Managing Editors:
J.P. Ancot, Netherlands Economic Institute, Rotterdam, The Netherlands
A.J. Hughes Hallet, University of Strathclyde, Glasgow, United Kingdom
Editorial Board:
F.G. Adams, University of Pennsylvania, Philadelphia, U.S.A.
P. Balestra, University of Geneva, Switzerland
M.G. Dagenais, University of Montreal, Canada
D. Kendrick, University of Texas, Austin, U.S.A.
J.H.P. Paelinck, Netherlands Economic Institute, Rotterdam, The Netherlands
R.S. Pindyck, Sloane School of Management, M.I. T., U.S.A.
H. Theil, University of Florida, Gainesville, U.S.A.
W. Welfe, University of Lodz, Poland
The titles published in this series are listed at the end of this volume.
Econometrics of Health Care
edited by
G. Duru
and
J. H. P. Paelinck
"
~
021093100 ts
Preface
G. Duru and J. H. P. Paelinck Vll
v
vi Table of contents
G. DURU, Lyons
J. H. P. PAELINCK, Rotterdam
Y. SAILLARD
CEPREMAP, 142, rue du Chevaleret, F75013  Paris
The need for a macroeconomic approach arises from an old debate, the crisis
revived, on the connection between the development of tertiary or non
market sectors and the economic growth slowdown. Characteristics of the
health sector (especially the effects of technical progress on costs of pro
duction) appear to many authors to be among the causes of economic crisis
or at least to be an indicator of contradictions in contemporary economic
evolution.
Theoretical foundations of this characterization of the health system's
macroeconomic dynamics can be found in Baumol (1967), Bacon and Eltis
(1976), Lorenzi et at. (1980) ... We don't intend here to give answers to this
general problem. Instead we propose, following the usual macroeconomic
method, to study the effects of health activities' development. The main
constraint here is the macroeconomic modelling which does not introduce
variables and relations which would permit one to estimate health risks or
effects of care activities on productivity in market sectors.
Health system
'Health system' is a very useful term but its frontiers must be defined. Here,
we use the French Health Accounts assumptions. So, activities included in
the health system are those 'economic activities which enter directly into the
provision of health care'.3 This function is ambiguously defined according to
the criterion of 'bringing specific means into play, these means being more or
less tied to medical techniques'.3
Such a criterion implies a sometimes difficult distinction between medical
and lodging activities. It excludes direct actions on morbidity risk factors (or
primary prevention); only the delivery, as opposed to the production, of
medical and pharmaceutical products is included. In fact, we restrict our
attention to medical activities, excluding other activities, which are induded
in French Health Accounts: teaching and training of medical and nonmedical
manpower, medical research, sanitation, management of the health system.
Institutional units 'exercising by some main or secondary means one health
activity' 4 belong to the health system. As for medical activities, these units
are mainly hospitals and physicians' practices.
More precisely, using usual macroeconomic models leads one to study
only the 'care system' as opposed to the 'health system'. This means that the
variable 'population health condition' is excluded.
Previous modelling approaches 5 told us that, with available statistical
information, a more complete approach introducing care services consump
tion (prescribed or directly bought), sociocultural factors, care supply, and
population health condition, environment ... above all displays the pre
eminence of environment on care services consumption in explaining varia
tions in health condition.6 But the factors which appear the most relevant are
precisely those which are not usually included in macroeconomic models.
Even if they were, we would have to propose explicit relations between
environment, care supply and other somewhat elusive indicators like 'socio
cultura1' indicators. Last, we may wonder if a general indicator like 'popula
tion health indicator' has even a sense. We don't pursue this debate, except to
insist on the limitation of the present approach.
of care demand will lead us to focus on the supply side of the Care system. In
connection with this general approach, we use projection models'? The
general method of such models consists of collecting variables which
describe the care system, to select strategic ones, to distinguish endogeneous
and exogeneous variables. Then projections are achieved, using coherence
relations. Past trends play a prominent role in such models. Then the quality
of projections measures more or less the degree of inertia of the health
system (at least in real terms but relative price estimates remain uncertain).
In order not to give an excessive weight to past trends, an alternative
method is to describe behaviour of Care system factors, assuming that the
care system can be viewed as a set of several markets.8 One can wonder if
such a method is relevant, even in the American institutional framework (for
instance, the estimation of usual consumption functions with net prices as
explanatory variables). Moreover the problem remains to aggregate the
microeconomic behaviour (which, according to recent micro economic analysis
could be much more complex than the usual microeconomic models) to get
macroeconomic relations. Lastly retrospective simulations, in the approach
we quote, lead to corrections on the parameters of the supply of health
services (for instance on the distribution between different kinds of physi
cians' services or physicians' specialities). So medium long term evolutions
must be exogeneously introduced, showing the limits of even sophisticated
microeconomic modelling.
Previous remarks and the French institutional framework (without markets
in the microeconomic sense) explain why our approach to the care system is
very close to that of projection models, yet linked with a macroeconomic
model.
Every quantitative model of the care system, even the simplest, must make
assumptions on the links between this system and the whole economy: in
particular concerning the wage rate and prices outside the care system and
often by default, selecting only thin relations.
The first benefit one can expect from a linkage between a macroeconomic
model and a health model is to get a better coherence among the economic
variables and the levels of activity and spending in the care system. The
coherence we can really attain depends on the quality of the linkage but also
on the features of the macroeconomic model.
Here, we propose a very simple linkage between a model which describes the
6 Y. Saillard
l.a. "Parallel" working l.b. Feedback at the wage 1.c. Full linkage
price block level
Figure 1. Possible linkages between Keynesian macroeconomic model and a health submodel.
Health expenditure growth and macroeconomic model 7
Here, we only illustrate the first two linkages, using an already existing simple
model (AGORA).9 In order to interpret some quantitative results which are
summarized in the next paragraph, we first proceed to a very brief presenta
tion of the macroeconomic model and of the health model in AGORA.
 Relative prices are those which make consistent the production struc
ture and the distribution of value added. They arise from hypotheses on the
stability of the wage share in value added, on the general wage rate and on
endogeneous labour productivity for each market subsector. 10
 Other hypotheses are made to simplify the macroeconomic model (at
least in its basic framework):
varying capacities of production are not introduced
exports are exogeneous and imports vary according to elasticities with
respect to final consumption expenditure or to domestic production
money creation is endogeneous without effect on real flows and the
interest rate does not affect firms' choices
nonmarket sectors (hence health activities) grow independently of each
other
the rate of growth of all benefits (except health benefits) is a fixed
multiple of the inflation rate. Yet, unemployment benefits vary with the
basic wage and the level of unemployment.
fiscal pressure is constant in all its components.
1 !
Physician and Nonphysician Manpower
hospital medical services per patient
 Hospital Care
mean stay duration
Parameters admissions in hospitals (rate per head)
=
describing
health Ambulatory Care { Physician Manpower
system Acts per physician
activities
Medical and Pharmaceutical Products
/
 Labour supply Public sector
\
Households
 Aggregate demand  Public financing of health  Final consumption (for health services)
expenditure  Primary income (income received
from health sector)
 Disposable income (after collective
financing)
Figure 2. Linkages between the health submodel and the macroeconomic model in AGORA.
ments for this consumption. So there are no more countereffects on the non
health final consumption side even without fast adjustment of the savings
rate.
The final effect in production is the same as in the previous case but with
a smaller effect on the consumption structure (that is to say weak substitution
between health and nonhealth final consumption). The Budget deficit dete
rioration is very close to what we get in the first case: collective financing
grows faster but activity, and therefore resources, also grows faster. The only
aggregate indicator, that fares worse in the second case is the balance of
trade, which is still almost the same as before.
A third simulation assumes that an acceleration in the private payments
due to faster growth in health expenditure can be damped by household
saving. So we establish an ex post macroeconomic link between the savings
rate and the growth of consumption, whose level is not determined by
traditional household behaviour but in accordance with a largely institutional
dynamic. It amounts to the same thing to assume that the savings rate is
defined, for households, out of health consumption and benefits. With our
simple macroeconomic model, such a hypothesis implies that the positive
multiplier effects can expand by all channels: increase in final health
consumption, increase in incomes received in market and nonmarket
sectors. A faster economic growth rate implies (according to our mOdel), a
better government budget balance but a worse trade balance.
A more complex use of the model is obtained when adding constraints on
the budget deficit. In such a case a limit is fixed for the level of budget deficit.
Health expenditure growth and macroeconomic model 11
V\
CareProduction
IStructure of health
expenditure financing
Structure of
care production
Taxation
_...J " '       . Labour
Cost

Price
Household
disposable income
rHousehold savingl
behaviour r
Household
final consumption
expenditure
Structure of household
final consumption
expenditure Profits and
wages shares
Health Non health in value added
•
consump tion expenditure
/
'\..Aggregate _ Production
Multiplier effects Demand
Internal
financing
profit rate
Figure 3. Effects of a variation in the level of health activity when feedbacks between the health
submodel and the macroeconomic model only work through the wageprice block.
ture affects labour cost: its structure or rate of growth. For us, changing the
structure of labour cost means increasing contributions to social security
(from employees and employers) to the prejudice of the net wage. Except for
effects of new labour cost structure on low wages,12 we can suppose that for
employers, the total labour cost is more important than its structure. On the
other hand, if the share of the net wage is reduced, it will have appreciable
effects on final consumption. Wellknown consumption functions do not take
into account the structure of disposable income.
On the whole, household disposable income varies in our simulations
according to three combined effects: a growth in wages in the health system
and in the market economy (due to more activity), a growth in health benefits
and a slowdown in net wages (new structure of labour costs). The final
variations depend on the constraint on the budget deficit (and on how it is
satisfied) and obviously on the way the macroeconomic model works
(especially the production functions in each sector and the division of
consumption between domestic products and imports).
So, health activities' development can generate cumulative effects, distor
tion in labour cost structure reinforcing the increase of nonmarket goods in
household consumption.
The second case (acceleration of labour costs) uses a new group of
interdependances which are linked to inflationary effects. A mediumterm
loop is added to the previous ones: it acts through operating surplus, level of
internal financing, rate of profit in market sectors. Then everything depends
(still assuming no productivity earnings coming from health activities) on the
way employers recover labour cost increases in price acceleration (with
constraints on domestic and international trade). A too tight constraint on
prices involves reducing investment projects and recessionary consequences.
A weaker constraint may imply more inflationary pressure so that the
problem reverts to consumers' choices and the financial sector.
Conclusion
Most methods used to describe health system dynamics, including ours, lead
one to focus on the autonomous development of these activities, as if they
were impervious to economic contingencies. It seems mainly due to statistical
or methodological constraints. To take into account all the links between the
health system and the whole economy would first imply an analysis of an
epidemiological type. Thus, we could display how socioeconomic variables
are responsible for health conditions. But such an analysis must integrate the
study of access to the health system and the evaluation of efficacity according
to several points of view. Then we face problems like relations between risk
factors, levels (individual and social) of analysis, indicators of health,
measures of effects of health conditions on productivity ... Attaining such
objectives is still a utopia.
Health expenditure growth and macroeconomic model 15
Notes
* I express my most profound thanks to Professor Mike Jerison (of NewYork and
CEPREMAP) for correcting my English. However the responsability for all other errors
that remain in the text are mine alone.
1. Group which is not homogeneous.
2. Remark which is not particular to our area and we could also make for household
equipment goods or some services. When made clear, it permits here to render an
account of social functions of medical profession: Levy et al. (1982).
3. Comptes de la Sante  Methodes et series 19501977 (1979).
4. Comptes de la Sante  Methodes et series 19501977 (1979).
5. For France, see Letourmy (1976).
6. Very often, in terms of variations of rates of mortality. It also shows the cumulative effect
of sociocultural and supply factors.
7. For France, see Couder et al. (1972).
8. For instance, Yett et al. (1979).
9. Peaucelle et al. (1983). AGORA is a mediumlong term model with a macroeconomic
model and two submodels on Education and on Health. At the beginning of the building
of this model (1975), there was no French minimodel. So the team had to build its own
macroeconomic model in addition to the two submodels. The main objective of this study
was to analyze the links between market and nonmarket sectors.
10. Depending on the level of production in each market subsector.
11. So it is very near to a projection model.
12. Due in France to the (less and less important) ceiling of contributions to social security.
References
Bacon, R. and Eltis, W. (1978), Britain's Economic Problem: Too Few Producers, The
MacMillan Press Ltd, 2nd edition.
Baumol, W. J. (1967), Macroeconomics of Unbalanced Growth: The Anatomy of Urban Crisis,
American Economic Review, Juni.
Comptes de la Sante (1979), Methodes et series 19501977, CREDOCINSEE, Ministere
de la Sante et de la Famille, Rapport CREDOC.
16 Y. Saillard
Couder, B., Sandier, S. and Tonnellier, F. (1972), Recherche de projections coherentes pour
des variables interdependantes, Consommation nO 3, JuilletSeptembre.
Letourmy, A. (1976), Sante, environnement et consommations medicales, Discussion autour
d'un modele, Revue Economique nO 3.
Levy, E., Bungener, M., Dumenil, G. and Fagnani F. (1982), La croissance des depenses de
sante, Economica.
Lorenzi, M., Pastre, O. and Toledano, 1. (1980), La crise du XXe siecle, Economica.
Peaucelle, I., Petit, P. and Saillard, Y. (1983), Depenses publiques: structure et evolution par
rapport au PIB. Les enseignements d'un modele macroeconomique, Revue d'Economie
Politique nO l.
Yen, D. E., Drabek, L., Intriligator, M. D. and Kimbell, L. J. (1979), A Forecasting and Policy
Simulation Model of the Health Care Sector. The HRRC Prototype Microeconometric
Model, Lexington Books.
PART ONE
Introduction
During the last decade a huge progress in econometric studies has been made
in developing models with socalled latent or unobservable variables. I In this
paper we apply the latent variable technique in the field of the health care
sector. Health, being one of the most predominant explanatory variables for
health indicators like the use of health care facilities or sickleave, is an
outstanding example of a latent variable. During the last years some studies
have appeared specifying health as a latent variable. 2 One of the profits of
these MIMICmodels 3 is that it yields an 'index of the health status'. This
MIMICHealth Status Index (MIMICHSI) is fully characterized by its causes
(health determining factors) and by its indicators (health indicators).
In this paper we will give a definition of the MIMICHSI and we will
interpret its numerical properties. One of the advantages of the MIMICHSI
over the classical health status indexes like e.g. a weighted sum of health care
utilization and sickleave, is that factors like moneyprice, timeprice and
supply of health care facilities (influencing health care utilization) or the level
of the sickpay and other insurance conditions (influencing sickleave) have
no longer any influence on the health status index.
where HIt is the tth Health Indicator, X k is the kth explanatory variable,
TCu is an unknown parameter indicating the influence of X k on HIt and Vt is
a disturbance term, representing all explanatory variables not explicitly
accounted for in the equation. 4 In order not to complicate the notation, we
will suppress the index n indicating the nth observation (individual, region,
etc.), n = 1,2, ... N.
One of the recommendations of the Conference on health status indexes
(Tucson, Arizona, 1972) was to make a difference between a health
indicator and a health index. The first term is used for rather specific
measurements such as those mentioned above. The latter term should be
used for derived measures that combine several health indicators (Berg,
1973, p. 253). Often the index is taken to be weighted sum of indicators:
L
where HSI stand for Health Status Index and wt indicates the weight that HIt
carries. We refer to model (1) as the Classical Model.
A major problem that may occur when using a health status index HSI like
(lb) is that factors X k which do not influence health, but do have an
influence on HIt> are influencing HSI via HIt. For instance, if in an economic
recession the sickleave rate (HIt) decreases as a result of fear for losing ones
job or as result of a lower sickpay, then the value of HSI will increase and
incorrectly be associated with a better health status. Another example is the
number of hospital beddays per capita, which is largely determined by
hospital capacity. Closing some 25 hospitals, as the Dutch government
intends to do, will lower the total bed capacity with about 12% and thereby
the number of hospital beddays (HIt) with approximately the same per
centage. This lower use of hospital facilities will incorrectly be associated
with a better health status. Building a new hospital or decreasing the
copayment rate will increase the health indicator 'hospital consumption' and
will in the same way incorrectly lower the value of HSL
Other problems which arise when using a health status index HSI like (1)
are: Which weights W t have to be assigned and who determines the weights?
(see e.g. Berg, 1973, pp. 254255).
The MIMICmodel
With the aid of present day computer facilities and statistical methods it is
The MIMIC health status index 21
possible to develop a health status index that is not bowed down by these
problems. In constructing a new health status index we start from the view
that health is a theoretical concept which is not directly measurable, that
there exists an ordering in different health statuses and that health is fully
characterized by its causes and by its indicators. Following the terminology
used in the statistical literature (Joreskog and Goldberger, 1975) we call the
index based on the above mentioned starting points the MIMICHealth
Status Index, where MIMIC stands for Multiple Indicator MultIple Causes.
Our approach is in the spirit of Robinson and Ferrara (1977), who stated:
'The approach we take seems new, for we model health as unobservable link
between observable causes and observable effects' (p. 139). In formula the
MIMICmodel is defined as follows:
K[ K2
HIl = L alktXlk +L awX2k + OtMIMICHSI + Ut (2a)
t = 1,2, ... L,
KJ KJ
where
Schematically the difference between Classical model (1) and the MIMIC
model (2) may be illustrated as in Figure 1.
What are the main advantages of the MIMICHSI as defined in (2) over
the HSI as defined in (I)?
First, Xlvariables which do not influence health have no longer any direct
influence on the value of the health status index.
Second, the total effect of :TCt of X2variables on HIt has been disentangled
in the direct effect au and the indirect effect ddJ2 (via MIMICHSI). As
an illustration we mention the results of Vande Yen and Vander Gaag
(1982) who found evidence that income has both a positive direct
(income)effect (au > 0) and a negative indirect effect via MIMICHSI
(dtfJ2 < 0) on medical consumption. The latter is the result of a positive
effect (fJ2 > 0) of income on health (better lifestyle, hygiene, attitude
to health etc.) and a negative effect (d t < 0) of health on medical
consumption.
Third, the weights W t need no longer to be predetermined e.g. by intuition
or by some Delphimethod, but all unknown parameters, including those
needed for calculating MIMICHSI, may be estimated using an appropriate
database and an appropriate estimation method.
t = 1,2, .. . L,
where Jr is decomposed in Jr l , Jrz and Jr3 according to Xl' X z and X3 and
with
(4a)
(4b)
(4c)
(4d)
(33k = Jr3ktn
. (4e)
C
24 W. P. M. M. van de Ven and E. M. Hooijmans
K, K,
where f32k and f33k are estimates of the unknown parameters {32k and {33k'
Substituting i32k and (33k from (4d) and (4e) in (5a) yields
(5b)
Applications
Province MIMICHSI
NoordBrabant 0.27
Utrecht 0.23
Limburg 0.22
Gelderland 0.21
Drente 0.19
Overijssel 0.18
ZuidHolland 0.07
NoordHolland 0.05
Groningen 0.02
Friesland 0
Zeeland 0
about 28 guilders (in half a year) on medicine prescribed by the GP etc. (see
Table2b).
The costs of increasing health can then be compared with the benefits in
terms of reduced health care costs.
The results in Table 2 should only be considered as an illustration of the
possible applications of the MIMICHSI because the data base on which
these results are based, was not primarily collected for the performed
MIMICanalysis. Nevertheless, it offered sufficient possibilities to illustrate
the use of the MIMICHSI at the individual level.
Table 2b. Total (reduced form) effect of the MIMICHSI on the use of some health care
facilities b •
Notes
* A previous version of this article has been presented at the European Meeting of the
Econometric Society, Dublin, September 1982.
1. E.g. Zellner (1970), Goldberger (1972), Joreskog and Goldberger (1975) and Aigner and
Goldberger (1977).
2. E.g. Robinson and Ferrara (1977), Lee (1979), Van de Ven and Van der Gaag (1982),
Wolfe and Van der Gaag(1981) and Hooijmans and Vande Ven (1982).
3. Multiple Indicator MultIple Causes (Joreskog and Goldberger, 1975).
4. Equation (la) may be considered as the reduced form of a more complicated, simulta
neous model. Here we will only be concerned with the parameters Jrk( and not with the
underlying structural parameters.
5. F= 1.8 C + 32.
6. Th = Y3;3 in their model should be substituted in the 1]! and 1]4equation; 1]1 and 1]4
should be substituted in the 1]6equation. The correlation of the reduced form disturbance
terms as a result of the latter substitution does not affect the discussion about the
interpretation of the MIMISHSI in the previous section.
7. Note: the MIMIChealth care model on which these results are based, is linear in the
logarithms of the variables. So we are dealing with the logarithm of the MIMICHSI.
References
Aigner, D. J. and Goldberger, A. S. (eds.) (1977), Latent Variables in SocioEconomic
Models, North Holland.
Berg, R. L. (1973), Health Status Indexes: Proceedings of a Conference Conducted by Health
Services Research, Tucson, Arizona, Chicago Hospital Research and Educational Trust.
Goldberger, A. S. (1972), 'Maximumlikelihood Estimation of Regressions Containing Un
observable Independent Variables', International Economic Review 13, 115.
Grossman, M. (1972), The Demand for Health: A Theoretical and Empirical Investigation,
Columbia University Press, New York.
Hooijmans, E. M. and van de Ven, W. P. M. M. (1982), 'Implementing a Health Status Index
in a Structural Health Care Model', In J. van der Gaag, B. Neenan and T. Tsukahara (eds.),
Economics of Health Care, Praeger, pp. 302330.
Joreskog, K. G. and Goldberger, A. S. (1975), 'Estimation of a Model with Multiple Indicators
and Multiple Causes of a Single Latent Variable', Journal of the American Statistical
Association 70,631639.
Lee, LungFei (1982), 'Health and Wage: A Simultaneous Equation Model with Multiple
Discrete Indicators', International Economic Review 23,199221.
Robinson, P. M. and Ferrara, M. C. (1977), 'The Estimation of a Model for an Unobservable
Variable with Endogenous Causes', In D. J. Aigner and A. S. Goldberger (eds.), Latent
Variables in Socioeconomic Models, NorthHolland Publ. Co., Amsterdam, pp. 131142.
Siegel, S. (1956), Nonparametric Statistics for the Behavioral Sciences, McGraw Hill.
Van de Ven, W. P. M. M. and van der Gaag, J. (1982), 'Health as an Unobservable: A
The MIMIC health status index 29
1. Introduction
The data used in this study were collected in 1975 by the Rochester
Community Child Health Survey. A 1% sample of families with children
under 18 years of age in Monroe County (where Rochester, New York, is
situated) were interviewed in 1975. Observations on 972 adults and 1191
children aged 118 within 514 households are used in this work. The data
are rich in health information, both in terms of health status and medical care
utilization, and in information regarding the households' attitudes towards
the seeking of professional care. To the data we have added provider availa
bility, by matching resident location to physician location and calculating the
travel distance to hospitals and health centers. For more detail on the data
see Wolfe (1980).
Health factors
As stated earlier, the data contain many health measures. For everyone in the
sample, we observe a subjective evaluation of health (HSTAT) given by the
respondent, the presence of a handicap (HCAP), whether the individual's
activity is limited in any way (LIM), and over the past year the number of
days ill (DAYS ILL), the number of days in bed (DAYS BED), and whether
or not the family member has been ill (ILL). Table 1 presents the relative
frequency distribution of these measures for adults.
We see that although almost half the sample is rated as having excellent
health, nearly 60% have been ill during the year. Persons report 4.49 days in
bed on average, a figure slightly below that reported on the Health Interview
Survey (HIS) of the civilian noninstitutionalized population of the United
States for 1980. More days ill than days in bed are reported, as expected.
(No comparable data are available in the HIS.) The percentage who report
handicaps is similar to those reporting 'with limitation in major activity'
(5.2%) in the HIS, while the percentage reporting LIM is similar to those
reporting 'with activity limitations' (8.6%) in the HIS Survey.3
While it is clear from these data that most adults are relatively healthy, it is
not an easy matter to decide which health variables best represent the health
status of the adults. Some variables contain overlapping information (DAYS
BED, DAYS ILL) while other variables seem to convey conflicting informa
tion (HSTAT, ILL).
For children, the picture is even more complicated. First we observe the
same health information as for adults. Table 2 presents the relative frequency
distribution of these health measures for children. Again, over half the
sample is reported to be in excellent health, yet over 75% have been ill
during the year. In terms of comparisons to national statistics, these children
have fewer days ill and days in bed than those reported in HIS. The handicap
percentages are closer: 2.0 for HIS, 2.1 for this sample; 3.8 for activity
limitation for HIS, 5.45 for this sample.
In Table 3 we display the incidence of seventeen specific health distortions
for children. Taken one by one, the data merely provide frequencies; e.g.,
nearly a quarter of the sample have allergies other than asthma or hay fever.
It is likely, however, that there is overlap; we can expect that the information
on breathing problems (24.22%) at least partly contains the same informa
tion as that on certain allergies.
The main purpose of this paper is to assess which aspects of health are
relevant to health care utilization. This task would be greatly simplified if the
large number of health measures available could be reduced to a smaller set
of independent variables. We construct such a set of health factors by
calculating the principal components of the correlation matrices of the health
measures. Tables 4 and 5 show the rotated factor matrices for the health
measures of Adults and of children.4 In Table 4 we see that the five variables
for adults reduce to two independent components: factor 1, with large
loadings on the handicap measures (HANDICAP), and factor 2, with large
2 3 4
Excellent Good Fair Poor HCAP LIM ILL DAYS ILL DAYS BED
Factor 1 Factor 2
(HANDICAP) (ACUTE)
" Only factors with an eigen value exceeding 1.00 are shown.
Table 5. Rotated factor matrix for children's health measures (Varimax Rotation)."
a There were eight factors with an eigen value exceeding 1.00. Only the first four are shown.
DAYS ILL, DAYS BED). The other two factors presented in Table 5 are
also easily interpretable. Factor 2 scores high on all measures of respiratory
diseases (RESPIRATORY), while factor 4 relates to diseases with a large
behavioral content (BERAVIOR). The four factors contain almost 35% of
the total variation of the 22 original health measures.
We will use the two factors obtained for adults and the four factors for
children in the analyses that follow, and interpret them as suggested above.
As pointed out in the Introduction, it is likely that some or all of the health
Table 6. Average values of attitude variables for families.
1. Guaranteed access 1.08 7. Reasonable fees 1.38 13. MD careful 1.16 18. Quality of care 1.43
2. Convenient hours 1.47 8. Fast appointments 1.35 14. MD concerned 1.34 19. Satisfied 1.13
3. Convenient location 1.70 9. Short office wait 1.56 15. MD listens 1.16 20. Relative care 1.50 ~
....
4. Recommended by friend 2.25 10. Friendliness of staff 1.47 16. MDtime 2.78 21. FindMD 2.38 §.
l:l
5. 24 hr emerg. care 1.23 11. Type of patients 2.81 17. MDinfo 2.54 ....
6. Comprehensive services 1.92 12. All see 1 MD 2.07 ~.
Values: 1 (very important) to Values: 1 (not enough) to Values: 1 (excellent) to
~
3 (not important) 3 (enough) 4 (poor) ~
l:l
;:s
l:l..
~
"I
~
C1:>
l:l..
§.
"
2
~
VJ
...)
38 J. van der Gaag and Barbara L. Wolfe
" There are six factors with eigen values exceeding 1.00. Only 3, those easily interpretable, are
shown.
HSTAT
Factor 1 Factor 2 (excellent = 1,
(HANDICAP) (ACUTE) poor = 4) HCAP LIM DAYS ILL DAYS BED ILL
AGE 0.011 (2.79) 0.003 (0.71) 0.012 (4.01) 0.002 (2.08) 0.005 (3.32) 0.009 (0.08) 0.018 (0.29) 0.003 (1.33)
FEMALE 0.179 (2.79) 0.156 (1.56) 0.004 (0.06) 0.028 (1.18) 0.088 (2.19) 2.29 (0.84) 1.51 (0.91) 0.111 (1.81)
EDUCATION 0.027 (1.83) 0.015 (1.06) 0.014 (1.24) 0.005 (1.47) 0.011 (1.92) 0.223 (0.56) 0.018 (0.08) 0.016 (1.83)
FAMSIZE 0.030 (1.12) 0.001 (0.03) 0.007 (0.35) 0.004 (0.71) 0.017 (1.63) 0.524 (0.73) 0.222 (0.51) 0.011 (0.67)
NONWHITE 0.077 (0.62) 0.082 (0.66) 0.209 (2.13) 0.030 (1.02) 0.013 (0.26) 2.47 (0.72) 0.949 (0.46) 0.113 (1.49) t;l
MARRIED 0.217 (1.41) 0.112 (0.73) 0.205 (1.69) 0.032 (0.88) 0.105 (1.70) 1.66 (0.39) 1.73 (0.68) 0.200 (2.13) §.
