (partial) internalization of patients’ health
benefits allows for negative monetary payoffs. The functional form In an economic equilibrium the utility level obtained by type-
of (9) and the parameters ˛ and ˇ are common knowledge. ij agents can be expressed by their indirect utility function Vijn (),
As already noted above, the expected health care expenses, where
HCE =  [c(q, e) + ], can either be financed by optimal income
taxes, proportional income taxes, or insurance premiums. To bal- VijT () = yi − Tij () + j [b(q()) − L] , (17)
ance the public budget, public revenues need to be equal to the
expected health care expenses. We have Vijt () = (1 − t())yi + j [b(q()) − L] , (18)
p
ij ij Tij ≡ T = HCE, (10) Vij () = yi − [(1 − ϕ)j + ϕ]p() + j [b(q()) − L] . (19)

ti i yi ≡ ty = HCE, (11) The indirect utility function represents an individual’s preferen-
ces over the cost-sharing parameter, . Inspection of Eqs. (17)–(19)
pj j [(1 − ϕ)j + ϕ] = p = HCE, (12) reveals that these preferences depend on the mode of health care
financing. This already points to the different distributional proper-
ties of the three financing regimes. With individualized lump-sum
10
A strictly positive reservation utility would simply add to the fixed costs of being transfers the government can perfectly redistribute between indi-
active in the market. viduals so as to equalize their (marginal) utilities. With proportional
 R. Nuscheler, K. Roeder / Journal of Health Economics 42 (2015) 197–208
income taxation, income can still be redistributed from rich to poor
using health care as a vehicle (see, e.g., Gouveia, 1997). While any
form of redistribution is ruled out when premiums are risk-based
(ϕ = 0), partial pooling, ϕ ∈ (0, 1], implies redistribution from low-
is the Lagrangean multiplier on the budget constraint.
risk to high-risk agents with the extent of redistribution being
increasing in ϕ. In Sections 3 and 4 we carefully analyze how these
differences affect the allocations in normative and political econ-
omy environments.
3. Optimal financing and funding
In this section we study the optimal interaction between
health care financing and funding considering four different allo-
cations: first-best, second-best, and two third-best allocations.
To determine the first-best allocation we consider quality and
effort contractible and let optimal income taxation be feasible,
that is, health care financing is through individualized lump-sum
transfers. By maintaining the optimal income taxation assump-
tion and introducing non-contractibility for both quality and effort
we obtain the second-best allocation. This allocation is then com-
pared to two third-best allocations where quality and effort are still
considered non-contractible but alternative health care financing
regimes are considered, namely, proportional income taxation and
insurance premiums.
Throughout our analysis we assume a HA or government who
may have redistributive concerns, that is, who may aim at redis-
tributing to the double disadvantaged in society, that is, to high-risk
low-income individuals. We incorporate redistributive consider-
ations into the analysis by letting the HA’s objective function be

W= ij (Uij ),
ij
where (·) is a strictly increasing and weakly concave function of
individual utility levels, i.e.  (·) >0 and  ( · ) ≤ 0. This formulation
comprises a utilitarian HA as limiting case,  ( · ) = 0. Due to quasi-
linearity every income distribution would then be optimal. In other
words, all redistributive concerns of the HA are contained in . An
immediate implication is that a necessary condition for a difference
between the second-best and the two third-best allocations is the
presence of redistributive concerns. In the following, we investigate
how these concerns shape optimal health policy.
3.1. Normative benchmarks — optimal income taxation
3.1.1. First-best
The HA maximizes (20) with respect to Tij , q and e subject to the
HCP’s participation constraint (9) and the public budget constraint
(10). The optimization problem is11

max W ∗ (Tij , , q, e) = ij yi − Tij + j [b(q) − L]
Tij ,,q,e
ij
s.t. (13).
Inserting the participation constraint into the budget constraint
ij = 0 ⇒ ij =
∀ ij, (26)
and taking the derivative with respect to Tij , q and e yields
∂W ∗ (Tij , q, e)
= −ij ij +
ij = 0 ⇔ ij =
∀ ij, (22)
∂Tij
∂W ∗ (Tij , q, e)
= −cq + [1 + ˛ˇ]bq = 0,
∂q
11
Chalkley and Malcomson (1998b) argue that this welfare function includes
double counting of the patient’s health care benefits. Our results do not change qual-
itatively when the altruistic term is dropped from the HCP’s participation constraint,
that is, when ˇ = 0 (see also Kaarbøe and Siciliani, 2011).
201
∂W ∗ (Tij , q, e)
= ce + ve = 0, (24)
∂e
where
The first condition states that individualized lump-sum transfers
should be chosen such that marginal utilities are equalized. For
 < 0 this implies that, for a given risk type, high-income indi-
viduals have to pay more taxes than low-income agents: Trj > Tpj .
Similarly, for a given income type, low-risk agents have to pay more
taxes than high-risk agents as the former suffer the utility loss from
illness with a higher probability than the latter: Til > Tih . We get
the following ordering of individual lump-sum taxes: Trl > max {Tpl ,
Trh } ≥ min {Tpl , Trh } > Tph . Whether Tpl ≶ Trh depends on the rel-
ative inequity between income and risk. Without redistributive
concerns,  = 0, all Tij that balance the public budget would be
optimal.
Eq. (23) states that first-best quality, q* , should be expanded
until the marginal benefit to patients, bq , is equal to the marginal
costs of providing health care services, cq − ˛ˇbq . Obviously, the
costs of expanding quality are lower when the participation
constraint of the HCP includes an altruistic component (ˇ = 1)
as compared to a situation where the HCP has to break even
in monetary terms (ˇ = 0). As a result, the first-best efficient
quality level is higher in the former case and the quality differ-
ential is increasing with ˛. The first-best cost reduction effort, e* ,
equates the marginal disutility of exerting effort, ve , to the marginal
expected treatment cost savings, −ce ; see Eq. (24).
3.1.2. Second-best
(20) The government can still impose optimal income taxes but both
quality and effort are non-contractible.12 Using the reimbursement
system the HA can steer the HCP’s incentives to invest in quality and
to contain costs. We focus on linear contracts as given in Eq. (2)
that specify a lump-sum transfer per patient  and a cost-sharing
parameter .
Again, the HA maximizes the welfare function subject to the
participation constraint (9) of the HCP and the public budget con-
straint (10). The HA, however, no longer optimizes over Tij , q and e,
but over Tij and . In other words, she loses one degree of freedom
as compared to the first-best problem. In determining the optimal
cost-sharing parameter the HA has to consider the HCP’s optimal
quality and effort responses to cost-sharing arrangements as given
by Eqs. (5) and (6). Specifically, the optimization problem is given
by

