Content uploaded by Arie Kapteyn
Author content
All content in this area was uploaded by Arie Kapteyn on Aug 20, 2015
Content may be subject to copyright.
R
Health, Wealth, and the
Role of Institutions
Michael Hurd and Arie Kapteyn
DRU-3006
March 2003
Labor and Population Program
Working Paper Series 03–09
The RAND unrestricted draft series is intended to transmit
preliminary results of RAND research. Unrestricted drafts
have not been formally reviewed or edited. The views and
conclusions expressed are tentative. A draft should not be
cited or quoted without permission of the author, unless the
p
reface
g
rants such
p
ermission.
RAND is a nonprofit institution that helps improve policy and decisionmaking through research and analysis.
RAND’s publications and drafts do not necessarily reflect the opinions or policies of its research sponsors.
Health, Wealth, and the Role of Institutions
Michael Hurd
and
Arie Kapteyn
A positive relationship between socio-economic status and health has been observed over many
populations and many time periods. One of the factors mediating this relation is the institutional
environment in which people function. We consider longitudinal data from two countries with
very different institutional environments, the U.S. and The Netherlands. To structure the
empirical analysis, a theoretical model is developed relating changes in health status to income
and changes in income to health status. We show that income or wealth inequality is closely
connected with health inequality. We empirically estimate counterparts to the theoretical
relationships with generally corroborative results.
Support from the National Institute on Aging is gratefully acknowledged. Many thanks to Megan Beckett for
helping us access the literature.
Michael D. Hurd is Senior Economist, RAND, Santa Monica, California. His E-mail address is mhurd@rand.org.
Arie Kapteyn is Senior Economist, RAND, Santa Monica, California. His E-mail address is kapteyn@rand.org.
Hurd and Kapteyn 2
I. Introduction
A positive relationship between socio-economic status (SES) and health has been
observed over many populations and many time periods.1 SES can be assessed in many ways
including occupation, social class, education, income, and wealth. Health can be measured as a
reduced level of mortality, morbidities, health-related functional limitations, mental and
emotional problems as well as in other ways, and, broadly speaking, the positive relationship still
obtains. The literature has identified a number of causal mechanisms, and their relative strengths
vary over the life course, over populations, and over the level of economic development. In
broad generality causality could flow from SES to health, from health to SES, or from a third
latent factor to both SES and health. A major object of investigation has been to find the
dominant flow of causality and to quantify the causes of the relationship between SES and
health.
Until recently the main contributions to the literature have been from the disciplines of
sociology, epidemiology and public health, and in this literature the dominant flow of causality
has been thought to be from SES to health (Robert and House 2000). An obvious example
would be access to health care services that greater economic resources would purchase, but
many other mechanisms have been proposed. A prominent view in the literature is that higher
SES leads to reductions in psychosocial and environmental risk factors. Examples of risk factors
are unstable marriage, smoking, excessive alcohol consumption, stress, work-related pathogens,
chemicals and dangers, neighborhood effects, and a lack of social support networks. SES acts to
reduce these risks in various ways. Education could induce better health behaviors such as less
smoking; better, less physically demanding occupations could lead to safer, healthier work
environments; income can be used to purchase housing in clean, quiet neighborhoods.
Hurd and Kapteyn 3
This theory has been used to explain the finding that the relationship between SES and
health (the SES gradient) seems to reach a maximum in late middle or early old age: at least as
operating through occupation the cumulative effect builds over the work life, but with retirement
many SES related psychosocial mechanisms no longer operate. Were the main causal pathway
relating SES to health to operate through psychosocial and environmental risk factors, policy to
increase incomes or education, or to improve the structure of occupations would also eventually
lead to an improvement in health.
The sociological literature uses the terms selection, mobility selection, and reverse
causality to address in a limited way the effects of health on SES (Robert and House 1994;
Goldman 2001). In its simplest form it supposes that those with better underlying health will be
upwardly mobile in social class. For example, someone with better health will receive better
education, which will lead to a better occupation and, hence, higher income. It is unclear
whether health is causal in the sense that altering health after the completion of education would
change economic outcomes, or whether it is purely selection.
Selection is said to have only a minor affect on the SES gradient (Wilkinson 1999).