WORKFULL 0.231 (1.80) 0.079 (0.62) 0.019 (0.19) 0.041 (1.33) 0.110 (2.13) 1.94 (0.55) 2.11 (0.99) 0.080 (1.02) ~
WORKPART 0.128 (0.85) 0.037 (0.24) 0.005 (0.04) 0.031 (0.86) 0.048 (0.80) 2.01 (0.48) 2.81 (1.12) 0.013 (0.14) :::.
OCCUPATION 0.001 (0.37) 0.001 (0.33) 0.002 (1.12) 0.000 (0.39) 0.000 (0.33) 0.007 (0.10) 0.026 (0.62) 0.000 (0.03) ~
FAMINC 0.006 (1.10) 0.012 (2.05) 0.010 (2.12) 0.001 (0.97) 0.003 (1.34) 0.373 (2.34) 0.116 (1.21) 0.004 (1.15) 1}
MEDINC 0.025 (1.92) 0.016 (1.26) 0.021 (2.04) 0.006 (2.05) 0.009 (1.66) 0.349 (0.99) 0.343 (1.62) 0.001 (0.14) ~
~
CONSTANT 0.886 (2.84) 0.366 (1.18) 1.97 (8.03) 0.228 (3.08) 0.464 (3.72) 1.83 (0.21) 2.42 (0.47) 0.216 (1.14) ~
R2
:::..
0.051 0.016 0.103 0.016 0.063 0.017 0.012 0.027 '0'
....
~
N = 755. <"Il
(statistics in parentheses.
:::..
1')'
~
"
C")
~
~
VJ
1.0
40 J. van der Gaag and Barbara L. Wolfe
AGE 0.001 (0.07) 0.006 (0.81) 0.010 (1.19) 0.003 (0.29) 0.011 (2.22)
FEMALE 0.076 (1.14) 0.057 (1.00) 0.024 (0.38) 0.287 (4.42) 0.013 (0.35)
FAMINC 0.015 (2.09) 0.003 (0.44) 0.004 (0.51) 0.002 (0.34) 0.005 (1.35)
MEDINC 0.012 (0.86) 0.003 (0.27) 0.012 (0.89) 0.002 (0.15) 0.019 (2.46)
BIRTHORD 0.006 (0.13) 0.082 (2.19) 0.007 (0.18) 0.086 (2.01) 0.011 (0,48)
LMAGE 0.100 (0.55) 0.179 (1.15) 0.16 (0.09) 0.067 (0.38) 0,017 (0.18)
FAMSIZE 0.018 (0.51) 0.025 (0.82) 0.092 (2.67) 0.035 (1.01) 0.003 (0.15)
NONWHITE 0.060 (0.48) 0.029 (0.27) 0.143 (1.18) 0.150 (1.22) 0.041 (0.60)
MARRIED 0.525 (2.50) 0.040 (0.22) 0.065 (0.32) 0.190 (0.92) 0.003 (0.03) ~
FFULL 0.295 (2.07) 0.022 (0.18) 0.001 (0.01) 0.131 (0.94) 0.082 (1.07)
§.
$::>
FPART 0.208 (0.53) 0.633 (1.88) 0.461 (1.22) 0.444 (1.15) 0.109 (0.51) :::to
FOCC 0.005 (1.40) 0.000 (0.14) 0.003 (0.81) 0.007 (2.12) 0.002 (1.08) ~
MFULL 0.113 (0.71) 0.185 (1.34) 0.038 (0.25) 0.195 (1.25) 0.228 (2.64) ~
MPART 0.039 (0.25) 0.062 (0.46) 0.049 (0.32) 0.073 (0.47) 0.223 (2.61) ~
$::>
MOCC 0.002 (0.70) 0.001 (0.58) 0.000 (0.07) 0.002 (0.85) 0.003 (2.10) ~
MEDUC 0.046 (2.88) 0.027 (1.91) 0.033 (2.15) 0.026 (1.67) 0.005 (0.53) >:>..
CONSTANT 0.110 (0.40) 0.405 (1.72) 0.286 (1.08) 0.832 (3.10) 1.93 (13.05) ~
""'I
R2 0.036 0.026 0.038 0.046 0.059 ~
~
~
N=999. £5
"'
(statistics in parentheses. £5
~
+>
>'
42 1. van der Gaag and Barbara L. Wolfe
N = 514.
Estimating demand for medical care 43
In order not to complicate the analysis too much, we will, in the next
section, always include these taste factors in the demand analysis. Thus, we
should bear in mind that if a significant impact of one of the taste factors
(especially factor 3) on utilization is found, the coefficients for income, race,
and mother's education show only partial effects, 'holding taste constant.'
In general, the analysis of health care utilization is hampered by the fact
that no generally acceptable unidimensional health measure exists. As shown
above, principal component analyses or factor analytical techniques can
successfully be employed to reduce the sometimes large number of corre
lated measures into a smaller set of independent ones. But this approach is
quite mechanical and still does not yield one unidimensional measure.
It is probably fair to say that one unidimensional measure of health status,
representing all facets of health, and usable for a variety of purposes, simply
does not exist. However, in Section 5 we will show how a single, comprehen
sive health measure can be obtained, once the purpose of that measure is
specified. But first we will present an analysis of the demand for medical
care, including taste factors and using the health factors derived in the
previous section as proxy measures for health. We will provide comparisons
with results obtained when the longer list of the health measures is used and
when the health measures are completely omitted.
health care utilization, except for hospital outpatient care. All three measures
are important in explaining the total number of visits. Clearly the HSTA T
measure contains information that is not contained in the two more objective
health factors.
With respect to the other explanatory variables, we find only a few signifi
cant coefficients for HOSPERVS. Individuals with Medicaid insurance have
more visits to a hospital emergency room than privately insured individuals.
There are also slight racial differences. The overall explanatory power of this
equation is low, R2 = 0.070.
For HOSPOPVS we find significant racial differences: nonwhites seek
care more often in a hospital outpatient clinic than whites. Individuals with
Medicaid coverage also show more visits to an outpatient clinic. Being
married and being employed fulltime reduces the number of these visits. We
note, finally, that high scores on the 'Quality' and 'Convenience' (ATTITl,
ATTIT3) factors show a negative impact on HOSPOPVS. Apparently this
type of health service does not stand in high esteem for the quality conscious.
Our regression results explain 22% of the variation in HCORCLVS and
27% of the variation in OFFHMVS. Nonwhites with Medicaid coverage or
HMO insurance show a relatively high number of visits to health centers or
clinics. Whites from high income families and 'rich' neighborhoods, and with
private health insurance show more visits to the physician's private office.
These racial and incomerelated differences are less pronounced for the total
number of visits. The variables NONWHITE and F AMINC show no signifi
cant effect, but median income in the neighborhood is positively related to
overall utilization. Adults with Medicaid coverage or HMO insurance show a
higher number of visits than do the privately insured. The total number of
visits of adults scoring high on ATIITI ('Quality') is slightly below average,
but the other attitude factors show no effect.
We also find two familiar results: women show higher utilization rates than
men, and individuals living in large families show a lower number of visits
than members of small families. We finally note that our availability measures
do not show any significant impact on utilization. The measurement errors
inherent in the way we constructed these variables might have caused this
result. Or the differences in availability in the relatively small area from
which we obtained the data are simply so small that no effect on utilization
can be observed.
The above results appear to be somewhat sensitive to the use of alterna
tive variables 'to control for health.' Table 12 gievs some selected regression
results for the case where no health variables are included (column 1), only
the two health factors (column 2), only HSTAT (column 3) and, finally, in
column 4, the two health factors plus HSTAT (as in Table 11). The regres
sion coefficients of the variables not included in the table appear to be not
sensitive to the changes in health variables.
From Table 12 we learn that it does matter whether or not one controls
for differences in health status. For instance, no income effect and no signifi
46 J. van der Gaag and Barbara L. Wolfe
Table 12. Selected regression results for adults, using various health measures.
HOSPOPVS
FAMINC 0.009 (0.39) 0.005 (LlO) 0.008 (1.73)" 0.005 (1.09)
NONWHITE 0.157 (1.54) 0.200 (2.10)b 0.138 (1.36) 0.199 (2.08)b
ATIIT1 0.048 (1.14) 0.071 (1.80)" 0.063 (1.50) 0.072 (1.81)"
ATIIT2 0.016 (0.43) 0.021 (0.60) 0.010 (0.25) 0.021 (0.59)
ATIIT3 0.057 (1.82)" 0.051 (1.75)" 0.042 (1.34) 0.050 (1.69)"
HCORCLVS
FAMINC 0.008 (1.41) 0.007 (1.19) 0.007 (1.24) 0.006 (1.12)
NONWHITE 0.377 (2.90)b 0.339 (2.99)b 0.352 (2.72)b 0.364 (2.80)b
ATIlT! 0.010 (0.19) 0.004 (0.07) 0.008 (0.15) 0.009 (0.17)
ATIIT2 0.084 (1.78)" 0.085 (1.81)a 0.075 (1.60) 0.077 (1.64)"
ATIlT3 0.064 (1.63)" 0.063 (1.59) 0.048 (1.20) 0.049 (1.24)
OFFHMVS
FAMINC 0.011 (0.92) 0.023 (1.97)b 0.017 (1.44) 0.025 (2.22)b
NONWHITE 0.453 (1.58) 0.373 (1.41) 0.597 (2.19)b 0.498 (2.30)b
ATIlT 1 0.015 (0.12) 0.051 (0.46) 0.126 (1.10) 0.122 (1.12)
ATIIT2 0.055 (0.52) 0.059 (0.61) 0.003 (0.03) 0.019 (0.21)
ATIIT3 0.030 (0.34) 0.019 (0.23) 0.082 (0.97) 0.060 (0.75)
TOTAL
FAMINC 0.007 (0.44) 0.011 (0.82) 0.001 (0.10) 0.014 (1.08)
NONWHITE 0.012 (0.03) 0.132 (0.43) 0.203 (0.62) 0.040 (0.13)
ATIITI 0.049 (0.34) 0.115 (0.91) 0.201 (1.47) 0.204(1.66)"
ATIIT2 0.174 (1.37) 0.184 (1.65)" 0.097 (0.82) 0.129 (1.20)
ATIIT3 0.165 (1.55) 0.147 (1.58) 0.022 (0.22) 0.053 (0.59)
cant racial differences are measured for OFFHMVS if no health variables are
included, but both variables f>how a significant effect in column 4, when the
two health factors and HSTAT are added to the equation.6
The choice of the health variables is also relevant. If only HSTAT is
included, we find no significant racial differences for HOSPOPVS but a
significant income effect. If only the two health factors are included, we find
Estimating demand for medical care 47
MFULL 0.013 (0.36) 0.047 (0.76) 0.075 (0.91) 0.088 (0.59) 0.198 (1.14)
MPART 0.014 (0.43) 0.088 (1.51) 0.170 (2.17)" 0.403 (2.89)b 0.301 (1.85)"
HOSP 0.003 (1.45) 0.004 (1.02) 0.022 (1.70)'
GPPED 0.157 (0.91) 1.81 (0.03) 0.463 (0.55)
HMO
XHMO 0.000 (0.06) 0.001 (0.04)
FACTI 0,015 (1.19) 0.017 (0.82) 0.036 (1.27) 0.011 (0.21) 0.051 (0.85)
FACT2 0.036 (2.49)" 0.036 (1.45) 0.117 (3.50)" 0.187 (3.14)1> 0.305 (4.40)"
FACT3 0.063 (4.80)" 0.087 (3.88)b 0.106 (3.51)" 0.4 77 (8.80)" 0.738 (11.67)b
FACT4 0.006 (0.50) 0.052 (2.39)1> 0.Q38 (1.27) 0.039 (0.73) 0.045 (0.73)
HSTAT 0.012 (0.50) 0.119 (2.89)" 0.180 (3.23)b 0.558 (5.59)b 0.873 (7.50)b
R2 0.053 0.063 0.221 0.275 0.280
Table 14. Selected regression results for children, using alternative health measures.
HOSPOPVS
FEMALE 0.056 (1.26) 0.078 (1.74)" 0.054 (1.23) 0.072 (1.61)
FAMSIZE 0.008 (0.47) 0.018 (1.00) 0.010 (0.55) 0.017 (0.96)
NONWHITE 0.365 (4.12)b 0.378 (4.31)b 0.358 (4.07)b 0.370 (4.24)b
MPART 0.101 (1.71)" 0.096 (1.66)" 0.088 (1.51) 0.088 (1.51)
ATIITl 0.039 (1.96)b 0.024 (0.62) 0.019 (0.47) 0.013 (0.33)
ATIIT2 0.014 (0.42) 0.017 (0.50) 0.013 (0.39) 0.016 (0.47)
ATIlT3 0.002 (0.06) 0.002 (0.06) 0.009 (0.29) 0.004 (0.13)
HCORCLVS
FEMALE 0.057 (1.94)". 0.087 (1.46) 0.053 (0.89) 0.078 (1.31)
FAMSlZE 0.062 (2.58)b 0.045 (1.89)" 0.060 (2.51)b 0.046 (1.95)b
NONWHITE 0.242 (1.97)b 0.259 (2.15t 0.235 (1.94)b 0.251 (2.10)b
MPART 0.155 (1.94)b 0.156 (1.98)b 0.177 (2.24)b 0.170 (2.17)b
ATIlT 1 0.111 (2.04)b 0.083 (1.54) 0.079 (1.46) 0.065 (1.22)
ATIlT2 0.086 (1.84)" 0.083 (1.80)" 0.088 (1.90)" 0.085 (1.85)"
ATIlT3 0.082 (1.98)b 0.92 (2.27)b 0.071 (1.75)" 0.084 (2.07)b
OFFHMVS
FEMALE 0.044 (0.39) 0.079 (0.72) 0.033 (0.30) 0.050 (0.47)
FAMSlZE 0.194 (4.33)b 0.132 (3.08)b 0.187 (4.31)b 0.135 (3.20)b
NONWHITE 0.588 (2.62)b 0.557 (2.62)b 0.621 (2.85t 0.593 (2.82t
MPART 0.452 (3.02)b 0.444 (3.14)b 0.393 (2.71)b 0.403 (2.89t
ATIlT! 0.068 (0.66) 0.156 (1.61) 0.162 (1.62) 0.210 (2.18)b
ATIIT2 0.060 (0.68) 0.086 (1.02) 0.056 (0.64) 0.082 (0.99)
ATIIT3 0.286 (3.75)b 0.231 (3.19)b 0.317 (4.28)b 0.257 (3.59)b
To identify all parameters in the model, we fix the constant term of the
HOSPERVS equation to be equal to its value obtained from the regression
analyses in the previous section. HEALTH is dimensioned by setting its
impact on HCORCLVS equal to 1.0. Thus a one unit increase in health
results in one less visit to a health center.
At the bottom of Table 15, we see that a one unit increase of HEALTH*
decreases HSTAT by 2.9, the number of days ill by 1.6, and the number of
days in bed by 8.4. Furthermore, it reduces the probability that HCAP = 1
or LIM = 1 by 0.09 and 0.14, respectively.
Each column in the top part of Table 15 represents an equation in our
modeL With respect to HEALTH* we find a significant positive effect of
family income. This, of course, is a 'summary' of the findings presented in
Table 8. HEALTH* in fact can be viewed as a weighted sum of the health
indicators used. The HEALTH* equation indicates a negative effect of
FEMALE and a slight negative effect of NONWHITE on health. The utiliza
tion equations show that, 'holding health constant,' nonwhites have fewer
visits to private offices and hospital emergency rooms, and more visits to
health centers.
The other variables included in the health equation show no impact on
health. Given the analyses in the previous section, this result is not surprising.
As we have seen, various socioeconomic variables have positive, negative, or
no effect at all on health, depending on which health measure we use.
Consequently, when we obtain a unidimensional health measure, based on
the various health measures previously employed, the effect of the socio
economic variables can be expected to be small at best. A similar result is
obtained for children (see Table 16): FAMINC is the only significant variable
in the health equation, apart from the familiar age and sex effects.
HEALTH*, in the model for children, correlates highly with HSTAT, but
does not show much relationship with the four health factors. In fact, the
coefficient for the second factor (RESPIRATORy) has the 'wrong' sign
(bottom of Table 16). When interpreting both Tables 15 and 16 it is useful to
remember the scale we used for the children's and adult's health variables.
For both children and adults, we scaled the health measure so that a one unit
increase on health results in one less visit to a health care center. The corre
sponding reduction for hospital outpatient visits is also nearly one less visit
for adults, approximately onehalf for children, and for hospital emergency
room visits is slightly greater than onethird for adults but only approxi
mately onehundredth for children. These results seem plausible and give
some basis for our claim that the unobserved HEALTH* factor can serve as
a comprehensive unidimensional health index. However, the results for
private office visits (OFFHMVS) imply a very high response to a one unit
increase in HEALTH*  3.07 and 6.16 for children and adults respec
tively.I2 Given the average values of OFFHMVS (approximately 1.6 for both
children and adults), these results seem implausible. On the other hand, many
of the results look quite reasonable and are consistent with those based on
the regression analysis in the previous sections.
Table 15. Estimation results of the structural model of demand for medical care (adults).
HEALTH* HOSPERVS HOSPOPVS HCORCLVS OFFHMVS
MEDUC 0.034 (1.36) 0.004 (0.68) 0.012 (0.74) 0.041 (1.57) 0.171 (2.33)b
MFULL 0.214 (1.37) 0.020 (0.54) 0.154 (1.55) 0.269 (1.61) 0.642 (1.38)
MPART 0.167 (1.18) 0.017 (0.50) 0.D18 (0.20) 0.334 (2.20)b 0.066 (0.16)
FAMSIZE 0.D25 (0.66) 0.002 (0.16) 0.021 (0.84) 0.031 (0.74) 0.102 (0.90)
LMAGE 0.080 (1.04)
MARRIED 0.083 (1.34)
MEDINC 0.004 (0.63)
FAMINC 0.031 (1.90)a 0.002 (0.92) 0.016 (1.68)" 0.027 (1.59) 0.091 (1.87)"
Constant 4.391 (2.08) 0.111 () 2.245 (1.88) 3.711 (1.75) 12.28 (1.97)
HSTAT FACTI FACT2 FACT3 FACT4
3.81 (2.17) 0.003 (0.31) 0.D17 (1.58) 0.020 (1.58) 0.011 (1.08)
 = value fixed.
a Significant at 10% level.
b Significant at 5% level.
Estimating demand for medical care 55
6. Conclusion
It has been common practice to add one or more proxy variables for health
in demand equations for medical care 'to control for variation in health
status.' The choice of these proxy variables is almost always guided by the
availability of the data.
In this paper we show that this habit is not as innocent as it seems. Health
measures should obviously be included in demand equations for medical
care. But the choice of the variables representing health will have an impact
on the estimation results regarding various socioeconomic variables.
As shown in Section 3, health should be treated as endogenous. But doing
this does not solve the problem presented by the fact that not one of the
available health measures is by itself a sufficient proxy for health. A variety
of proxy measures must therefore be used.
In Section 5 we showed how these proxy measures can be used as
indicators for an unidimensional health measure. This measure, which is
unobservable, is introduced in a structural model for health care demand.
Thus HEALTH becomes a latent variable in a Multiple CausesMultiple
Indicators (MIMIC) model.
As indicators we use the health proxy measures and utilization of health
care. The latter can be used as an indicator for health once we adequately
control for income, insurance, availability, and taste differences. As causal
factors in a health production function, the socioeconomic variables that
were correlated with one or more of the health measures analyzed in Section
3 were included.
The results are encouraging, especially for children. The model yields
reasonable estimates, as compared to the unrestricted OLS regressions on
56 1. van der Gaag and Barbara L. Wolfe
utilization, and the latent variable HEALTH does have the impact one would
expect if it represents a true measure of health status.
Some caveats, however, should be mentioned. We did not solve the
question of how to choose among various healthproxy measures. We merely
pushed the problem one step back by including one latent variable,
HEALTH, in the demand equations and by stating that the proxy measures
were proportional to this overall measure. Thus the expost interpretation of
HEALTH is conditional upon the choice of the health indicators used and
the weight they get in the estimation process.
The estimation assumes normality of the disturbances. For adults we use
various 01 dummy variables as health indicators, which makes the nor
mality assumption less plausible. For children we transformed many discrete
healthproxy variables into a small set of continuous factor scores, which is
an important improvement over our earlier work. But the problem remains
with respect to the healthcare utilization data which are bounded from
below by zero, and usually take only a few discrete values. This problem
seems particularly severe with respect to private office visits. This variable
has a large concentration of zeros and (other than HOSPOPVS) correlates
strongly with a number of other variables. This might explain our implausible
results with respect to OFFHVS, both for adults and children.
Finally, we should mention that part of the model is constructed in an ad
hoc manner, with a little a priori knowledge and without a firm theoretical
base. The utilization module can easily be shown to be derived from a
general demand framework. But the 'production function of health' should be
viewed as a first attempt to show the impact of various socioeconomic
variables on a comprehensive measure of health status. The formulation fits
within Grossman's theory of the demand for health. But the analyses lack the
input of other disciplines, e.g., epidemiology. A further understanding of the
causal relationships between, say, income or family size or education and
health is needed to improve the specification of the health production
function in the model.
Notes
* The research reported in this paper was supported in part through funds granted to the
Institute for Research on Poverty by the Department of Health and Human Services
pursuant to the provisions of Economic Opportunity Act of 1964.
1. There are, of course, major exceptions: wellbaby care, immunizations, some gyneco
logical care, some screening tests, care during pregnancy.
2. Manning et al. (1981) takes a similar view.
3. The HIS numbers are for adults aged 1744 years, both sexes. See U.S. Department of
Health and Human Services (1981), p. 24.
4. In order to assess what information is contained in the variable HSTAT, we performed
the principal component analyses with and without this variable. In both cases the same
health factors were obtained. HSTAT correlates with two of these factors, HANDICAP
Estimating demand for medical care 57
References
Colle, A. and Grossman, M. (1978), 'Determinants of Pediatric Care Utilization', Journal of
Human Resources 13 (Supplement), 11558.
Davis, K. and Reynolds, R. (1976), 'The Impact of Medicare and Medicaid on Access to
Medical Care', In R. Rosett (ed.), The Role of Health Insurance in the Health Services
Sector, New York, National Bureau of Economic Research.
Edwards, L. and Grossman, M. (1980), 'Children's Health and the Family', In R. Scheffler
(ed.), Annual Series of Research in Health Economics, Vol. 2, Greenwich, Conn., JAI
Press.
58 1. van der Gaag and Barbara L. Wolfe
Goldman, F. and Grossman, M. (1978), 'The Demand for Pediatric Care: A Hedonic
Approach', Journal of Political Economy 86, 25980.
Grossman, M. (1972), 'On the Concept of Health Capital and the Demand for Health',
Journal of Political Economy 8012, 22355.
Hooymans, E. M. and Van de Ven, W. P. M. M. (1982), 'Implementing a Health Status Index
in a Structural Health Care Model', In van der Gaag, J., Neenan, B. and Tsukurhara, T.
(eds.), Economics of Health Care, New York, Praeger.
Hyman, J. (1971), 'Empirical Research on the Demand for Health Care', Inquiry 8, 6171.
Joreskog, K. G. and Sorbom, D. (1978), Estimation of Linear Structural Equation Systems by
Maximum Likelihood Methods, Chicago, International Educator Services.
Lee, L. F. (1979), 'Health and Wages: A Simultaneous Equation Model with Multiple Discrete
Indicators', Working Paper No. 79127, Department of Economics, University of
Minnesota.
Manning, W. G., Newhouse, J. P. and Ware, J. E. Jr. (1981), 'The Status of Health in Demand
Estimation: Beyond Excellent, Good, Fair and Poor', NBER Conference Paper 86,
January.
Newhouse, J. P. (1981), 'The Demand for Medical Care Services: A Retrospect and Prospect',
In J. van der Gaag and M. Perlman (eds.), Health, Economics, and Health Economics,
Amsterdam, North Holland.
Newhouse, J. P. and Phelps, C. E. (1974), 'Price and Income Elasticities for Medical Care
Services', In Mark Perlman (ed.), The Economics of Health and Medical Care, New York,
John Wiley and Sons.
Robinson, P. M. and Ferrara, M. C. (1977), 'The Estimation of a Model for an Unobservable
Variable with Endogenous Causes', In Aigner, D. J. and Goldberger, A. S. (eds.), Latent
Variables in Socioeconomic Models, Amsterdam, NorthHolland.
Roset!, R. and Huang, L. P. (1973), 'The Effect of Health Insurance on the Demand for
Health Care', Journal of Political Economy 81, 281305.
Shakotko, R. A. (1980), 'Dynamic Aspects of Children's Health, Intellectual Development,
and Family Economic Status', New York, NBER Working Paper No. 451.
U.S. Department of Health and Human Services. (1981), Current Estimates from the National
Health Interview Survey: United States, 1980, DHHS Publication No. (PHS) 821567,
December.
Van de Ven, W. P. M. M. and van der Gaag, J. (1982), 'Health as an Unobservable: A
MIMICModel for Health Care Demand', Journal of Health Economics 1, No.2, August.
Wolfe, B. L. (1980), 'Children's Utilization of Medical Care', Medical Care 18 (December):
pp. 11961207 (Institute for Research on Poverty Reprint No. 419).
Wolfe, B. and van der Gaag, J. (1981), 'A New Health Status Index for Children', In van der
Gaag, J. and Perlman, M. (eds.), Health, Economics, and Health Economics, Amsterdam,
NorthHolland.
An empirical model of the demand for
health care in Belgium
I. Introduction
Beneficiaries
the health insurance scheme for blue and white collar workers. The first
subgroup includes the active persons and the persons in their charge (e.g.
children). The second subgroup comprises widows, orphans, pensioners,
invalid persons and the persons in their charge (henceforth referred to as
WOPI). Note that the total number of beneficiaries has increased by
1368357 from 1966 to 1980. In 1980, the number of beneficiaries
amounted to 8491479.
The sources of revenue for the HIS consist of employers' and workers'
contributions on the one hand and government subsidies and taxes on the
other. More specifically, they are:
(a) The contributions of employers and workers that are calculated on total
wages; there is no special ceiling for the calculation of these contribu
tions. The employers' and workers' contribution rates are 3.75 and 1.8%
respectively.
(b) A government subsidy which is equal to 95% of the costs of the
treatment of the socalled social diseases 1 plus 27% of the normal
expenditures for medical treatment.2
(c) A special government allowance that finances the health insurance of the
unemployed. 3
(d) An excise taX on tobacco. 3
Table 1. Expenditures of the HIS for medical care and their share in national income,
196680.
Source: Various issues of Algemeen Verslag van het Rijksinstituut voor Ziekteen Invaliditeits
verzekering (RIZIV), Voornaamste financiiile en statistische uitkomsten van de verplichte
verzekering tegen ziekte en invaliditeit (Brussels, 1 July 1981) and own computations.