max W T (Tij , ) = ij (yi − Tij + j [b(q) − L])
Tij ,
ij (25)
s.t. (5), (6) and (13).
(21) After some rearrangements, the first order conditions with respect
to Tij and  can be written as
∂W T (Tij , )
= −ij ij +
∂Tij
12
While this assumption is rather plausible for cost reducing effort it is at odds
with the increasing importance of pay-for-performance schemes. Nevertheless,
(23)
the assumption can be justified on a number of grounds. First, the assumption is
fairly standard in the earlier literature on provider payment (see, e.g., Chalkley and
Malcomson, 1998a; Chalkley and Malcomson, 1998b). Second, even the more recent
contracting literature acknowledges that there are unobservable quality dimensions
(see, e.g., Kaarbøe and Siciliani, 2011) and we concentrate on those. Third, measur-
ing quality is a complicated endeavor and creates problems in its own right (as a
response to report cards Dranove et al. (2003), for instance, find selection behavior
on the side of providers).
202 R. Nuscheler, K. Roeder / Journal of Health Economics 42 (2015) 197–208

  
∂W T (Tij , )
  y
ij ij ij i
−  
ij ij ij
 y
ij ij i
= −cq + [1 + ˛ˇ]bq q − (ce + ve )e = 0. (27) y ≡   . (32)
∂ 
ij ij ij
  y
ij ij i
Like in the first-best solution individualized lump-sum transfers
should be chosen such that marginal utilities are equalized. Eq. (27) These measures amount to the standardized covariances between
yields the optimal degree of cost-sharing,  T , which, using the HCP’s the welfare weights the government attaches to a particular ij-
first order conditions (5) and (6), dictates qT and eT : type and health risk and income, respectively. For a utilitarian
HA,  = 0, all individuals have the same welfare weight irrespec-
˛bq (qT ) = (1 −  T )cq (qT ) and − (1 −  T )ce (eT ) = ve (eT ). (28) tive of health and income, that is, ij = > 0 ∀ ij. The absence of
Implementation of the first-best allocation is generally impossible redistributive concerns is reflected in the distributional character-
as the HA has only one instrument, the cost-sharing parameter , istics both assuming the value zero,   =  y = 0. Similarly, without
but two margins, namely, q and e. It can easily be verified that a agent heterogeneity the distributional characteristics would be
cost-sharing parameter  = 0 implements the efficient effort level zero: l = h implies   = 0 and yp = yr implies  y = 0.13
as the provider is the residual claimant for his cost savings. Quality Using Eqs. (31) and (32), Eq. (30) can be rewritten as
provision, however, will then be inefficient unless ˛(1 − ˇ) = 1. As  

−cq +  + ˛ˇ bq q − (ve + ce )e = 0, (33)

˛ ∈ [0, 1] and ˇ ∈ {0, 1} first-best quality requires perfect altruism
˛ = 1 and ˇ = 0 when  = 0. If ˛(1 − ˇ) = / 1 we get  =/ 0 and with it
where  ≡ (1 +   )/(1 +  y ). The following proposition states how
a distortion of the HCP’s incentives to exert effort. The HA can, thus,
the third-best allocation with proportional income taxes relates to
either implement the first-best cost reducing effort or the first-best
the second-best outcome.
level of quality but not both. The optimal cost-sharing parameter,
 T , optimally trades off the inefficiencies along the two dimensions. Proposition 1. (Optimal policy with proportional income taxes)

3.2. Third-best — the optimal financing and funding interaction (i)   ≥  y ⇔  t ≥  T ⇔ qt ≥ qT and et ≤ eT .
(ii) The third-best allocation with proportional income taxes is identi-
In addition to the resource constraints and the non- cal to the second-best allocation if and only if  y =   = 0.
contractibility of quality and effort the HA now faces financing
constraints. When optimal income taxation is no longer feasible
The intuition for part (i) is as follows.14 Whenever  > 1, which
the HA has to resort to alternative financing regimes, propor-
is equivalent to   >  y , the marginal benefit of quality provision
tional income taxation and insurance premiums being the most
is higher under proportional income taxes than under optimal
prominent arrangements. We investigate the resulting third-best
income taxation. The reason being the distributional consequences
allocations in turn.
of improvements in quality. High-risk individuals are more likely to
benefit from higher quality care but, ceteris paribus, do not pay more
3.2.1. Proportional income taxes
taxes than low-risk individuals. Additionally, rich individuals con-
With proportional income taxes the policy instruments are
tribute more to the financing of better quality without benefitting
given by the tax rate t and the cost-sharing parameter . The opti-
more than poor individuals. The increase in quality, thus, implies
mization problem reads as
 more distribution towards poor and high-risk individuals. These
max W t (t, , ) = ij ((1 − t)yi + j [b(q) − L]) redistributive benefits have to be weighed against the additional
t, distortion of effort incentives and the third-best optimal degree
ij (29)
s.t. (5), (6) and (14). of cost-sharing does so optimally. For  < 1 cost-sharing in the
third-best is less pronounced than in the second-best and qual-
Inserting the constraints, the optimization problem simplifies to ity and effort levels compare accordingly. Optimal cost-sharing in
the third-best and second-best optima is identical if and only if
 
