Vous êtes sur la page 1sur 18

Subjective mortality expectations

and consumption and saving behaviours


among the elderly
Martin Salm

CentER, Tilburg University

Abstract. Life expectancy is an important factor that individuals have to take into account
for saving and consumption choices. The life-cycle model of consumption and saving
behaviour predicts that consumption growth should decrease with higher mortality rates.
The aim of this study is to test this hypothesis based on data about subjective longevity
expectations from the Health and Retirement Study merged with detailed consumption
data from two waves of the Consumption and Activities Mail Survey. This study nds
that an increase in subjective mortality by 1% corresponds to an annual decrease in
consumption of non-durable goods of around 1.8%. JEL classication: D81, D91
Esperances de vie subjectives et comportement de consommation et depargne des vieilles
personnes. Lesperance de vie est un facteur important dont les personnes doivent tenir
compte dans leurs choix de consommation et depargne. Le mod`ele de comportement
de consommation et depargne au cours du cycle de vie predit que la croissance de la
consommation devrait decrotre a` mesure que le taux de mortalite augmente. Cette e tude
met au test cette hypoth`ese a` laide de donnees sur lesperance de vie subjective tirees
des resultats dune e tude sur la sante et la retraite arrimes aux resultats de deux vagues
denquetes postales sur la consommation et les activites qui ont produit des donnees
detaillees sur la consommation. Cette e tude montre quun accroissement de un pour cent
dans lanticipation subjective de mortalite correspond a` un declin d`a peu pr`es 1.8% dans
la consommation annuelle de biens non durables.

I thank Han Hong, Michael Hurd, Ahmed Khwaja, Daniel McFadden, Daniel Schunk, Frank
Sloan, Alessandro Tarozzi, Curtis Taylor, and two anonymous referees for helpful advice. I also
thank seminar participants at Duke U, U Conn, Baylor U, UNC Greensboro, HEC Lausanne,
the Netspar Workshop in Utrecht, and the RAND Workshop on subjective probabilities in
Jackson for valuable suggestions. Email: m.salm@uvt.nl
Canadian Journal of Economics / Revue canadienne dEconomique, Vol. 43, No. 3
2010. Printed in Canada / Imprime au Canada
August / aout

0008-4085 / 10 / 10401057 /  Canadian Economics Association


C

Subjective mortality expectations 1041


1. Introduction
This study examines how subjective mortality expectations affect the saving and
consumption behaviours of the elderly. Life expectancy is an important factor that individuals have to take into account for their saving and consumption choices. The life-cycle model, which goes back to the pioneering work
of Modigliani and Brumberg (1954) and Yaari (1965), predicts a specic relationship between mortality expectations, consumption growth, and time and
risk preference parameters. At the core of the life-cycle model is the idea that
forward-looking agents hold the ex ante expected marginal utility from consumption constant across periods. This implies that agents with higher expected
mortality rates, who are less likely to benet from consumption in future periods,
should allocate fewer resources for future as opposed to present consumption.
Therefore, the growth rate of consumption should be lower (or the decline in
consumption faster) for individuals with higher expected mortality rates. This
study tests this hypothesis by estimating Euler equations that relate consumption growth to subjective mortality rates. My empirical approach controls for
agents who are credit constrained or buffer-stock savers, and it allows for a
precautionary savings motive by controlling for the predicted variance of outof-pocket medical expenditures. The data I use merge information about subjective mortality expectations from the Health and Retirement Study with detailed
consumption data from two waves of the Consumption and Activities Mail
Survey.
Several previous studies examine the effect of life table mortality rates on
consumption and saving behaviours (Skinner 1985; Hurd 1989; Palumbo 1999;
Bloom, Canning, and Graham 2003; De Nardi, French, and Jones 2005). However, individual mortality probabilities differ from life table mortality rates. If
differences between individual mortality probabilities and life table mortality
rates are correlated with other determinants of consumption choice, such as time
preferences and risk aversion, then estimation using life table mortality rates
might lead to biased results.1 For example, smokers tend to have higher mortality probabilities than non-smokers, and they also tend to differ in their risk
aversion and time preferences (Khwaja, Sloan, and Salm 2006). An alternative
to using life table mortality rates is to use subjective mortality expectations.
Previous studies nd that subjective longevity mortality probabilities vary with
known predictors of mortality such as smoking, income, and education, and are
on average remarkably good predictors of actual mortality (Hamermesh 1985;
Hurd and McGarry 1995, 2002; Smith, Taylor, and Sloan 2001; Khwaja, Sloan,
and Chung 2007).2 Gan et al. (2004) use subjective mortality probabilities to
estimate a structural model of saving and consumption that includes a bequest
motive. They nd that estimates using subjective expectations t the data better
1 Skinner (1985) and DeNardi, French, and Jones (2005) adjust mortality rates for occupation
and wealth, respectively.
2 Subjective probabilities also tend to be good predictors of events other than mortality (see the
survey by Manski 2004).