However, it is important to distinguish the measure of SES. If it means social class as it is given
at birth, health would have little effect on the gradient. If it means income, it would seem
obvious that health would have an important effect particularly in economies in which most
income is from earnings rather than from public transfer programs: even holding occupation and
education constant more robust individuals will be able to work longer and more intensively,
leading to greater incomes in the future. Furthermore, in a dynamic setting the flow of causality
from health to SES is directly observed, at least in the U.S.: health events such as a heart attack
lead directly to worsened health and directly to income loss because of labor market
Hurd and Kapteyn 4
interruptions; the income loss in turn leads to reduced wealth accumulation over a lifetime
(Smith 1999; McClellan 1998). In this example the health shock will increase the cross-sectional
correlation between health and SES as measured by income or wealth. This mechanism can
explain the increasing SES gradient with age until the age of retirement: as health shocks
accumulate for some individuals, their health levels will increasingly fall below the average, and
at the same time their income and wealth levels will decline relative to the average.
Despite this obvious and observable explanation for at least part of the correlation
between SES and health, the sociological literature, with its focus on selection as the only way in
which health can influence SES, has lacked investigation of this direct causal mechanism. A
possible explanation for this focus is the poor quality of income data and the complete lack of
wealth data in data sets such as the National Health Interview Survey and in the Americans’
Changing Lives study, which are often used by sociologists for SES-health research. Not having
good economic measures may have led to an emphasis on education where selection is a
plausible and possibly important mechanism.
A third mechanism for the positive relationship between SES and health is based on a
latent model of health. Individuals have unmeasured variation in fitness, tastes, attitudes, the
childhood environment, and so forth. These unmeasured factors produce individuals who both
have good health and the ability to succeed in life. Possibly because of the complexity of this
mechanism, little attention has been paid to it in the sociological literature, even though it would
seem to be a good explanation for some of the leading findings. For example, a common finding
is that SES is positively related to health throughout the range of SES, not just at the lower end
of the SES scale: in the Whitehall studies health increases with grade in the British civil service
even at the highest grades (Marmot et al. 1991). A model based on latent fitness would explain
Hurd and Kapteyn 5
this outcome as those with better latent fitness are more successful in their working careers and
they are healthier.
A more structural model, which in principle is amenable to testing, is an economic model
which emphasizes the role of the subjective time rate of discount (Grossman 1972; Fuchs 1982).
An individual with a low time rate of discount will invest in health and in human capital, and
later in life these investments will produce better-than-average health and higher income. In
addition such an individual will save more out of income leading to even higher wealth than
would be produced from the income alone.
Although we may not have good measures of many of the components of latent health, an
intervention could still change some components, leading to an alteration in some SES outcomes
such as income and wealth and even social connections. The fact that such an intervention
would lead to a change in SES, even rather late in the life cycle, distinguishes the latent health
model from the model of selection.
An important complication in assessing the quantitative importance of the flow of
causality is that the importance is likely to vary across populations. In less developed
populations giving people additional economic resources is likely to improve their health via
improved nutrition and access to health care services, whereas such effects are likely to be much
smaller in developed economies. Within a population, giving economic resources to those who
are economically deprived may similarly improve their health but it would not do so for those
who are better off. Yet, improvements in health could affect income over the entire range of the
income distribution. Thus, there could be nonlinearities in the relationship between SES and
health depending on the measures that are used and on the main direction of the causal flow.
Because of the complexity of the problem it is not surprising that there are no widely agreed-
Hurd and Kapteyn 6
upon estimates of the relative importance of the three broad explanations of the correlation
between SES and health.
As for data requirements, it is not realistic to expect that these explanations can be
separated in cross-sectional data. Even in panel data, it is very difficult to separate them because
the data requirements are substantial. At a minimum one needs good data on a number of
indicators of health status, on economic status and on other SES measures such as education,
occupation, and social network.
At a conceptual level, it seems reasonable to say that SES causes health if individuals
with higher SES have improvements (or slower deterioration) in health compared with
individuals who have lower SES, and to say that health causes SES if individuals in better health
have greater increases (or slower declines) in SES than individuals in worse health. Of course
the empirical implementation of these concepts is complex, but if we have data on transitions in
health status and economic status that are functions of health and economic status, it may be
possible to attribute causality. An example is the work of Adams et al. (forthcoming) who test
for Granger causality based on data from three waves of the AHEAD. They use 19 health
conditions such as cancer, heart attack, and self-rated health to explain wealth change; and
wealth, income and education to explain mortality and incidence of the 19 health conditions.
They cannot reject the null hypothesis of no causal link from SES to mortality and to the onset of
most acute conditions; but there is some evidence of a possible link from SES to the gradual
onset of chronic conditions. They find some evidence for a causal link from health to changes in
SES as measured by wealth change.