  = 1. If   =  y =
/ 0, then health care provision is second-best effi-
W t () = ij 1− [c(q(), e()) cient. As distributional concerns matter, however, the allocation is
y
ij not second-best efficient. As a consequence of suboptimal health
F
  care financing the income redistribution is inefficient. Only in the
+v(e()) − ˛ˇb(q()) + yi + j [b(q()) − L] . absence of redistributive concerns,  = 0, or when there is no need

for redistribution, l = h and yp = yr , the third-best and second-best
The optimal cost-sharing parameter,  t , is implicitly determined by allocations coincide (ii).
the following first order condition One can certainly speculate about how the distributional char-
acteristics compare to one another. We conjecture that the HA
 
dW t ()    attaches a higher welfare weight to the disadvantaged, that is,
= ij ij − yi (cq − ˛ˇbq )q + (ce + ve )e to individuals with high health risks, Cov( ij , j ) > 0, and low
d y
ij income, Cov( ij , yi ) < 0. This is equivalent to   > 0 and  y < 0

which implies  > 1 and with it more cost-sharing, higher qual-
+j bq q = 0, (30) ity and less effort under proportional income taxation as compared
to optimal income taxation.
where q and e are defined as in Eqs. (7) and (8). To gain a better
understanding of how redistributive concerns affect optimal cost-
sharing under proportional income taxation  t and thus optimal 13
Interpretation of the distributional characteristics is not straightforward. Note,
quality qt and effort et , we introduce the distributional character- that we made no assumptions about the correlation between income and risk.
istics of health risk and income (see Feldstein, 1972): This implies that any assumption on the sign of the distributional characteristics
   implicitly includes assumptions on the sign and size of this correlation.
  
ij ij ij j
−  
ij ij ij
 
ij ij j
14
Analytically, the result follows — like in all other propositions and corollaries —
 ≡  
 , (31) from a simple comparison of the coefficients on quality in the respective first order

ij ij ij
 
ij ij j conditions, here Eqs. (27) and (33).
 R. Nuscheler, K. Roeder / Journal of Health Economics 42 (2015) 197–208
3.2.2. Insurance premiums
With insurance premiums the optimization problem of the HA
is given by
 leading to higher expected health care expenses in the former sys-
max W p (p, , ) = ij (yi − [(1 − ϕ)j + ϕ]p + j [b(q) − L])
p,
ij
s.t. (5), (6) and (15).
Employing the same approach as in the previous section and using
Eq. (31), the first order condition for the optimal cost-sharing
parameter  p can be written as
  
income taxation (first-best and second-best) and contrasted them
1 + 
−cq + + ˛ˇ bq q − (ce + ve )e = 0.
1 + (1 − ϕ) 
This condition allows us to state our next proposition where we
compare the results with insurance premiums to the second-best
allocation.
Proposition 2. (Optimal policy with insurance premiums)
(i) When insurance premiums entail some redistribution, ϕ > 0, then:
 ≥ 0 ⇔ p ≥ T ⇔ qp ≥ qT and ep ≤ eT .
(ii) For purely risk-based premiums, ϕ = 0, health care provision is
second-best efficient while health care financing is not.
(iii) The third-best allocation with insurance premiums is identical to
the second-best allocation if and only if  y =   = 0.
The intuition is similar to the one of Proposition 1. Suppose
that insurance premiums imply some redistribution from low-risk
to high-risk individuals as stated in part (i) of the proposition.
Then an increase in quality triggered by more cost-sharing ben-
efits high-risk individuals with a higher probability than low-risk
individuals. The resulting increase in insurance premiums, how-
ever, is not shared accordingly when premiums are partially pooled.
High-risk individuals disproportionately benefit from an increase
in quality making such improvements desirable whenever the HA
attaches a higher welfare weight to high-risk individuals than to
low-risk individuals, i.e., whenever   > 0. In this case it pays off
to distort quality and effort away from its second-best levels as the
resulting improvement in the distribution of income outweighs the
associated losses in efficiency. There is no such distributional ben-
efit when premiums are purely risk-rated, (ii). This result carries an
important policy message: second-best optimal health care provi-
sion can be achieved by both financing regimes, optimal income
taxation and risk-rated insurance premiums. It emphasizes that an
isolated (second-best) analysis of provider payment rests on strong
assumptions regarding the financing of health care — unless, (iii),
the HA has no redistributive concerns.
Combining Propositions 1 and 2 we find that the optimal policy
is sensitive to the mode of health care financing when redistribu-
tive concerns matter. Only in the special case of no redistributive
concerns is the redistribution of income, implied by a change from
one health care financing regime to the other, welfare neutral. We
emphasize this in the following corollary.
Corollary 1. (Optimal policy: proportional income taxes vs. insur-
ance premiums)
(1 − ϕ)  ≥  y ⇔  t ≥  p ⇔ qt ≥ qp and et ≤ ep .
In general, the comparison of cost-sharing across financing
regimes depends on the welfare weights the HA attaches to the
different types. We already suggested above, that the HA may wish
to redistribute towards the disadvantaged in the society, namely,
to high-risk low-income agents. This implies   > 0 and  y < 0 and,
following the corollary, less cost-sharing with insurance premiums
203
independent of the extent of pooling ϕ. As a result, health care qual-
ity in tax based systems is expected to exceed quality in insurance
financed systems. It is the other way round for cost reducing effort,
tems than in the latter.
(34)
4. Political implementability of optimal policies
In the previous section we assumed a normative perspective
and investigated the optimal health care policies under optimal
(35) with alternative financing regimes (third-best). Implementing an
optimal cost-sharing parameter, however, may not be feasible
politically. Under majority voting, policy makers will commit to
policies that maximize the number of votes rather than the social
objective. Taking a median voter approach, this section derives the
political equilibria under proportional income taxation and insur-
ance premiums and contrasts them with the respective third-best
and second-best policies.
With a continuum of individuals each voting agent has zero
mass, so that no individual vote can change the outcome of the
election. We let all agents alive cast a ballot over the cost-sharing
parameter . Although it appears more natural to let voters decide
on the parameter which determines the financing of health care
rather than on one parameter of the reimbursement scheme, this
approach is without loss of generality. Remember that in an eco-
nomic equilibrium only one of the three policy instruments can
be set freely. The other two are then residually determined by the
HCP’s participation constraint and the HA’s budget constraint. It is
thus irrelevant whether voters decide on , t/p or . We opt for
the former as it allows for a direct comparison of the equilibrium
allocations of the two approaches, normative and positive.
4.1. Proportional income taxes
Individuals maximize their indirect utility function (18) with
respect to the cost-sharing parameter  subject to the constraint
on the tax rate t given by Eq. (14). The first order condition of an
ij-type amounts to15
dVijt ()
= (−ıi cq + [ıj + ˛ˇıi ]bq )q − ıi (ce + ve )e = 0, (36)
d
where we defined ıi ≡ yi /y and ıj ≡ j /. The above equation
implicitly determines the most preferred cost-sharing parame-
ter ijt of a type-ij individual.16 To see how the most preferred
cost-sharing parameter depends on income and risk we apply the
implicit function theorem which, observing the first order condi-
tions of the HCP, Eqs. (5) and (6), yields
dijt (cq + ˛(1 − ˇ)bq )q + ce e
= < 0 and (37)
dıi SOC ij
dijt bq q
=− > 0. (38)
dıj SOC ij
These inequalities allow us to order the four types according to their
most preferred cost-sharing parameter and with it to identify the
median voter.
15
For convenience we neglect the constraint  ∈ [0, 1]. Both, the ordering of types
with respect to their most preferred cost-sharing parameters as well as the com-
parison of allocations, normative and positive, remain valid.
16
In the following, we assume that the second order condition is satisfied,
2
d Vijt /d 2 ≡ SOC ij < 0, implying that preferences are single-peaked (see Appendix
A.1.2).
204 R. Nuscheler, K. Roeder / Journal of Health Economics 42 (2015) 197–208