1042

M. Salm

than estimates that are based on life table mortality rates. In contrast to their
study, I include a precautionary savings motive for health care expenditures, and
I use detailed consumption data instead of predicting wealth levels.
The shortage of high-quality longitudinal consumption data has long presented a major difculty for studying saving and consumption behaviours. Many
previous studies either use information on food consumption only, which is included in some commonly used panel data sets, or calculate consumption from
differences in wealth levels between periods (see the survey by Browning and
Lusardi 1996). However, food consumption might not be a good proxy for overall consumption (Attanasio and Weber 1995; Browning and Lusardi 1996), and
changes in assets can be an imprecise measure of consumption. Other studies
create pseudo-panels from cross-sectional data (Parker and Preston 2005). In this
study I employ a measure of annual consumption spending on non-durable goods
based on two waves of the Consumption and Activities Mail Survey, which was
administered to a sub-sample of the Health and Retirement Study population in
2001 and 2003.
The main result of this study is that consumption growth is indeed lower for
individuals with higher subjective mortality rates, as predicted by the life-cycle
model. I nd that a 1% higher subjective mortality rate corresponds with an
annual decrease in consumption of non-durable goods of around 1.8%. My ndings are robust for the inclusion of additional covariates such as health, income,
wealth, and the predicted variance of medical expenditures, and for alternative
methods to control for focal answers about subjective longevity expectations. Estimation results for subjective mortality rates are similar to estimation results for
life table mortality rates, but results for subjective mortality rates are estimated
more precisely. In view of increasing life expectancy and longer retirement periods, it is a reassuring nding that individuals take life expectancy into account
for decisions about consumption and saving. This study also contributes to a
debate about whether the saving and consumption behaviours of the elderly are
consistent with the life-cycle model. Some studies cite the lack of (or slow pace
of) asset decumulation among the elderly as evidence against the life-cycle model
(Hurd 1987, 1990; Attanasio and Hoynes 2000). Whether consumption growth
decreases with higher subjective mortality rates is an alternative and more direct
test of the life-cycle model.

2. Data description
This study combines data from waves ve and six of the Health and Retirement
Study (HRS), which were collected in 2000 and 2002, with information from the
Consumption and Activities Mail Survey (CAMS) from 2001 and 2003. The HRS
is a national panel study, which is representative for the elderly and near-elderly
population of the United States.3 The sample in the year 2000 survey includes
3 See the description at www.hrsonline.umich.edu.

Subjective mortality expectations 1043

TABLE 1
Selection of the estimation sample
Sample description

Number of observations

Participants in HRS in year 2000


Of which answered both CAMS 2001 and CAMS 2003
Of which are age 65+ and live in single-person households
Of which have non-missing information on subjective longevity
expectations
Of which have non-missing observation on consumption growth
Of which have non-missing observations on other covariates
(such as race and education)
Excluding respondents with 0% longevity expectations
Excluding persons who are credit-constrained or buffer-stock
savers (sample for baseline estimation)

19,597
2,989
583
503
481
478
417
387

about 19,600 respondents. The HRS contains detailed information on health,


income, assets, and expectations about the future, as well as questions about
attitudes and preferences. One shortcoming of the HRS as well as of other large
U.S. household panel surveys is the lack of detailed information about household
consumption. The only information about household consumption included in
the main HRS survey concerns food consumption. The Consumption and Activities Mail Survey remedies this decit and includes detailed information on
household consumption spending, and also spending intentions. A description
of the CAMS survey is provided in Butrica, Goldwin, and Johnson (2005). The
2001 CAMS questionnaire was sent to 5,000 households randomly drawn from
the HRS population. Of these households 4,157 also received the 2003 CAMS
questionnaire. In 2001 and 2003, 2,989 households completed both surveys.
In most specications, I restrict the sample to 583 respondents who are above
age 65 and live in single-person households. Restricting the sample to singleperson households allows disregarding difculties in modelling intra-household
decision making, and at age 65 most respondents are retired and are also eligible
for Medicare.4 After excluding 105 respondents with missing information about
consumption change, subjective longevity expectations, or other covariates, 61
observations with 0% longevity expectations, and 30 respondents who are identied as being either buffer-stock savers or credit constrained, the estimation
sample in the baseline specication consists of 387 observations (see table 1 for
an overview on how the baseline sample was selected).

4 In alternative specications, the sample is restricted not by age, but by self-reported retirement
status (for different denitions of retirement in the HRS see Gustman and Steinmeier 2000), and
I also include a specication in which the sample includes both single and married respondents.
Results for these specications are shown in table A3, columns 24 of the on-line appendix,
which is linked to this article at the CJE journal archive (www.economics.ca/cje/en/archive.
php).