As far as we know, the work of Adams et al. is the most extensive and systematic study
of causality based on a large nationally representative panel data set. Yet, the study has the
Hurd and Kapteyn 7
limitations of any study based on non-experimental data. In our view a natural extension is an
international comparison where social programs may alter the relationship between SES and
health that we observe in U.S. data. Our goal in this paper is to find whether variation in
institutional structure as measured at the national level could help us understand more about the
flow of causality. The Netherlands has an institutional structure that aims to shield, at least in
the short run, individuals from the economic consequences of a decline in health. In the U.S.,
while there are programs to reduce the severity of the consequences of a decline in heath, the
consequences are certainly not eliminated. In The Netherlands access to health care is universal
whereas in the U.S. the greater use of health care services is associated with higher income.
We will use data from the American Health and Retirement Study (HRS), from the Dutch
Socio-Economic Panel (SEP), and from the Dutch CentER Savings Survey (CSS) to find
qualitatively whether the institutional structures have the expected effect on the relationships
between SES and health; and quantitatively whether the effects are important.
We will use the panel nature of these data sets in conjunction with the differences in
institutional environment to shed light on the positive relationship between health and wealth
that exists in both countries. The panel nature of the datasets allows us to address causality
issues, whereas the differences in institutional settings make it possible to assess some common
explanations for the observed relationship.
In 1992 the HRS surveyed about 12,600 persons approximately aged 51-61 and their
spouses with subsequent waves in 1994, 1996, 1998, and 2000. We use data from the five
waves of HRS. The SEP is a longitudinal household survey representative of the Dutch
population, excluding those living in special institutions like nursing homes. We will use the
Hurd and Kapteyn 8
years 1994-1997 of the SEP. The CSS is an annual panel of about 4000 persons. We use waves
1993 through 1998.
Our measures of SES will be both income and wealth because of the possible differing
relationships between them and health as a function of age.
Institutional differences should affect some of the following explanations for the
observed positive relationship between wealth and health:
• Out-of-pocket health expenses: In The Netherlands such expenses are on the order of 1 or
2 percent of total expenditures, with no discernible relation with age (Alessie et
al.1999b). In the age range of HRS out-of-pocket expenses are rather skewed: for
example, between waves one and two about 33 percent had no out-of-pocket expenses
whereas 7 percent had from $1,000 to $5,000 in out-of-pockets expenses and 2 percent
had more than $5,000 (Hill and Mathiowetz 1998).
• The role of earnings interruptions: The U.S. and The Netherlands differ in their income
maintenance provisions, and hence earnings interruptions may be expected to have
different effects on wealth accumulation in the two countries. In The Netherlands
generous income maintenance provisions aim to mitigate any adverse effect of health
related earnings interruptions.
• Differential access to health care: The Netherlands has essentially a universal health care
system. Thus in The Netherlands, such an explanation would be of limited importance.
We investigate this issue by estimating equations that explain subsequent health on the
basis of past wealth. The extent to which we find differences in this relationship between
the U.S. and The Netherlands can be seen as an indication of the importance of
differential access to health care. Our results indicate that conditional on baseline wealth
Hurd and Kapteyn 9
and health, there is a significant effect of wealth and income on subsequent health status
both in the U.S. and The Netherlands, but in The Netherlands the relation is weaker than
in the U.S. This lends some credence to this explanation.
• Mortality risk: Individuals (or couples) with a higher life expectancy have more reason to
save (see for example, Hurd 1987, 1989, 1998). Hence, we expect healthier individuals to
save more, other things being equal. In The Netherlands, however, annuity income is the
dominant source of income among the elderly, more so than in the U.S. This should lead
to a weaker relationship between health and saving (and hence wealth) in The
Netherlands than in the US.
The remainder of the paper is organized as follows. In Section II we present a small
conceptual model that will help to organize the empirical analysis and which facilitates
interpretation of empirical results. The model consists of three differential equations: one
equation relating income changes to health levels, one equation relating health changes to
spending on health, and one equation relating spending on health care to income. The solution of
the differential equations allows us to characterize in broad terms the relation between dynamic
and cross sectional correlations between health and income and wealth. In Section III we
describe the data, present several descriptive statistics, and document the strong cross sectional
relationship between health and income and wealth. Sections IV and V present the empirical
counterparts of the differential equations presented in Section II. In Section IV we consider
income and wealth changes as a function of health, whereas in Section V we consider health
transitions as a function of income and wealth. In the concluding Section VI, we provide an
interpretation of the empirical results in the light of the theoretical model and draw some general
conclusions.
Hurd and Kapteyn 10
II. An Illustrative Model
To motivate the empirical analysis, we analyze a three-equation model of earnings,
spending on health, and health status. The model is at the individual level, which will allow us to
make comparisons within a population as well as across populations. The model specifies that
health spending depends on income but the dependence can vary across institutional settings.