Lemma 1. (Median voter: proportional income taxes)
<  ⇔ ıl < 1 +   . High-risk individuals disproportionately benefit
The most preferred cost-sharing parameters can be ordered as fol- from quality improvements. As the median voter is a low-risk agent,
lows: rlt < min{pl
t ,  t } ≤ max{ t ,  t } <  t . The median voter is
rh pl rh ph
he aims at limiting the redistribution towards high-risk individuals
t = t .
a pl-type agent, i.e., m and he does so by implementing a low cost-sharing rate. As the dis-
pl
tributional preferences of the HA will usually point in the opposite
Eq. (37) implies rjt < pj t . For a given risk type high-income
direction,   > 0, cost-sharing is more pronounced in the third-best
agents contribute more to the financing of the health care scheme allocation than in the political outcome. These arguments extend to
than low-income agents. As the risk type is given, the bene- the case with inequities in both, income and risk.
> : the cost-
fit from more cost-sharing in terms of higher quality is uniform
across income types making cost-sharing less attractive for high-
income individuals. Similarly, Eq. (38) implies ilt < iht . For a given
income, the expected benefits from quality improvements trig-
gered by more cost-sharing are higher for high-risk individuals
than for low-risk individuals. As financing costs are uniform across
risk types, high-risk agents prefer more cost-sharing than low-risk
sharing benefit accruing to the median voter is larger than in the
social objective as, due to the sufficiently low health risk inequal-
ity as compared to income inequality, the benefits of better quality
care are distributed more evenly. In that sense the redistribution is
better targeted towards the median voter. The HA’s redistributive
concerns are lower so that her cost-sharing incentives are weaker.
= : although this is a knife-edge case it is noteworthy that
agents. Although the comparison of the most preferred cost-sharing
parameters of pl-types and rh-types generally depends on the
distribution of types, identification of the median voter is straight-
forward: as neither the high-risks nor the rich can form a majority
= 1,
( h < 0.5 and  r < 0.5, respectively), the median voter is a type-pl
agent who either forms a majority with low-income high-risk indi-
viduals or with high-income low-risk individuals.
Inserting the median voter’s type into Eq. (36) and dividing by
ıp allows a direct comparison with the third-best policy under pro-
portional income taxation.
(−cq + [
+ ˛ˇ]bq )q − (ce + ve )e = 0,
where
the distributional preferences of the median voter and the HA are
aligned so that the political outcome is identical to the third-best
allocation.
If the inequality in health risk and income is identical,
then the political outcome with health care financing through pro-
portional income taxation leads to second-best efficient health care
provision. The reason is that the median voter’s benefits from cost-
sharing through improvements in the distribution of income are
exactly compensated by the worsening of the distribution of health
benefits. Finally, this allocation is second-best efficient if   =  y = 0.
We argued above that the HA likely attaches a higher welfare
(39)
weight to high-risk or low-income individuals than to low-risk or
≡ ıl /ıp . While ıl ≤ 1 and ıp ≤ 1 measure the inequity in high-income ones. But then  > 1 and the third-best can only be
health and income respectively, their fraction,
, is a measure of implemented through a majority voting process if the inequality
the relative inequity between the two dimensions. in health risk is smaller than the inequality in income such that
 =
> 1. In general, however, the second-best cannot be achieved
Proposition 3. (Political outcome with proportional income taxes) as a political equilibrium.