1044

M. Salm

FIGURE 1 Mean expenditures on non-durable good categories in 2001 and 2003 (measured in real
2001 USD)

The dependent variable is the real annual percentage change in consumption of non-durable goods. This includes spending on food, gas, clothing, dining
out, vacations, tickets to events, and hobbies. I calculate the yearly percentage
change in consumption by taking the difference of the logarithms of consumption spending on non-durable goods in 2003 and 2001, divided by two. I compute
real consumption growth rates by adjusting for the increase of the consumer price
index for all goods. I exclude purchases of durable goods such as cars. Expenditures on durable goods do not coincide with the consumption ows received from
them. Adjusting consumption ows from durable goods is also costly for consumers. The consumption variable also excludes medical expenditures. Medical
care typically does not provide direct utility to consumers, but is an investment
in health. For studying changes in consumption, the change in the consumption of non-durable goods is one of the best available measures (Browning and
Lusardi 1996). Alternatively, I also include a specication that is based on food
consumption only. Figure 1 shows sample means for total non-durable-goods
consumption and for non-durable-goods consumption categories in the years
2001 and 2003. On average, total consumption of non-durable goods was higher
in 2001 than in 2003.
Among the explanatory variables in my estimation is the subjective annual mortality rate. The HRS does not directly ask about subjective mortality

Subjective mortality expectations 1045


probabilities in the following year, but it includes questions about subjective
longevity probabilities. Specically, the HRS asks about the percentage chance
that a respondent would live to age A, where A depends on the respondents
current age and is between 11 and 15 years above the respondents current age.
Previous studies have shown that subjective longevity probabilities are in general
very good predictors of actual longevity (Hurd and McGarry 1995, 2002; Smith,
Taylor, and Sloan 2001; Khwaja, Sloan, and Chung 2007). However, the high
frequency of focal answers raises concerns about the validity of self-reported
longevity probabilities (Hurd and McGarry 1995; Hurd, McFadden, and Gan
1998; Basset and Lumbsdaine 2001). In the 2000 HRS survey, 9.5% of respondents stated that their subjective longevity probability was 0% and 10.7% stated
it was 100%. Several studies such as Delavande, Perry, and Willis (2006), Bloom
et al. (2006), and Gan, Hurd, and McFadden (2003) suggest procedures for adjusting stated probabilities. For the main specication, I decided against correcting stated probabilities, because even somewhat unrealistic expectations might
still be what agents base their decisions on. However, I also include a specication, which follows the approach by Delavande, Perry, and Willis (2006), Bloom
et al. (2006), and Fang et al. (2007), who use information on parental longevity
as instrumental variables for subjective mortality expectations in order to adjust
for focal answers. The exact denition of variables for parental longevity in my
study follows Delavande, Perry, and Willis (2006).5
For calculating annual subjective mortality rates based on subjective longevity
probabilities, I follow Gan, Hurd, and McFadden (2003) in assuming that subjective annual mortality rates mi,t are the product of annual life-table mortality
rates m0,t and an individual specic mortality factor i :
mi,t = i mo,t .

(1)

I use life table mortality rates for 1998 separately for men and women of
each year-of-birth cohort, which are provided by the Center for Disease Control
(www.cdc.gov/nchs/nvsr48_18.pdf). Given equation (1) the subjective probability si,a,A of individual i to survive from age a to age A can be written as

si,a,A =

A1

t=a

(1 mi,t ) =

A1


(1 i m0,t ).

t=a

5 Instrumental variables include indicators for mothers age at death 76 or younger, mothers age
at death 77 to 84, and mothers age at death 85 or older, with mother still alive as omitted
reference group, and indicators for fathers age at death 72 or younger, fathers age at death 73 to
80, and fathers age at death 81 or older, with father still alive as omitted reference group.

1046

M. Salm

Individual mortality factor i can then be calculated as approximately

i ln si,a,A


A1

m0,t .

t=a

However, one shortcoming of this approach is that it does not allow calculating subjective annual mortality probabilities for persons who state that their
subjective longevity probability is zero. It is not clear what the subjective survival
probability for the next year should be for agents who dont expect to live for
another 11 to 15 years. I employ two alternative approaches to this problem. One
approach is to change the answer from 0% to 1% (similar to Khwaja, Sloan, and
Chung 2007). The other approach is to omit the observations with a subjective
longevity probability of 0%. I also test if estimation results change if observations with a subjective longevity expectation of 100% are excluded. The distributions of subjective annual mortality rates and of life table annual mortality rates
in the baseline estimation sample are shown in gure 2. The mean subjective
mortality rate is 3.6% ,compared with 4.3% for life table mortality rates (see
table 2).
Further variables employed in the analysis include net nancial wealth and
income, which is the sum of social security payments, employer pensions, and
capital income. A binary variable for whether or not the respondent is in good
health is set to one if self-reported overall health is excellent, very good, or good,
and it is set to zero if self-reported health is fair or poor. The number of limitations
in activities of daily living ranges from 0 to 6, and indicates whether respondents
are able to independently walk, dress, bathe, eat, get into bed, and use the toilet.
I also include binary indicators for male respondents, for black respondents, and
a variable for years of schooling. Further included are two indicator variables
for medium nancial planning (more than one and up to ve years), and for
a long nancial planning horizon (more than ve years). A nancial planning
horizon of one year or less it the omitted reference category. Respondents with
different levels of risk aversion are identied based on answers to a question on a
hypothetical lifetime income gamble. Information on nancial planning horizon
and risk aversion is imputed for some observations.6 Sample statistics are shown
in table 2.