The dependence on income should be thought of more broadly than out-of-pocket spending, at
least also incorporating work-related health care insurance and perhaps other spending with a
health effect (for example buying a house in an area with good air quality). The evolution of
health is assumed to depend on the amount of spending for health care. Again, spending should
be thought of broadly as spending on nutrition, housing and other attributes that are thought to
influence health. The model specifies that income growth depend on health status, which
incorporates the idea that the healthier will be able to work harder and so have greater income
growth.
Let yt be earnings, ht health status, and st spending on health, all at age t. Then consider
the following simple system of differential equations:
h
+b=
y
=
dt
dy
t
t
t
α
′ (1)
s
+a=
h
=
dt
dh
tt
t
δ
′ (2)
y
+c=
st
t
τ
(3)
Hurd and Kapteyn 11
By substituting (3) into (2) we can write the system in terms of y and h only:
h
+b=
y
=
dt
dy
t
t
t
α
′ (1)
y
+d=
h
=
dt
dh
t
t
t
γ
′ (4)
with d=a+δc and γ=δτ.
The generic solutions to the differential equations are
(5)
l
+
ek
=
hh
t
ht
θ
(6)
l
+
ek
=
yy
t
y
t
θ
By differentiating the generic solutions with respect to t and comparing parameters we obtain
that the solution can be written in terms of the original parameters as follows:
αγ
θ
b/-
e
k=
ht
t (7)
γα
θ
d/-
ke
=
yt
t (8)
with θ=π(αγ) and k a scale parameter, essentially fixing the origin of the t-axis. To fully
determine the solution, we have to specify initial conditions for ht and yt. Let the initial values for
ht and yt at time zero be equal to h0 and y0. We then have five parameters in (7) and (8), α, γ, b, d,
k. If we choose one initial condition, this fixes k and thereby removes any indeterminacy.
The value of the other initial condition is then determined as well. Somewhat arbitrarily, we fix
the initial condition for yt at y0. This implies:
Hurd and Kapteyn 12
α
γ
d/
0+
y
=k (9)
(7) and (8) can then be written as:
-d
e
)d/+
y
(=
yt
0t
θ
γ
(10)
αγ
α
γθ
b/-
e
)d/+
y
(=
ht
0
t (11)
Let us now use this framework to investigate the nature of the relation between income
and health in the population. To do so, we introduce individual heterogeneity. Assume that y0
varies randomly across individuals with mean zero and variance σ2. Furthermore, let b and d be
random variables with mean zero and covariance matrix Σ, both uncorrelated with y0. The
parameters α and γ are assumed to be fixed, i.e. the same for all individuals.
Consider the regressions of ht on yt, y
=
t on ht and h
=
t on yt. Write:
αγ
α
γ
α
γ
b/-d/
u
,
u
+
ytt
t
t≡=
h (12)
Let observations on yt, ut, and ht be stacked in n-vectors y, u, and h. Then the least squares
coefficient in the regression of ht on yt is
yy
yu
+=
yy
yh ′
′
′
′
α
γ
(13)
Hurd and Kapteyn 13
We have
′),( dbCov-
V(d)
1)-
e
(=uy
n2
t
γ
α
γ
θ
1
plim (14)
where V(.) denotes variance and Cov(.) denotes covariance.
)
1-
e
(
V(d)
+
e
)
y
V(=yy
n
2
t
2
t2
0
θθ
γ
′
1
plim (15)
For positive θ and t sufficiently large, this implies that
α
γ
′
′
yy
y≈
h (16)
Similarly, for large enough t the regression of y on h will give
γ
α
≈
′
′
hh
yh (17)
A similar analysis shows that a regression of h
=
t on yt will approximately estimate γ,
whereas a regression of y
=
t on ht will approximately estimate α.
We can make a number of observations on the system of differential equations and their
solution. First of all, we note that the scales of the variables Αearnings≅ and Αhealth≅ have not
been defined yet. So for instance health can be defined in deviation of some Αaverage≅ age
trajectory in a population. The exponential form of the solutions (10) and (11) therefore does not
imply that health will get exponentially better or worse over the life cycle, but rather that health
Hurd and Kapteyn 14
paths diverge with increasing age. Somewhat similarly, we can define earnings in levels, logs or
some other transformation that may fit the data.
Secondly, note the tight connection between income inequality and health inequality. The
regression coefficient in the regression of ht on yt is approximately π(γ/α). The parameters α and
γ represent the strength of the feedback between earnings and health in the dynamic equations
(cf. (1) and (4)). Thus we could have either large or small regression coefficients in the cross
section, when both dynamic feedbacks are small or when both dynamic feedbacks are large.