1974 +113.3
1975 +693.9
1976 2823.2
1977 +227.9
1978 4436.7
1979 5979.1
1980 2625.4
Cumulative deficit end of 1980 14829.3
specified. The last six categories are exogenous in the model. To give the
reader an idea of the relative importance of the different categories, we
present in Table 3 the expenditures of each category and their share in health
insurance expenditures for 1980. One can verify that the endogenous part of
the model captures 88% of the HIS expenditures. It can also be seen that the
expenditures for hospital beddays, special care and laboratory tests account
alone for 40% of total HIS expenditures. In Table 4 we present the share of
the two subgroups in the expenditures related to each above mentioned
category of medical care.
la 4943.6 3.94
Ib 5937.6 4.73
2a 601.9 0.48
2b 4300.3 3.43
3a 2490.3 1.98
3b 1727.9 1.38
4 1476.9 1.18
5a 9020.4 7.19
5b 15610.5 12.44
6 3069.5 2.45
7a 14028.3 11.18
7b 10043.1 8.01
8a 4948.6 3.95
8b 1095.7 0.87
9 22327.4 17.80
10 5474.6 4.36
11 2949.9 2.35
12 7566.1 6.03
13 807.5 0.64
14 147.9 0.11
15 108.7 0.09
16 1618.7 1.29
17 4891.8 3.90
1a 71.6 28.4
Ib 38.0 62.0
2a 95.9 4.1
2b 66.5 33.5
3a 89.5 10.5
3b 59.8 40.2
4 54.0 46.0
5a 48.4 51.6
5b 31.1 68.9
6 53.1 46.9
7a 64.9 35.1
7b 61.3 38.7
8a 58.1 41.9
8b 68.3 31.7
9 48.3 51.7
10 38.0 62.0
11 20.4 79.6
12 38.7 61.3
13 92.0 8.0
14 55.1 44.9
15 97.6 2.4
16 96.4 3.6
17 49.7 50.3
explained by the fact that parents frequently dislike transporting a sick child
to a physician's office because the transport itself may enhance the child's
illness. Furthermore examinations (in the physician's office) of sick children,
having caught an infectious disease, may be discouraged by the physician
himself in order to limit the transmittal of the disease.
Fourthly, medical care may be affected by climatic conditions; for instance
severe winters may boost the demand for consultations with physicians due
to the widespread occurrence of colds, influenza, etc. We have therefore
introduced average winter temperature (in centigrade) as a determinant in
the equations explaining consultations.
Let us now include the additional determinants in Equation (1) and
present the completely specified equations. For Category 1 we have:
In qfO = a 1+ (31 In( YIP), + Yll In( pgoI pgh), + Y12 In( pgoI pas), +
+ YJ31n(pgoIPpe),+ 01 In Rr + clin TEMPI +
+ ~I In CHILD t + CPlln OLD t (2)
In qrh = a 2 + .82 In(YIP), + Y21 In(pgh/ pgO)1 + Y2Zln(pgh/pos), +
+ Y231n(pgh/Ppe)t+ O2 In Rf + c2in TEMP, +
+ ~21n CHILD + CP21n OLD,. (3)
The superscripts go, gh, pe, and os refer to medical care categories la, Ib, 2a
and 2b respectively. Rg is the ratio of general practitioners to population
covered, TEMP indicates the average winter temeprature 6 in centigrade,
while CHILD and OLD are the ratios of the child population (of less than 15
years) and adult population (between 40 and 60) to the total population
(under 60), respectively. Two kinds of substitution effects are introduced in
the theoretical specifications: first, substitution between GP's and specialist's
care and, secondly, substitution between GP's consultation at the patient's
home and GP's consultation at the doctor's office.
The equations for Category 2 are the following:
In qfe = a 3 + .83ln( YIP), + Y31 In( ppe / pgO)t + Y32 In( ppc/ pgh)t +
+ 03ln Rre + c3ln TEMPt + ~31n CHILD t (4)
In q~S = a 4+ .841n(Y/P)t + Y41In(pos/pgO)t+ Y42In(pos/pgh), +
+ 04 In R~ + c4ln TEMPt + ~41n CHILD t +
(5)
In these equations, we also introduce substitution between GP's and special
ist's care. Rpe and RS denote the ratios of pediatricians and specialists
(excluding pediatricians) to the total amount of beneficiaries, respectively.
Given that pediatricians only deliver medical care to children, the variable
OLD has been omitted from Equation (4).
An empirical model of the demand for health care in Belgium 67
Basic specification
The medical care categories discussed here are different from those treated
before. They are not the immediate result of the patients' own demand.
Rather, they are the result of prescriptions by general practitioners and
specialists. Some types of medical care are also closely linked to hospital
stays. The determinants of these medical categories will therefore include
demand for medical care of Categories 1 and 2 and, wherever appropriate,
hospital stays (expressed in beddays). In other words, these determinants
reflect that one first needs to consult with doctors or to stay in a hospital in
order to be referred to more specialized forms of medical care.
Another determinant is likely to be the price of medical care relative to
the price of a substitute. The sign of the price coefficient is expected to be
negative. Indeed, there may be a negative price effect on medical care caused
by patients who consider the price as being too high and consequently refuse
or postpone a particular prescription. Note that in the equations for hospital
beddays, special care, laboratory tests and surgery, no relative price variable
will appear in view of the fact that over the sample period the coinsurance
rate has been zero for these types of medical care.
The basic specification can therefore be written as follows:
(8)
is the patient's price of qi; and P is the consumption price deflator. The
coefficients {3, y and 0 can be referred as to prescription elasticities whereas
e is the price elasticity. In the following subsections, we will treat the
different medical categories in somewhat greater detail and adjust the basic
specification (Equation (8» wherever necessary.
Prostheses (qpr)
We reason here that prostheses are prescribed by general practltloners
and/or specialists. Relative prices are also supposed to have an effect upon
qpr. The specified relationship is therefore
In qfr = a7 + .87 In q~ + Y7ln q~ + 07 In (ppr/ P)t. (9)
and
In q; = a 12 + f312ln qr + Y12 ln q~ + 0 12 In q~ + e12 t, (14)
where t refers to the time trend.
Drugs (qdr)
The consumption of qdr is explained first by qh in order to account for the
fact that hospital inpatients are important consumers of drugs. Secondly,
drugs are prescribed to outpatients by general practitioners and specialists,
whence qg and qS are introduced in the specification for qdr. Thirdly, a
relative price variable may influence the drug consumption.
The specification is therefore
In q~r = a lS + f31sln q~ + Ylsln q~ + olsln q~ + eJsIn(pdrI P)t. (17)
Physiotherapy (qk)
Physiotherapy is performed mainly upon prescription by doctors and as a
complementary service to hospital inpatient care. Therefore qh, qg and qS are
included as determinants in the qk equation. Furthermore, we may again
have a relative price effect from (PkIP). In addition, we will investigate
whether there is supplierinduced demand by introducing the ratio of
physiotherapists to beneficiaries (R k) in the equation. The specification is
now
In q~ = a 16 + f3161n q~ + Y16ln qf+ 016 ln q~ +
+ E 16 In(pkIP)t + ¢161n R~. (18)
qS and qg are included as determinants. We will also test whether the relative
price is an explanatory factor in this case. The specification is
In q~ = a 17 + f317ln q~ + Y17ln qr + 017 ln q~ + E17In(pnj P)t. (19)
Identities
E/ = L E:, (24)
i=1
The data
Data from 1966 to 1980 were used to estimate the equations. The data on
medical care are taken from the RIZIVstatistieken published by the Minis
terie van Sociale Voorzorg. The other data are taken from various issues of
the Statistisch laarboek van de Sociale Zekerheid and the Statistisch laarboek
van Belgii!.
Note that the prices of medical care that are directly available in the
statistics are those that are reimbursed by the HIS to the patient. In view of
the rather stable relationship between the patient's price and the reimbursed
costs for medical care over the sample period, we decided to use the latter as
proxies for the patient's price, without having to fear large approximation
errors. The estimation technique used was basically ordinary least squares.
The results for the active and the WOPI are presented in Tables Sa and b
respectively. The first general remark is about the crossprice elasticities. In
Table 5a. Estimation results of medical care Categories I to 3  subgroup of the active. :t.
~
l1:>
Category Dependent Explanatory variables Numbers of OW S.E.E. if> ~
variables observations
Constant WAGE/P pi/p pgh/p'" pgh/ PP' Pgo/po' Ri CHILD OLD TEMP ~:
.
1a qgl) 1.4242 0.7236 0.1959 0.8063 4.0460 14 2.21 0.014 0.977 g
(1.0178) (0.0677) (a) (0.3401) (0.9204) f}
q,h
.
Ib 4.1034 0.21 0.0659 0.0571 1.5045 3.3374 0.0176 13 1.83 0.028 0.621
.Q,
(2.2146) (a) (a) (a) (0.3263) (1.4980) (0.0119)
~
2a qPC 4.8268 0.7701 0.2315 0.3118 1.6395 15 1.97 0.022 0.962 l1:>
(1.9 I (9) (0.0829) (0.0990) (0.1668) (0.3887) f}
2b qm 2.4248 0.7157 0.0552 0.2210 1.0858 3.0554 15 2.43 0.017 0.989
(0.5730) (1.7696)
(1.3269) (0.2516) (0.0411) (0.1642)
~~
qpd ~
3a 8.7403 1.3163 0.6306 0.4059 15 1.67 0.053 0.956
(2.6980) (0.3385) (0.3831) (0.1094)
'0'
"'I
3b quJ 0.3375 0.1661 0.4481 15 1.61 0.015 0.916
(1.2520) (0.0386) (0.1776) ~
~
~
Notes: 1. All variables are expressed in natural logarithms. 2. The figures below the coefficients are standard errors. 3. The superscript j refers to the demand category 2
~
j estimated. 4. OW, S.E.E. and j[' are the DurbinWatson statistic, the standard error of regression and the cofficient of determination (corrected for degrees of
freedom) respectively. 5. The symbol a indicates that the coefficient value has been assigned.
s·
~
eRr
~.
.J
......
.J
tv
0
Q
g.
I:l
;:
Table 5b. Estimation results of medical care Categories 1 to 3  subgroup of the WOPI. ~
~
Notes: 1. See notes of Table 5a. 2. p is the coefficient of firstorder autocorrelation between the residuals. An iterative technique was used to calculate p; note that the
DWtest may be inconclusive.
An empirical model of the demand for health care in Belgium 73
(25)
where the subscript c indicates that the crossprice effects are compensated
price effects. We now define the crossprice elasticities
(26)
and
C nm = ( :q:)
p
p:
q
. c
(27)
C mn = ( :q:)
p c
~
q
(28)
(29)
medical care. Since it exceeds unity, it would mean that preservative dental
care is a lUxury good. Notice also that the income elasticities of the active
exceed those of the WOPI. According to us, the latter indicates that income
is less of a constraining factor for the WOPI due to the fact that medical care
is delivered at a very low average price to this subgroup. The same reason
can be used to explain the rather high unexplained variance of qgO of the
WOPI. Considerable stochastic movements in qgO are the cause of a low R2:
they are likely to arise more frequently here because income and prices only
constitute weak constraints on the WOPI's demand for medical care.
Concerning the price elasticities, we were unable to find significant
substitution effects between GP's home and office consultations! Substitution
effects were obtained between medical care offered by GP's and specialists.
For instance, in the case of qOS for both subgroups, the price effects indicate
that patients are more inclined to demand specialist's care if the GP's medical
care becomes relatively more expensive. The results also convey that patients
demand more pediatric care if the difference in prices of GP's and pediatric
care is narrowing. Notice that the price elasticities in the case of dental care
are also very significant.
Inspecting Table Sa, it is interesting to note that the demographic variables
CHILD and OLD has a statistically significant impact on consultations of the
active with GP's and specialists. The population group between 40 and 60
years old has definitely a higher medical care consumption pattern than the
younger generations; the marginal effect of OLD is at least twice as high as
that of CHILD. Notice that CHILD has a stronger influence on the demand
for pediatric care and the demand for home visits than on the demand for
office consultations and specialist's care. Our estimation results also reveal
that the demographic variables exert no special influence on dental care.
The results further show that a supplierinduced demand effect seems to
be present in the case of pediatric and dental care for the active. The effect
of climatic conditions is not to be neglected in the demand of both subgroups
for home visits by GPs.
The results in Tables 6a and b clearly show that there are strong linkages
among the various medical care categories. The variables qg and qS are
codeterminants in most equations. The variable qSU has a special impact upon
qhs and qa. Hospital stays qh has significant effects in the equations for ql, qdr,
qk and qn. Comparing the results of the WOPI with those of the active, we
notice especially that the prescription elasticities in the case of qpr and qho for
the WOPI exceed those for the active. This reflects a higher medical need by
the WOPI for these particular types of medical care. Furthermore, we see
that the elasticity of qSU on qhs is also higher for the WOPI, conveying their
need for longer hospital stays due to surgery.
The price elasticities indicate that prices seem to matter in the allocation
Table 6a. Estimation results of medical care Categories 4 to 11  subgroup of the active.
of medical care. They show that patients or their doctors, being their patients'
agents, have a demand for medical care that is price sensitive. The elasticities
are especially high in the case of prostheses and physiotherapy and indicate
that strong substitution takes place if these types of medical care get dearer
visavis other commodities.
The wage effects in the equations for the active explaining qhO and qhs
show that there is a tendency to shorten one's hospital stay as the oppor
tunity cost in terms of lost wages in the market increases. The influence of
supplierinduced demand could be detected in the case of surgery and
physiotherapy. Its impact is especially strong in the equations for qk. Note
that the improvement in medical technology, as captured by a time trend, has
an impact on surgery, special care and laboratory tests.
It has to be recognized that the equations describing the demand for drugs
are doing less well in terms of R 2 than other equations. The estimated price
elasticities are also not significant statistically. These results can be explained,
firstly, by the highly aggregated nature of the variable drugs which makes it
very difficult to find an ideal specification. Secondly, the reimbursement
system that was valid up to November 1980 did not give any incentive to
doctors and their patients to choose a costeffective treatment. Indeed,
patients paid a fixed amount per drug regardless of the drug's cost. Thus, the
relative absence of hard economic constraints on prescription behaviour
contributes to the explanation of both the rather stochastic nature of drug
consumption, leading to a low R 2, and the statistical insignificance of the
price coefficients.
V. Concluding remarks
The model explained above is estimated using macrodata for all beneficiaries
of the insurance scheme for workers. We have been able to show that real
income and prices of medical care matter in the allocation of medical care. In
some cases we also found effects on demand generated by suppliers of
medical care. Furthermore, the progress in medical technology, captured by a
time trend, creates extra demand for certain types of medical care. Climatic
conditions and the population structure also seem to matter in a number of
equations.
Earlier, in Table 1, we presented the share of medical care expenditures in
national income. This share increases from 1966 to 1980. Yet, it is stable
from 1978 to 1980. Note now that no major policy changes were introduced
during those particular years. We argue that the reason for this apparent cost
containment can be found in the evolution of some important explanatory
variables revealed by our model. According to us, the variables that are most
responsible for this stable share are real wages and real pensions. Indeed, the
growth in those real incomes during the years 1978 to 1980 has been
remarkably modest compared with the growth in the pre1978 period. This
78 G. Carrin and J. van Dael
lower growth is likely to have slowed down the growth in the demand for
medical care.
In general the explanatory power of the equations is rather high. We do
find high standard errors of estimate, however, in the case of qtrnt for the
active and of qtrnt, qk and qn for the WOPI. A major problem encountered
while building the model was that the sample size was rather small. There is
the risk that the estimates may not always be robust. In a number of cases,
explanatory variables are subject to too low a variance, contributing to the
statistical insignificance of certain estimates. Such insignificant coefficients
can be tolerated, however, when they have a theoretically correct sign. It is
also our hope that as more data information becomes available, more precise
estimates will be found.
It is granted that the present paper represents only a first step in the
construction of a comprehensive model of the health sector in Belgium. In
addition, one could model the markets for various categories of health
personnel and the determination of medical care prices. Studies of the
macroeconomic type like this one can also be complemented by models
using microdata on patients and suppliers of medical care. Finally, it is
evident that the present model can and will be used in forecasting health care
expenditures and in simulating alternative government policies.
Acknowledgement
Thanks are due to the Belgian Fonds voor Kollektief Fundamenteel Onder
zoek for financial support. Comments and suggestions by A. P. Barten, H.
Deleeck, L. Delesie, D. Deli(~ge, J. Kesenne, W. Nonneman, N. Van Belle and
an anonymous referee on previous drafts of this paper are gratefully
acknowledged. We also thank participants of the Public Economics Seminar
of Namur and the 9th International Conference of Applied Econometrics
(Budapest, March 1982) for useful comments. Remaining errors are the
authors' sole responsibility.
Notes
Introduction
(1)
Because the morbidity data may have only been available for a larger spatial
unit than the study area, it is usually necessary to modify the f'ik values in the
base period by a small common correction, defined as llb to ensure that, for
each specialty k, the sum of the actual observed admissions T~k from zones i
to hospitals j equals the sum of expected demand from zones i, that is,
where we now define the adjusted relative demand (1 + llk) f'ik as W;k.
For the simplest class of demand/supply interaction model, estimates of
the expected or relative demand W;k for care are all that is required for input.
However, such a model implies that a uniform relative increase of capacity in
each hospital would induce a correspondingly uniform relative increase in
admissions from each zone, irrespective of the existing relative levels of
unsatisfied demand in such zones. In order to be able to support a demand/
supply interaction model with elastic demand, it is necessary to determine the
need for care, rather than just the relative demand. The procedure used is to
seek the zone i which has the maximum value (1 + r) of the ratio (~j T~k)/
W;k of observed admissions to expected admissions based on the morbidity
of its population. This zone may in a general sense be defined as that most
'accessible' to care for the given specialty. It is then hypothesised that if new
supply were provided such that all zones were made as accessible as this best
zone, each of them would also exhibit this same ratio of observed admissions
to expected admissions. Furthermore, if strong queuing exists in the system
for the specialty, even this best zone will not have produced its full latent de
82 J. R. Roy and M. Anderson
mand for care, and one may perform sensitivity analyses with trial values e of
a further uniform increase in demand. Thus, from the above, the need for
care W;k may be defined as:
~k = (1 + e)(1 + r) W;k (3)
representing a uniform increase of the expected demand values W;k' Note
that, if the incidence of day clinics varies greatly in different parts of the
study region, the values of W;k would need to be correspondingly corrected
before being used as inputs to the hospital demand/supply model. In
addition, if the availability of primary care varied markedly within the region,
projections from the hospital model would only be expected to be reliable if
the proposed hospital supply policies were accompanied by policies to en
courage general practitioners to locate in deprived areas.
Having evaluated the need for care in our excess demand system, the next
task is to determine the expected pattern of hospital admissions under
alternative scenarios of hospital construction, partial or complete closures,
changes in hospital casemix and changes in the throughput characteristics
for each hospital for each specialty. Several approaches are possible to
develop such a model, and HOSPIM allows the user the choice of the
approach perceived as most relevant for his region.
Firstly, the entropy maximisation procedure, introduced into spatial
modelling by Wilson (1967), is extended to handle elastic demand (Roy et
al., 1987). Such a model is simpler to calibrate than the analogous utility
model presented by Mayhew and Leonardi (1982). Defining the microstates
of the system as either distinct potential patients not being admitted for
treatment or as distinct patients being admitted to any of the particular
hospitals in the study area, the task is to maximise the number of microstates
to determine the most probable macrostate I;j' that is the number of patients
admitted from i to j, and the unsatisfied demand (W;  l: j I;j) in i called Vi'
Note that, as such an analysis must be performed for each specialty in tum,
the specialty index k is omitted for convenience. The number of microstates
NT is given as the number of ways (combinations) that each distinct potential
patient of group W; in the base period may be assigned either to the
unsatisfied group (~  l:j I;J or to admitted groups I;j in each hospital j,
yielding
(4)
(5)
are enforced, where Dj is the nominal case load capacity of hospital j for the
given specialty (computed from data on available beds, occupancy rate and
average length of stay). If a travel time (distance) constraint for each hospital
is also added, with tij the travel time (distance) between zone i and hospital j,
the entropy maximisation problem is given as
(7)
(8)
where A = exp  (Pj lij) and Bj = exp  (AJ, which may be obtained
iteratively after substitution of (8) into (5). The objective (7) may then be
modified, converting the original Lagrange multipliers Pj to parameters, via
the following Legendre transform (Lesse, 1982)
(9)
where ~ is the base period travel time (2:i T~ tiJ to hospital j. The Legendre
transform switches the role (i.e. as knowns or unknowns) between certain
nominated Lagrange multipliers and the righthand sides of the correspond
ing constraints, whilst leaving unchanged the optimal solution of the original
problem. In our case, the original problem, with average trip times Cj to each
hospital given and the corresponding multipliers Pj as unknowns, is trans
formed to a problem where Pj is given and the corresponding average trip
times ~ to each hospital are unknowns, which can then be evaluated anew in
the forecast period as the spatial distribution of demand and supply changes.
Thus, as trip times here represent the internal behavioural variables of the
model, the transformed problem can evaluate the behavioural response of
84 J. R. Roy and M. Anderson
wher~ Dj represent the forecast (or planned) hospital case load capacities
and W; the forecast need values from (3). The solution for (10) is of a similar
form to (8), except that tV; is replaced by tv;, and Dj by Dj in (5) to reflect
new hospital capacities.
An alternative formulation of (8) can be made in terms of the unknown
unsatisfied demand levels Vi' yielding
(11)
(12)
(13)
lem (7). It is seen that this approach has decomposed the interaction effect
into a time (distance) component and a bias component. The user is free to
choose an impedance parameter f3j for each of the new hospitals from
knowledge of those for 'comparable' hospitals. Also, average bias terms can
be obtained for the new hospitals consistent with any errors in patient
admissions from origin zones in the existing system (Roy, 1987).
Finally, in some hospital systems experiencing overall excess demand,
certain outmoded or less accessible facilities may well be running below their
nominal capacity, while the rest are, as expected, subject to significant
queuing. For this case, the problem in (7) should be revised by interpreting
Dj as the actual utilised capacity in each hospital j, which the admissions
have to satisfy via (5) at the base period. This will be less than the nominal
capacity for such hospitals. Then, in forecasting, the problem in (10) is
modified to give
(15)
(16)
(17)
which ensures that both {3j and Aj are treated as parameters. Inclusion of the
latter (AJ enables the implied hospital attractiveness terms Bf to be projected
for forecasting in cases where the facilities are running below capacity in the
base period and possibly also in the forecast period. The solution of (15)
comes out as
(18)
(20)
Returning to the entropy derivation in (7), removing the entropy term for
unsatisfied demand and adding constraints on all demand being satisfied
from each zone as
(21)
(22)
where fij is as before and Bj is exp  (A. j ), with A. j the Lagrange multiplier on
(5), where Dj must here represent the actual utilised capacity, which in the
excess supply case will usually be less than that available. Also, in this excess
supply case, it is advisable to obtain .f;j as exp  (f3i tij ), with f3i defined for
each zone, thus avoiding the potential multicollinearity with the hospital
attractiveness parameters Bj if f3j were used. Now if the Legendre transform
of (17) is applied to the excess supply problem, the model structure for
forecasting becomes
(23)
where 0; is the new satisfied demand obtained from (1) and Ej has the same
interpretation as in the analogous Equation (18). Note that, whereas in the
excess demand case most facilities run at their nominal capacity for a given
specialty with the possible exception of a few less favoured facilities running
below capacity, the converse is true for the excess supply situation. In this
case, the inequality hospital capacity constraints in (16) will usually be
Reconciling spatial demand/supply imbalances in acute care 89
inactive, leading to Ej = 1 in most cases. Note that, the model (23) may be
regarded as a hybrid form between the classical originconstrained and
doublyconstrained entropy models of Wilson (1967).
All the above entropy models for excess supply can be converted to
information theory models in a similar way as shown for the excess demand
case in (13) and (14). For the case of no new facilities being constructed, the
interaction term tij simply equals the observed admissions T~. For situations
where new facilities are being considered, hj in (14) is now given as
(24)
Finally, the reader may wonder why the excess demand model of (8) is not
interchangeable with the excess supply model of (22) by merely reversing the
demand indices i and the supply indices j. The reason is that in the original
definition of microstates in (4), the potential patients within each zone i were
treated as distinct, but not the beds within each hospital j. This is because (i)
the 'urgency' of different patients for admission can vary widely within a
region and (ii) even if some particular wards within a hospital are regarded as
'better' than others, the patient cannot usually choose a particular ward, but
must accept what is immediately available within the category covered by his
hospital insurance. In other words, the simple inequality constraints (18) on
hospital capacity in the excess supply case differ fundamentally from the
elastic demand 'constraints' on potential patients in the excess demand case,
introduced via the entropy on unsatisfied demand (~  ~ j Tij) in (8). The
latter terms ensure that, so long as overall excess demand exists in the
system, all zones will have some unsatisfied demand.
Sequence of operation
After reading in the data common to all specialties (e.g. travel time, hospital
catchment specifications, etc.), HOSPIM computes the demand and case load
capacities and calibrates the spatial models for each specialty in tum, using
base period data. A redundancy index is computed, representing the explana
tory power of the model. In addition, the root mean square error between the
modelled admissions from each zone and those observed is evaluated. For
each specialty,the calibration is followed by an update phase, which usually
represents the 'no build' situation in the forecast time period. The update
phase incorporates changes which are not potentially controllable by the
health care planner, such as population and morbidity changes, planned
facilities coming into service by the forecast period, planned case mix
changes, planned closures, expected changes to average lengths of stay and
occupancy rates, etc. The results of the update phase are then regarded as
the new base case, adjusted to the forecast pe~od. Finally, one enters the key
90 J. R. Roy and M. Anderson
module of HOSPIM, the interactive forecast phase. In this phase, the planner
can interactively try alternative hospital development policies, including
integrated casemix changes, closing wards, hospital extensions, hospital
closures and new hospital construction. The forecast phase can also be used
as a vehicle for sensitivity analyses about uncertain data, such as population
changes, morbidity changes, reductions in average lengths of stay, etc.
For each run through the forecast phase, HOSPIM evaluates the changes
to the key results and indicators with respect to the update phase. Some of
these results are for each specialty in turn, and others are with respect to the
whole system. Successive changes can be accumulated, after which the system
may be returned to the update phase condition using a single command. This
feature is useful for comparing the independent effects of individual changes
with the coupled effects of a set of integrated changes. In this regard,
HOSPIM warns the user if a set of integrated casemix changes exceeds the
bed capacity of the hospital, as well as indicating the likely presence of idle
beds. It is the planner's responsibility to estimate if the planned casemix
changes, etc., can be handled by the expected availability of medical, nursing
and support staff.
Output of model
Although the main results of HOSPIM are the expected admissions 1!jk
between each zone i and hospital j for each specialty k, the planner is more
directly interested in the various performance indicators which can be
evaluated from these flows and other available data. At this stage, HOSPIM
does not compute costbased indicators  these can always be added when
better cost data by specialty becomes available.
Several indicators are populationbased, and require the determination of
standardised populations Pik by zone and specialty, which are obtained as
follows
(25)
where P is the total study area population, 0ik the relative leakages to outside
hospitals defined in (1), and W;k the expected relative demand for care
computed earlier from age, sex and (optionally) other characteristics of the
population. A typical populationbased indicator nik is given as
(26)
(27)
(28)
used to access hospitals out of each zone i. This is clearly influenced by the
level of spatial aggregation adopted in aggregating the origins of patients.
A level of service index Sik for each zone i and specialty k may be defined
as the relative level of satisfied demand, given as
(29)
Similarly, the equity of the level of service can be given as the coefficient of
variation of the above indices over all zones i. Zones with poor accessibility
to facilities will usually be expected to have relatively low values of the above
satisfied demand index.
Another important criterion is zonal selfsufficiency for health care, which
indicates the proportion Pik of a zone's patients in specialty k admitted to
hospitals within their own zone, given as
(30)
where ji is the subset of hospitals (if any) located in zone i. This criterion is
again dependent on the level of zonal aggregation.
In the excess supply situation, hospital capacity indicators are very
92 1. R. Roy and M. Anderson
(31)
where Djk is the case load capacity for specialty k in hospital j. In addition,
HOSPIM indicates the proportion of a hospital's case loads coming from its
own zone, as well as (optionally) the proportions coming from up to two
userdefined rings of zones around the hospitals' zone. Also, it allows the
user to (optionally) aggregate the zone system used for analysis up to the
health region or health district level, with certain results then being output at
this more policyrelevant level. Finally, the relative attractiveness 1jk of a
hospital j for specialty k can be defined as
(32)
where Bjk is the destination factor from (8) etc. This represents the hospitals
'pulling power' above and beyond its size and average accessibility to
patients.
In addition to the above specialtyspecific indicators, global indicators
representing the average performance of the system as a whole can also be
obtained. Also, if HOSPIM is later combined with operating cost and
maintenance models within a unified data base, costbased indicators can be
readily obtained for the integrated patient/activity/building system, which
would provide the planner with a more comprehensive decision aid package.