6 I use IVE ware imputation and variance estimation software, which follows a sequential
regression imputation method described in Ragunathan et al. (2001). For the baseline sample,
information on nancial planning horizon and risk aversion is imputed for 180 observations.
However, most missing observations are not due to respondents refusal, but because the
question was asked only to a subsample of respondents, which make the assumption of
randomly missing observations for imputation more credible.

Subjective mortality expectations 1047

Density / kdensity subj_mort


10
20
30

40

A. Subjective annual mortality rates

.1
.2
subjective annual mortality rates
Density

.3

kdensity subj_mort

Density / kdensity lifetab_mort


10
20
30

40

B. Life table annual mortality rates

.1
.2
life table annual mortality rates
Density

.3

kdensity lifetab_mort

FIGURE 2 Distribution of subjective mortality rates and life table mortality rates

1048

M. Salm

TABLE 2
Descriptive statistics

Consumption growth
Food consumption growth (N = 350)
Subjective mortality rate
Life table mortality rate
Predicted variance of medical expendituresa
Age
Male
Black
Years of education
Income (in $1,000)
Financial assets (in $1, 000)
Good health
Number of ADL limitations
Low-risk aversion
High-risk aversion
Short nancial planning horizon
Medium nancial planning horizon
Long nancial planning horizon
Intention spend all (N = 417)
Number of observations (baseline estimation)

Mean

Standard deviation

0.059
0.030
0.036
0.043
1
75.294
0.198
0.067
12.390
28.812
105.573
0.816
0.173
0.276
0.724
0.307
0.346
0.346
0.071
387

0.502
0.503
0.037
0.029
9.436
6.454
0.399
0.250
2.739
37.767
194.514
0.387
0.574
0.447
0.447
0.462
0.476
0.476
0.258

a The mean of this variable is normalized to one.

3. Identication strategy
My identication strategy follows directly from a standard life-cycle model. I
consider a single retired agent, who chooses consumption and saving in each
period in order to maximize expected lifetime utility. I assume that utility is
additively separable between periods, and that future utility is discounted with
factor . The subjective probability of survival from age t to age j is denoted as
st,j (with j t). Then the maximization problem can be summarized as
max Et

T


jt st,j u(cj ),

j=t

where T is the maximum age a person can live, and Et is the expectations operator,
based on information in period t. I further assume that within-period utility
depends on the level of consumption ct in period t, and is given by a constant
relative risk aversion (CRRA) utility function:
1

u(ct ) =

ct
,
1

(2)

where is the parameter of relative risk aversion. The intertemporal elasticity of


substitution is given by 1/ , the inverse of risk aversion. In each period agents

Subjective mortality expectations 1049


receive income yt from Social Security and pensions. This income is nonstochastic. Social security payments increase with ination (cost-of-living adjustment) and are constant in real terms. Agents face uncertain out-of-pocket
medical expenditures t in each period. Out-of-pocket medical expenditures are
treated as exogenous and are not part of consumption. I assume that there is one
asset that yields a risk-free real return of Rt between periods. Assets in period
t +1, at+1 , are determined by the following asset accumulation equation:
at+1 = Rt (at + yt ct t ).
Social Security entitlements cannot be used as collateral for loans and it is
difcult to borrow against employer pensions. This credit constraint imposes the
following restriction on consumption:
ct at + yt t .

(3)

If the credit constraint is not binding, then the rst-order condition requires
that the marginal utility from consumption in period t is equal to the expected
marginal utility from consumption expenditure in period t +1:
u (ct ) = Rt st,t+1 Et [u (ct+1 )].
Substituting the CRRA utility function into equation (4) yields

 
ct+1
= Rt st,t+1 .
Et
ct

(4)

(5)

Uncertainty about future consumption derives from stochastic out-of-pocket


medical expenses. Under the assumption that consumption changes are lognormally distributed, equation (5) can be transformed into the following Euler
equation:
Et ( ln ct+1 ) = 1/ (rt mt ) + ( /2)Vart ( ln ct+1 ),

(6)

where ln ct+1 = ln ct+1 ln ct is the growth rate of consumption, rt =


ln Rt is the real interest rate at time t, = ln is the time discount rate, and
mt = ln st,t+1 is the subjective mortality rate in period t. Equation (6) postulates
that expected consumption growth should increase with higher real interest rates
and decrease with higher time discount rates and higher expected mortality rates,
and that these effects should be smaller if agents are more risk averse. Expected
consumption growth should also increase with a higher variance of consumption
growth. A Euler equation very similar to equation (6) can also be derived without the assumption of log-normally distributed consumption growth rates from
a second-order Taylor approximation of equation (5) (Carroll 2001; Ludvigson
and Paxson 2001).