(i) The relationship between cost-sharing arrangements in the third-

best optimum and the political equilibrium depends on how the 4.2. Insurance premiums
relative inequity along income and risk relates to the distribu-
tional characteristics of cost-sharing with respect to these two When health care financing is through insurance premiums,
dimensions. More precisely, we have:  
⇔  t  m t ⇔ individual preferences over the cost-sharing parameter are given
t t t
q  qm and e  em . t
by Eq. (19). The corresponding first order condition for  is given
(ii) Health care provision is second-best efficient if and only if
= 1. by
Second-best efficiency of the allocation requires the absence of
redistributive concerns,   =  y = 0. p
dVij ()  
=− [(1 − ϕ)j + ϕ] cq q + ce e + ve e − ˛ˇbq q + j bq q = 0
d
Comparison to the first order condition for the cost-sharing (40)
   
parameter in the third-best, Eq. (33), shows that the difference j
⇔ −cq + + ˛ˇ bq q − (ce + ve )e = 0.
between these two equations depends on
and . We discuss (1 − ϕ)j + ϕ
the three possible cases in turn.
< : the benefits of the quality
of care receive a lower weight in the median voter’s first order con-
dition than in the HA’s first order condition. Since quality provision
can be stimulated by cost-sharing, q > 0, the cost-sharing parame-
ter will be larger in the normative equilibrium than in the political
one.17 Suppose there is no inequity in risk, then
First, we note that for purely risk-based premiums, ϕ = 0, all
individuals have identical most preferred cost-sharing rates so that
there is no conflict in the electorate on how to shape health care
funding. If premiums include a redistributive element, ϕ > 0, then
<  ⇔ ıp > 1 +  y . high-risk individuals benefit more from health care provision and
This condition can only be satisfied if the HA aims at redistributing vote for a higher cost-sharing parameter than low-risk agents:
towards the poor, that is, if  y < 0. The smaller  y the more pro- p
dij /dj > 0. This allows us to order the four types by their most
nounced are the redistributive concerns of the HA. If the inequality
preferred cost-sharing parameter and, as a result, to identify the
in income is sufficiently small as compared to the distributional
median voter.
preferences of the HA, ıp > 1 +  y , so are the benefits accruing to
the median voter from improvements in quality. This, in turn, Lemma 2. (Median voter: insurance premiums)
implies less cost-sharing in the political outcome than in the opti- The most preferred cost-sharing parameters can be ordered as fol-
mal one. Similarly, suppose there is no inequity in income, then p p p p
lows: rl = pl ≤ rh = ph . The median voter is an l-type agent, i.e.,
p p
m = il for i = r, p.
17
To see this formally observe As rich low-risk individuals share their cost-sharing preferences
t
dm bq q with poor low-risk individuals there will always be a majority for
=− > 0.
d
SOC ij their most preferred cost-sharing parameter. Inserting l into Eq.
(40) we arrive at our next proposition.
 R. Nuscheler, K. Roeder / Journal of Health Economics 42 (2015) 197–208
Proposition 4. (Political outcome with insurance premiums)
(i) When insurance premiums entail some redistribution, ϕ > 0, then
the relationship between cost-sharing arrangements in the third-
best optimum and the political equilibrium depends on how the
inequity in risk relates to the distributional characteristic of risk.
p
More precisely, we have: 1 +   ≥ ıl ⇔  p ≥ m ⇔ qp ≥
p p p
qm and e ≤ em .
(ii) Health care provision is second-best efficient if and only if ϕ = 0.
Second-best efficiency of the allocation requires the absence of
redistributive concerns,   =  y = 0.
If the HA aims at redistributing towards high-risk individuals,
  > 0, then the condition formulated in part (i) of the proposi-
tion always holds. High cost-sharing parameters imply generous
redistribution towards high-risk individuals. Other than the HA
the median voter, a low-risk agent, dislikes redistribution towards
high-risk types. Accordingly, cost-sharing is less pronounced in the
political outcome than under the third-best policy. This ordering
can only be reversed if the HA has a sufficiently strong preference to
redistribute towards low-risk type agents. Remarkably, for purely
risk-based premiums, ϕ = 0, the incentives of voters are aligned
with those of the HA in both the second-best and the third-best
environment.18 Although this result appears somewhat surprising,
the intuition is relatively straightforward. The optimal allocations
trade off inefficiencies along cost reduction effort and quality and
this tradeoff is not blurred by redistributive motives. Similarly,
in the political game redistribution plays no role as risk-based
premiums preclude any form of redistribution. The third-best allo-
cation can thus be implemented through a majority voting process.
In contrast to the proportional income taxation case, this result
holds independent of agent heterogeneity. Finally, note that for
purely risk-based premiums health care provision is second-best
efficient. The resulting income distribution may still be inefficient
unless there are no redistributive concerns.
Combining Propositions 3 and 4 we arrive at the following
corollary that facilitates comparison of political outcomes under
alternative financing regimes.
Corollary 2. (Political outcome: proportional income taxes vs. insur-
ance premiums)
t ≥ p p implemented. With progressive income taxation their net marginal
ıp − ıl ≤ ϕ(1 − ıl ) ⇔ m m ⇔ qtm ≥ qm and em t ≤
p benefit of health care provision is larger than with proportional
em .
The comparison of cost-sharing across financing regimes hinges
on how risk pooling, as measured by ϕ, relates to the relative
inequity between income and risk. Only in the knife-edge case,
ıp − ıl = ϕ(1 − ıl ), is health care provision under majority voting
invariant to a change from insurance premiums to proportional
income taxes.
5. Discussion
One can think of many extensions of the framework analyzed
in this article. Here we discuss two important ones: progressive
income taxation and endogenous choice of the health care financing
regime. We discuss these extensions in turn.19
5.1. Progressive income taxation
When considering tax financed health care we look at two
polar cases, namely, optimal income taxation and proportional
18
Analytically this is the case because Eqs. (27), (35) and (40) coincide.
19
We thank two anonymous reviewers for suggesting to extend the analysis along
these lines. Technical details are available as an online supplement to this article.
205
income taxation. Although somewhat unrealistic, the former finan-
cing arrangement offers a valuable benchmark (the second-best
allocation) against which the allocations under alternative finan-
cing regimes can be compared. The latter financing arrangement,
proportional premiums, are the rule rather than the exception
in the financing of social health insurance systems. 14 out of 16
OECD countries with such a system use proportional contributions
to finance health care.20 The only exceptions are Hungary and
Switzerland. While the former country relies, in part, on progressive
income taxes, Switzerland is an example for a regressive finan-
cing scheme that collects partially pooled insurance premiums.
Redistribution across risk types is limited insofar as individuals
can choose from a set of contracts that differ in both, premiums
and deductibles. As individual selection is — among other things
— driven by risk, pooling is incomplete, ϕ < 1. Income is relevant
insofar as individuals below a canton specific threshold receive
premium subsidies.21
Tax financed national health service systems tend to use pro-
gressive taxation to finance health care. Such a system is more
redistributive than a proportional financing regime and likely less
redistributive than a system that applies optimal income taxes.
Therefore, the resulting allocation under progressive taxation will
be some sort of mixture between the allocations under proportional
income taxation and optimal income taxation. In the following, we
investigate the normative and positive allocations for a simplified
progressive income tax scheme.
Suppose that there are two proportional income tax rates and
that the one for the rich exceeds the one for the poor. Such a scheme
is more redistributive than a proportional income tax system miti-
gating the role of health care in redistributive policies. This implies
that the HA would optimally implement a lower degree of cost-
sharing.
For the positive analysis of progressive income taxes we first
have to identify the median voter. Rich low-risk individuals
are still the ones who stand to benefit the least from publicly
financed health care and poor high-risk individuals are the ones
who benefit the most. As rich high-risk individuals can never
form a majority with risk types or income types of their kind,
poor low-risk individuals remain pivotal. It is, thus, the most
preferred cost-sharing level of the type-pl agent that is being
income taxation as high-income individuals now contribute more
to the financing of the health care scheme. This implies more
cost-sharing under progressive than under proportional income
taxation.
We can now compare the normative to the positive outcome.
Like with proportional income taxation the comparison hinges
on the marginal incentives for redistribution. As argued above,
in the normative setting, these incentives are likely to be smaller
under progressive income taxation than under proportional income
taxation. This is the other way round in the positive setting. Conse-
quently, the set of parameter values for which the positive outcome
entails more cost-sharing than the normative outcome is larger
under progressive income taxation (see Proposition 3).
20
Austria* , Belgium, Czech Republic* , Estonia, Germany* , Greece, Japan* , Korea* ,
Luxemburg, The Netherlands* , Poland, Slovakia* , Slovenia, and Turkey. The aster-
isk indicates that income thresholds exist above which no additional premiums are
being charged. In that sense, these schemes are proportional only up to the thresh-
old and then regressive (Source: HiT country profiles of the European Observatory
on Health Systems and Policies, available at http://www.euro.who.int/en/about-us/
partners/observatory.)
21
Kifmann (2005) argues that the Swiss arrangement is similar to, e.g., the German
setting where premiums are proportional up to an income ceiling.
206 R. Nuscheler, K. Roeder / Journal of Health Economics 42 (2015) 197–208