An illustrative example
c
2. BULLA
8. WHITTLESEA
c
eAustin
combined, rather than a separate analysis for each specialty. As not all of the
hospitals possessed all specialties, this aggregation inevitably produced some
distortions.
The calculation of need for care was based on metropolitanwide morbid
ity rates by age and sex being applied to the age/sex distribution of the
population in each municipality, corrected for leakages to hospitals outside
the eleven included. These demand values were further enhanced to bring the
ratio of observed admissions to expected admissions up to be 'best' value for
all zones. Then, sensitivity analysis was made with respect to a further
increase of this demand, with uniformly good results being found for
increases in the range of 15 to 35% (see (3».
In the calibration of the model (8) on 1983/84 data, the R.M.S. error of
modelled vs. observed admissions from each zone was 17.5%, reducing to
15% for the 1986/7 forecasts when model (8) was used with the corrected
interaction terms k given from (14). On the other hand, the simplified model
in (12) produced a 26% R.M.S. error on the 1983/84 calibration data,
reducing to 16% for the 1986/7 forecasts when model (12) was used with
the corrected interaction terms k given from (24). Thus, the elastic demand
feature in (8) greatly increases the ability to calibrate compared with the
conventional approach in (12), whilst having little effect on its forecasting
accuracy. Because of the major change to the hospital insurance system
94 1. R. Roy and M. Anderson
References
JOEL W.HAY
Senior Research Fellow, The Hoover Institution, Stanford University,
Stanford, California 94305, U.S.A.
1. Introduction
graduates work shorter hours and earn less than their U.S. counterparts.
They found net income to vary with the size of practice, and with the volume
of radiology and laboratory services provided by the practice.
Perhaps a less obvious flaw in the modelling approach used by both Sloan
and Hadley is the treatment of specialty income as an exogenous explanatory
variable in the specialty choice equation. If specialty income and specialty
choice are jointly determined by the medical students' background and
environment, then an estimation technique that takes this simultaneity into
account would seem to be more appropriate.
is being used both to estimate this class of models and also to provide
consistent initial starting values for FIML parameter estimation. However,
even the Mill's ratio method requires a probit maximum likelihood estimator
in the first stage, and is thus computationally expensive. Olsen (1980) has
shown that by dropping the underlying assumption of jointnormal distur
bances, a linear probability modification of the Mill's ratio method allows
standard OLS regression techniques to be used to correct for nonrandom
sample bias in this class of models.
Hay (1980) proposes a logistic modification of the Mill's ratio method.
While this modification has only a minor computational advantage over the
Mill's ratio method in the case of a dichotomous sampling rule, it may be the
only computationally feasible selectivity bias correction in the case of a
polytomous sampling rule, where the number of nonrandom samples in the
data is large. The Mill's ratio method can be extended theoretically to include
polytomous models of this type, but in practice it requires computationally
burdensome estimation of an nvariate firststage probit model. While the
Olsen linear probability modification is quite tractable in the dichotomous
sampling rule case, there are no clear theoretical extensions of it to the
polytomous case.
Moreover, as will be shown below, the logistic modification may be
embedded in a wellknown stochastic utility model of choice  the condi
tional logit model  allowing the statistician to examine hypotheses regard
ing the behavioral structure of the nonrandom sample selection process. This
is particularly useful in cases where selectivity bias is due to selfselection, as
in the union participationunion wages analysis, for example.
(1)
Ii = 1 iff Vi < Zi Y
(2)
where Xi' Zi are row vectors of exogenous observable variables, and /3, yare
column vectors of unknown coefficients. Equation (2) can be rewritten as:
E(uiuj ) = o~ i = j
= 0 otherwise
E(ViVJ = o~ i= j
= 0 otherwise
E(uiVj) = pOuov i = j
= 0 otherwise
100 1. W.Hay
(4)
We assume additionally,
Var (Yi IX;, Vi < ZiY)  P 2 au2 Var (V; IVi < ZiY) +
av
(5)
Up to this point, the analysis parallels that of the Mill's ratio method and
the Olsen modification. Defining
where fO and F(.) are respectively the pdf and cdf of Vi' we may write:
E(Yi IX;, Vi < ZiY) = XJ3 + AiD. (6)
Given an estimate of Ai, Ai it is possible to estimate the parameters {3, 15 in
(6) using standard regression techniques. It is not necessary to make any
additional assumptions about the distribution of U i • When Vi is distributed as
standard normal,
Ai = f(Ziy)[F(ZiY)rl
is termed the Mill's ratio, where fO and F(.) are the standard normal pdf and
cdf respectively. When Vi is uniformly distributed on the [0,11 interval, Ai =
(1  ZiY) is the Olsen (1980) linear probability modification of the Mill's
ratio. 3
Physicians' specialty choice and specialty income 101
r
V.  a )
exp
( 'r
f(v;) ~ r[ 1 + exp (_ ~ a V; )
(7)
F(v;) = [ 1 + exp (  V; ~a ) r 1
= 1/2 [ 1 + tanh ( V; ;r a ) ]
k =
fco
Zi)' v·e
I
Vi dv·
(1  e
I
Vi)2
I
F(Z;y)
(8)
= ([log(1 + e ziY )](1 + e ZiY ) + Z;ye ZiY }.
Defining the logistic probability F(Z;y) == Pi' (8) may be rewritten as:
(9)
= liP; [Pi 10g(P;) + (1 P;) 10g(1  Pi)]
=H/Pi •
(10)
102 J. W. Hay
(11)
ZiY 2 Vi
Ai= f
vie dVi
<Xl
(1 + eVi )2'
(12)
= Jt2
6
+C r
nI
(If [ (I ZiYI 2 + 21 ZiyI + 2)enIZiYI _
n n
2],
n
Iij= 1 iff
v}j = en  eij < W;jOj  W;) 0) = Zijy
v~ = ei2  eij < WijOj  W;202 = Z~y
(17)
This is to say,
The set of points for which Vij = Vij' for any j, j' is of probability measure
zero.
It is assumed that,
=0 i ~ i',
E(uij vt) = pp~2av i = i',
= 0 otherwise
1
aEJJ
= au (1 P,~
JJ
L Pj),
v
1
1 1
Var(YijIXij , Vij < ZijY) = auii L COV(Vij IVij < ZijY) L Pj+
v v
1
+ au}l  pi L Pj),
v
(18)
where
1
OJ ;: a~12 Pl· L
11
V
At = F(ZijYrl J J 4r
00
Z~r
00···
J Z;jr
00 vtf(vij) (}vij
( 1+ I
h 1
e Z;;y ) J
~ X,{J! +
6
,,' lf(Jl)j loge Pij ) + (20)
t 10g(Pjk )
( 1 ~'p, }i1Y+',
)
kI J
k¢J
where (l.
1
= pja,~I2.
II
Physicians' specialty choice and specialty income 107
(21)
and this equation is estimated using OLS on the sample for which
nonzero data is observed on Yij.
5. An empirical example
Auxiliary Statistics.
LOG LIKELIHOOD 1891.0 1889.0 1890.0
SUM OF SQUARED RESIDUALS 4397.0 4234.0 4390.0
DEGREES OF FREEDOM 4226.0 4225.0 4225.0
PERCENT CORRECTLY PREDICTED 62.89 62.99 62.89
Goodness of Fit Statistics.
LIKELIHOOD RATIO INDEX 0.1885 0.1894 0.1888
LIKELIHOOD RATIO STATISTIC 878.6 882.7 879.6
Notes
* Research support from the National Center for Health Services Research (NCHSR)
(DASH Grant # 03150) is gratefully acknowledged. I would like to thank the Center for
Health Policy Research of the American Medical Association (AMA) for providing
physician survey data. The' views expressed here are not necessarily those of either
NCHSR or the AMA.
1. See, for example, Lee (1976); Olsen (1975); Heckman (1974, 1977); Lee and Trost
(1978); Nelson (1977); Nelson and Olson (1978).
2. While a stochastic utility framework may also be developed in the case where Vi is
assumed to be jointnormal, it should be pointed out that this type of behavioral motiva
tion is not applicable in the case where Vi is uniform. As Domencich and McFadden
(1975) point out, a uniformly distributed random variable cannot be written as the
difference between two identically independently distributed random variables.
3. In this cale Ai includes a location shift due to the fact that E(Vi) = 112 # o.
112 J. W. Hay
References
PAUL 1. GERTLER
The RAND Corporation, 1700 Main Street, P.O. Box 2138, Santa Monica,
California 90406, U.S.A.
I. Introduction
pay' patients. Homes reduce 'private pay' demand and operating costs by
lowering qUality. These quality differences can be quite large. Specifically, in
our sample, we find that homes who recieve high Medicaid 'plus' factors
provide hundreds of thousands of dollars less in goods and services than
homes who receive average Medicaid 'plus' factors, ceteris paribus.
CON attempts to control nursing home expenditures by limiting the
supply of beds with capacity constraints and entry barriers. CON policy
makers are forced to trade off containing the size of the industry (and
therefore total Medicaid payments) against quality and access. Specifically,
we find that the capacity constraints and the reduced competition from the
entry barriers lead to lower quality and fewer Medicaid patients receiving
care.
The theoretical analysis is based on Gertler (1985a, b).l The model relies
on the notion that there are quantity and quality aspects to production, and
both quantity and quality are endogenous. 2 Specifically, nursing homes
produce a series of commodities, such as medical care, room and board, and
social activities. The quality of nursing home care is the utility patients derive
from consuming this package. Nursing home output, then, is characterized by
the total number of patients and average qUality. Proprietary homes are
assumed to maximize profits by choosing quality and 'private pay' price
(which determines the mix of 'private pay' and Medicaid patients) subject to
the CON capacity constraint. 'Not for profit' homes are specified as utility
maximizers subject to CON capacity and break even constraints.
The empirical analysis estimates a linear specification of the reduced form,
but is complicated because quality is not directly observed. Instead of
directly estimating the reduced form quality equation, we use the theory of
production to specify a model in which quality is a latent variable, but the
parameters of the reduced form quality equation are identified. The model
has a Multiple Indicator Multiple Cause (MIMIC) interpretation, where the
indicators are conditional input demands which have been filtered to remove
all sources of variation except quality and random disturbances. 3
We discuss the regulatory environment in Section II, the theoretical model
in Section III, the empirical specification in Section IV, the data in section V,
the empirical results in Section VI, and summarizes in Section VII.
The nursing home industry has expanded from approximately S190 million
in 1950 to over S18 billion in 1980.4 Most of the expansion took place after
1966, the year in which the Medicaid program began. As of 1980, the public
share of nursing home expenditures was over 65%. Health care regulators
have the task of trying to control this expansion, while simultaneously
providing the poor with access to nursing home care and promoting a high
Subsidies, quality, and regulation in the nursing home industry 119
standard of quality. The major forms of government intervention are the
Medicaid patient subsidy program and the CON cost containment program.
Medicaid is an entitlement program established under the Social Security
Act to provide the poor with a minimum floor of health services. Through
direct subsidies, the Medicaid program makes health care available to
individuals who otherwise could not afford it. It is jointly financed by State
and Federal governments, but administered on a State basis. The Medicaid
program reimburses nursing homes a set fee for the care of Medicaid
patients. Typically, States pay nursing homes via 'costs plus' reimbursement,
although a few States have opted for a prospectively set flat reimbursement
rate.
The CON cost containment program was passed into law in response to
the rapid growth of the health care industry during the late 1960's and early
1970's. It requires that, in order to expand an existing nursing home or build
a new one, the government must certify that the proposed facility is indeed
+'needed'. Effectively, CON limits the existing capacity of existing nursing
homes and new entry into the market. s It was thought that the expansion
could be contained by limiting the available supply of nursing home beds.
In essence, government regulation has turned nursing homes into price
discriminators. The Medicaid program creates a second market for nursing
home care, and CON restricts supply so that there is excess Medicaid patient
demand. Therefore, Medicaid demand is perfectly elastic at the Medicaid
reimbursement rate. The excess Medicaid demand hypothesis is supported
empirically in Scanlon (1980).6 Hence, nursing homes compete with each
other for 'private pay' patients knowing that they can always admit Medicaid
patients at the Medicaid reimbursement rate if they have excess capacity.
Nursing homes face 'private pay' and Medicaid demand. 'Private pay'
demand is given by X(P, Q), where X is the number of 'private pay' patients,
P is the price charged 'private pay' patients, and Q is average quality. 'Private
pay' demand is increasing in Q, and decreasing in P. In contrast, Medicaid
demand is perfectly elastic at the Medicaid reimbursement rate.
CON imposes a capacity constraint on each home. Since there is excess
Medicaid demand, the constraint is binding. It is specified as
X+M=X, (1)
where M is the number of Medicaid patients and X is the CON allowed
capacity.
Nursing homes are required to provide all patients with the same level of
120 P. J. Gertler
1. Equilibrium
Homes choose 'private pay' price and quality so as to maxmuze profits
subject to the CON capacity constraint. The profit function is
Tl = PX(P, Q) + RIX  X(P, Q») ce.K, Q). (3)
The first order conditions are
PXp+X=RXp (4)
PXQ=(XfX)CQ +RXQ. (5)
'Private pay' price is chosen in (4) such that marginal 'private pay' revenue
equals the opportunity cost of Medicaid revenue, and quality is chosen in (5)
such that marginal 'private pay' revenue equals the marginal cost of quality
plus the opportunity cost of Medicaid revenue. Since the cost of caring for
Medicaid patients is recovered via Medicaid reimbursement, the marginal
cost of quality is weighted by the proportion of 'private pay' patients.
c. CON entry policy. Entry increases competition for 'private pay' patients,
and thus reduces 'private pay' demand and marginal revenue to each nursing
home. In response, homes lower quality. The fall in quality decreases both
'private pay' marginal revenue and R. Hence, 'private pay' price rises (falls) if
marginal 'private pay' revenue rises less (more) than the rise in R.
1. Objectives
Unlike proprietary nursing homes, the objectives of 'not for profit' homes are
not well defined. Economists typically model 'not for profit' firms as utility
maximizers.8 The arguments of a 'not for profit' firm's utility function are
debatable and depend upon the institutional setting. In the nursing home
industry, where the religious institution dominates, we assume that not for
profit homes are basically altruistic in nature. Therefore, we expect these
homes to be concerned with quality, and with providing care to the poor.·
Hence, the 'not for profit' nursing home's utility function is assumed to be an
increasing function of quality and the number of Medicaid patients. 'Not for
profit' nursing homes are assumed to choose 'private pay' price and quality
so as to maximize utility subject to the CON capacity constraint and subject
to a break even constraint.9
2. Equilibrium
Let G( Q, M) be the nursing home's utility function, and A the Lagrange
multiplier on the break even constraint. Then, assuming that the Kuhn
Tucker conditions are satisfied at an interior solution, the first order condi
tions are
PXp +X(lIA)GMXp =RXp . (6)
PXQ+ (lIA)(GQ  GMXQ) = (x&)CQ + RXQ • (7)
'Private pay' price is chosen in (6) such that marginal 'private pay' revenue
122 P. 1. Gertler
plus the marginal utility of Medicaid patients equals the opportunity cost of
Medicaid revenue, and quality is chosen in (7) such that the marginal 'private
pay' revenue plus the marginal utility of quality equals the marginal cost of
quality and marginal opportunity cost of Medicaid revenue.
J K
where the Zy's include exogenous demand variables, the Medicaid 'plus'
factor, and the market concentration level implied by CON policy. The W;/s
are exogenous supply variables: input prices, capital stock, and the CON
Subsidies, quality, and regulation in the nursing home industry 123
capacity constraint. Finally, the f3;/s are unknown parameters, and the e;/s
are independently distributed random variables with zero mean.
A binding CON capacity constraint implies several restrictions on the
reduced form. If the constraint is binding, then X; and M; sum to X; for all i,
implying that Equations (9) and (10) sum to X; for all i. Let W;I be X;. Then,
a binding CON capacity constraint implies
f3Ij = f32j for j = 0,1, .. . ,J,J + 2, ... , K (13)
f31,J+ I = 1  f32,J+ I' (14)
and a singular covariance matrix. These addingup restrictions imply that
Equation (10) is redundant. Therefore, we exclude Equation (10) from the
estimation. The remaining equations have the same right hand side variables
and no cross equation restrictions, which allows us to efficiently estimate (9)
and (11) by least squares, and then recover the coefficients in (10) from the
restrictions in (13) and (14).
where the a's are unknown coefficients and the vi/s are random distur
bances. Total quality is just average quality times the total number of patients
(i.e. X;Q;). Therefore, from (12), the reduced form total qualitY equation is
J K
cp; = f340X; + L f34jZ;jX; + L f34,J+ k W;kX; + e4 ;X;, (16)
jI kl
Substitution of (16) into (15) gives the reduced form input demand equations
K J
where
(18)
YOt = all + acpti34o, (19)
Yjt = arpti34j for j~ O. (20)
V. Data
The data are constructed from New York State's 1980 survey of Long Term
Care Facilities. The sample consists of 455 nursing homes chosen from 798
possible cases. Excluded were government homes, hospital attached homes,
and nonreporting homes. In the sample are 288 proprietary and 167 'not for
profit' nursing homes. Unless otherwise specified, the variables are daily
averages, with the unit of observation being the nursing home. Descriptive
statistics are presented in Table 1.
In the reduced form, the dependent variables are 'private pay' price, the
number of Medicaid patients, and the number of private pay patients. The
dependent variables in the input demand equations are 100's nursing labor
hours, 1OO's of other labor hours, and a supplies quantity index. I I
The exogenous supply variables are the input prices and capital stock. The
input prices are the hourly nursing wage rate, the hourly other labor wage
126 P. J. Gertler
rate, and a supplies price index. Since, the majority of capital owned by a
nursing home is the facility itself, capital stock is measured as total area of
the facility in 1 OO,OOO's of square feet.
The exogenous demand variables are the per capita income of the people
living in the nursing home's market area, the population over age 65 in the
nursing home's market area, the proportion of 'private pay' patients in the
nursing home whose last residence before entering the nursing home was
located in the same county in which the nursing home is located, and an
index of health status of the patients in the nursing home. Income is
Subsidies, quality, and regulation in the nursing home industry 127
VI. Results
The models developed in Section lIT were estimated separately for the
proprietary and 'not for profit' samples. The estimated proprietary reduced
form equations are reported in Table 6, the 'not for profit' reduced form is
reported in Table 7, and the structural input demand equations are reported
in Tables 8 and 9. The estimates are quite reasonable. As expected,
hypothesis tests, reported in section A, uniformity reject pooling the propri
etary and 'not for profit' samples, and accept the restrictions that identify the
reduced form quality equation. In addition, the coefficients on the policy
variables are consistant with theory, and the signs on other independent
variables such as own price in the input demand equations are generally as
128 P.l. Gertler
one would expect. The policy results are discussed in detail in section B, and
are summarized in Tables 2 through 5.
Table 2. Estimated coefficients and tstatistics on medicaid and CON policy variables.
Policy variable
We begin by testing the hypothesis that the proprietary and 'not for profit'
samples can be pooled in the reduced form 'private pay' price and 'private
pay' patients equations. The test statistics are 4.03 and 2.66, respectively, and
are distributed F(12, 431). The corresponding critical value at the 0.05
significance level is 1.77. Consequently, we reject pooling.
Before testing to see if the proprietary and 'not for profit' reduced form
quality and structural input demand equations can be pooled, we test to see if
restrictions discussed in the identification section are valid. The test statistics
are 20.08 for the proprietary sample and 16.68 for the 'not for profit'
sample, and are distributed X2(27). The corresponding critical value at the
0.05 significance level is 40.11. Consequently, we accept the hypothesis that
the restrictions are valid. Under the assumption that the restrictions are valid,
we test the hypothesis that the proprietary and 'not for profit' samples can be
pooled for the reduced form quality and structural input demand equations.
That test statistic is 99.74, and is also distributed X2 (27). Consequently, we
reject pooling.
Subsidies, quality, and regulation in the nursing home industry 129
B. Comparative statics
Proprietary homes
Total # of patients 124.61 124.61 149.61 199.61
+20.06% +60.19% +60.19%
Quality index 2.27 2.25 2.23 2.21
0.85% 1.70% 2.54%
Total expenditures' $2009 $1992 $1975 $1,958
$17 $34 $51
Private pay price $59.89 $60.19 $60.49 $60.79
+0.50% +1.00% +1.50%
Patient mix ratio 3.90 4.42 4.87 5.26
+13.38% +24.86% +34.79%
Private pay patients 25.44 27.59 29.74 31.89
+8.45% +16.90% +25.35%
Medicaid patients 99.17 122.00 144.82 167.65
+15.76% +46.03% +40.85%
Proprietary homes
Concentration index 0.12 0.13 0.17 0.22
+8.33% +41.67% +83.33%
Quality index 2.27 2.26 2.21 2.15
0.52% 2.60% 5.22%
Total expenditures" $2009 $1999 $1957 $1904
$10 $52 $105
Private pay price $59.89 $59.70 $58.96 $58.03
0.31% 1.55% 3.11%
Patient mix ratio 3.90 3.96 4.24 4.63
+1.62% +8.73% +18.82%
Private pay patients 25.44 25.11 23.78 22.12
1.31% 6.53% 13.06%
Medicaid patients 99.17 99.50 100.83 102.49
+0.33% +1.67% +3.34%
'Not for profit' homes
Concentration index 0.11 0.12 0.16 0.21
+9.09% +45.45% +90.91%
Quality index 2.77 2.77 2.78 2.78
+0.05% +0.25% +0.51%
Total expenditures a $2432 $2433 $2438 $2444
+$1 +$6 +$12
Private pay price $59.86 $60.05 $60.82 $61.79
+0.32% +1.61% +3.22%
Patient mix ratio 4.80 4.90 5.30 5.90
+1.94% +10.39% +22.74%
Private pay patients 21.40 21.06 19.70 18.01
1.58% 7.92% 15.84%
Medicaid patients 102.81 103.15 104.51 106.20
+0.33% +1.65% +3.28%
equilibria for small, medium, and large differences in the Medicaid 'plus'
factor. Table 3 reports predictions of the endogenous variables at the mean
of the data, and for nursing homes whose Medicaid 'plus' factor is five
dollars, ten dollars, and fifteen dollars greater than the mean, respectively.
These differences are observed within the sample, as the standard deviation
of the Medicaid 'plus' factor is 15.21. Deviations from the mean are also
reported.
Consider the differences in quality for proprietary homes. The 'mean'
home provides 2.27 units of quality per patient, the 'mean plus five' home
produces 2.22 units, the 'mean plus ten' home produces 2.17 units, and the
'mean plus fifteen' home produces 2.12 units. To get a more interpretable
measure of the magnitude of the quality differences, we translate these values
132 P. J. Gertler
Table 7. 'Not for profit' reduced form estimated coefficients and Tstatistics in parentheses.
provides 61 thousand dollars less in goods and services than the 'mean'
home, the 'mean plus ten' home provides 123 thousand dollars less, and the
'mean plus fifteen' home provides 184 thousand dollars less.
Not surprisingly, there is almost no difference in the proprietary 'private
pay' prices. On the other hand, the 'not for profit' prices show substantial
differences; The 'mean plus five' home charges 'private pay' patients 5% less
than the 'mean' home, the 'mean plus ten' home charges 10% less, and the
'mean plus fifteen' home charges 15% less.
Finally, consider the differences in patient mix as represented by the ratio
of the number of Medicaid to 'private pay' patients. For proprietary homes,
the 'mean plus five' home has a ratio 13% higher than the mean home, the
'mean plus ten' home has a ratio 30% higher, and the 'mean plus fifteen'
home has a ratio 50% higher. For 'not for profit' homes, the 'mean plus five'
home has a ratio 8% higher than the 'mean' home, the 'mean plus ten' home
has a ratio 16% higher, and the 'mean plus fifteen' home has a ratio 26%
higher.
134 P. J. Gertler
at the mean plus 25 patients, at the mean plus 50 patients, and at the mean
plus 75 patients. Again these differences are observed in the sample, as the
standard deviation of the 'total number of patients' variable is 94.04.
Deviations from the mean are also reported beneath the predicted value.
Consider the differences in quality. The 'mean' proprietary home provides
2.27 units of quality per patient, the 'mean plus 25' home provides 2.25
units, the 'mean plus 50' home provides 2.23 units, and the 'mean plus 75'
home provides 2.21 units. Translating these differences into total expendi
tures on goods and services, we find that the 'mean plus 25' home provides
17 thousand dollars less in goods and services than the 'mean' home, the
'mean plus 50' home provides 34 thousand dollars less, and the 'mean plus
75' home provides 51 thousand dollars less. Turning to 'not for profit'
homes, the 'mean' home provides 2.77 units of quality, the 'mean plus 25'
home provides 2.74 units, the 'mean plus 50' home provides 2.70 units, and
the 'mean plus 75' provides 2.62 units. In terms of expenditures on goods
and services the 'mean plus 25' home provides 30 thousand dollars less than
the 'mean' home, the 'mean plus 50' home provides 60 thousand dollars less,
and the 'mean plus 75' home provides 90 thousand dollars less.
Homes of all sizes charge the same 'private pay' prices, but larger homes
have higher Medicaid to 'private pay' patient ratios. In the proprietary case,
the 'mean plus 25' home has a ratio 13% higher than the 'mean' home, the
'mean plus 50' home has a ratio 25% higher, and the 'mean plus 75' home
has a ratio 35% higher. In the 'not for profit' case, the 'mean plus 25' home
has a ratio 27% higher than the 'mean' home, the 'mean plus 50' home has a
ratio 55% higher, and the 'mean plus 75' home has a ratio 84% higher.
136 P. 1. Gertler
3. Entry
Entry into the market reduces the market's concentration. Hence, CON entry
policy can be analyzed by looking at the coefficients on the market con
centration index. Theory predicts that an increase in concentration causes
nursing homes to lower quality, and has an indeterminate effect on 'private
pay' price and patient mix. The empirical results in the proprietary model are
consistent with theory. In the proprietary reduced form, the coefficients on
the market concentration index are negative and significantly different from
zero in the quality, 'private pay' price, and 'private pay' patients equations,
and is positive and significant in the proprietary Medicaid patients equation.
Hence, entry causes proprietary homes to raise quality, raise 'private pay'
prices, and to substitute Medicaid for 'private pay' patients.
On the other hand, all the coefficients on the concentration index in the
'not for profit' reduced form equations are positive and not significantly
different from zero. Alternatively, it may be the case that potential 'not for
profit' patients do not view proprietary and 'not for profit' nursing homes
care as close substitutes. This would suggest that proprietary and 'not for
profit' nursing homes do not compete in the same markets even if they are
located geographically close to one another. Hence, a joint index of concen
tration heavily weighted towards the proprietary homes market would be
relevant for proprietary endogenous variables, but not for the 'not for profit'
homes.
Again, we gauge the magnitude of these effects for proprietary homes by
comparing predictions of the endogenous variables at the mean of the data to
predictions for small, medium, and large differences in market concentration.
Table 5 reports the predictions at the mean concentration level, at the mean
plus 0.01, at the plus 0.05, and at the mean plus 0.1. As usual, these
differences are observed within the sample. The 'mean plus 0.01' home
provides 10 thousand dolalrs less in goods and services than the 'mean'
home, the 'mean plus 0.05' home provides 52 thousand dollars less, and the
'mean plus 0.1' home provides 105 thousand dollars less. On the other hand,
increases in concentration appear to have little effect on 'private pay' price
and patient mix.
VII. Conclusions
Health care regulators are concerned with assuring a high standard of quality,
providing the poor with access to care, and controlling the expansion of the
industry. With these goals in mind, this paper has analyzed effects of
Medicaid and CON policy on proprietary and 'not for profit' nursing home
behavior.
We find that Medicaid policy makers are faced with a qualityaccess trade
off. Specifically, nursing homes whose Medicaid reimbursement rates include
higher Medicaid 'plus' factors care for more Medicaid patients, but provide
Subsidies, quality, and regulation in the nursing home industry 137
lower quality. The empirical work suggests that homes with high 'plus' factors
provide substantially lower quality than homes with mean 'plus' factors.
These quality differences when translated into nursing home expenditures on
goods and services provided patients are observed to be in the hundreds of
thousands of dollars. Further, high 'plus' factor homes have Medicaid to
'private pay' patient ratios possibly 50% higher than mean 'plus' factor
homes.