1050

M. Salm

Equation (6) can be estimated by the following linear regression model:


 ln ci,t+1 = a0 + a1 mi,t + a2 hi,t + i,t ,

(7)

where a0 , a1 , and a2 are regression coefcients, mi,t is the subjective annual


mortality rate of agent i at time t, and hi,t is the predicted variance of out-ofpocket medical expenditures for agent i at time t. i,t is an error term, which
reects health cost shocks and measurement errors of consumption growth. The
estimated relative risk aversion parameter can be calculated as = 1/a1 .
I use the predicted variance of health cost risk as a proxy variable for the
expected variance of consumption growth, which cannot be easily observed.
While for the elderly some sources of nancial risk such as unemployment or
wage risk are less important, out-of-pocket medical expenditures tend to be high
and very variable (French and Jones 2004). Thus, differences in the variance of
out-of-pocket medical expenditures can explain differences in the variance of
consumption growth. I calculate the individual-specic health cost risk based
on out of pocket medical expenditures of HRS respondents in the two years
preceding the 2002 interview. Out-of-pocket medical expenditures oopi,t include
expenditures for hospitals, nursing homes, doctor visits, dentist visits, outpatient
surgery, home health care, special facilities, and the average monthly expenditures
for prescription drug costs. I compute the predicted variance of out of pocket
medical expenditures by the following two-stage procedure. The rst-stage regression equation is
oopi,t+1 = b0 + b1 xi,t + i,t ,

(8)

where xi,t is a vector of covariates from the 2000 HRS survey. Covariates include
previous out-of-pocket medical expenditures; information on health insurance,
age, years of education, gender, self-reported health of good or better, total nancial wealth, and total household income; the number of limitations in activities
of daily living; and previous diagnoses of chronic conditions such as diabetes,
cancer, lung diseases, heart diseases, stroke, and mental health conditions. The
second-stage estimation regresses the squared error term of the rst-stage regression on the same covariates as above:
2
= c0 + c1 xi,t + i,t .
i,t

(9)

The predicted variance of medical expenditure, h i,t , is then computed by


h i,t = c 0 + c 1 xi,t .
This approach allows identifying agents with varying risk of medical expenditures. h i,t is included in regression equation (7) as a proxy for overall consumption
risk.

Subjective mortality expectations 1051


There are several caveats about regression equation (7). The rst caveat concerns the validity of Euler equation estimates in the presence of credit-constrained
agents and buffer-stock savers. If the credit constraint in equation (3) binds, then
the rst-order condition in equation (4) might not hold with equality, and utility
parameters estimated from regression equation (7) are misleading. Carroll (1998,
2001) points out that a similar argument can also hold for households with positive wealth that are buffer-stock savers. I identify constrained agents using the
answer to the following survey question about spending intentions of a windfall
gain: Suppose next year you were to nd your household with 20% more income
than normal, what would you do with the extra income? For the estimation of
utility parameters I exclude all agents from the sample who answer that they
would spend the entire windfall gain. This leaves a sample of agents who are not
credit constrained or buffer-stock savers.
A second caveat is that subjective mortality could be a proxy for other factors
that are related to both mortality expectations and consumption growth, for
example, health or wealth. As a robustness check, I test whether estimation
results are sensitive to the inclusion of additional variables for race, education,
income, wealth, health, the number of limitations in activities of daily living, and
the nancial planning horizon. If estimation coefcients for subjective mortality
rates do not change after the inclusion of these additional variables, this adds
credibility to the results of this study.
The identication strategy discussed above does not explicitly account for a
bequest motive. However, a bequest motive would affect the levels of consumption in all periods, but not necessarily the changes in consumption. Since this
study examines changes in consumption, the identication strategy can still be
valid in the presence of a bequest motive.

4. Results
Estimation results are shown in table 3. The estimation sample for the baseline
regression includes all persons above age 65, who participated in both waves
of the Consumption and Activities Mail Survey and lived in a single person
household at the time of the year 2000 HRS interview. Persons with subjective
longevity probabilities of zero are excluded from the sample (they are included
in the estimation in column (3) of table 4). The baseline regression also excludes
individuals from the sample who intend to immediately spend a windfall gain.7
The dependent variable is the growth rate of real annual consumption expenditures on non-durable goods. Columns (1) to (3) show the relationship between
consumption growth and subjective mortality rates, adding successively more
7 Estimation results for this sample are shown in column 1 of table A4 of the on-line appendix,
which is linked to this article at the CJE journal archive (www.economics.ca/cje/en/archive.
php).

1052

M. Salm

TABLE 3
Regression results

Subjective mortality

Consumption
growth
(no
covariates)
(1)

Consumption
growth
(add
covariates)
(2)

Consumption
growth
(baseline
regression)
(3)

1.780
(0.659)

1.932
(0.678)

1.805
(0.700)

0.002
(0.013)
0.167
(0.127)
0.0000
(0.0001)
0.0002
(0.0005)
0.080
(0.089)
0.016
0.063

0.003
(0.009)
0.178
(0.103)
0.0000
(0.0001)
0.0002
(0.0007)
0.073
(0.072)
0.016
(0.046)
0.074
(0.062)
0.081
(0.065)
0.006
(0.014)
0.082
(0.139)
0.034
387
3SLS

Life table mortality


Years of education
Black
Financial assets
(in $1,000)
Income (in $1,000)
Good health
ADL limitations
Medium nancial
Planning horizon
Long nancial planning
Horizon
Predicted variance of
medical expenditures
Constant
R-squared
Observations
Estimation methoda