Assuming, for example, that in The Netherlands both feedbacks are small, while they are larger
in the U.S. we would still see similar cross section coefficients in both countries. If the cross
section regression coefficients are of similar magnitude in both countries then a larger income
inequality in the U.S. will translate into a larger health inequality as well. Several papers have
documented a positive relation between health inequality and income inequality across countries.
See for instance, Van Doorslaer et al. (1997).
Thirdly, the model assumes that health is a continuous and observable variable, yet health
is actually a latent unobservable variable. What we observe are health status indicators
corresponding to the interval in which latent health falls. We shall base estimation on a logistic
model of the determination of health categories and transitions between health categories. To see
how this will affect the interpretation of the results, suppose that latent health is given by the
following equation:
u
+
x
=
httt
λ
β
(18)
Hurd and Kapteyn 15
where ut is a logistically distributed error term, xt is a vector of explanatory variables and λ a
scaling parameter to account for the fact that the variance in health, conditional on observables,
could be different in two populations. If we have an indicator variable It=0 when ht
#
l, then
e
+1
1
=l)
u
+
x
Pr(=0)=
I
Pr( x
+
1
-
tt
t
λ
β
λ
λβ
≤ (19)
Thus the parameter on x has the interpretation β/λ. So if for instance λ is larger in the U.S. the
estimate of β is attenuated relative to The Netherlands.
Fourthly, much of the literature is concerned with the relationships between health and
both income and wealth, particularly at older ages. Wealth is an accumulation of savings over a
lifetime, so that variation in wealth across individuals will reflect variation in income across
individuals over a lifetime. We can show how health might be related to wealth by augmenting
our model with a simple model of consumption behavior.
Suppose that consumption is proportional to income: ct=ηyt. Over the short run there will be
deviations in this relationship due to unexpected events and transitory income. But over longer
periods this may be a reasonable approximation to consumption behavior. In the context of our
model we then obtain
ν
η
t
t
rt
t+
y
e
)-(1
w≈ (20)
where wt is wealth and r is the constant real rate of interest; vt depends on θ and r and on d/γ, so
that it will vary across individuals. This relationship shows that variation in income will induce
Hurd and Kapteyn 16
variation in wealth, and that the variance in wealth will grow at a faster rate than the variance in
income. We would expect that populations with high variance in wealth will have high variance
in health.
III. Data Description
We will be using three datasets: HRS/AHEAD for the U.S. and CSS and SEP for The
Netherlands. The Health and Retirement Study is a panel survey of individuals born from 1931
through 1941 and their spouses or partners. At baseline in 1992 the HRS had 12,652 respondents.
It was nationally representative of the target cohorts, except for over-samples of blacks,
Hispanics and Floridians (Juster and Suzman 1995). This paper uses data from waves one
through five that were fielded in 1992, 1994, 1996, 1998, and 2000. The limitation of the birth
cohorts to 1931-1941 naturally limits the age range in the sample. To have a simple cut off rule
in choosing the age range for all three samples used in the analysis, we only retain individuals in
the sample who are between 51 and 65 years of age.
Household income is income of an individual and spouse or partner. Its components as
measured in the HRS are earnings, asset income, pensions, Social Security, SSI, workers
compensation, unemployment, other government income (veterans' benefits, welfare, food
stamps). Wealth is financial wealth, business and real estate wealth, and housing wealth.
The CentER Savings Survey (CSS) derives from annual interviews with participants in
the so-called CentERpanel. The CentERpanel is run by CentERdata, a subsidiary of CentER at
Tilburg University. The CentERpanel comprises some 2,000 households. These households
have a computer at home, either their own or one provided by CentERdata, and the respondents
in the CentERpanel answer questions that are downloaded to their computer every weekend.
Hurd and Kapteyn 17
Typically the questions for the CSS are asked in May of each year, but in some years the timing
of the CSS has deviated considerably from this. In particular in the first year (1993) technical
difficulties delayed the survey to the extent that some parts of the questionnaire were
administered in early 1994. As a result some parts of the 1994 questionnaire were not
administered at all, including the health questions used in this paper. Initially, the CSS had two
parts: a representative panel of about 2,000 households and a so-called high-income panel of
about 1,000 households, but in 1997 the distinction was abandoned.
The total questionnaire of the CSS is quite long. To reduce respondent burden the
questionnaire is split up in five “modules” that are administered in five separate weekends. The
modules are: demographics and work; housing and mortgages; health and income; assets and
liabilities; and economic psychology. We will be using the waves 1993 and 1995-1998.