We also find that CON policy makers must trade off containing the size of
the industry (and therefore total Medicaid payments) against quality and
access. Specifically, we find that the capacity constraints and the reduced
competition from the entry barriers lead to lower quality and fewer Medicaid
patients receiving care.
Finally, we observe differences in proprietary and 'not for profit' re
sponses to regulatory policy. Both types of homes adjust quality and patient
mix in the same direction, but in different magnitudes. 'Not for profit' homes
tend to have larger quality responses. On the other hand, they have opposite
'private pay' price responses.
Notes
cost minimizers. This case is discussed in Frech and Ginsburg (1980). The altruism
assumption suggests cost minimization.
9. The supplies quantity index is calculated as the total expenditures on supplies divided by
an index of the price of supplies. The supplies price index is a weighted average of the
prices of the commodities that constitute nursing home supplies. These commodities are
drugs, other medical supplies, food, energy, and other supplies. The prices of these
commodities are national price indices in 1977 dollars, and are reported in the Depart
ment of Health and Human Services report, HEAL TH CARE FINANCING TRENDS,
1980. The weights are the proportion of a home's total supplies expenditures accounted
for by the particular commodity.
10. The health status index is computed from disability scores assigned patients. The
disability level of each patient in eight functional areas is reported in the survey. The
disability level in each functional area takes on one of three values: self care, partial care,
and total care. The functional areas are walking, transferring, wheeling, eating, toileting,
bathing, dressing, and breathing. Walking, transferring, and wheeling were treated as
mutually exclusive categories. Each home reports the number of patients in each cell,
where a cell is defined by functional area and disability level. The cells are then
aggregated by disability level. After that, the self care aggregate is multiplied by zero, the
partial help aggregate by seven, and the total help aggregate by fourteen. The scores are
then summed and divided by the total number of patients times 100. The result is an
index of the average illhealth of the patients in a facility. This index is used for the
purposes of quality control (see Ullmann, 1983).
11. For each home, the market population is computed as a weighted sum of the number of
persons over age 65 in each county, using the home's proportions of private patients from
the counties as weights. Similarly, the per capita income of the population in a home's
market area is computed as the weighted sum of the counties' per capita incomes, using
the same weights.
12. Since each county is treated as a separate market, and each home participates in several
county markets, the concentration of a home's private pay patient market depends upon
its degree of participation in the various county markets. Therefore, the concentration
level of a home's private pay patient market is a weighted sum of the county market
concentration levels, using the home's proportion of private pay patients from each
county as weights. The concentration of a county private pay patient market is computed
using a Herfindahl index, which is based on each home's share (proportion) of a county's
private pay patients. Specifically, it is the sum of squared shares (see Scherer (1980) p.
58). Entry reduces the value ofthis index.
References
Aigner, D., Hsaio, C., Kapteyn, A. and Wansbeek, T. (1984), 'Latent Variable Models in
Econometrics', In Z. Griliches and M. Intriligator (eds.), the Handbook of Econometrics,
Vol. 2, NorthHolland, Amsterdam, pp. 13211393.
Bishop, C. E. (1980), 'Nursing Home Cost Studies and Reimbursement Issues', Health Care
Financing Review 2, 4764.
Chamberlain, G. (1982), 'Multivariate Models for Panel Data', Journal of Econometrics 18,
546.
Frech, H. E. and Ginsburg, P. B. (1980), The Cost of Nursing Home Care in the United
States: Government Financing, Ownership, and Efficiency', In J. Van der Gaag and M.
Perlman (eds.) Health, Economics, and Health Economics, North Holland, Amsterdam,
pp.678l.
Gertler, P. J. (1985a), 'Subsidies, Quality, and Regulation in the Nursing Home Industry',
Subsidies, quality, and regulation in the nursing home industry 139
BERNARD LARCHER
Ecole Superieure de Commerce et d'Administration,
des Entreprises de Bourgogne Franche Comte, 29 rue sambin 21000 Dijon 80.72.12.40
1. Introduction
only varying at each period. The process of the counts does not stay
stationary and the results of Poisson's processes theory in continuous time
cannot be applied. The variance of the process is changing at each period of
time. The process of the error remains unknown; but their autocorrelations
can be indirectly estimated by the sample autocorrelations of another non
stationary process. This other process is the one of the difference between
the count and its unconditional expectation. Its knowledge increases the
predictability of the Poisson process and the coefficients of the unknown
process can be estimated.
Empirically, the hypothesis of independence of the residuals is often
inadapted. The error is either a stationary process or a non stationary one.
We will show situations where the process is an ARMA 4 stationary process
or an ARMA non stationary one, the later being called a process with time
changing coefficients and noted ARMA t • We introduce a non stationary
error and we think that the relationship between the errors depends on the
day in the week; the process is then said to be cyclostationary.5 For conven
ience, this new process is called CARMA. The cyclostationarity appears
when the memory of the system is linked to days. All the CARMA processes
have a VARMA representation under constraints of some elements of two
matrices. The interpretations made for multiple series can be adapted to the
case of the cyclostationary processes.
Under the hypothesis where the process of the error is a CARMA, the
exhibited new process is a model whose coefficients are varying with time
and present some similarity with the Kalman filter [24, 26]. In another way,
the variance of this new process depends on past errors but the model differs
from the ARCH model of Engle [9] and the GARCH model of Bollerslev [3].
We use the asymptotic framework of Gourieroux et al. [12, 13, 14]. For
convenience, we choose the mean square that is equivalent to the maximum
likelihood under the hypothesis of normality. Asymptotically, the choice of
the normal law leads to estimators equivalent to those established by all
other linear exponential laws. Gourieroux et al. [12, 13, 14] apply the pseudo
maximum likelihood method to the basic Poisson model and quasi gener
alised pseudo maximum likelihood to the Poisson model, the parameter of
which contains an error. The law of the error is unknown but the first two
moments are supposed to exist. In their approach, the errors are independ
ent. Under the hypothesis of the dependence of the errors, the estimator of
quasi pseudo maximum likelihood has to be adapted to and a new procedure
of estimation has to be constructed. We show how to adapt the procedure of
the quasi generalised pseudo maximum likelihood in this particular case.
At the end of each period, the cumulated realizations of the Poisson process
A Poisson process of which the parameter contains a nonstationary error 143
are observed: Yt is the number of events that appear at the time t; At is the
parameter of the Poisson law at time t and P(Yt) is the probability of
apparition of Yt events in t.
IF'
P(Yt) = exp(A t)  I . 2.1:1
Y
2.1:2
At is a random variable composed of a deterministic part exp(X;p) and a
random multiplicative non observable error ~ t = exp( ct ). The mean of the
error is one and its variance is 'YJ 2.
E(A t) = exp(X;p) 2.1:3
»
V(At) = exp(X;p)2 V(exp(c t = exp(2X;P)'YJ2 2.1:4
with
V(exp(C t» = 'YJ2 2.1:5
E(exp( ct» = 1. 2.1:6
In this section, we suppose that the process of the errors is a white noise with
a variance equal to one and the covariance between the errors ~, is null for
each lag h.
C(exp( ct) exp( Cth» = 'YJt" h = O. 2.1:7
The use of the exponential function ensures that At is positive. The random
error allows us to introduce the imperfection in the determination of the
parameter of the Poisson law. In those conditions, the law of Yt conditional to
X; and c, is a Poisson law [L(y/X;, c,)]; but the law of Yt conditional to X;
alone is not a Poisson law. The law of Yt can easily be specified in the case
when ct follows a gamma law. The particular specification was introduced by
Greenwood and Yule in "accidentproneness" (in Cox and Lewis [7],
Johnson and Kotz [23], Weber [35]).
We take the formulation of Gourieroux et at. [13] and [14] in the univariate
case. In the univariate Poisson process, the asymptotic covariance is directly
deduced from multivariate formulations of asymptotic variance of the esti
mators ofPML.
v (f3) = E ( of
as x of3
~I lof
op
)1 E ( opof ~I
x
Q ~I 10f)
op
The variance covariance of the true but unknown law of y, is equal to:
2.2:5
In the univariate case, ~(X, f3) is the variance YI according to the chosen
pseudo law. The asymptotic covariance may be estimated by sampling.
However, it is possible to define estimators with minimum variance that are
named estimators of quasi generalised pseudo maximum likelihood (QGP
ML). This procedure is defined [13] and [14] when the conditional law of y,
A Poisson process of which the parameter contains a nonstationary error 145
The method of QGPML proposed by Gourieroux et al. [12, 13, 14, 34]
give strongly consistent estimators, asymptotically gaussian and with minimum
variance (theorems 4 of [14]). The first step gives an estimator of the
coefficients; the second step gives an estimator of the variance of Yt according
to the estimations of coefficients. The asymptotic variance of coefficients
does not depend on the variance of the chosen pseudo law.
v p_
as( ) Ex
(Of g_1 tOf
ob ~
)1 2.3:1
For estimation, the two steps can be repeated until the numerical conver
gence of the algorithm.
In the doubly stochastic Poisson model, we chose the normal law. The
maximum of likelihood consists in maximising the following expression at the
first step.lO
T
T (y, _ exp(X,(3»2
Maxzlz= L 2.4:4
fJ ,= I exp(X;p) + exp(2X;p)1j2 .
146 B. Larcher
The count data are modelled by a doubly stochastic Poisson process and the
error of the parameter may be stationary or not. This model is used to model
the series of deaths in a large hospital. The number of daily deaths in the
hospital appears to be the realization sum of the indicator functions of all the
people present on that day. The probabilities for each one are unknown and
also the class of risk. We consider different classes of risks. The number of
deaths in the class with high risk is represented by the Binomial law (B(n!,
PI» with random parameters (nl' pJ and the number of deaths in the class
with low risk is modelled by a Poisson law with a random parameter [P(A I )].
Moreover, in the high risk class, the parameters PI can be time dependent
and this dependence can be different for different days of the week. This
decomposition in classes of risk is unknown and the relationship between
parameters cannot be deduced. We are using Poisson's law whose parameters
are stochastically non stationary. We think that this model is well adapted to
these data.
3.1. Meaning ofa Poisson law whose error of the parameter is an ARMAI
This is the case of daily series the parameters of which depend on other
random events whose present and future effects differ according to the day
when the events appear. 11
where
D = the number of periods in the cycle (7 for a week)
d( t) = a cyclical function which gives the day of t in the week.
<P d(t)( B) = the polynomial of the lag operator of the autoregressive part.
e d(llB) = the polynomial of the lag operator of the moving average part.
A Poisson process of which the parameter contains a nonstationary error 147
Pt = I WjEI _ j 3.1:5
j=O
where
Wj = a matrix of order D X D
E t _ j = a vector of a white noise of order D.
It can be decomposed in an autoregressive part and in a moving average part;
this decomposition is named vectorial ARMA or VARMA.
p q
The constraint is the ,po is strictly triangular inferior (,po. i,j = 0 for each j >
i). The autoregressive part corresponding to ,po represents instantaneous
causality in multiple series. PI represents days of a week, the realization of
the day before the day d of the week t is the day d  1 whose realization is
in PI if d > 1. We have to introduce that ¢o and ()o which do not correspond
to d  1, are necessarily equal to zero. ,po; i, j = ()o; i, j = 0 for each j > i.
PI = the vector of the realizations of a cycle (a week for instance).
¢k = a matrix of order D X D corresponding to autoregressivity of lag k
()k = a matrix of order D X D corresponding to moving average coefficients
of lag k.
A cyclostationary process may be analysed as a multivariate process P(I),
P(2), ... , P(d), ... , P(D). In the case of day data, P(I) is the realization of
Monday and P(2) is the realization of Tuesday and so on until Sunday. Each
univariate process is stationary and a multivariate process is called a
VARMA process (see Hannan [17], Granger and Newbold [16]).
The approach of the maximum likelihood is due to Akaike [11 and the
exact maximum likelihood to Hillmer and Tiao [20]. The procedure is
explained by Tiao and Box [33]. For an example of application, we may refer
to the analysis of pig market prices made by Cordier and Indjehagopian [211.
gives the network of these relationships. But, without any additional informa
tion, it is impossible to derive further explanations.
At a lag different from zero, the covariance C(yl' Yt h) between the random
variables Yt and Yt _ h is a function of the covariance between the errors of the
parameter of Poisson's law (Annex 1). C(yl' Yth) may be estimated by a
similar method to the one which is used to estimate the variance in QGP
ML.
Yt = X;{3 + ct + v t
ct = 'Pt(B)ct 3.2:1
w, = 'P,(B)et + Ut
where
'Pt(B)e, = the non stationary process of the error ofthe parameter.
u, = the error of the Poisson lottery of law I. L(O, X;{3 + 'PrCB)et); the
variance of u, is equal to the variance of Yt with e, the unknown
E(9;"(h» = g;"(h) ( 1  ~) 
1 nl
 2 L (n  k) [y~(h  lk) + r~(h + lk)]
n k 1 3.2:2
y;,,(h) = E[(E/ + uJ (E/ h + u/ h )] = r:(h).
The variance of this estimator is greater than the variance of the sample of
errors; unfortunately, it is impossible to observe a sample of errors. The
sample variance is not an asymptotically unbiased estimator of the variance
of error.
E(9;"(0» = y;"(O) ( 1  ~) 
1 n1
 2 L (n  k) [y~(lk) + y~(lk)]
n kl
polynomials (1  <I> tC B» and [1  E> t( B)] lie outside the unit circle.
et = 'lft(B)et = [1  <I>~(B)rl [1  E>~(B)let = rr~CB)et
and
3.2:5
The process et has an ARt representation and this process will be approached
by the process wt = IP'(B)w[' In these conditions, we have to evaluate the
variance of wtCy;"CO» by the method used in annex 1.
Y( = X;{3 + rr;"(B)wt + e;" + U t
y;"(O)=x;{3+rr;"(B)w(+ V(e;") 3.2:6
V(u,) = V(y/I;"_l) = X;p + rr;(B)wt + V(en.
With yE the vector of the autocovariances at lag h different from zero, <pE
the vector of the coefficients and P the matrix of the autocovariances at lags
oto p 1, the YuleWalker equations are:
yE = P <pE and <pE = (prl yEo 3.2:10
The coefficients <pE of the process e" can also be estimated by the resolution
of the YuleWalker equations. For process of W,' the equations of Yule
Walker are:
p w
yW(1)= L <P7,jE(yw(j1».
j= 1
3.2:12
A part of V(Et) has been explained by the process <PtCB) Wt. The true
variance of U t is V( ut) = V(Yt) = X,f3 + ~ <p jE, _ j + a; and depends on
unknown variables. The variance of ut conditional on the process W t is
V(u7) = V(Yt) = X;f3 + <Pt(B) wt + V(E7).
The generalised least square cannot be used because the variance of
exogeneous variables WtI' wt  2 , ••• , w,_p changes over time. The process
Yt  X tf3  <I> rC B) wt is then an orthogonal process. The matrix of variance
covariance changes with time. In a linear model with a homoskedastic
autoregressive error (Pierce) [27], OLS may be used to evaluate the
coefficients of the AR(p); the errors are recursively computed and intro
duced in the model and the YuleWalker equations would give an equiva
lent result. If the errors are heteroskedastic, OLS cannot be used to
evaluate the coefficients <I> W of the process.
If the variance of the error is changing with time, the matrix of covariance
of the exogogeneous variables <I>;" is changing with time and <I>;" has to be
computed separately by inverting f;" at each t. If we write the relationship for
A Poisson process of which the parameter contains a nonstationary error 153
t and t 1, then we derive the relationship between cl>7 U'llb cl>71'
yW = yE = r7c1>7 = f714>71
nW 
'V t 
(rW)1
/
rwt!'V/!
nw 3.2:13
3.2.4. Process with time changing coefficients for an ARt and comparison
with Kalman filter
Furthermore, when the process of the unobservable error is an ARt and in
particular a cyclostationary process, the autocovariances at a lag different
from 0 are cyclical, the coefficients of the process of the unobservable error
are cyclical but not the coefficients of the process of wf • All the terms of r;V
depend now on t. If the procedure is described in a sequential form and if we
use a bayesian formulation, then the system is similar to that of the Kalman
filter [24, 26). In its bayesian formulation, the model with changing coeffi
cients is that of the Kalman filter. In a matrix form:
Wt=F;cI>;"_l + vt and F;=[Wt\,Wt2, ... ,Wtp] 3.2:14
cl>7 = Ptcl>;V1 + 'l']t·
The procedure of estimation takes into account the different properties that
we have exposed earlier and is presented in annex 2.
4. Concluding remarks
The analysis of the series of deaths gives an example of using the non
stationarity in a doubly stochastic Poisson model. Non stationarity is met in
many phenomena that are observed daily and must be correctly modelled by
using non stationarity and more precisely the cyclostationary properties. For
this reason, we think that the CARMA should be used in the future.
These processes may be used to define parameters of doubly stochastic
processes. In particular, for the Poisson process, the doubly stochastic
Poisson process has the advantage of giving us a flexible structure to model
phenomena, whose Poisson processes in continuous time are unknown and
whose parameters depend on unknown variables. The introduction of a non
stationary error allows us to improve predictability but lays down an
estimation of a model whose parameters are time dependent.
Practically, the estimation and identification procedures of stationary
processes, before or after transformation, are well known; its promotion is
principally due to Box and Jenkins [2). The underlying statistical foundations
can be easily extended to the cyclostationary case. Moreover, the set of
ARIMA model becomes a particular case of a model with error CARIMA. If
the exogeneous variables are dummy variables which represent time, the
process is equivalent to a CARIMA.
154 B. Larcher
The different results concern the univariate case and may be extended to
multivariate series of Poisson whose parameter errors follow a vectorial
CARIMA; this new model can take into account the relationships between
the parameters of different Poisson series.
However, further investigations and simulations have to be made to test
the situations where there is significant and increased predictability.
Annex 1
O~(O) = iA
(A2)
0;(0) = i 2 A + l 2 A2
E(Yt) = exp(X;f3)
E(y~) = E(A + A2) = exp(X;f3) + exp(2X;f3) (17 2 + 1) (A3)
V(Yt) = exp(X;f3) + exp(2X;{3)172.
Yt=X;+ et
V(Yt) = X;{3 + 17 2 • (A4)
In the case where the model is additive, the covariance does not depend on
X/{3:
(A6)
A Poisson process o/which the parameter contains a nonstationary error 155
Annex 2
O. INITIALISATION
0.1. a(Yt) = 1
0.2. <P7(B)wt = 0
0.3. rr(r7)1 rf = o.
1.1. Estimation of ,B
2.1. Estimation of V( e n
V(e7) is the residual variance of e/ and is equal to V(et)  E(explained by
the autoregressive process).
(A7)
Notes
1. Theoretical justifications are based either on the theory of waiting line models or on the
sum's convergence of indicator functions.
2. A non homogeneous Poisson model is a process with a parameter changing with time.
3. The function can be linear or not; if not, the non linear optimization model has to be
used.
4. An ARMA process is an autoregressive moving average process the use of which is
widely due to Box and Jenkins [2J.
5. We have taken the word cyclostationarity from the signal theory, (Papoulis [29J, Gardner
and Franks [11 J "characterization of the cyclostationary random signal").
6. We refer to the convergence in probability or to the convergence almost surely or to the
convergence almost everywhere of integration theory (Serfling) [32J.
7. The main results of linear exponential law and models are in Monfort [28J.
8. These results are proved in annex 1.
9. These results are proved in annex 1.
10. At the first iteration, the variance is equal to one beacuse it is not possible to take into
account the variance of Yt that depends on unknown parameters f3 and 1]2.
11. The characteristics of the non stationary processes may be found in Priestley [3 I] p. 816
and the ARMAt is Hassairi [18]. The generalisation of the theorem of Wold to the non
stationary case was made by Cramer [8].
12. The concept of causality may be applied for the relations between the elements of a
univariate series; this use is found in particular in Brockwell [5].
13. The procedure as shown in the following part is not orthodox because the variance of this
process changes at each time.
14. n is conventionally the number of observations used to perform the variances.
15. If the mean is non significantly different from zero; the variances and the covariances are
performed from moment of order two and the estimators are unbiased.
16. For this, see Brockwell [5] or Gourrieroux and Monfort [IS].
References
[1 J Akaike, H. (1973), 'Maximum Likelihood Identification of Gaussian Autoregressive
Moving Average Models', Biometrika 60(2), 255265.
[2] Box, G. E. P. and Jenkins, G. M. (1976), 'Time Series Analysis, Forecasting and
Control', San Francisco: Holden Day.
[3] Bollerslev, T. P. ~1986), 'Generalised Autoregressive Conditional Heteroskedasticity
with Applications in Finance, Ph.D University of San Diego.
[4J Breslow, N. E. (1984), 'ExtraPoisson Variation in Loglinear Models. Applied
Statistics', Royal Statistical Society, 33(1) 3844.
[5J Brockwell, P. and Davis, R. (1987), Time Series, Theory and Methods. Springer Verlag
ed.,1987.
[6] Burguette, J. Gallant, R., and Souza, G. (1982), 'On Unification of the Asymptotic
Theory of non Linear Econometric Models', Econometrics Reviews 1(1): 151190.
[7] Cox, D. R. and Lewis, P. A. (1966), 'The Statistical Analysis of Series of Events',
London: Methuen and Co Ltd.; New York. Wiley, J. and Sons Inc.
[8) Cramer, H. (1961), 'On the Structure of Purely nondeterministic Stochastic Processes',
Arkivfor matematik (4), 249266.
[9] Engle, R. F. (1982), 'Autoregressive Conditional Heteroskedasticity with Estimates of
the Variance of U.K. Inflation', Econometrica (50), 9871008.
[10J Frome, E. L. Kutner, M. H. and Beauchamp, J, (1973), 'Regression analysis of Poisson
distributed Data', Journal of the American Statistical Association 68(344),935940.
A Poisson process of which the parameter contains a nonstationary error 157
1. Introduction
level) or admitted into a hospital (third level). For longterm nursing a patient
will be admitted into a nursinghome (fourth level). This division into four
levels of medical care recurs in MEDIKA. However, no distinction could be
made between outpatient and inpatient treatments of specialists.
In MEDIKA are distinguished four types of variables: capacity, consump
tion, employment and cost variables. This division is shown in Figure 1.1.
1. capacity. A measure for capacity is defined for each of the four levels of
medical care in Section 2. The capacity of the third level of care, i.e. the
supply of beds in hospitals, is supposed to be endogenous.
2. consumption. Equations for the flows of patients between the different
levels of health care provision and for the consumption within these levels
Primary care
(1st level) G.P.'s
t.=:==;:==~+I.I
Specialist care
(2nd3rd level) Operations
/+
Hospitals
(3rd level)
Nursinghomes
(4th level)
D Endogenous variable(s)
with a timelag
Direction dependency
Figure 1.1. Structure of the somatic health care according to MEDIKA (except the part about
cost and financing).
The construction of a model for medical cost and labour (MEDlKA) 161
2. Capacity
1st2nd level
In the previous section four levels of the somatic health care system are
described. The capacity in the first and the second level is defined respective
as the "number"4 of general practitioners per 1000 population (GPP) 5 and
the number of specialists per 1000 population (SPP). Both variables are
exogenous. How many specialists work in policlinics is not known. So the
variable SPP is also a measure for the capacity of the third level.
3rd level
level. The density of beds in hospitals has increased to 0.56 beds per 100
population in 1973, after which a decrease occurred (Diagram 5.3). Quality
aspects, like the number of functions in hospitals, are not taken into account
with this variable. Although the government controls the number of beds in
hospitals in principle, other factors, especially the consumption in the past
appear to determine the development of the number of beds (Vanden Broek
[3]). Therefore we specify the equation for the number of beds in hospitals
(BED) in the form of a stockadjustment model:
On the one hand the number of beds is dependent with a timelag on the
occupancy ratio (OC) in the past. We distinguish a shortterm effect
(OC90_2t ) and a longterm effect (OC90_6). The following definition holds:
OC90_2t = 10/9 . t(OC_2 + OC_3) (2.2.a)
OC90_6 = 10/9 '1/5(OC_4 + OC_5 + OC_6 + OC_7 + OC_s). (2.2.b)
Both variables are defined in such away, that there is no influence of
these variables on the number of beds, when 90% of the beds is occupied.
On the other hand the relative change in the number of beds depends with a
timelag on a standard for the density of beds (BEDXP), in proportion to the
density of beds (BEDP). Two exogenous variables determine BEDXP:
(2.3)
The realization of a certain density of beds will be dependent on the
economic possibilities, represented by the Gross National Product per capita
(GNPP). After establishing the AWBZ6 in 1968 a bed in a nursinghome
(BEDN) might form a substitute for a bed in a hospital. We take this into
account with the index mEDNA. Substituting (2.3) in (2.1) results in:
i{2' 0.8112
Nb 25
4th level
3. Consumption
1st level
2nd level
The flow of patients from the first to the second level is approximated by the
number of referrals of publicly insured? The referral rate of publicly insured
(REFPU) increased over the period 19531979 (Diagram 5.1).
Several crosssection analyses are performed in the Netherlands to
quantify the influence of exogenous variables on the referral rate. The
following influences, relevant for a dynamic analysis, are found (Hoeksma
[15], pp. 3063):
a positive influence of the density of specialists;
different results about the influence of the density of general practitioners;
a small positive influence of the percentage of aged persons.
164 R. J. A. M. van den Broek
E S  bl (3.2.a)
(SPI P)  SPISP  b2
The elasticity between the referral rate and the density of the number of
specialists e(SPISP) can be derived:
iP 0.9752
N 26
respectively 0.44, 0.39 and 0.35. In this last study one can perceive a
decreasing trend of the elasticity. The results of Hooijmans are comparable
to ours.
2nd3rd level
(3.3)
166 R. 1. A. M. van den Broek
iP 0.9892
N 26
3rd level
(3.4)
LS = b E (AGEP) . [
365. BED ]b 13
. SP ]b .
[
14
18 2 ADM BED
(3.5)
Morbidity
An important development, that has changed the morbidity in the past, is the
ageing population. For that we introduce in the model the variable the
number of aged persons per capita (AGEP). It is assumed, that a change of
AGEP, ceteris paribus, does not influence the admission rate of aged persons
as well as the admission rate of persons under 65. The same is assumed for
the average length of stay of aged persons and persons under 65. When is
given that an aged person is admitted bI9 times more in hospitals and stays
b20 times longer than a person under 65, then the following relations hold for
the functions E;(AGEP) under the assumptions (see Appendix III).
EI(AGEP) = 1 + (b I9 1) AGEP (3.6.a)
E2(AGEP) = (1 + (b 19 b20 1) AGEP/EI(AGEP). (3.6.b)
Based on data of the consumption of aged persons and persons under 65
we choose as fixed values b19 = 1.8 and b20 = 2.0.
Insurance system
Cost for clinical care in hospitals are covered almost completely by an
insurance. Therefore prices hardly influence the consumption of clinical care.
Only the maximum length of stay, that is covered by the publicly insurance
companies, was increased twice. From 42 to 70 days in 1955 and from 70 to
365 days in 1964. We account for these changes in the insurance system with
two dummies DPU1 and DPU2.
Capacity
The capacity of hospitals to admit patients is determined by the available bed
density given the length of stay (BEDP/LS). The average length of stay is
limited by the number of beds per admitted patient (BED/OPN). The
variable specialists per bed in hospitals (SP!BED) is a measure for the
clinical as well as the policlinical capacity. Therefore it is difficult to predict
the influence of the number of specialists per bed on the admission rate. It
may be expected that an increasing number of specialists per bed has a
negative influence on the average length of stay.
168 R. J. A. M. van den Broek
Substitution
The general practitioner decides how may patients will be referred to a
higher level of care. As a matter of fact, in principle a patient can be
admitted into a hospital only after a referral by the general practitioner.
Therefore we add the referral rate (REFPU) in the equation for the admis
sion rateP It might be possible, that an increase of the supply of beds in
nursinghomes will result in a shift of a relatively small number of patients
with a long average length of stay from the hospital to the nursinghome. This
has been taken into account by introducing the variable IBEDNA in the
equation for the average length of stay.
Successively we discuss the estimation results in Tables 3.3 and 3.4 of the
Equations (3.4) and (3.5). The equations are estimated with a two stage least
squaresmethod.