0.006
(0.034)
0.017
387
OLS

0.124
(0.185)
0.027
387
OLS

Consumption
growth
(life table
mortality)
(4)

1.998
(0.917)
0.005
(0.010)
0.171
(0.104)
0.0000
(0.0001)
0.0002
(0.0007)
0.037
(0.073)
0.002
(0.047)
0.090
(0.062)
0.095
(0.064)
0.013
(0.013)
0.099
(0.145)
0.001
387
3SLS

Consumption
growth
(instrumented
mortality)
(5)
1.157
(1.705)
0.001
(0.010)
0.103
(0.104)
0.0000
(0.0001)
0.0003
(0.0007)
0.114
(0.076)
0.024
(0.046)
0.076
(0.064)
0.120
(0.068)
0.002
(0.012)
0.025
(0.182)
0.026
374
3SLS

NOTES: Huber-White standard errors are in parentheses. Signicant at 10%; signicant at 5%;

signicant at 1%.
a Estimation results for the predicted variance of medical expenditures for the baseline estimation are
shown in table A2 of the on-line appendix; rst-stage estimation results for instrumented mortality
are shown in table A3 of the on-line appendix.

covariates. In column (1) the subjective mortality rate is the only explanatory
variable. An increase in the subjective mortality rate by 1% is associated with a
consumption decline of 1.78 % per annum. Thus, consumption growth is lower
for persons with higher subjective mortality rates, as predicted by the life cycle
model. This effect is statistically signicant at the 1% level. The specication in
column (2) includes additional covariates for education, race, income, wealth,
health, and the number of limitations in activities of daily living. The estimation
coefcient for subjective mortality rates is almost unaffected by the inclusion of
these covariates, and none of the coefcients for the additional covariates differs
signicantly from zero. The baseline specication, which is shown in column (3),

Subjective mortality expectations 1053

TABLE 4
Robustness checks
Consumption
growth
(food
consumption)
(1)
Subjective mortality

Consumption
growth
(no 100%
answer)
(2)

Consumption
growth
(with 0%
answer)
(3)

Consumption
growth
(include risk
aversion)
(4)

0.382
(0.803)

1.719
(0.749)

0.874
(0.381)

0.019
(0.010)
0.142
(0.120)
0.0002
0.0001
0.0006
0.0007
0.064
(0.078)
0.071
(0.059)
0.008
(0.068)
0.084
(0.069)
0.010
(0.014)
0.189
(0.146)
0.022
350
3SLS

0.008
(0.010)
0.266
(0.108)
0.0000
(0.0001)
0.0002
(0.0008)
0.067
(0.075)
0.005
(0.047)
0.121
(0.065)
0.101
(0.068)
0.008
(0.014)
0.111
(0.156)
0.044
342
3SLS

0.002
(0.009)
0.138
(0.098)
0.0000
(0.0001)
0.0002
(0.0007)
0.058
(0.067)
0.021
(0.041)
0.070
(0.060)
0.075
(0.061)
0.005
(0.013)
0.039
(0.134)
0.024
441
3SLS

1.103
(0.810)
2.154
(1.519)
0.189
(0.085)
0.004
(0.009)
0.161
(0.102)
0.0000
(0.0001)
0.0001
(0.0007)
0.062
(0.072)
0.022
(0.046)
0.085
(0.062)
0.093
(0.064)
0.010
(0.013)
0.034
(0.139)
0.040
387
3SLS

Subjective mortality
low risk aversion
Low risk aversion

Years of education
Black
Financial assets
(in $1,000)
Income (in $1,000)
Good health
ADL limitations
Medium nancial
Planning horizon
Long nancial planning
Horizon
Predicted variance of
medical expenditures
Constant
R-squared
Observations
Estimation methoda

NOTES: Huber-White standard errors are in parentheses. Signicant at 10%; signicant at 5%;

signicant at 1%.
a In a three-stage least squares model, the regression equation for consumption growth is jointly
estimated with the regression equation for the predicted variance of medical expenditures for the
corresponding sample.

includes further variables for the nancial planning horizon and for the predicted
variance of out-of-pocket medical expenditures. This specication is estimated
with a three-stage least squares regression model, which allows us to estimate
correct standard errors for the variance of medical expenditures, which is predicted with uncertainty.8 The coefcient for subjective mortality rates is 1.805,
8 In the three-stage least square model, regression equations for the predicted variance of medical
expenditures (as specied in equation (9)) and for consumption growth are simultaneously
estimated. Results for the regressions which predict level and variance of medical expenditures
are shown in table A2 of the on-line appendix.