Although the technology might suggest otherwise, the CSS is not restricted to households
with (initial) access to the Internet. Respondents are recruited by telephone using random digit
dialing and if they do not have Internet access (or a computer) they are provided with Internet
access (and if necessary a computer) by CentERdata. Panel members are selected on the basis of
a number of demographics so as to match the distribution of these demographics in the
population. The presence of the high-income panel in the earlier years, of course led to an over-
representation of high income/ high wealth households in those years. Whenever presenting
sample statistics, we will therefore employ household weights.
The Socio-Economic Panel (SEP) is conducted by Statistics Netherlands. The SEP is a
longitudinal household survey representative of the Dutch population, excluding those living in
special institutions like nursing homes. The first survey was conducted in April 1984.
Information collected includes demographics, income, labor market participation, hours of work,
Hurd and Kapteyn 18
and (since 1987) assets and liabilities. In some years (1994-1997) respondents were also asked to
assess their own health status. We will use the years 1994-1997 of the SEP.
An evaluation of the quality of the SEP data and a comparison with macro statistics or
other micro data sets is reported in Alessie, Lusardi and Aldershof (1997). We can briefly
summarize their findings as follows: the data on some major components of wealth, such as
housing, mortgage debt, and checking accounts are well reported in the SEP and compare
reasonably well with aggregate statistics. However, some other components, in particular stocks,
bonds, and savings accounts seem under-reported in the SEP, and the level of measurement error
may also change over time. This problem is typical of wealth surveys and can be found in other
similar data sets.2
We have deleted from the sample those cases with missing or incomplete responses in the
assets and liabilities components and in the demographics.3 We have also excluded the self-
employed from the sample, since wealth data for the self-employed are not available after 1989.
Due to these selections, we find that both low and high wealth households have a tendency to
drop out of the sample. Also for the SEP we will use household sample weights when presenting
sample statistics.
A. Descriptive Statistics
To facilitate comparability across the three datasets, we have restricted observations in
the HRS, SEP, and CSS to individuals older than 50 and younger than 65 (for example the
highest age is 64). Table 1 presents a number of descriptive statistics for the three datasets. We
observe that the Dutch samples exhibit a somewhat lower average and median age than the HRS.
One should note that given the way the original HRS cohort has been drawn, average age in the
Hurd and Kapteyn 19
HRS should increase over time, as no new younger individuals are added to the sample we are
using if time progresses. Household size is somewhat higher in the HRS.
Income in the HRS is measured before tax, while in the Dutch data income is measured
after tax, all in 1998 currencies. Although at first sight the different treatment of income in The
Netherlands and in the U.S. may seem to cause problems, we believe that the way we use income
in the quantitative analyses (by constructing quartiles of the distribution) is robust against the
different treatment of taxes in both countries. For the use of quantiles it is only necessary that the
ranking of incomes before and after tax is the same, which would appear a reasonable
approximation in most cases.
Income in CSS is somewhat higher than in SEP. This is consistent with a suspicion of
underestimation of income in SEP (comparison with external sources suggests an
underestimation in SEP by about 10 percent on average). More strikingly, net worth and assets
are much higher in CSS than in SEP. To a fairly large extent this can be ascribed to
underestimation in SEP as well, as noted above. The CSS-questionnaire is much more detailed
than the SEP questionnaire, which makes it likely that more components of wealth are picked up
than in the SEP. Furthermore, self-employed are included in CSS, but not in the SEP. Alessie
and Kapteyn (1999a) compare the SEP wealth data with external data published by Statistics
Netherlands and find that average net worth in SEP may be underestimated by about 20 percent.
Comparing this to sample means and medians reported for CSS in Table 1 would then suggest
some overestimation in the CSS (even taking into account the omission of the self-employed
from the SEP). This may point to a less than perfect reweighting of the data by means of the
sample weights used in the CSS, resulting in an over-representation of high wealth households.
Hurd and Kapteyn 20
Altogether, one may surmise that CSS and SEP provide respectively an upper and a lower bound
on the wealth holdings of Dutch households.
Finally, with respect to Table 1, we should note that the definitions of education used in
the two Dutch datasets differ, so that the distributions across education levels are not
comparable. In the empirical analyses using education, we will therefore always distinguish
between the education definitions in CSS and SEP.
Table 2 provides the distribution of self assessed health. The verbal labels associated with
the categories are given in the table as well. Clearly, the definitions of the health categories vary
by dataset. The frequency distribution of health categories is more similar between the two
Dutch datasets than between the HRS on the one hand and the Dutch datasets on the other hand.
Taken at face value, the distribution of health levels is more dispersed in the American data.
These differences may reflect true differences in health dispersion across the two countries or
just be the effect of different wordings or different meanings attached to the verbal labels in
different cultures (cf. for example Finch et al. 2002). These possibilities will have to be kept in
mind when analyzing the different patterns across socio-economic groups in the two countries.