It can be deduced, that the increase of the percentage of aged persons
results in an increase of the elasticity between the admission rate and the
percentage aged persons from 0.06 in 1953 to 0.08 in 1979. The elasticity of
the capacity to admit (0.17) is lower than the value 0.39 Van der Gaag [14]
p. 92 finds for publicly insured. Even higher elasticities are found in cross
section analyses of the public sector (Rutten [26], p. 95; Hooijmans [16], p.
22). An explanation is given by Hooijmans and Rutten [18], p. 1217. They
ascertain, that the clinical consumption of privately insured in hospitals is
less dependent on the capacity than the consumption of publicly insured.
This happens perhaps, because privately insured in general value their time
more than publicly insured. An increase of the number of specialists per bed
causes a small increase of the admission rate, ceteris paribus. This may not
seem to coincide with the insignificant or even negative coefficient found for
the public sector. Hooijmans and Rutten ([18] pp. 1314) however, do find
a positive value for the private sector. They give the following explanation.
"Here we may be confronted with an effect caused by differences in fees, that
specialists acquire by treating publicly or privately insured patients. The fees
in the private sector are considerable higher than those in the public sector.
AGEP b l9 1.8
365 • BEDP/LS bs 0.171 3.2
SP/BED b9 0.051 2.0
REFPU blO 0.464 16.8
DPUI bll 0.012 4.0
Constant b12 1.289 3.7
R2 0.9981
N 26
The construction of a model for medical cost and labour (MEDlKA) 169
In a region with a high number of specialists per bed, other things being
equal, the specialists will have to treat more privately insured patients in
order to reach the same level of income as in other regions." The positive
elasticity 0.46 between the referral rate and the admission rate is confirmed
by the result 0.40 of Van der Gaag [14], p. 92. Crosssection analyses
produce lower elasticities.
According to Table 3.4 the length of stay in hospitals depends more on
the capacity than the admission rate does. The elasticity 0.69 is higher than
the value 0.37, found by Van der Gaag [14], p. 92 for publicly insured. Also
for publicly insured, Hooijmans [16] p. 22 finds an elasticity between 0.5 and
0.6 for the years 1974, 1975 and 1976. We expected a priori a slightly lower
elasticity in our analysis, because Equation (3.5) is estimated for publicly and
privately insured together. The coefficient b14 has as expected a negative sign,
but the value is insignificantly different from zero. According to Table 3.4
there has been substitution between hospitals and nursinghomes in the
seventies. Rutten [26], p. 96 draws the following conclusion in his cross
section analyses for the years 1971 and 1973. "In contrast with the results
for 1971 we find for 1973 a significantly negative impact of the number of
beds in nursinghomes. This may be related to the fact that development
towards stimulating substitution between hospital care and nursinghome
care started in the late sixties, so realization of substitution of both forms of
care could not have occurred until 1973".
A quantity, that has not been accepted as explanatory variable in Equa
tions (3.4) and (3.5), is the number of nurses employed by hospitals. In some
of the studies mentioned earlier a significantly positive coefficient has been
found between this variable and the admission rateP In these studies is
assumed, that the number of beds, given the length of stay, will be used more
R2 0.9947
N 25
p' 0.57
intensely, when more nursing staff is available (Van der Gaag [14], p. 93). We
think it is plausible, that the relation holds the other way around: a relatively
high number of admissions will be the cause of a high demand for nurses. 14
This reasoning is elaborated in Section 4.
The number of occupied beds per 100 population in hospitals is defined
by the admission rate and the average length of stay.
OBEDP = ADMP . LS/365. (3.7)
The elasticities between the explanatory variables of Equations (3.4) and
(3.5) and the density of occupied beds in hospitals can be calculated easily.
The most interesting result is the high elasticity of 0.71 between the number
of beds and the number of occupied beds. Feldstein [11], pp. 8587 states,
that the relation does not need to be a causal one. For instance it might be
possible, that the planned capacity is adjusted to the expected needs for the
future. In MEDIKA the supply of hospitals is endogenous. It is influenced by
the consumption in the past (see Section 2). It might also be possible, that the
physician has a builtin preference for a clinical treatment, what makes him
dependent on the available capacity (Fuchs [12], p. 96). In the Netherlands
this preference would be enforced by the existing financing system. Accord
ing to this systeem hospitals have to maintain a certain level of beds occupied
to keep their finance balanced.
4th level
4. Employment
Policlinical staff and clinical staff can not be distinguished. We divide the
staff in hospitals into three categories: (para)medical staff, nursing staff and
general staff.
A. (Para)medical staff
There is a lack of information about the development in medical treatment.
The construction ofa modelfor medical cost and labour (MEDlKA) 171
One gets an impression with the rapid increase of the number of operations
by specialists (see Diagram 5.2). It is clear, that the increase of paramedical
staff (PMS) and medical staff15 in hospitals (6%) over the years 19621979
has caused an intensification of treatment. This has resulted in a detectable
influence on the development of the number of patient days in hospitals.1 6
The increase of the paramedical staff is possible because of the considerable
inflow from the trainingcourses. The variable is assumed to be exogenous in
this incomplete version of MEDlKA.
B. Nursing staff
We distinguish three types of nursing staff: qualified nursing staff (QNS),
student nursing staff (SNS) and a small remaining category other nursing staff
(ONS). The last category is kept exogenous. Diagrams 5.5 and 5.6 show, that
the number of inservice student nurses has decreased in the seventies. The
training of nurses outside the hospital was started at the same time. The
work, that remained to be done because of the departure of inservice
student nurses had to be carried out by qualified nurses.
tion: the number of occupied beds (OBED) and admissions (ADM). Part of
the work is performed by student nurses (SNS) and other nurses (ONS),
which makes substitution 18 possible. This results in:
DQNS = c6 + c3 OBED_ 1+ c4 ADM_I + cs(SNS_ I + ONS_ I). (4.3)
It should be noticed, that hospitals cannot attract unlimited numbers of
staff. This limitation might influence the coefficients of Equation (4.3).
Substitution of Equations (4.2) and (4.3) into Equation (4.1) gives:
(QNS  QNS_ I) = co(CAND . T  c2 QNS_ I) +
+ cd[c6 + c3 OBED_ I + c4 ADM_ I +
+ cs(SNS_ I + ONS_1)1  QNS_ I}
with Co + cI ~ 1. (4.4)
According to the estimation results in Table 4.1 the relation Co + cI ::::. 1
applies. This means, that the number of qualified nurses is adjusted to a
combination of supply and demand. The demand for nurses depends, as
expected, positively on consumption (the variables OBED and ADM). A
positive coefficient is also found by Van Aert and Montfort [11, p. 17 in a
crosssection analysis. The coefficient Cs indicates, that because of a lower
production 10 student or other nurses can be replaced by 7 qualified nurses.
Based on the number of working days and an assumption of the productivity
of students of each class Legerman and Van der Hoef [211, p. 10 estimate,
that the production of 10 student nurses equals the production of 6 qualified
nurses.
Table 4.1. Estimation results of Equation (4.4): the qualified nursing staff in
hospitals (QNS).
C. General staff
The general staff can be divided into two categories: domestic staff (DS) and
administrative staff (AS). It is assumed, that the number of general staff
changes proportionally to the need for support because of other activities in
the hospital (medical treatment, nursing).
R2 0.9624
N 19
iP 0.9989
N 19
4th level
The major part of the staff in nursinghomes consists of nurses and general
staff. A set of equations similar to those for hospitals explains in MEDIKA
the development of the qualified nursing staff (QNSN), the student nursing
staff (SNSN), and the general staff (DSN + ASN) 19 in nursinghomes. The
specification and the estimation results of these equations are not given in
this paper.
5. Simulations
In this section we try to gain some insight in the efficacy of MEDIKA. First
of all we have simulated the period 19531979. The most important results
are shown in Diagrams 5.15.8. Especially the differences between realiza
tion and simulation of the number of beds in hospitals are considerable. This
happens, because differences between the simulation and realization of the
number of beds and consumption in the past have an important influence on
the prediction of the number of beds (Table 2.1). The error made with the
prediction of the number of beds effects consumption directly (Tables 3.3
and 3.4). The important turningpoint in the development of the number of
beds in the seventies however is simulated correctly by MEDIKA.
The next four variants are simulated in order to get acquainted with
MEDIKA:
variant Al SP(IS) +10%;
variant A2 GP +10%;
variant A3 (I)BEDN +10%;
variantA4 lSI +10%.
The construction of a model for medical cost and labour (MEDlKA) 175
Table 5.1. Consequences of the variants AIA4 for the most important endogenous
variables of MEDlKA after 1 year (20 years) in percents.
Al A2 A3 A4
SP(IS) + 10% GP+ 10% (I)BEDN + 10% lSI + 10%
Endogenous
variable lY 20Y lY 20Y 1Y 20Y lY 20Y
that given a constant number of beds in hospitals at short notice, the average
length of stay increases with 1.4%. This increase will be reduced after 20
years (0.8%) as a consequence of the adjusted number of beds in hospitals.
An increase of the number of beds in nursinghomes with 10% (variant
A3) causes an immediate increase of general and nursing staff in nursing
homes with respective 2.4 and 4.1%. On the other hand it reduces the
average stay in hospitals with 2.8%. This reduction will increase to 5.3% after
20 years. This will result in a significant decrease of the staff employed in
hospitals.
According to the results of variant A4 the inflow of students has two
consequences for the number of qualified nurses. On the one hand the
increase of the inflow of students with 10% will result after a few years in an
increase of 8.5% in the number of successful candidates for general nursing.
The number of qualified nurses will rise in consequence of this increase of
supply. On the other hand the need for qualified nurses diminishes, because
of the rising number of student nurses employed in hospitals. These two
effects result in a reduction of the nursing staff with 2.1 % after 3 years. The
longterm reduction is less (1.7%).
The possibilities of MEDIKA are illustrated by calculating the conse
quences of the variants AIA3 for cost and employment in the Dutch
health care sector, using the calculating scheme mentioned in the introduc
tion. Not all inputs of the scheme are determined by the structural equations
described in this paper. For this we use a few simple supplementary assump
tions, which we will not describe. In the introduction of this paper was
written that the model, described in this paper, is incomplete. This fact
influences the results of the variants. So the excercition serves for illustration
only. Table 5.2 shows the consequences for the employment of a change in
the number of specialists, general practitioners or beds in nursinghomes
according to the variants AIA3.
Other things being equal, 10 new specialists will initiate at short notice in
hospitals the employment of 2 members of the general staff and 5 qualified
nurses. On the other hand employment in hospitals decreases with 15
persons when 10 general practitioners start working. According to A3, a rise
of the number of beds in nursing homes will have no important consequences
for total employment after 20 years. The increase of employment in nursing
homes will by then be compensated by a decreased employment in hospitals.
In Table 5.3 the results are given of calculatings concerning the conse
quences of variants AIA3 for the cost of the health care sector. The results
only serve as illustration of the possibilities of MEDIKA just like the
calculations in Table 5.2.
According to variant A 1 every new specialist will generate at short notice
extra cost of f 56,000 in a hospital. The purchase of medical equipment,
paramedical staff and assistents, that may be induced by the specialist is not
included in this amount.
In variant A2 we assume, that the general practitioners will not save
The construction of a model for medical cost and labour (MEDlKA) 177
Table 5.2. Consequences for employment in the Dutch health care sector per specialist/
general practitioner/bed in nursinghomes after 1 and after 20 years in fulltime equivalents.
Al A2 A3
per general per bed in
per specialist practitioner nursinghomes
lY 20Y lY 20Y lY 20Y
Table 5.3. Consequences for the cost of the Dutch health care sector per specialist/general
practitioneribed in nursinghomes after 1 and after 20 years in Dutch florins (X 1000).
Al A2 A3
per general per bed in
per specialist practitioner nursinghomes
lY 20Y lY 20Y IY 20Y
expenses even at a long range. The same holds for a reinforcement of the
fourth level, because only a part of the new patients in nursinghomes would
otherwise have been admitted into hospitals. In a complete analysis homes
for aged persons and care for the elderly should also have to be implicated.
•0
realisatIon
simulation
38
36
3.
32
30
28
26
24
22
20
18
16
,.
12
10
i i i t j i i t i i i i
1952 195. 1956 1958 1960 1962 196. 1966 1968 1970 1972 1974 1976 1978 1980
20
realilatlcn
simulation
18
16
14
12
10
i i
1952 1954 1956 1958 1960 1962 1964 ~966 1968 1970 1972 1974 1976 1978 '9~O
The construction of a model for medical cost and labour (MEDlKA) 179
'.60
reallsQtlon
.Imulatl."
O.SS
0.50
0.4S
0.40
j I i i i
1952 1954 1956 19S8 1960 1962 1964 1966 1968 1970 1972 1976 1978 1980
0.55
realisation
~imulQtfon
0.50
0.45
~
0.40 "
O.J5
J.JO
i i • i i , ' i i i
1952 1954 1956 \9Sa 1960 1962 1964 1966 1968 1970 1972 1974 1976 1978 1980
180 R. I. A. M. van den Broek
r.aIl301'fon
3lmulatlon
32
30
26
cT
<II
"E 26
:l: 24
" 22
0
0
9
x 20
16
16
14
1 i i j i i i i i
1952 1954 1956 1958 1960 1962 1964 1966 1968 1970 1972 1974 1976 1975 1960
26
real!.sotlon
sfmulCitlon
26
24
cT
"
"
E 22
o
o
9
x 16
14
The construction of a model for medical cost and labour (MED lKA) 181
28
reall.ailon
simulation
26
...
24
&
E 2.2
=
:b
::l
20
0
0
52
X 18
16
14
[ j j i i ! i i i i i i t i
1952 1954 1956 195a 1960 1962 1964 1966 1968 1970 1972 1974 1976 1978 1980
1952 1954 1956 1958 1960 ~962 1964 :966 ~96B ~970 1972 1974 1975 :972 1980
182 R. 1. A. M. van den Broek
Notes
1. The author was employer of the Central Planning Bureau till November 1984 and after
that the Department of Welfare, Public Health and Culture. Presently he is working at the
Department of Justice.
2. "Arbeidsvraag en arbeidsaanbod in de gezondheidszorg op lange termijn". Project
personeelsvoorziening kwartaire sector. Bulletin no. 5. Centraal Planbureau, Sociaal en
Cultureel Planbureau. Den Haag, juni 1984.
3 Algemene Wet Bijzondere Ziektekosten.
4. In the paper we always correct for differences in working time over the years.
5. For the signification of the symbols see Appendix II.
6. Algemene Wet Bijzondere Ziektekosten. Cost for nursinghome care will be paid
according to this bill out of a collective premium.
7. There are no data available about privately insured.
8. Assuming a constant ratio between the referral rate of aged persons and persons under
65, the referral rate has changed over the years 19531979 only with 1.6% as a
consequence of the ageing population.
9. General practitioners do not refer patients to specialists in the field of anaesthesiology,
bacteriology, clinical chemistry, pathological anatomy, radiology and radiotherapy.
10. Minor poJiclinicai operations are an exception. In the Netherlands the specialist is paid a
fixed amount for each publicly insured, that is referred to him. This amount depends on
the duration of the policlinical treatment. Only major policlinical operations are paid
separately.
11. We only use data about specialists for internal diseases and surgeons.
12. It is assumed, that the referral rate for private insured has the same development.
13. Different results are found in relation to the average length of stay.
14. This assumption is supported by the scheme of directives for staff in hospitals. According
to these directives the allowed number of nursing staff depends on the number of beds
occupied.
15. Among whom specialists.
16. The influence ofthe variable SPIBED in Equations (3.4) and (3.5).
17. Only those candidates are counted, who have graduated from a trainingcourse relevant
for hospitals.
18. The substitution between qualified nurses, and student or other nurses is a linear process.
For that we specify Equation (4.3) linear.
19. General staff = domestic and administrative staff.
20. Also of specialists, that are on the payroll of hospitals.
21. Excluding eyesurgery.
Appendix I. References
[1) Aert, J. H. van, Montfort, A. P. W. P. van (1977), Basisonderzoek Kostenstructuur
Ziekenhuizen, deel 5. Personeelscategorieen en kostensoorten. Nationaal Ziekenhuisin
stituut, Utrecht. 73 pp.
[2) Altman, S. H. (1970), The Structure of Nursing Education and its Impact on Supply.
Empirical Studies in Health Economics, The Johns Hopkins Press, London, pp. 335
352.
[3) Broek, R. J. A. M. van den (1979), Eerste aanzet tot een simultane verklaring van
gebruiksen capaciteitsontwikkeling van ziekenhuisbedden in Nederland. Centraal Plan
bureau, notitienr. 11, hoofdafd. IV, Den Haag, 25 pp.
[4) Broek, R. J. A. M. van den (1981), Middellange termijnraming intramurale gezond
heidszorg, Centraal Planbureau, notitienr. 1, hoofdafd. IV, Den Haag, 81 pp.
[5) Broek, R. J. A. M. van den (1982), Het aantal verwijzingen door huisartsen. Centraal
Planbureau, notitienr. 4, hoofdafd. IV, Den Haag, 14 pp.
The construction of a model for medical cost and labour (MED [KA) 183
[6J CentraaI Bureau voor de Statistiek (1981), Inkomens vrije beroepen. Een methodolo
gische studie, M13, 98 pp.
[7J Davis, K. and RuseH, L. B. (1972), The Substitution of Hospital Outpatient Care for
Inpatient Care', The Review of Economics and Statistics 54, 109120.
[8] Es, J. C. and van Pijlman, H. R. (1970), 'Het verwijzen van ziekenfondspatienten in 122
Nederlandse huisartsenpraktijken', Huisarts en Wetenschap 13,433449.
[9] Feldstein, M. S. (1967), Economic Analysis for Health Service Efficiency, North
Holland Publishing Company, Amsterdam, 332 pp.
[10J Feldstein, M. S. (1973), Econometric Studies of Health Economics, Discussion paper
291. Harvard Institute of Economic Research, Cambridge, 94 pp.
[11] Feldstein, P. J. (1979), Health Care Economics, John Wiley & Sons, Inc. U.S.A., 457
pp.
[12J Fuchs, V. R. (1974), Who Shall Live? Basic Book Inc., New York.
[13J Fuchs, V. R. and Kramer, M. J. (1972), Determinants of Expenditures for Physicians
Services in the United States 19481968, National Bureau of Economic Research,
Occasional Paper 117, DHEW Publication No. (HSM) 733013, 63 pp.
[14] Gaag, J. van der (1978), 'An Econometric Analysis of the Dutch Health Care System',
Proefschrift, Rijks Universiteit Leiden, 160 pp.
[15] Hoeksma, B. H. (1978), 'De polikliniek als schakel in de gezondheidszorg', Scriptie,
College voor Ziekenhuisvoorzieningen. Utrecht, 97 pp.
[16] Hooijmans, E. M. (1976), 'Estimates of a Model of the Dutch Health Care System over
the Years 1974, 1975, 1976', Report 82.01b Center for Research in Public Economics,
Leiden, 31 pp.
[17J Hooijmans, E. M. (1982), 'Het aantal klinische en poliklinische verrichtingen volgens
tarief 3', Center for Research in Public Economics, Leiden, 40 pp.
(18) Hooijmans, E. M. and Rutten, F. F. H. (1982), 'The Impact of Supply on the Use of
Hospital Facilities; Differences Between High and Low Income Groups in the Nether
lands', Report 82.04. Center for Research in Public Economics, Leiden, 21 pp.
[19] Hurd, R. W. (0000), 'Equilibrium Vacancies in a Labor Market Dominated by non
Profit Firms: The "Shortage" of Nurses', The Review of Economics and Statistics 55, 2.
[20] Jacobs, P. (1974), 'A Survey of Economic Models of Hospitals', inquiry 11,8397.
[21J Legerman, A. and Hoof, J. A. M. van den (1977), 'Schatting van de macroeconomische
effecten van de substitutie van inservice opleidingen voor de verpleegkunde diploma'S
A, B en Z door MBOv. VoMil', Stafbureau Raadsadviseur Lange Termijn Planning,
Leidschendam, 2 del en.
[22J Lempers, F. B. (1974), 'Een model voor de gezondheidszorg in Nederland', Centraal
Planbureau, notititienr. 12, Hoofdafd. IV, Den Haag, 43 pp.
[23] Monfort, A. P. W. P. van (1980), 'Production Functions for General Hospitals',
Proefschrift Katholieke Hogeschool Tilburg, 184 pp.
[24] Nieuwenhuis, A. (1975), 'Het verwijzen van ziekenfondspatienten door de huisarts',
Centraal Planbureau, notitienr. 15, Hoofdafd. IV, Den Haag, 16 pp.
[25] Posthuma, B. H. and Zee, J. van der (1977), Tussen eerste en tweede echelon 1',
Nederlands Huisartsen Instituut, Utrecht, 104 pp.
[26] Rutten, F. F. H. (1978), 'The Use of Health Care Facilities in the Netherlands',
Proefschrift, Rijks Universiteit Leiden, 225 pp.
[27J Yett, D. E. (1970), The Chronic "Shortage" of Nurses: A Public Policy Dilemma.
Empirical Studies in Health Economics, The Johns Hopkins Press, London, pp. 357
389.
The addition of P/PU/A indicates, that the variable is divided by the population/publicly
insured/number of aged persons in the Netherlands;
Variables with an i as subscript are i years accelerated;
184 R. J. A. M. van den Broek
Endogenous variables
Exogenous variables
ILLP index of the average number of cases of illness per male worker (1970 = 1)
IMSP index of the quantity of medical supply in hospitals per capita
lSI the inflow of students for general nursing in the first class
P the population ofthe Netherlands
The construction of a model for medical cost and labour (MEDlKA) 185
and
LS 65 + = b20 LS 65 · (l.b)
By definition holds:
ADM = ADM lis + + ADM lis  (2.a)
ADM· LS = ADM6S+ • LS 65 + + ADM6s  • LS 6S . (2.b)
The combination of (l.a) and (2.a) gives
ADM lis  (3.6.a)
ADMP = P _ AGE [1 + (b 19 1) AGEP].
The admission rate of persons under 65 is explained by the other exogenous variables. The
equation for the average length of stay is a bit more complicated.
From (l.a), (2.b) and (3.6.a) can be derived:
PETER ZWEIFEL
lnstitut flir Empirische Wirtschaftsforschung, Universitiit Zurich
Introduction
monthly premiums if the insured abstains from filing a claim during a year; C
offers a bonus amounting to two monthly premiums (as of 1982) in the first
year with no claims, three in the second year, and four starting from the third
consecutive year with no claims. Using a simple microeconomic model,
predictions are derived concerning the policies' respective ability to induce
an insured to forego ambulatory medical care given a minor impairment of
his health. These predictions are subjected to empirical tests in the third part
of the paper. Rebates and bonuses also create incentives for not submitting a
claim although medical care was consumed. Particular care is taken to
eliminate this submission effect because it amounts to a mere costshifting
between the insured and the insurer. Social savings only accrue if the insured
changes his behavior, i.e. if he reduces his utilization of medical care services.
But by doing so, he might jeopardize his future health. Therefore, the
empirical analysis is extended to cover two consecutive years in an attempt to
detect traces of a toothsaw pattern in medical utilization. The concluding
section of the paper contains a discussion and evaluation of the results, with
particular reference to the ongoing debate about the reform of social health
insurance.
At first sight, a rebate offer might seem to be nothing but a gimmick on the
part of the insurer for cashing in higher premiums earlier, with the insured
foregoing interest income he could have if his money was not tied up with the
insurance. This conception of a rebate or a bonus option presupposes two
things, however: First, there would have to be a full commitment on the part
of the insured to leave this earmarked money in the account, thus making
absolutely sure that cash will be available for covering the net cost of medical
treatment accruing under a conventional policy. Second, the degree of risk
aversion must be the same regardless of the state of health. By way of
contrast, Figure 1 below shows an individual that has a stronger degree of
risk aversion when ill than when healthy. There is at least some indirect
empirical evidence to support such an assumption, see Fuchs (1982). For
simplicity, a conventional cost sharing policy containing a deductible of D
Dollars is compared to a financially equivalent rebate offer. An insured
covered under the conventional policy has initial wealth of Wo  PD' with PD
symbolizing the premium of this policy. If he falls ill, the insured has to pay
D Dollars out of his own pocket, resulting in final wealth Wo  PD  D.
Using a probability of 50% for falling ill, his expected wealth amounts to
E(W)D·
The rebate offer it. selected such that it is financially equivalent to the
conventional alternative, i.e. E(W)R = E(W)D. With no discounting for
present value and 50% probability of illness, this simply means that the
premium of the rebate contract (PR) costs as much as the premium PD of the
Bonus systems in health insurance 189
,,
. ,
,,'
W
W"PR E(W)D
=W,,PDD = E(W)R
deductible contract plus one half of the deductible. Despite this financial
equivalence, the two alternatives are not equivalent as soon as the insured
value gains and losses differently, depending on health status (see Dionne
and Eeckhoudt, 1983, for some implications of statusdependent Von
NeumannMorgenstern utility functions). If risk aversion is stronger given
bad health than given good health, the utility function Ui (W) exhibits a
higher degree of concavity from below than does the utility function Uh( W).
Under a conventional deductible plan, the individual starts at Ub when
healthy and ends up at U"D when ill, resulting in expected utility E (U)D.
If covered by a policy with a rebate offer, the same individual has initial
wealth of only Wo  PR , associated with utility U"R. But he has a 50% chance
of cashing in a rebate. Since he is healthy then, he moves up along utility
function Uh( W) to reach point utin Figure 1. Due to that function's low
degree of concavity, expected utility amounts to E (U)R. Since E (Uh >
E (U)D' the rebate offer dominates a policy with cost sharing despite the fact
that the two plans are financially equivalent.
It should be emphasized that this argument does not prove that traditional
contracts (as well as selfinsurance) are always inferior to rebate offers or
190 P. Zweifel
Conclusion 1
For individuals whose degree of risk aversion is higher when ill than when
healthy, a rebate option will tend to dominate both a financially equivalent
policy with cost sharing as well as selfinsurance.
Given that individuals are insured by the same company, they are exposed to
much the same financial incentives. Therefore, generating predictions con
cerning one type of policy only is not sufficient. Rather, differential predic
tions among the three insurers must be derived, e.g. comparing insurer B
(offering a fixed rebate of three monthly premiums) to insurer C (offering an
experiencerated bonus of two to four monthly premiums). In order to be
able to deal with insurer C in the sequel, the planning horizon is two years
throughout. In Figure 2, a simple two goods model is shown with hours of
medical service (M) on the horizontal and all other goods (X) on the vertical
axis. In an attempt to mirror limited consumer sovereignty in health care,
choice is restricted between 0 and MJ physician hours (which are assumed to
cost 150 OM or 50 U.S. Dollars).
The straight line AA' depicts the budget constraint of a member of
insurance A with full coverage. It runs close to the origin because such a plan
has a high price. By way of contrast, BB' mirrors the lower premium of the
rebate policy written by insurer B. If a member of B accepts the treatment
offer comprising M] hours of care, he attains utility level 1; at point B. If he
saves his rebate by foregoing ambulatory care, his new budget constraint
originates from point B+ on the X axis. However, he would have to face the
full price for every physician minute, making him move along the steep
budget constraint towards point R. The two budget constraints intersect at
point R in Figure 2, indicating a submission threshold: For an annual bill
falling short of R (about 4 hours' work of ambulatory care), it is better to pay
it out of one's own pocket. Beyond point R, the insured fares better
submitting his medical bill.
Bonus systems in health insurance 191
x
B++
\
\
\
\
\
\
\
\
\
\
\S '\
~  B+'
\ B+
\
\ '\
\ \
\
\
\
,,
\
\
\
Br __~~~~~~~~~
Ar____________~~~A'
M.
2 3 4 5 6 7 8 11 12 13 (physician hours,
9 10
at 50 $)
Conclusion 2
C3+"
\ C"
\ \
\1" \
\ \
\ \
"~~
C3'\
~
C3"
\
\ J"
A A'
2 3 4 5 6 7 8 9 10 11 12 13 (physician hours
at 50 $)
the fact that insurer C's plans have a deductible of 250 DM (1.67 physician
hours) throughout. If he saves his bonus, the constraint is given by the
straight line starting from point C3+. Should he decide to see the physician in
the first period, he ends up at point C3  , which is by assumption on the same
level of utility as C3+.