1054

M. Salm

very similar to the results in columns (1) and (2). Hence, the estimation results
are robust to the inclusion of variables for health, income, wealth, nancial planning horizon, and the predicted variance of medical expenditures. Coefcients
for the medium and long nancial planning horizons are positive, as would be
predicted by the model if longer nancial planning horizons point towards lower
time-discount rates. But they are not signicantly different from zero. The coefcient for predicted variance of medical expenses is also not signicantly different
from zero. Thus, I nd no evidence for a precautionary savings motive related
to the predicted variance of out-of-pocket medical expenditures. One possible
explanation is that only around 4% of the variance in medical expenditures can
be accounted for by observed characteristics included in equation (9).9
The coefcient for subjective mortality rates of around 1.8 corresponds
to a parameter of relative risk aversion of around 0.55. This estimate is lower
than previous estimates. Based on subjective mortality rates, Gan et al. (2004)
estimate a relative risk aversion parameter of 0.98, while studies based on life
table mortality rates estimate parameters for relative risk aversion that range
from 1.08 (Hurd 1989), 2.1 (Skinner 1985), and 3 (Palumbo 1999) to 8.2 (De
Nardi, French, and Jones 2005). One explanation of why my estimates for risk
aversion are lower than those of previous studies could be that I estimate risk
aversion for single elderly persons ,who constitute a different demographic group
from those examined in most previous studies.10
Column (4) of table 3 shows results for life table mortality rates instead of subjective mortality rates. The point estimate of the coefcient for life table mortality
rates is 1.998, which is close to the coefcient for subjective mortality rates in
the baseline regression. However, the standard error of the coefcient for life table
mortality rates is around 30% larger than the standard error of the coefcient for
subjective mortality rates in column (3). Thus, subjective mortality rates allow
us to estimate the relationship between mortality rates and consumption growth
more precisely than life table mortality rates do.
Column (5) of table 3 shows estimation results, which correct for focal point
answers about subjective longevity expectations by using an instrumental variables approach. Indicators for parental longevity are used as instrumental variables for subjective mortality rates. The exact denition of these instrumental
variables follows Delavande, Perry, and Willis (2006).11 The coefcient for subjective mortality rates is 1.157, compared with 1.805 for estimation without
instrumental variables. In the instrumental variables specication, the coefcient
for subjective mortality rates is measured imprecisely and is not signicantly different from zero. A Haussman test does not reject the hypothesis that the OLS
estimation results and the instrumental variables estimation results are identical
9 Regression results for the predicted variance of medical expenditures are shown in table A2 of
the on- line appendix.
10 Gan et al. (2004) also examine single, elderly persons.
11 First-stage regression results are shown in table A3 of the on-line appendix.

Subjective mortality expectations 1055


(p-value = 0.99). The rst-stage F-statistic is 1.9, which is an indication for weak
instruments.
Table 4 shows regression results for some additional specications.
Column (1) replicates the baseline estimation for a different measure of consumption growth, the percentage change of at-home food consumption. Owing
to a lack of better data, previous studies have often resorted to food consumption
as a proxy for total household consumption (Browning and Lusardi 1996). The
effect of subjective mortality rates on at-home food consumption growth is a
positive value of 0.382, as opposed to a negative value of 1.805 in the baseline
regression. This result adds further evidence to the argument that food consumption is not additively separable from other non-durable consumption goods and
is therefore not a good proxy for non-durable goods consumption.
Estimation results in columns (2) and (3) examine whether estimation results depend on the inclusion or exclusion of focal point answers about subjective longevity expectations. Column (2) excludes 45 observations with subjective longevity expectations of 100% from the sample. Estimation results barely
change, compared with the baseline specication. Column (3) includes 54 additional observations with a subjective longevity probability of zero. As discussed
in section 2, subjective annual mortality rates cannot be easily calculated for
such respondents. So far, these observations were excluded from the sample. In
column (3), I assume that respondents who state that their subjective longevity
probability is zero have, in fact, a subjective longevity probability of 0.01, which
allows calculating subjective annual mortality rates. Consequently, the estimated
coefcient for subjective mortality rates changes from 1.805 in the baseline
regression to 0.874, which is still signicantly different from zero at the 5%
level. Several possible explanations could account for this change in estimation
results. First, focal answers might not reect respondents actual expectations.
The average calculated annual subjective mortality rate for agents who report a
zero longevity probability is 20.7%, compared with 3.6% for the baseline sample.
Actual mortality expectations of focal respondents might be lower. Another possible explanation is that respondents who give focal answers to survey questions
differ in their risk aversion and in their consumption and saving behaviour from
other agents, which could explain different estimation results.
In column 4 subjective mortality rates are interacted with stated risk aversion,
based on answers to a survey question about whether respondents are willing to
accept a lifetime income gamble. Theory predicts that the negative relationship
between subjective mortality rates and consumption growth is weaker for persons with higher risk aversion. According to estimation results, a 1% increase is
associated with a decrease of 1.1% for respondents with high-risk aversion, and
it is associated with a decrease of 3.25% for respondents with low-risk aversion.
However, the difference between these coefcients is not signicantly different
from zero.
Further specications conrm the estimation result that there is a negative
relationship between subjective mortality rates and consumption growth also

1056

M. Salm

for alternative samples, which include married respondents, single retired respondents (which includes retired persons below age 65), and single respondents
before retirement.12

5. Conclusions
The results of this study imply that consumption growth decreases with higher
mortality expectations, as predicted by the life cycle model. An increase in subjective mortality rates by 1% corresponds with an annual decrease of consumption
of non-durable goods of around 1.8%. This result is robust for the inclusion of
additional covariates such as health, income, wealth, and the predicted variance
of medical expenditures, and for alternative methods to control for focal answers
about subjective longevity expectations. Estimation results for subjective mortality rates are similar to estimation results for life table mortality rates, but results
for subjective morality rates are estimated more precisely.