5.2. Endogenous choice of the health care financing regime

Our model considers health care financing arrangements to

be exogenous and shows how these arrangements interact with
provider payment. One may argue that this approach is incomplete
as neither the HA nor the voters have a say on the mode of health
care financing. In what follows we look at the optimal financing
regime and compare it to the one that would be implemented by
majority voting.
We start with the normative analysis be assuming that a fraction
of health care expenses are financed through proportional income
taxes and the remaining part through insurance premiums.22 Sup-
pose that health insurance contracts are (partially) pooled, i.e.,
ϕ > 0. Then, health care financing via insurance premiums implies
redistribution across risk types but not across income types. If
  > 0 the redistributive effect of insurance premiums is valued
by the HA.23 This redistribution is still intact when health care
financing is through proportional income taxation: taxes do not
depend on risk but income. Proportional income taxes not only
redistribute from low-risk to high-risk agents but also from high-
income to low-income ones. If   > 0 and  y < 0 (as we conjectured
in Section 3.2.1) the HA values both redistributive effects rendering
proportional income taxation unambiguously more attractive than Fig. 1. Allocations under optimal policies and their political implementation under
majority voting.
insurance premiums. From the perspective of the HA this implies
that health care should exclusively be financed through income
taxes. Cost-sharing incentives are, thus, governed by , see Eq. (33) in the two frameworks. With risk rated premiums in the political
and Proposition 1. outcome and proportional income taxation in the normative one
For the political economy analysis we suggest sequential voting
where individuals first cast a ballot on the health care financing
regime (stage 1) and then on the cost-sharing parameter (stage 2).
With backward induction, voting behavior at stage 1 will take the
voting outcome of stage 2 into account. Suppose that health care
financing is (partially) through insurance premiums. As the median
voter is a type-pl agent, there will be no risk pooling as type-l indi-
viduals are not prepared to subsidize type-h individuals. But then
the decision at stage 1 boils down to a comparison between propor-
tional income taxation and risk-rated insurance premiums. While
the latter scheme precludes any form of redistribution, the former
comes along with redistribution across both income and risk. The
median voter benefits from the redistribution across income but
is negatively affected by the redistribution across risk types. If the
inequity in income is sufficiently large as compared to the inequity
in risk, then the type-pl agent is in favor of proportional income
taxation. In this case, type-ph agents would also be in favor of pro-
portional income taxes as they benefit from both redistribution
channels. The positive analysis then predicts proportional income
taxation. If, however, the inequity in income is sufficiently low as
compared to the inequity in risk, then the type-pl agent will be in
favor of risk-rated premiums. He would team up with type-rl agents
so that the political outcome predicts risk-rated premiums.24
Comparison of equilibria is now straightforward. With income
taxation in both settings, normative and positive, the difference in
allocations hinges on how the relative distributional characteristics
compare to the relative inequity in income and risk. More precisely,
there will be more cost-sharing in the normative outcome than in
we get more cost-sharing in the latter if and only if  > 1 (see the
discussion following Proposition 1).25
Finally, this analysis demonstrates that — for political economy
reasons — the prevailing health care financing regime may differ
from the optimal one. This is the case when the median voter
prefers risk-rated insurance premiums but the HA proportional
income taxes.
6. Conclusion
Usually health care financing and health care funding are stud-
ied in isolation. This paper simultaneously analyzed the financing
of health care and provider payment and thereby filled a gap in the
literature. Considering a multi-task agency framework we derived
the optimal allocations under varying constraints. These included
alternate financing regimes yielding the first-best, second-best and
third-best allocations. We then investigated whether the optimal
policies can be implemented as political equilibria under majority
voting. A pairwise comparison of allocations allowed us to iden-
tify conditions under which allocations coincide. In other words,
we showed when the mode of health care financing is irrelevant
and when optimal policies can be implemented as a median voter
equilibrium (Fig. 1 in Appendix A.3 summarizes the respective con-
ditions).
More precisely, when the quality of care and cost-reducing
effort are non-contractible but optimal income taxation is feasible,
the optimal allocation is second-best efficient. While the income
distribution is optimal, the inefficiencies along quality and effort
the political one if and only if  >
(see Proposition 3). More inter- are, using supply-side cost-sharing, optimally traded off against
esting is the case where the mode of health care financing differs one another. If optimal income taxation is not feasible, and if the
health authority has redistributive concerns, health care financing
has redistributive consequences which, in turn, will affect optimal
provider payment.
22
Optimal income taxation is considered infeasible. So we are in the third-best With proportional income taxation health care financing implies
environment. a redistribution from high-income to low-income individuals and
23
For a social health insurance system an immediate implication is that the HA
would optimally choose flat premiums that is full pooling (ϕ = 1).
24
These results are similar to the ones obtained in the premium risk framework
25
by Kifmann (2005), who identifies conditions under which the electorate will unan- Recall that risk-rated premiums imply second-best health care provision, see
imously vote for proportional income taxes. Proposition 3 (ii).
 R. Nuscheler, K. Roeder / Journal of Health Economics 42 (2015) 197–208 207