Budget constraints in the second year depend on the decision made in the
first year. If the individual took advantage of the physician's treatment offer
after all, his point of departure in the second year is C3  in Figure 3. Should
he fall back on insurance once more, the boundary defined by points C3 ,
C3 ' , and C3  " applies. On the other hand, he could earn a bonus amount
ing to two monthly premiums if he managed to refrain from consuming
medical care in the second period. But this incentive will not suffice to make
him go without care because the same individual under the same conditions
was just indifferent when the bonus at stake was as high as three monthly
payments. Thus, consuming Ml hours of medical care will rank higher than
the prospect of saving the bonus in the second year, given that ambulatory
care was preferred to the bonus in the first year.
Alternatively, the individual might have saved his bonus in the first year.
In that case, the attainable bonus in the second year will amount to four
monthly premiums, corresponding to the point of departure C4+ on the X
axis of Figure 3. This is to be compared to consuming Ml of care in the
second year, symbolized by point C4  in Figure 3. Since the individual
previously was indifferent between medical care and the bonus of three
monthly premiums, he will certainly save his bonus of four monthly pre
miums in the second year. Thus, the optimal decision in the second year is to
refrain from consuming medical care, implying that the total bonus at stake
in the first year is not three but as much as seven monthly premiums. Under
these circumstances, the individual insured by C will refrain from calling on
the physician in both years. This argument leads up to
Conclusion 3
The prospect of saving an even higher bonus in the subsequent year may
induce members of insurance C to refrain from consuming ambulatory care
in the current as well as in the subsequent year whereas a member of
insurance B would have demanded medical care in both periods.
It is important to see that the statement holds true only in the case of minor
illness. Should the physician deem the health problem severe enough as to
warrant 2Ml hours of care in the first period, then even an enrollee of
insurer B would want to take advantage of this offer. In Figure 3, the newly
attainable point is in the neighborhood of C3+" (at M = 12), which certainly
ranks higher than point C4 + on the X axis, according to indifference curve
fl. Alternatively, the insured himself might rate his health problem as serious,
Bonus systems in health insurance 195
Empirical results
As a rule, insureds will submit their medical bills only if their total exceeds
the value of the rebate (plus a deductible if applicable), cf. intersection points
Rand U in Figures 2 and 3. Beyond this threshold, the billings distribution
should be complete. Since conclusions 2 and 3 do not refer to the filing
decision but to impacts on actual demand for ambulatory care, analysis of the
billings distribution must focus on values at or above this submission
threshold.
Submission thresholds are determined for a twoyear horizon since only
billings for the years 1981 and 1982 are jointly available for all three
insurers. This threshold is highest for enrollees of insurer C: In 1981, the
bonus at stake could be as high as seven monthly premiums (undiscounted).
For these insureds, the threshold amounts to 'seven monthly premiums plus
deductible 250 DM'. This is a conservative estimate because members who
had not attained maximum bonus in 1982 or did not count on saving their
bonus in subsequent years would have submitted smaller billings to their
insurance as well.
Members of insurances A and B must be assigned virtual submission
thresholds that would be in effect if they were enrolled by C. First, in each
agesexcell of insurer C's population, the maximum value of the submission
threshold is determined. Second, these values are assigned to members of
insurers A and B in the same agesexcell.
For empirical estimation, a binary dependent variable D82 is defined as
follows: If the billing for ambulatory medical care exceeds the submission
threshold, it takes on the value of 1, otherwise it is O. Table 1 below
documents the construction of this dependent variable, along with some
distributional information. Table 2 contains analogous information concern
ing explanatory variables. Among other things, it shows that the three
insurers have a different age structure in that the share of members aged 45
to 54 varies between 12% with A and 27% with C.
196 P. Zweifel
Table 1. Dependent variable and auxiliary variables used for their construction, 19S2.
A B C (*)
Note. (*) Total sample, used in the estimate shown on the left hand side of Table 3.
Bonus systems in health insurance 197
Table 3. Probability for ambulatory care costs to exceed the THRESHOLDS defined 
Table 1, 1982.
Note. Intercepts not shown. Coefficients are estimated partial impacts on probability, derived
from multiplying the parameters of a logit regression by p(1  p), with p = average probability
(Theil, 1972, p. 169). *(**, ***) denote 0.05 (0.01,0.001) levels of statistical significance; t
values to be interpreted asymptotically. df: degrees of freedom. N: number of observations.
CONC: share of concordant pairs between predicted and actual observations.
198 P. Zweifel
Conclusion 4
Conclusion 5
Conclusion 6
Conclusion 7
Concluding remarks
The starting point of this contribution is the fact that there is an alternative to
the negative sanctions contained in conventional health insurance policies.
This alternative takes the form of rebate options and experiencerated
bonuses as offered by German private health insurers. From the consumer's
point of view, these new options may well be preferable to plans featuring
Bonus systems in health insurance 201
deductibles and coinsurance (conclusion 1). At the same time, rebates and in
particular bonuses are predicted to have a dampening impact on demand for
ambulatory medical care (conclusion 2). Moreover, the experiencerated
bonus is predicted to continually reduce demand even more than a roughly
comparable rebate offer (conclusion 3).
These predictions are subjected to empirical tests using insurance files
from three German private health insurers for the years 1981 and 1982.
Great care is taken to eliminate effects on accounted billings that merely
reflect the insured's decision not to submit a bill. Rather, estimated effects on
the billings distribution should reflect modifications of behavior, as studied in
the theoretical model. Insureds exposed to a rebate offer are found to exceed
a threshold value of almost 1400 DM (465 U.S. Dollars) for ambulatory
medical care with a lower likelihood than insureds having a conventional
policy, with age, sex, and supplementary hospital insurance held constant.
This dampening effect is even more pronounced among members of an
insurance that offers an experiencerated bonus (conclusion 4). While
seeking to save their rebate or bonus, insureds might jeopardize their future
health. Statistically, such behavior would give rise to a toothsaw pattern in
medical care cost (conclusion 5). On the other hand, these estimated impacts
could be due to a mere selfselection of risks. Under this premise, insureds
having either a rebate or a bonus option should have a particularly stable
pattern of utilization over time. Alternatively, a bonus might educate insureds
to become permanently good risks, resulting in very high intertemporal
stability (conclusion 6). An analysis of billings from two consecutive years
leads to rejection of the toothsaw pattern and selfselection hypotheses while
yielding some evidence in favor of an 'educational effect' of bonuses
(conclusion 7).
Thus, these insurance options with their positive rather than negative
economic incentives can be commended for their dampening impact on
health care cost. In view of the notorious financing problems of almost all
Western social health insurance schemes, experiences made by private
insurers with their innovative plans may well be of relevance to social health
insurance in countries such as Austria, Belgium, France, Germany, and the
Netherlands. Since selfselection always plays a certain role in insurance
markets where consumers have a choice, a conclusive comparison of plans in
terms of their impact on medical care costs would require an experiment of
the type of the health insurance study initiated by the Rand Corporation
(Manning et aI., 1981). Ever since Beck's (1974) study of the effects of cost
sharing on the poor, the notion that copayment may restrain demand for
medical care much more at lower than at higher income levels has been a
major concern to policy makers. Although this study is based on a high
income segment of the population, the stability of estimated relationships was
investigated by dividing the samples of insurers A and B into three broad
socioeconomic groups (Zweifel, 1988). Coinsurance rates, deductibles, and
rebates all turned out to have minimum effect on utilization among enrollees
202 P. Zweifel
of the uppermost group and maximum effect in the lowest group, which is
still very much middle class, of course. However, rebates and bonuses are
defined relative to an insurance premium, and to the extent that lower
income groups buy less insurance, financial incentives for refraining from
medical care consumption are scaled down along with income, which is not
true of deductibles and coinsurance rates. Moreover, these effects are small
compared to primary determinants of demand for medical care such as age
and sex. In conclusion, these innovations in health insurance merit serious
attention in the debate about future reforms of social health insurance.
References
Beck, R. G. (1974), The Effects of Copayment on the Poor', Journal of Human Resources
9(1),129142.
Bellmann, R. (1957), Dynamic Programming, Princeton, N.J.: Princeton University Press.
Brook, R. H. et al. (1983), 'Does Free Care Improve Adults' Health?' Results from a
Randomized Controlled Trial', New England Journal of Medicine 309(23), 14261434.
Cairns, J. A. and Snell, M. C. (1978), 'Prices and the Demand for Care', In Cu1yer, A. J. and
Wright, K. G. (ed.), Economic Aspects of Health Services, London: Martin Robertson, pp.
95122.
Dionne, G. and Eeckhoudt, L. (1983), Risk A version and Statedependent Preferences, Paper
submitted to the 10th Seminar of the European Group of Risk and Insurance Economists.
Rome.
Feldstein, M. S. (1973), 'The Welfare Loss of Excessive Health Insurance'. Journal of Political
Economy 81(1), 251280.
Fuchs, V. R. (1982), 'Time Preference in Health', In Fuchs, V. R. (ed.), Economic Aspects of
Health, Chicago: University of Chicago Press, pp. 93120.
Harrel, F. E. (1980), 'Procedure Logist', In SAS Supplementary User's Guide. Raleigh. N.C.:
SAS Institute Inc.
Intriligator, M. D. (1971), Mathematical Optimization and Economic Theory, Englewood
Cliffs: PrenticeHall.
Manning, W. G. et al. (1981), 'A Twopart Model of the Demand for Medical Care:
Preliminary Results from the Health Insurance Study", In Van der Gaag, J. and Perlman,
M. (eds.), Health, Economics, and Health Economics, Amsterdam: North Holland, pp.
103124.
Newhouse, J. P., Marquis, K. H. and Morris, C. N. (1981), 'Some Interim Results from a
Controlled Trial of Cost Sharing in Health Insurance', New England Journal of Medicine
305(25),15011528.
Theil, H. (1972), Statistical Decomposition Analysis, Amsterdam: North Holland.
Uleer, C. (1981), 'Erfahrungen der privaten Krankenversicherung mit der Selbstbeteiligung
(Experiences with Cost Sharing madc by Private Hcalth Insurcrs),. Pharmazeutische
Industrie 43(11),10701076.
Zweifel, P. (1985), 'Experiences with Rebates for No Claims in Health Insurance', In Goppl,
H. and Henn, R., 3. Tagung fiber Geld, Banken und Versicherungen, Vol. II, Karlsruhe:
Verlag VVW, pp. 15191534.
Zweifel, P. (1988), 'Premium Rebates for No Claims: The West German Experience', In
Frech, H. E. III (ed.), Health Care in America, San Francisco: Pacific Research Institute,
pp.323352.
Microsimulation of the costs of the health system in
the Federal Republic of Germany
RALPH BRENNECKE
lnsfifutfiir Soziale Medizin, Freie Universitiit Berlin
During the last years the development of the cost structure of the German
health care system is not quite different from that of other European
countries. Likewise one finds in all countries tendencies to reduce or stabilize
the increase rates of the costs and tendencies to get more information about
what's going on within the field of health care.
Unlike the American or British health system, the German system can be
characterized by two general restrictions: there exist detailed legal provisions
with regard to the scope, structure and type of care of the health service, and
the bulk of the services is taken up by payment in kind.
Two books, the Social Law Code (Sozialgesetzbuch) and the Imperial
Insurance Code (Reichsversicherungsordnung) according with a lot of
administrative regulations, determine:
who has to be insured at the Statutory Health Insurance scheme (SHI,
Gesetzliche Krankenversicherung)
which types of services may be provided and to what extent
according to which criteria the remuneration for the services should be
calculated
who is permitted to provide ambulatory (outpatient) and stationary
(inpatient) services
It is also ruled that the patient receives services as a consequence of his
being insured in the SHI and that the physician, the hospital and the
pharmacist must settle their accounts through interconnected collective
institutions with the SHI. The patient is usually kept in ignorance of the costs
which have arisen from his treatment.
This constellation results in considerable deficiencies in information,
which become especially noticeable in times of general economic stagnation
or recession. Due to the construction, that the costs of the health services
arise from the individual use made of those services, both the SHI and the
politicians only become aware of the exact extent of the total expenditure of
the SHI some considerable time after the end of the year. On the other hand,
the insurance contributions are levied in advance from the insured. Secondly,
the transfer income distribution for groups of people selected according to
certain criteria, cannot be simply conveyed since the patients do not know
the costs of their services. Thus, direct econometric analyses of the use made
of the health services, with prices and services, are not feasible.
Nevertheless, considerable interest is being shown in the development of
models of the health system both by the administration and politicians as well
as by academics. This is due not only to the difficulties with regard to the
planning of the health expenditure and the definition of adequate contribu
tions but also because of other problems which cannot be gone into in any
detail here. The Ministry of Work and Social Affaires (Bundesministerium
fUr Arbeit und Sozialordnung) engaged a group of experts in 1970 to
envolve a macromodel for the calculation of health expenditure. The study,
with the data and methods then at its disposal, was unable to come to a
satisfactory conclusion. In 1975 the SRI commissioned the German Institute
for Economic Research (Deutsches Institut fUr Wirtschaftsforschung) to
examine the possibilities of creating a model for the ascertainment of the
SRI's expenses. The Institute concluded that a great number of influential
factors, especially on an individual level, had an effect on expenditure and for
that reason the development of a model seemed, at that time, to be impos
sible. In 1976, a regional model of demand, developed on the system
dynamicbasis, was published by Klimpke (1976). However, the model was
not directly suited to the estimation of costs. In the meantime an other
macromodel was constructed by Camphausen (1983).
One might say that the quintessence of previous studies reveals that there
is almost no point in developing a model of the German health system based
on macrodata. Changes in the law and shifts in structure can be taken into
consideration only with great difficulty. A microanalytically oriented formu
lation on the other hand, would appear to be more promising. Here however,
the quality of the results depends decisively on the available data. Research
in the Federal Republic has to face several problems in this respect, in
connection with the data protection laws. Problems, which lead one to search
for second best solutions.
Interme
I<_Idiate stor.
education
population
period t employment
14.1 Income
tax and
transfers
heal th care
1++""*+1 financing
determine
heal th estimate determine
costs 01
r      + i in~~~~unsce length of
Incapaci ty
duration
J
yes
household determine household
with all typo 01 wltn the
variables health incapacity var iables
of all insurance 01 work? no 01 all
persons funds persons
at por iod at periOd
t.~
r
II estimate
length of
stay and
no
costs
yes
no
estimaw estimate
frequencies costs ot
of vlsi 18, prescrip
and costs tions
01 care
previous period, it is necessary to simulate the type of sickness fund and the
insurance status, because the treatments and the type of fee vary among
different insurances.
The health insurance scheme of the FRG consists  with respect to costs
 of five parts. The most important part is the SHI, which covers about 90%
of the population (Stone, 1980). The SHI is very dismembered and com
posed of about 1200 independent sickness funds (RVOfunds) and 15
substitute funds (Ersatzkassen). Each fund calculates their contributionrate
on the basis of their own situation of expenditure. Within the RVOfunds
there are four types of funds: local funds (Ortskrankenkassen), factory funds
(Betriebskrankenkassen), guild funds (Innungskrankenkassen) and agricul
tural funds (Landkrankenkassen). Miners (Bundesknappschaft) and mariners
(Seekasse) have entirely seperate funds. The four types of RVOfunds have
linked together to State Associations of the funds (Landesverbiinde, 11 for
each type), the State Associations to one Federal Association for each type
of fund (Bundesverbiinde).
About 7% of the population, mostly civil servants and selfemployed, but
also some employees, are covered by private health insurance. Only less than
1% are not covered at all. The remaining population is covered by special
schemes, e.g. armed forces, civil servant state subsidy or government grant.
On the other hand two additional systems are present, the Statutory Accident
Insurance (SAl) and the pension insurance funds. All health care, which is
caused by an accident will be payed by the SAl, rehabilitation mostly is
financed by the pension insurance funds.
Statutory Health Insurance and the pension insurance fund are financed
by contributions in the form of earmarked payroll taxes (Geissler, 1980).
50% of the contributions are payed by the employer, the rest by the insured.
Depending on their working status employees are obliged to be insured in
the SHI. Blue collar workers are generally insured, white collar workers up to
an annually changing wage ceiling. Family members of the insured persons
are free of charge as long as they are not employed. Until 1982 pensioneers
were covered free of charge, now they have to contribute to the SHI with an
increasing proportion of their pension. The contributions to the SAl are
totally payed by the employers, each employee is automatically insured.
After having determined the insurance for each person, the second
estimation sequence will cover strongly aggregated illness patterns, for each
individual. From the combination of complaints and socioeconomic vari
ables we estimate the number of visits to various consultants or specialists, as
well as to the family doctor and to the dentist (compare Section 4).
The introduction of costinformation on the basis of received treatments is
not that good as it should be. When one studies the structure of the data of
the German health system it becomes clear why we are forced to select a
second best solution for the estimation of costs. The SHI data are available in
computerized form only for some sickness funds, where the conversion to
E.D.P. has already been completed. Nevertheless, the data contain detailed
208 R. Brennecke
Secondly it is well known that different factors will influence the primary
contact with a physician and the following visits. Our data base contains a lot
of variables, from which the used are shown in Table 1. The types of
variables are very similar to comparable researches (Wan and Soifer 1974,
1975; Manning et aI., 1981) and comments at this point are not necessary.
For testing the hypothesis of different influences we used a binomial logit
estimator to calculate contacts and regression analysis to estimate the
number of visits.
A next problem was to define sexspecific reference groups with a satisfac
tory number of cases and a meaningful combination of variables. After
various tests we found for men the group
married and living in a twopersonshousehold in a region with up to
50,000 inhabitants, fully employed and insured at the local funds, low
degree of education and average state of subjective health and no illness
patterns of the list.
Accordingly the reference group for women was found as
married and living in a twopersonshousehold in a region up to 50,000
inhabitants, too, but out of work, low degree of education and insured in
local funds due to their husbands (without own contributions to the
Abbrev. Meaning
Table 1. (continued)
Abbrev. Meaning
Variable GENPAR INTERN EAR NOS DERMAT RADIOL ORTHO UROLO OCULI DENT
CASES 18916 18916 18916 18916 13357 7126 7126 7126 18916
REFGRP 146 146 146 146 83 37 37 37 124
RH0 2 0.12 0.12 0.06 0.25 0.08 0.11 0.23 0.05 .18
DF 38 38 38 38 38 38 38 38 39
Table 2. (continued) ~
t:J;j
Variable GENPAR INTERN EAR NOS DERMAT RADIOL ORTHO UROLO OCULI DENT ~
~
~
("I>
t")
WHICOL 0.13 2.2 0.11 1.1 0.05 0.3 0.29 1.3 0.20 1.3 0.13 0.5 0.20 0.6 0.36 1.6 0.30 3.7 8""
BLUCOL * * * * * * * * *
PENSION 0.25 1.2 0.08 0.2 1.84 2.9 1.95 2.4 0.18 0.3 0.70 0.8 12' 0.0 0.01 0.0 0.16 0.5
OTHER SO 0.10 1.1 0.26 1.1 1.07 2.8 0.05 0.2 0.19 0.8 0.05 0.1 13. 0.0 0.03 0.1 0.29 2.2
Enabling factors
SHICOM * * * * * * * * *
SHIPEN 0.08 0.5 0.12 0.5 0.83 1.7 1.48 2.0 0.57 1.3 0.82 1.4 0.62 0.9 0.03 0.1 0.03 0.1
SHIVOL 0.10 1.6 0.14 1.4 0.23 1.5 0.20 0.9 0.22 1.2 0.30 1.1 0.14 0.4 0.36 1.6 0.13 1.6
SHIFAM 0.53 3.2 0.38 1.3 0.01 0.0 0.21 0.5 0.06 0.1 1.93 2.6 0.43 0.4 0.24 0.5 0.14 0.7
PRIVIN 0.34 3.6 0.43 3.0 0.08 0.3 0.56 1.8 0.02 0.1 0.83 2.2 0.34 0.6 0.38 1.1 0.34 2.9
NOT IN 0.63 3.0 0.51 1.7 0.20 0.4 0.61 0.5 0.34 0.5 0.60 0.6 1.19 0.8 0.50 0.8 0.12 0.4
LOG FUN * * * * * * * * *
FACFUN 0.09 1.9 0.12 1.4 0.01 0.1 0.12 0.7 0.18 1.4 0.17 0.8 0.15 0.6 0.12 0.7 0.06 0.9
OTHFUN 0.06 1.0 0.38 3.4 0.20 1.3 0.26 1.3 0.24 1.7 0.31 1.2 0.16 0.4 0.10 0.5 0.09 1.1
SUB FUN 0.03 0.6 0.24 3.0 0.08 0.7 0.14 0.9 0.05 0.4 0.03 0.2 0.18 0.7 0.26 1.6 0.18 2.6
COM SMA * * * * * * * *
COM AVE 0.15 3.8 0.48 6.2 0.50 4.3 0.55 3.6 0.10 0.8 0.62 3.1 0.12 0.5 0.24 1.6 0.14 2.4
COMLAR 0.30 7.3 0.80 11.0 0.52 4.4 0.70 4.6 0.24 2.1 0.74 3.6 0.14 0.6 0.44 2.9 0.26 4.3
Variable GENPAR INTERN EAR NOS DERMAT RADIOL ORTHO UROLO OCULI DENT
NEURO 0.08 2.0 0.13 2.1 0.01 0.0 0.11 0.9 0.31 3.1 0.17 1.1 0.21 1.1 0.37 3.0 0.16 2.9
~
CIBCU 0.61 16.0 0.53 8.3 0.10 1.0 0.16 1.2 0.37 3.6 0.34 2.1 0.07 0.3 0.13 1.0 0.19 3.3 r:)'
RESDIR 0.44 13.0 0.07 1.2 0.89 9.8 0.02 0.2 0.13 1.4 0.18 1.3 0.13 0.7 0.05 0.5 0.Q1 0.1
DIGEST 0.39 9.9 0.42 7.1 0.23 2.5 0.09 0.7 0.61 6.5 0.13 0.8 0.07 0.4 0.22 1.8 0.07 1.2 ....'"~
~
KIDNEY 0.33 4.5 0.42 4.7 0.10 0.7 0.16 0.8 0.73 5.5 0.22 0.8 2.75 14.0 0.33 1.8 0.08 0.8 0:::
CUTAN 0.09 1.1 0.22 1.7 0.19 1.1 3.38 30.0 0.25 1.3 0.12 0.4 0.11 0.3 0.21 1.0 0.25 2.5 i:i'
SKEMUS 0.23 6.2 0.10 1.7 0.18 1.9 0.09 0.7 0.30 3.0 1.51 9.6 0.03 0.2 0.09 0.7 0.10 1.9 6'
;:::
TEETH x x x x x x x x 2.34 46.0
V GOON 0.33 5.9 0.32 2.5 0.27 1.6 0.22 1.1 0.42 2.1 0.77 2.2 1.09 1.8 0.22 1.0 0.06 0.8 .Q,
01< 01< 01< 01< So
GOODH * * * * * ~
AVERAH 0.38 9.8 0.55 7.4 0.28 2.6 0.18 1.3 0.25 2.1 0.73 4.2 0.02 0.1 0.20 1.5 0.08 1.3
BADH 0.79 13.0 0.84 8.9 0.53 3.5 0.22 1.1 0.65 4.4 1.13 4.7 0.30 1.0 0.36 1.9 0.04 0.5 ~
1;:1'
VBADH 1.04 11.0 1.00 8.4 0.77 4.0 0.19 0.7 0.80 4.2 1.46 5.0 0.27 0.7 0.29 1.2 0.40 2.9
.Q,
So
~
* Variable of the reference group. ;:::r.
x Variable not defined. ~
, Coefficient results from small cell frequencies. §:
Source: Logit estimations from 47,000 questionaires of 24 surveys during the years 1970 to 1977. The analysis was done by Pia Zollmann.
~
toO
~
~
tv
.....
W
"{'Ihle 3. Results of the logitestimation of contacts with physicians. women older than 13 years. N
.......
.j;:.
Variable GENPRA INTERN GYNACO EARNOS DERMAT RADIOL ORTHO UROLO OCULI DENT
;:t1
t:l:l
CASES 27H60 27H60 27H60 27860 27H60 19740 10490 10490 10490 27H60 ~
REFGRP 140 140 140 140 140 93 41 41 41 126 ~
;::
RHO' 0.11 0.11 0.14 O.OH 0.24 0.07 0.10 0.19 0.06 0.16
~
DF 3H 3H 3H 3H 3H 38 38 38 38 39 8"
COEFFTV COEFFTV COEFFTV COEFFTV COEFFTV COEFFTV COEFFTV COEFFTV COEFFTV COEFFTV
CONST 0.53 5.3 5.6 31.0 3.45 24.0 4.61 17'.0 4.33 14.0 4.64 15.0 6.23 11.0 9.46 9.9 3.05 8.4 2.97 22.0
Time factor
JEAR 0,0\ 0.9 0.06 6.3 0.11 14.0 0.01 0.2 0.05 2.5 0.06 2.4 O.OS 1.5 0.24 2.9 0.01 0.4 0.05 5.6
Predisposing factors
AGE 0.01 3.0 0.05 6.7 0.08 11.0 0.01 0.9 0.02 1.6 0,0\ 0.4 0.04 2.1 0.08 2.5 0.03 2.4 0.03 4.9
AGE*2 0.02 5.0 0.04 5.6 0.13 15.0 0.01 0.3 0.01 0.4 0.02 1.4 0.06 2.6 0.09 2.4 0.03 2.7 0.05 6.9
SINGLE 0.02 0.6 0.04 0.7 0.16 3.0 0.09 0.9 0.02 0.1 0.10 0.9 0.05 0.3 0.21 0.8 0.08 O.S 0.21 3.7
MARRIED
OTHER II 0.06 1.3 0.09 1.1 0.26 3.9 O.OS 0.6 0.03 0.2 0.08 0.6 0.03 0.2 0.10 0.3 0.19 1.4 0.26 3.9
EDUCAL *
EDUCAA 0.14 4.0 0.54 10.0 0.17 4.0 0.27 3.1 0.36 3.4 0.20 2.1 0.71 5.3 0.24 1.0 0.44 4.6 0.45 10.0
EDUCA II 0.32 5.0 0.75 8.6 0.36 5.0 0.33 2.3 0.64 3.9 0.13 O.S 0.12 0.4 0.46 1.0 0.54 3.3 0.49 6.6
FULEMP 0.24 2.1 0.23 1.2 0.53 3.7 0.05 0.2 0.69 1.9 0.33 1.0 0.2S 0.4 0.98 1.0 0.55 1.4 0.07 0.5
PAREMP 0.22 2.0 0.16 0.8 0.25 1.8 (l.O9 0.3 0.44 1.2 0.24 0.7 0.57 0.8 0.86 0.9 0.63 1.5 OJ I 0.8
JOBLESS 0.13 0.7 0.12 0.4 0.30 1.4 0.16 0.4 0.30 0.5 0.04 0.1 0.21 0.2 0.42 0.3 0.22 0.4 0.03 0.1
NOTEMP
FARMER 0.33 2.3 0.44 1.3 0.34 1.3 14.' 0.1 0.45 0.6 0.18 0.3 0.21 0.2 13.' 0.0 0.04 0.1 0.63 2.6
SEL EMP 0.34 2.0 0.22 I.l 0.06 0.4 0.27 0.9 0.41 0.9 0.15 0.4 0.36 0.5 1.86 1.6 0.51 1.1 0.07 O.S
CIVILS 0.21 1.4 0.07 0.3 0.57 3.2 0.57 1.7 0.21 0.5 0.48 1.2 0.34 0.4 0.04 0.0 0.44 0.9 0.04 0.2
Table 3. (continued)
Variable GEN PRA INTERN GYNACO EAR NOS DERMAT RADIOL ORTHO UROLO OCULI DENT
WHI COL 0.18 2.1 0.12 0.7 (UI 2.7 0.08 0.3 0.01 0.0 0.19 0.7 0.10 0.2 0.94 1.6 0.32 1.0 0.05 0.4
BLU COL 0.10 1.1 0.02 0.1 (UO 0.8 0.02 0.1 0.04 0.1 0.28 0.9 0.02 0.0 1.0 1.6 0.13 0.4 0.03 0.3
PENSION 0.22 3.1 0.07 0.7 0.01 0.1 0.07 0.4 0.16 0.6 0.10 0.5 0.21 0.6 0.43 0.7 0.34 1.7 (J.09 0.8
~
OTHER SO * * r)'
~
Enabling factors '"§.