References
Attanasio, O., and H. Hoynes (2000) Differential mortality and wealth accumulation,
Journal of Human Resources 35, 129
Attanasio, O., and G. Weber (1995) Is consumption growth consistent with intertemporal
optimization? Evidence for the Consumer Expenditure Survey, Journal of Political
Economy 103, 112157
Bassett, W., and R. Lumbsdaine (2001) Probability limits: are subjective assessments
adequately accurate? Journal of Human Resources 36, 32763
Bloom, D., D. Canning, and B. Graham (2003) Longevity and life-cycle savings, Scandinavian Journal of Economics 105, 31938
Bloom, D., D. Canning, M. Moore, and Y. Song (2006) The effect of subjective survival
probabilities on retirement and wealth in the United States, NBER Working Paper
No. 12688
Browning, M., and A. Lusardi (1996) Household saving: micro theories and micro facts,
Journal of Economic Literature 34, 1797855
Butrica, B., J. Goldwyn, and R. Johnson (2005) Understanding expenditure patterns in
retirement, Working Paper, Urban Institute
Caroll, C. (1998) Buffer-stock saving and the life cycle / permanent income hypothesis,
Quarterly Journal of Economics 112, 155
(2001) Death to the log-linearized consumption Euler equation! (and very poor health
to the second order approximation), Working Paper
DeLavande, A., M. Perry, and R. Willis (2006) Probabilistic thinking and early Social
Security checking, Working Paper
DeNardi, M., E. French, and J. Jones (2005) Sex, income shocks, and the savings of
elderly singles, Working Paper
Fang, H., M. Keane, A. Khwaja, M. Salm, and D. Silverman (2007) Testing the mechanisms of structural models: the case of the Mickey Mantle effect, American Economic
Review 97, 539
12 Estimation results are shown in table A4, columns 24, of the on-line appendix.

Subjective mortality expectations 1057


French, E., and J. Jones (2004) On the distribution and dynamics of health costs, Journal
of Applied Econometrics 19, 70521
Gan, L., M. Hurd, and D. McFadden (2003) Individual subjective survival curves, NBER
Working Paper No. 9480
Gan, L., G. Gong, M. Hurd, and D. McFadden (2004) Subjective mortality risk and
bequests, NBER Working Paper No. 10789
Gustman, A., and T. Steinmeier (2000) Retirement outcomes in the Health and Retirement Study, Social Security Bulletin 63, 5771
Hamermesh, D. (1985) Expectations, life expectancy, and economic behaviour, Quarterly
Journal of Economics 50, 389408
Hurd, M. (1987) Savings of the elderly and desired bequests, American Economic Review
77, 298312
(1989) Mortality risk and bequests, Econometrica 57 (4), 779813
(1990) Research on the elderly: economic status, retirement, and consumption and
saving, Journal of Economic Literature 28, 565637
Hurd, M., and K. McGarry (1995) Evaluations of the subjective probabilities of survival
in the Health and Retirement Study, Journal of Human Resources 30, Supplement,
S26892
(2002) The predictive validity of subjective probabilities of survival, Economic Journal
112, 96685
Khwaja, A., F. Sloan, and S. Chung (2007) The relationship between individual expectations and behaviours: mortality expectations and smoking decisions, Journal of Risk
and Uncertainty 35, 179201
Khwaja, A., F. Sloan, and M. Salm (2006) Preferences and subjective beliefs of risk takers:
the case of smokers, International Journal of Industrial Organization 24, 66782
Ludvigson, S., and C. Paxson (2001) Approximation bias in linearized Euler equations,
Review of Economics and Statistics 83, 24256
Manski, C. (2004) Measuring expectations, Econometrica 72, 132976
Modigliani, F., and R. Brumberg (1954) Utility analysis and the consumption function: an
interpretation of the cross-section data, in Post Keynesian Economics, ed. K. Kurihara
(Piscataway, NJ: Rutgers University Press)
Palumbo, M. (1999) Uncertain medical expenses and precautionary saving near the end
of the life-cycle, Review of Economic Studies 66, 395421
Parker, J., and B. Preston (2005) Precautionary saving and consumption uctuations,
American Economic Review 95, 111943
Ragunathan, T., J. Lepkowski, J. van Hoewyk, and P. Solenberger (2001) A multivariate
technique for multiply imputing missing values using a sequence of regression models,
Survey Methodology Program Working paper Series 81, University of Michigan
Skinner, J. (1985) Variable lifespan and the intertemporal elasticity of consumption,
Review of Economics and Statistics 67, 61623
Smith, V., D. Taylor, and F. Sloan (2001) Longevity expectations and death: can people
predict their own demise? American Economic Review 91, 112634
Yaari, M. (1965) Uncertain lifetime, life insurance, and the theory of the consumer,
Review of Economic Studies 32, 13750