from low-risk to high-risk individuals. The extent to which the public health care program be invariant to health care financing
health authority distorts provider payment away from its second- and be independent of individual heterogeneity and redistributive
best (and with it health care provision) depends on individual preferences. More precisely, if no additional redistribution over and
heterogeneity in risk and income and on the welfare weights the above the redistribution via the transfer scheme was beneficial (the
health authority attaches to the different types. We show that transfer scheme is financed by optimal income taxes26 ) or if the
Feldstein’s (1972) distributional characteristics can be used to public health care program and the transfer scheme have identi-
characterize the third-best allocation. If the distributional charac- cal redistributive properties (e.g., if both programs are financed by
teristics of income and risk are identical, so are second-best and proportional income taxes and transfers within the income scheme
third-best health care provision. As income distributions might are perfectly correlated with risk).
differ, the third-best allocation is not second-best efficient unless Our results make very clear that an isolated analysis of provider
both distributional characteristics are nil. This is the case when the payment rests on very strong assumptions regarding the mode
health authority has no redistributive concerns, or in the absence of of health care financing (optimal income taxation or risk-based
individual heterogeneity. The political outcome is governed by the premiums) or the redistributive concerns of the health author-
preferences of the median voter who happens to be a poor, low- ity (none). This has a bearing on the applicability of the results
risk agent. With proportional income taxation the median voter derived in the optimal provider payment literature. The quality-
contributes less to the financing of the health care scheme than the cost containment tradeoff will generally differ internationally. In
average voter. But, at the same time, he needs health care with a principle, our framework could be used to derive testable hypoth-
lower probability than the average voter. The median voter’s pre- esis regarding the relationship between health care financing and
ferences for cost-sharing are thus driven by the relative inequity supply-side cost-sharing. Actual hypothesis testing would require
between risk and income. These preferences will likely differ from accurate data on the degree of cost-sharing, Feldstein’s distribu-
the ones of the health authority so that the third-best allocation tional characteristics, and individual heterogeneity. To the best of
can generally not be implemented as a median voter equilibrium. our knowledge, such data is currently not available so that we leave
Fig. 1 below gives the condition under which the two are identical. the empirical analysis for future research.
It should be noted that the political outcome may yield second-best
health care provision. This is the case whenever the inequality in Appendix A.
risk is identical to the inequality in income. Only in the absence of
individual heterogeneity or redistributive concerns is this alloca- A.1. Second-order conditions
tion second-best efficient.
When considering insurance premiums as a mode of health care A.1.1. The health care provider
financing, redistributive consequences are driven by the extent Since the two leading principal minors alternate in sign:
of risk pooling. In the extreme case of risk-based premiums (no
pooling) the third-best allocation yields second-best health care |H
 qq | = |˛bqq −(1 − )cqq | < 0 
provision. The reason is that risk-based premiums preclude any  Hqq Hqe   ˛bqq − (1 − )cqq 0 
 = >0
form of redistribution so that the health authority refrains from  Heq Hee   0 (1 − )cee + vee 
distorting health care provision away from its second-best. But
this also implies that there is no conflict in the electorate how due to bqq < 0, cqq > 0, cee ≥ 0 and vee > 0, the HCP’s chosen quality
to shape health care provision. The third-best allocation can thus level, q( ; ˛), and cost-reducing effort, e( ; ˛), are a maximum of
be implemented as a political equilibrium and health care pro- the objective function H(q, e).
vision is second-best efficient. As the income distribution might
not be optimal the resulting allocations will generally not be A.1.2. Individuals
second-best efficient. In the event of (partially) pooled premiums, The second-order condition of individual-ij amounts to
health care financing through insurance premiums involves a
2
redistributive element. Income is redistributed from low-risk to d Vijt ()
2
= (−ıi cqq + (ıj + ˛ˇıi )bqq )(q ) + (−ıi cq + (ıj + ˛ˇıi )bq )q
high-risk individuals. Redistribution from high- to low-income d 2      
individuals is ruled out. Consequently, the third-best allocation <0
2
≷0
−ıi (cee + vee )(e ) −ıi (ce + ve )e .
and the political outcome will depend on the distributional char-     
acteristic of risk and the inequity in risk, respectively. Only in the <0 <0
event of aligned cost-sharing preferences of the health authority
and the median voter can the third-best allocation be imple-
mented by majority voting. Again, the respective condition is given A.2. Comparative statics
in Fig. 1.
Finally, we derived conditions under which health care finan- Taking the total derivative of the HCP’s first order conditions (6)
cing is irrelevant in third-best and political economy environments. and (5) yields
It turns out that health care financing will generally matter unless
(˛bqq − (1 − )cqq )dq = −cq d − bq d˛ (41)
risk pooling relates to the distributional characteristics and indi-
vidual heterogeneity in the way laid out in Fig. 1. ((1 − )cee + vee )de = ce d (42)
The framework we adopted only considers public health care
and abstracts from additional public programs like, most impor- as we assume cqe = 0. The above system of equations can be written
tantly, an income transfer scheme. An immediate implication is as
that the former program implicitly assumes the redistributive role     
dq 1 (1 − )cee + vee 0 −cq d − bq d˛
of the latter program (Cremer and Gahvari, 1997). Our results are = , (43)
de D 0 ˛bqq − (1 − )cqq ce d
largely robust to adding an income transfer scheme in the sense that
individual heterogeneity, redistributive preferences, and health
financing would still matter. The reason is that, apart from risk-
based premiums, health care financing is inherently redistributive. 26
This is, of course, equivalent to assuming health care financing through optimal
Only in rather extreme and unrealistic cases would the size of the income taxes.
208 R. Nuscheler, K. Roeder / Journal of Health Economics 42 (2015) 197–208