The relation with education is qualitatively similar across the datasets (see Table 3) and
shows the familiar pattern that the distribution of health shifts in the direction of better health if
education goes up.
Tables 4, 5, and 6 show the relation between self reported health and BMI, smoking, and
alcohol consumption. These health behaviors are not available in SEP, so comparisons only
involve HRS and CSS. Table 4, suggests that the Dutch weigh less than the Americans and
confirm that a high BMI is bad for health (with minor non-monotonicities in some places).
Hurd and Kapteyn 21
The Dutch between fifty and sixty-five appear to smoke a bit more than the Americans
(Table 5). Smoking is bad for health in both countries. According to Table 6, the Dutch drink
more than the Americans. With some exaggeration, one could say that drinking more than four
glasses of alcohol is bad for health in the U.S. and good for health in The Netherlands.
There is a monotonic relationship between health and wealth in both countries (Table 7).
The same is true of the relation between health and income (Table 8), but the relation is less
steep in The Netherlands than in the U.S. Partly this simply reflects the more equal income
distribution in The Netherlands.
This point is even more clearly demonstrated by Table 9. Table 9 reports ordered logits of
health categories on income wealth quartiles, education, and a number of other controls, which
are not reported (but mentioned at the bottom of the table). To allow for nonlinearities in the
relation between health and income or wealth we define nine income-wealth categories: for each
of income and wealth we distinguish low, medium, and high, where low is the lowest quartile,
medium is either the second or third quartile, and high is the highest quartile. The coding of
income and wealth by quartiles is done separately for each year and for each dataset.4 The
purpose of the tables is not to suggest any causality, but rather to characterize the strength of the
relationship between health and SES in both countries. In the table we show results for The
Netherlands based on pooled data of CSS and SEP. When pooling CSS and SEP we retain
different education dummies for the two datasets in light of the differences in definition, as
discussed earlier. A test of equality of the income/wealth dummies across the two datasets does
not lead to rejection (Π2(8) 9.37 p=0.31), which justifies the pooling. Table 9 shows a much
steeper gradient of health with SES for HRS than for CSS/SEP (for instance the odds ratio of
income high and wealth high is 8.61 for HRS and 2.46 for CSS/SEP), although the relation is
Hurd and Kapteyn 22
statistically highly significant in all datasets. Education also has a highly significant relation with
health, but again in The Netherlands the relation is less steep.
IV. The Impact of Health on Wealth and Income Changes
We have established that in both the U.S. data and the Dutch data there is a strong
positive cross-section association between health and indicators of socio-economic status. In
keeping with the theoretical model we now consider the effect of health status on income and
wealth changes. The relation between health status and income changes would be similar to
equation (4) above.
Table 10 shows a strong effect of health level on percentage wealth changes in the U.S.
data. For instance, relative to people in excellent health, wealth changes of people in poor health
lag by 16 percent. The relation between wealth changes and health is much weaker in the Dutch
datasets. In the pooled CSS/SEP data we find statistically significant, but small effects. For
instance, relative to people in the healthiest category, the wealth change of people in the not so
good/ fair category lags by 3.6 percent. For the separate Dutch data we find some weak effects,
but these are only statistically significant in the SEP. Also note that a test for equality of health
effects on wealth in the two Dutch datasets leads to rejection at the 2 percent level. Thus the
pooled results are in principle based on a misspecified model. Hence we report both results based
on the separate datasets and results based on the pooled data.
The effects of health on income are statistically marginally significant in the HRS and
significant in the pooled SEP/CSS data. A test for equality of the health effects in CSS and SEP
leads to acceptance (p=.55). This outcome is somewhat surprising, in view of the relatively
Hurd and Kapteyn 23
extensive income maintenance programs in The Netherlands. The significance of the effects in
The Netherlands seems to be largely driven by the relatively large effect for poor/very bad
health. For the other health categories the effects are about equal to those for the U.S. data, if not
smaller.
V. The Impact of Wealth and Income on Health Transitions
We will quantify the effects of wealth and income on health changes via ordered logit
estimation of the rate of health transition from one wave to another. We consider transitions from
three initial health levels: from health being in the top-two categories, from health being in the
middle category and from health being in the two bottom categories. Tables 12 through 14
present the results. As in Section III, we also consider estimation results if we pool the CSS and
SEP data (while keeping separate education dummies). Since in all three cases we accept the null
that income and wealth effects are the same across SEP and CSS we only present the pooled
results for the Dutch data.