SHICOM 0.01 0.1 0.29 2.2 0.16 1.5 0.18 0.8 0.62 2.1 0.01 0.0 0.20 0.5 0.36 0.4 0.20 0.7 0.02 0.2 :;::
1.6 0.03 0.2 0.14 1.2 0.05 0.2 0.49 1.5 0.08 0.1 0.25 1.4 0.13 1.2 !:)
SHI PEN 0.08 1.2 0.16 0.7 0.30 ...
SHI VOL 0.20 2.1 0.23 1.7 0.04 0.4 0.38 1.6 0.47 1.7 0.13 0.6 0.08 0.2 0.09 0.2 0.01 (J.O 0.06 0.5 C·
;:::
SHIFAN
PRIV IN 0.56 8.7 0.68 7.3 0.23 2~ ~06 0.4 0.10 0.5 0.18 1.0 0.29 1.1 0.19 0.4 0.01 0.1 0.35 4.4
<Q.,
NOTIN 0.82 4.7 0.25 0.7 0.20 Oh ~25 0.6 14: 0.1 1.48 1.3 0.45 0.4 12.0 0.0 0.51 0.7 0.21 0.7
s.
('1\
(')
LOCFUN
FAC FUN 0.01 0.1 0.14 2.1 0.11 2.1 0.11 1.0 0.13 1.0 0.06 0.5 0.15 0.8 O.S2 3.1 0.02 0.2 0.n6 n.9 ~
t;;"
OTH FUN 0.06 1.2 0.02 0.3 OJ)7 I.l 0.19 1.4 0.09 0.5 OJl3 0.2 0.05 0.2 0.22 0.6 0.08 0.6 (LOS 0.7
0.3 O.2S 5.9
<Q.,
SUB FUN 0.10 2.8 0.46 8.2 0.37 8.3 0.37 4.0 0.06 0.6 0.19 1.9 0.28 1.9 0.53 2.2 0.03
COM SMA *
s.
('1\
COM AVE 0.17 5.3 0.50 8..1 0.27 5.9 0.31 3.1 0.29 2.6 0.18 1.8 0.20 1.2 O.OS 0.3 0.18 1.7 0.1 1 2.3 ::::
COM LAR 0.44 13.0 0.87 15.0 0.61 14.0 0.66 6.9 0.44 3.9 0.32 3.2 0.49 3.0 (US 0.7 0.30 2.8 0.19 4.2 ~
:::;:
::::
Morbidity (need factors) ~
METABO 0.67 8.5 0.37 4.5 (1.07 0.6 0.14 O./l 0.06 0.3 0.10 0.6 0.35 1.3 O.S5 1.2 0.34 2.4 0.10 0.9 ~
~
N
VI
N
I'
0'1
~
b::I
Table 3 (Comiflued) ~
;:s
;:s
~
("')
Variable GEN.PRA INTERN GYNACO EAR NOS DERMAT RADIOL ORTHO UROLO OCULI DENT
~
NEURO 0.07 2.3 0.19 3.9 0.21 5.3 0.20 2.5 0.17 1.7 0.49 5.3 0.08 0.6 0.01 0.0 0.35 4.0 0.19 4.7
CIRCU 0.61 21.0 0.59 12.0 0.11 2.8 0.26 3.1 0.13 1.4 0.12 1.3 0.13 1.0 0.28 1.4 0.23 2.6 0.11 2.7
RESPIR 0.33 12.0 0.10 2.3 0.Q3 0.9 1.01 13.0 0.01 0.1 0.28 3.7 0.19 1.6 0.14 0.8 0.01 0.1 0.11 3.0
DIGEST 0.21 6.9 0.35 8.0 0.19 4.8 0.05 0.6 0.13 1.4 0.43 5.5 0.14 1.1 0.01 0.1 0.26 3.2 0.10 2.5
KIDNEY 0.21 4.9 0.17 3.0 1.10 25.0 0.15 1.6 0.17 1.4 0.24 2.6 0.24 1.5 2.61 13.0 0.16 1.5 0.01 0.1
CUTAN 0.06 1.0 0.21 2.4 0.06 0.8 0.49 4.1 3.26 38.0 0.14 1.0 0.09 0.4 0.15 0.4 0.37 2.6 0.19 2.5
SKEMUS 0.12 3.8 0.19 4.1 0.06 1.5 0.38 4.8 0.09 0.9 0.35 4.2 1.27 9.1 0.27 1.3 0.42 5.0 0.17 4.2
TEETH x x x x x x x x x 0.02 50.0
V GOOD 0.49 8.9 0.33 2.8 0.12 1.9 0.54 3.0 0.21 1.3 0.33 1.6 0.32 1.2 0.36 0.8 0.56 2.8 0.09 1.4
GOODH * * * * * *
AVEPAH 0.35 11.0 0.38 6.3 0.07 1.7 0.11 1.1 0.07 0.6 0.25 2.3 0.50 2.9 0.46 1.7 0.13 1.3 0.09 2.0
BADH 0.71 16.0 0.73 10.0 0.01 0.2 0.38 3.2 0.17 1.2 0.79 6.4 0.98 4.9 1.00 3.3 0.10 0.8 0.11 1.8
VBADH 0.72 10.0 0.93 10.0 0.14 1.5 0.57 3.7 0.17 0.8 1.09 7.1 1.61 7.0 1.14 3.1 0.13 0.7 0.44 4.2
Variable GENPRA INTERN EAR NOS DERMAT RADIOL ORTHO UROLO OCULI DENT
CASES 9547 1883 703 481 668 277 172 448 3286
REFGRP 209 32 14 7 12 4 3 5 51
FSTAT 100.2 18.9 3.4 4.9 5.5 4.7 3.6 3.9 19.4
R2 0.25 0.17 0.14 0.12 0.18 0.21 0.11 0.06
Time factors
.;:s'"
>:::
YEAR 0.05 2.7 0.06 2.1 0.03 0.6 0.11 1.6 0.08 2.1 0.29 0.9 0.09 0.3 0.06 0.5 0.01 0.5 $S"'
....
QUART2 0.05 o.s 0.15 0.7 0.37 0.9 0.02 0.1 0.14 0.6 2.82 1.1 0.26 0.1 1.85 1.6 0.13 0.9 0·
;::s
QUART3 0.21 2.4 0.05 0.3 0.92 2.6 0.03 0.1 0.41 1.5 0.66 0.3 1.19 0.6 1.66 1.7 0.19 1.6
QUART4 0.22 2.5 0.49 2.5 0.56 1.6 0.90 2.3 0.01 0.1 1.40 0.6 0.70 0.3 1.60 1.4 0.14 1.2 ~
S.
~
Predisposing factors r,
C
AGE '"/;;'
AGE*2 0.01 2.5
LOG AGE 0.60 3.5 3.24 2.9 0.88 2.0 ~
SINGLE 0.37 1.8 0.25 2.0 S.
~
MARRIED * * * * * ~
* * * ~
OTHER II 0.52 2.3 0.24 1.7 1.69 2.7 1.19 1.8 ~
"lid,l" 4. (continlled)
?'
b:l
Variable GENPRA INTERN EAR NOS DERMAT RADIOL ORTHO UROLO OCULI DENT ~
::!
::!
n:.
HOTEMP 0.59 9.0 0.54 2.0 r"l
;>;
FARMER n:.
SELEMP 0.68 1.9 1.09 2.3
CIVILS 0.23 2.2 0.63 1.8
WHICOL
>I< >I< >I< >I< >I< >I<
BLUCOL * * *
PENSION 0.31 2.4 1.02 2.4 1.31 2.4
OTHESO 0.23 3.0 0.55 1.7 0.23 1.9 5.63 2.2 0.77 3.3
Enabling factors
>I< >I< >I< >I< >I< >I< >I< >I< >I<
SHICOM
SHIPEN 0.63 1.9 0.80 1.8
SHIVOL 0.24 3.1 0.56 2.1 0.33 2.4 0.94 2.1
SHIFAM 0.28 2.1 0.69 2.0 7.61 3.2
PRIVIN 0.47 4.5 0.31 1.9 0.60 1.9
NOT IN 2.36 3.0 0.69 1.9
>I< >I< >I< >I< >I< >I<
LOCFUN * * *
FACFUN
OTHFUN 0.15 2.0 0.33 1.7
SUB FUN 0.18 2.9 0.25 1.9 0.74 1.9
>I< >I< >I< >I< >I< >I<
COM SMA * * *
COMLAR 0.10 2.0 0.58 2.6 0.80 3.6 0.26 1.8
PHPOR 0.20 3.2 0.15 1.8
Table 4 (Continued)
Variable GENPRA INTERN EAR NOS DERMAT RADIOL ORTHO UROLO OCULI DENT
CONST 4.62 5.6 4.58 2.6 6.49 6.5 3.81 1.0 4.81 1.2 1.75 0.7 37.2 1.5 3.89 0.2 6.54 1.0 3.10 2.3
Time factors
YEAR 0.03 2.8 0.02 1.0 0.03 2.4 0.06 1.3 0.05 0.9 0.02 0.5 0.46 1.4 0.12 0.4 0.07 0.8 0.03 1.5
QUART2 0.12 1.4 0.09 0.5 0.17 1.8 0.64 1.8 0.70 1.7 0.09 0.4 3.14 1.2 2.71 1.1 0.67 0.8 0.08 0.6
QUART3 0.14 2.0 0.37 2.4 0.01 0.0 0.31 1.0 0.12 0.3 0.16 0.6 3.09 1.5 4.94 2.4 0.56 0.9 0.10 0.9
QUART4 0.28 3.7 0.43 2.5 0.01 0.1 0.46 1.4 0.21 0.6 0.03 0.1 3.73 1.7 6.51 2.7 0.33 0.5 0.01 0.1
Predisposing factors
AGE
AGE*2
AGE' 1 0.01 4.4 0.01 2.1 0.Ql 2.5 om 2.1 0.01 4.1 0.01 2.9
LOG AGE 1.54 5.7 2.05 2.6 0.98 2.0 1.16 2.7 4.12 3.6 0.44 2.6
SINGLE 0.23 2.7 1.36 3.2
MARRIED • * * * * * * *
OTHERH 1.23 2.3 0.30 1.7
ADULT 0.09 3.1 0.11 2.5 0.13 2.0
CHILD 0.04 1.8 0.36 1.8
EDUCAL * * * * * * •
EDUCAII
FULEMP 0.36 2.2 0.29 2.9 1.26 1.8
Table 5. (continued)
Variable GENPAR INTERN GYNACO EAR NOS DERMAT RADIOL ORTHO UROLO OCULI DENT

Table 5. Results of the regression offrequencies of visits women older than 13 years. N
N
N
Variable GENPAR INTERN GYNACO EAR NOS DERMAT RADIOL ORTHO UROLO OCULI DENT
?:l
t\;
NEURO 0.10 2.3 x
~
CIRCU 0.45 9.8 0.23 2.2 0.38 1.8 ;S
;S
0.68 1.8 x (1)
(")
PESPIP 0.13 2.8 0.87 2.5 x
~
DIGEST 0.19 4.3 x
KIDENY 0.13 2.2 0.55 11.0 1.31 2.7 0.81 2.2 x
CUTAII 0.50 1.8 0.85 4.5 1.37 2.0 x
SKEMUS 0.19 4.3 0.22 2.4 x
TEETH x x x x x x x x x 0.90 14.0
V GOOD 0.55 5.2 0.55 2.1 0.34 3.9 0.77 1.8 0.67 1.8 1.32 1.7 0.27 2.6
GOOD II 0.34 6.4 0.63 5.0 0.10 1.8 1.11 2.3 1.01 2.2
AVERA II * * * *
BAD II 1.00 19.0 1.05 9.5 0.21 3.0 0.56 2.1 0.40 3.7 0.73 1.7 0.32 3.5
VBADII 2.06 25.0 1.54 10.0 0.59 5.1 0.61 2.0 0.88 2.1 0.57 3.9 1.47 2.8 0.86 1.7 0.38 2.0 0.73 4.2
insurance). They should state their subjective health as average, too, and
should claim no illness patterns of the list.
The logit results of contacts with physicians of various specialities are shown
in Tables 2 and 3. As expected the strongest influence factors on primary
contacts to a physician are illness patterns and the subjective state of health.
Besides these factors the other variables are not so important, but one can
find additionally significant coefficients. Age for instance is  varying among
the different specialities  important, but the employment status scarcely.
The social status of men is nearly irrelevant, that of women seems to have an
influence on contacts with general practioners and gynacologists.
As Tables 4 and 5 show, the results of the frequencies of visits are a little
bit different from those of contacts. The correlation coefficients (multiple R2)
of all regressions are relatively low. Besides the effect of microdata regres
sion (Yett et al., 1979; Dworschak and Wagner, 1983; Galler, 1983) one
reason may be that variables, which represent physicianinduced demand, are
scarcely included at all. The R2 of the regression of frequencies of visits to
oculists with 0.11 could· show this particulary clearly. Normally one visit
is sufficient to establish the state of the eyes and to be issued with new
spectacles. Only diseases of the eyes necessitate subsequent visits, and the
oculist requests the patient to return for a further visit. We are unable to
include variables which represent this kind of demand.
For all specialists specific illness patterns are generally significant, the
subjective state of health only in some regressions. The age and the contact
with other specialists are significant factors, too.
An interpretation of all coefficients of our analysis would be beyond the
scope of the paper (but compare Zollmann and Brennecke, 1984). Our
ongoing work will be concentrated on introducing physicianpopulation
ratios into the logit analysis and on combining the results with average costs
per visit, as described below.
References
Synthesis
Segmentation and classification.
An application to patients' risk estimation
1. Introduction
2. Notation
Let X be the set of individuals, for simplicity (without loss of generality) let
us consider a problem of two classes WI and W2, with binary features Qi,
j= 1. . . p.
Thus, the endogenous variable Q, and the exogenous variables Qi are the
following mappings:
Q:X  {WI' w 2}
j= 1, .. . ,p,
where q{, q~, denote the two possible states of Qi.
In the following, for notation simplicity, we will not make any distinction
between the label W1 and the subset of X which are defined as follows:
{x E X: Q(x) = wd.
The training step consists in recording for a subset T of X the following
data:
(Qj(x),j = 1 .. . p; Q(x)), 'Vx E T.
The problem is to classify individuals y, y fE T, by observing as few variables
Qj as possible.
3. An arborescent procedure
3.1. Description
Figure 1.
Segmentation and classification 229
I Card( T) = 200
card(T n Wi) = 100, i = 1,2
! q{ is denoted by '0' Vj
q~ is denoted by '1' Vj
To each non terminal node we associate a feature Qi, and this association
aims to create new nodes s, such that one of the two numbers n;, n;
is near
to zero.
In that toy (ideal) example, the five terminal nodes contain only individ
uals belonging to one of the two classes.
Such a method clearly provides a description of the classes, for instance
class W 2 coincides on T with the individuals x such that:
or
In accordance with the previously defined objectives (3.1, 3.2) the terminal
nodes s of the tree should verify the following property:
{one of the two numbers n!, n;, is equal to zero, or near to zero.}
This suggests using all the exogenous variables Qj, j = 1 ... p, and thus
considering the tree with 2 P terminal nodes. Unless p is small or the size of
the training set is fairly large, this approach does not work. The 'quality' of
n~/(n! + n;) as an estimator of p(OJ/x,), as measured for example by a
confidence interval, is directly proportional to the (square root of the) sample
size (n! + n;), thus on average inversely proportional to the depth of s in the
tree.
Consequently, we have to prevent the exponential increase in the number
of nodes.
We propose a sequential procedure, which requires the choice of two
criteria:
we have to decide on the conditions which will define a terminal node,
 we have to decide on the way of associating an exogenous variable to a
non terminal node.
For the last choice, we will use Infonnation Theory. Let us consider an
exogenous variable Q, which has not yet been used between the root of
the tree and a given non terminal node s. We compute (see 3.4) a measure
I(Q, s) of the infonnation conveyed by Q at the node s (note that I(Q, s) is
always positive).
We choose for s the exogenous variable Q* which maximises I(Q, s)
among all the possible Q. We will say that a given node s is terminal if one
the two following conditions is fulfilled:
i) I(Q*,s) < e,where eis a given positive threshold,
ii) n! + n; ~ t, where t is a given integer.
These rules are clearly motivated by the concern of non increasing the
number of nodes without a significant infonnation gain.
2 1
h(s) = L f(w/Xs) log   
iI f(w/Xs)
Figure 2.
232 I.P. Auray et al.
The toy example of Figure 3a shows a difficulty which may accur with the
arborescent algorithm described in the previous section.
The nodes S3' S4' S5 are terminal (the parameter t was equal to 10), and the
decision risk about the class of an individual y belonging to Xs3 or Xs4 is very
high. This suggests 'fusing' the nodes S3' S4 (Figure 3b), and in so doing,
creating a new node G which is not necessarily terminal.
Thus we can imagine that there exists a feature Q operating the segmenta
tion of the node G (Figure 3b), which leads to terminal nodes G" G2 , the
associated decision risk of which are quite small with regard to the decision
risk associated to the nodes S3' S4. Thereby we have transformed the
arborescent structure into a lattice structure, and the profiles associated to
the terminal nodes will be described by means of the operators 'AND' and
'OR'. For instance the characteristic subset associated to the node G, (Figure
3) in T, is the following:
TGJ = {x E T: (Ql(X) = 0 AND Q(x) = 1) OR
(Ql(X)= 1 AND Q2(x)=OAND Q(x) = I)}
For consistency a unique information measure must be used for the two
following operations:
 creation of new nodes ('splitting'),
..
Q
Fig 3.a Fig 3.b
Figure 3.
Segmentation and classification 233
f(w/Xs ) =
n: + 1
n; + n; + 21
where 1 is a positive parameter [8].
5. Experimental results
The center for burnt patients of the Hopital E. Herriot (Lyon, France) has
requested a statistical approach to septic risk in burnt patients [6].
Three possible evolutions have been distinguished for burnt patients:
alive without infection (this type of evolution will correspond to class w),
 alive with infection (class ( 2 ),
S
1
S 5
 FUSING a
Figure 4.
234 f.P. Auray et al.
Figure 5.
Segmentation and classification 235
References
[I] Auray, J. P., Duro, G., Terrenoire, M., Tounissoux, D. and Zighed, A. Un logiciel pour
une methode de segmentation, Revue Inf. SciHum., n° 64, pp. 6478.
[2] Daroczy, D. (1970), 'Generalized Information Functions', Information and Control 16,
16.
[3] Devijver, P. and Kittler, J. (1982), Statistical Pattern Recognition, Prentice Hall.
[4] Duda, R. and Hart, P. (1973), Pattern Classification and Scene Analysis, Wiley, New
York. .
[5] Fu, K. S. (1982), Syntactic Pattern Recognition and Applications, Printice Hall.
[6] Marichy, J., Buffet, G., Zighed, A. and Laurent, Ph. (1984), Early Detection of Scepticemia
in Burnt Patients, Actes 3rd int. Conf. on Syst. Sci. in Health Care, Munich, pp. 505
508.
[7) Tounissoux, D. (1981), 'Processus sequentiel adaptatif de reconnaissance de forme pour
l'aide au diagnostic', These Lyon 1.
[8J Zighed, A. (1985), 'Methodes et outils pour les processus d'interrogation non arbores
cents', These Lyon 1.
A general equilibrium model of health care
MANAS CHATTERJI
State University of New York at Binghamton, School of Management, New York, U.s.A.
and
JEAN H. P. PAELINCK
Erasmus Universiteit, Rotterdam, The Netherlands
1. Introduction
2.1. Foundations
Lagrange fimction is
L = u[q; h(v,f, e); v,f, el l(q + P.v + prf+ Pee  r*) (3)
the price of other goods and services being normalised to 1.
Putting equal to zero derivatives with respect to q, v,! and c, one gets,
taking into account the fact that 'consumption' of ! and c are functions of v,
m(v) and n(v) (doctors' prescriptions): I
L'q = u'q  l =0 (4)
L~ = u~(h~ + hfm' + h~n') + u~ + ufm' + u~n'
 l(p. + Prm' + Pen') = O. (5)
Substituting (4) into (5), one obtains
, u~(P. + Pim' + Pcm')  (u: + u~m' + u~n')
uh = h: + h~m' + h~n' (6)
From (6) one sees that the marginal utility of health will be declining if the
marginal utility of visits increases, either through its health effect (denomina
tor) or its psychological effect (numerator), leading to more health expendi
tures in the 'normal' case (declining marginal utilities).
2.2. Insurance
It is also well known that the price of ethical drugs, visits and hospitals cost
are not paid completely by the individual: a portion, :re, is returned to the
patient. Let us assume that it is fixed and denoted by :re*, so budget Equation
(2) can be rewritten as 2
r** = r* + :re!v + :rei! + :re~c. (2 ')
With that new constraint (2 ') Equation (6) can be rewritten as
, _ u~[(P.  :re:)  (Pi  :r(7)m'  (Pc  :re:)n'l  (u: + u~m' + u~n') 6'
uh  h: + h~m' + h~n' ( )
of which the interpretation is obvious.
Again for each individual patient k, the budget constraint can integrate the
fact that his/her budget equals his/her income plus the amount he/she gets
back for his/her health expenses payment minus a total tax payment, i.e., a
fixed aportion he/she has to contribute for his/her contribution to health
insurance; with such a budget equation equalling r***, Equation (6') can be
easily extended by correcting each :re*  term with a factor (1  a), and the
new equation has again an obvious interpretation.
Let us now consider a welfare function for society as a whole, 1jJ, Let
A general equilibrium model of health care 239
Y:. ~ [ud with Uk as in (1). One now wants to maximize tJl(Y:.) subject to two
restrictions; first a global budget restriction
The left hand side in (7) denotes total net outlays for the nation as a whole
and the right hand side is total income; a second condition is
(at  a k ) (rk  r l ) ~ 0 (8)
which arises from the reasonable assumption that when r k > rl , a k ~ al.
The Lagrange function now is
2= l~. (lOb)
rk ,ulk
Equation (10) through (12) and assimilated can be solved to get the
demand functions of the patients which follow:
qZ = q(rk> Pv, PC' PI' n:, n~, nj, ak) (13a)
vZ = Vk(rk> Pv, PC' PI' n:, n~, nj, ak) (13b)
(13c)
(13d)
Notice that to patient k, ak is given, as is the set Pv, PC' PI' and in this
'social case', r k (not starred). Demand function (13a) will not interest us any
more; note that functions (13c) and (13d) are 'derived' demand functions (via
the functions m(v) and n(v».
So far one has assumed that the reimbursement parameters n:, nj and n~
are arbitrary determined by a central authority; however, the problem may
be how to choose values for which total social utility 1fJ(yJ is maximum.
In taking the derivative of (9) with respect to nv' e.g., one should take into
account the demand and 'derived' demand functions (13), a tedious exercise
which is not pursued further here.
To understand the difference between 2.1 and 2.3, let us consider the
following simple example; let the utility function of the kth individual be
Uk = uk(g) (15)
with his budget constraint
l!..*' 9.. = rk· (16)
The solution for maximum Uk with the corresponding budget cosntraint
A general equilibrium model of health care 241
will be
(17)
i.e., the classical result that the ratio of the marginal utility to the price of any
good i will be fixed for all k individuals in proportion to the marginal utility
of income of those individuals, Ak •
When one considers a preference function for society as a whole, as
before, and again denotes it as 1/J(11:.), where 11:. ~ [Uk], the Lagrangian for
maximizing total utility under a budget constraint becomes
(18)
(19b)
If one compares (17) with (19b) one sees that as long as P/1/Jk equals Ak ,
the conditions for individual and social optimizing will lead to the same
equilibrium result.
Equation (19b) can be rewritten as
(20)
with
(21)
(22a)
so
fif = pf  :n:f. (22b)
The utility of the physician depends on his income, which again depends
on the number of visits received multiplied by his unit fee, and the number of
hospital patients times income obtained from sending patients to the hospital;
furthermore medical ethics plays a part, depending on the number of visits,
the number of prescriptions and the number of patients sent to the hospital;
finally intervenes the general price level, p*.
The utility maximisation principle applied to (23) gives
du au ( av) au ak av (24a)
dpv = or v + Pv apv + ok ov apv = o.
or
du
dpv = v au or )1 au
or [ 1 + ( 1 + ( au aft
aft ov
1) E vp",]
Pv (24b)
where Evpv denotes the visiting price elasticity of the patients to the doctor's
office.
The solution of Equation (24b) gives the supply function of the physician;
and proceeding in a similar fashion for clinic visits and medicines, the
physicians 'supply functions' for private visits, clinic visits and medicines will
be given by
(2Sa)
(2Sb)
f Zs  fS( Pv,PoP *) .
v
(2Sc)
For office visits the demand and supply equations in partial equilibrium
now are
(26)
(27)
(28)
(29)
Solving the supply and demand function of and confronting the physician,
one gets the equilibrium price and quantity in a partial equilibrium setting;
one sees immediately that from fixing (centrally) Pjv' disequilibrium could
result.
A general equilibrium model of health care 243
4. Hospitals
Assuming that the objective of the hospitals is to maximize their net income,
i.e., to choose the price of hospital admission Pc and payments to doctors Pc
in (23) such that
(30)
with
(31)
i.e. x is a binary variable, (30)(31) being valid for each hospital h = 1, ... ,
H; specification (30)(31) allows of respecting the condition r ~ O.
The Lagrangian is
L = xr(pc' Pc)  a(x2  x). (32)
Taking derivatives with respect to Pc> Pc and x, and setting equal to zero,
results in:
dL dr
=x=O (33a)
apc apc
aL
=x=O
ar (33b)
afJc afJc
aL
ax = r  a(2x  1) = O. (33c)
straints
i ~ i* (36)
and
(37)
The Lagrangian is given by
L = xs(i, PI)  'fJ(i  i*)  O(X2  x) (38)
leading up to
aL as (39a)
ai=xa;'fJ=O
aL
=x=O
as (39b)
apf apf
aL s  O(2x  1) 0 (39c)

ax = =
or
1 s
x=+ (39b)
2 20
with a similar interpretation as above.
It is realised that the complexities of the drug and pharmaceutical
industries require a more sophisticated formulation, to be left for further
study.
(40)
the first argument representing premium receipts (Zik per definition), the
other ones disbursements. Consumer ik belongs to a subset of the K con
sumers, supposed to be known a priori (but see later on).
So the Lagrangian first order maximum conditions are
(41)
A general equilibrium model of health care 245
Conditions such as
t; ~ 0, 'Vi (42a)
:7li, ~ PVik (42b)
(42c)
(42d)
should be included, (42b) and (42d) taking into account the allocation of
consumers to doctors and clinics respectively. Again it is realised that the
above formulation for reimbursement is too simplistic: in some countries,
there is compelte national health insurance, in others it is partial; there
are institutional arrangements like HMO's, the possibility of copayments,
deductibles etc., which have to be incorporated in future developments.
7. Equilibrium
Neglecting qk, as said before, one can now draw a list of equations and
endogenous variables.
7.1. Equations
kEKj
L vi:2: vJ (44)
ch = C(Phc) (45)
p% = P(PhJ· (46)
I
There are IZH such equations.
(47)
OJ equation.
7.1.7. DS equilibrium
(50)
OJ equation.
7.1.8. Insurance supply and equilibrium
The conditions here are, consumers being supposed to be allocated to
insurance companies:
(51)
then being
equations).
m of these equations, the same applying to n iv ' nit' n ic ( [l[I
A general equilibrium model of health care 247
8. Conclusions
Notes
1. We skip the possibility of free purchase of drugs, which can be easily included in the
analysis.
2. For the sake of simplicity, we are not considering different possible types of health
248 M. Chatterji and f. H. P. Paelinck
insurance i.e. complete national health insurance, private insurance, copayment, deducti
bles, Health Maintenance Organisations, etc.
3. See also P. Zweifel, 1982.
4. This supposes that the set of patients with cardinal K is subdivided into J subsets of
cardinal Kj , ~ j Kj = K; we will corne back to that remark.
5. (J will be positive, as r < 0 implies x = O.
6. This has to be extended in later versions to multiple drugs and competition amongst firms
producing the same drug.
7. See J. H. P. Paelinck, 1985, pp. 35 ff; a possible treatment is through TinbergenBos
systems (ibidem, pp. 52 ff).
8. See J. H. P. Paelinck, 1984.
References