where D = (˛bqq − (1 − )cqq )((1 − )cee + vee ) > 0. Hence, we Chalkley, M., Malcomson, J.M., 2000. Government purchasing of health services.
have Handbook of Health Economics 1a, 847–890.
Cremer, H., Gahvari, F., 1997. In-kind transfers, self-selection and optimal tax policy.
dq −((1 − )cee + vee )cq −cq European Economic Review 41, 97–114.
= = >0 (44) Cremer, H., Pestieau, P., 1996. Redistributive taxation and social insurance. Interna-
d D ˛bqq − (1 − )cqq
tional Tax and Public Finance 3, 281–295.
de (˛bqq − (1 − )cqq )ce ce Dranove, D., Kessler, D., McClelland, M., Satterthwaite, M., 2003. Is more information
= = < 0. (45) better? The effects of “Report Cards” on health care providers. Journal of Political
d D (1 − )cee + vee Economy 111, 555–588.
Eggleston, K., 2005. Multitasking and mixed systems for provider payment. Journal
Additionally, we have of Health Economics 24, 211–223.
Ellis, R.P., McGuire, T.G., 1990. Optimal payment systems for health services. Journal
d2 q cq cqq of Health Economics 9, 375–396.
= >0 (46)
d 2 (˛bqq − (1 − )cqq )
2 Epple, D., Romano, R.E., 1996a. Ends against the middle: determining public service
provision when there are private alternatives. Journal of Public Economics 62,
297–325.
d2 e ce cee
= < 0. (47) Epple, D., Romano, R.E., 1996b. Public provision of private goods. Journal of Political
d 2 ((1 − )cee + vee )2 Economy 104, 57–84.
Feldstein, M.S., 1972. Distributional equity and the optimal structure of public prices.
American Economic Review 62, 32–36.
A.3. The health care financing and funding interaction Gouveia, M., 1997. Majority rule and the public provision of a private good. Public
Choice 93, 221–244.
Appendix B. Supplementary data Gravelle, H., 1999. Capitation contracts: access and quality. Journal of Health Eco-
nomics 18, 315–340.
Jack, W., 2005. Purchasing health care services from providers with unknown altru-
Supplementary data associated with this article can be found, in ism. Journal of Health Economics 24, 73–93.
the online version, at http://dx.doi.org/10.1016/j.jhealeco.2015.04. Kaarbøe, O.M., Siciliani, L., 2011. Multitasking, quality and pay for performance.
Health Economics 20, 225–238.
003 Kifmann, M., 2005. Health insurance in a democracy: why is it public and why are
premiums income related? Public Choice 124, 283–308.
Kifmann, M., Roeder, K., 2011. Premium subsidies and social health insurance:
References substitutes or complements? Journal of Health Economics 30, 1207–1218.
Ma, C.A., 1994. Health care payment systems: cost and quality incentives. Journal of
Blomqvist, A., Horn, H., 1984. Public health insurance and optimal income taxation. Economics & Management Strategy 3, 93–112.
Journal of Public Economics 24, 353–371. Ma, C.A., McGuire, T.G., 1997. Optimal health insurance and provider payment.
Breyer, F., Haufler, A., 2000. Health care reform: separating insurance from income American Economic Review 87 (4), 685–704.
redistribution. International Tax and Public Finance 7, 445–461. Nuscheler, R., 2003. Physician reimbursement, time consistency, and the quality of
Chalkley, M., Malcomson, J.M., 1998a. Contracting for health services with unmon- care. Journal of Institutional and Theoretical Economics, 302–322.
itored quality. The Economic Journal 108, 1093–1110. Zeckhauser, R., 1970. Medical insurance: a case study of the tradeoff between
Chalkley, M., Malcomson, J.M., 1998b. Contracting for health services when patient risk spreading and appropriate incentives. Journal of Economic Theory 2,
demand does not reflect quality. Journal of Health Economics 17, 1–19. 10–26.