Table 12 gives the estimated effects on the transition from excellent/very good health in
HRS or from excellent or good health (CSS) or from very good or good (SEP) to each of the
possible five destination states. A positive coefficient increases the chances of maintaining
health in the top-category or, in the event of a transition, that the transition will be to the middle
category rather than to the bottom categories. We have added a dummy for health being very
good (HRS) or good (CSS, SEP) to indicate where in the top categories one is at baseline.
Income and wealth have a significant influence on health transitions in both countries, but
the effect appears to be considerably larger in the U.S. than in The Netherlands. This would
Hurd and Kapteyn 24
suggest ( (cf. equation (4)) to be bigger in the U.S. than in The Netherlands. In both countries it
also appears that wealth is more important than income. In The Netherlands the combined effect
of income and wealth is dominated by wealth. In the U.S. we observe a non-linearity in the effect
of income and wealth: an increase in wealth at all income levels increases the odds of remaining
in the top health category; but an increase in income only increases the odds when wealth is low
or medium, not when wealth is high.
In Table 13 the baseline category is good health for HRS and fair health for both CSS and
SEP. As before, a positive coefficient increases the probability of a transition to better health or
reduces the probability of a transition to worse health. The wealth income interactions are
statistically significant in both countries. As before, the economic variables appear to have a
stronger effect in the HRS than in the Dutch data, but the differences are fairly minor.
In Table 14, the baseline category is fair or poor for HRS, not so good or poor for CSS,
and bad or very bad for SEP. The effects of income and wealth on the health transitions are now
totally insignificant in the Dutch data, possibly due to the modest number of observations.
VI. Interpretation and Conclusions
The conceptual model presented in Section 2 refers to the relationship between health and
income. The theoretical relationship between wealth and health is considerably more
complicated involving the propensity to consume out of income and the interest rate. We have
not fully developed that theory beyond the observation that in populations where there is large
variation in wealth there should be large variation in health. Therefore, when discussing the
results in relation to the model we will concentrate on the results relating income and health. We
Hurd and Kapteyn 25
should reiterate however the observation in the Introduction that healthier individuals have more
reasons to save for retirement, but that this reason is substantially less prominent in The
Netherlands, where most retirement consumption is financed out of annuity income. We do
indeed find that the effect of health on wealth is considerably smaller in the U.S. than in The
Netherlands.
Based on the cross-section estimations of the effects of income and wealth on health
status (Table 9) we can calculate the average change in relative risk holding constant wealth by
averaging the relative risk over each wealth category. For example, in the HRS the average risk
of being in a higher health category for someone in the top income quartile is computed to be
2.49 greater than the risk of someone in the lowest income quartile. In The Netherlands the
relative risk is 1.90. Relating this result back to our conceptual model (cf. equations (11) and
(16)), it indicates that (/∀ is somewhat greater in the U.S. That is, the change in health is more
strongly related to the income level in the U.S. than in the Netherlands.
Based on the three panel estimations (Tables 12, 13, and 14) we may calculate an
estimate of ( in a similar way, by assuming that these equations are the empirical counterpart of
(4). The estimations do not only hold wealth constant, but also baseline health constant in one of
the three health status categories. In the HRS among those in the top income quartile the average
relative risk of transiting to a higher health state can be calculated as 1.66 greater than the risk of
someone in the lowest income quartile. In The Netherlands the risk is 1.23 greater. Thus the
panel transitions indicate a higher level of ( in the U.S. than in The Netherlands.
By combining these findings we can estimate ∀ to be 0.27 for the U.S. and 0.34 for The
Netherlands. These estimates can be compared to the estimation results presented in Table 11.
According to our theoretical model, the coefficient on health for panel changes in income has the
Hurd and Kapteyn 26
interpretation of α. Given the categorical nature of the health measure and the definitional
differences between the categories in the two countries a comparison is somewhat tenuous, but
qualitatively it appears that α, as derived from Table 11, is greater in the Netherlands than in the
U.S., at least as measured by the largest of the coefficients (on poor health). Thus, we find
broadly consistent estimates of the underlying parameters of our theoretical model. However,
we clearly need better measures that can be compared with more confidence across the countries.
The current datasets do not provide such comparable measures. Also we note that these
calculations ignore possible differences in the variance of unobserved heterogeneity in both
countries (8 in equation (19)).
We began the cross-country comparison with the observation that if national policies alter
both the financing of health care services and other inputs into health production, and the
relationship between health and income, we should find predictable differences in the
relationship between health and income in cross-section and in panel. In The Netherlands health
care is universal and practically independent of income whereas that would not be the case in the
U.S. In The Netherlands income redistribution programs reduce the strength of the relationship
between health and income. We found differing relationships between health and income in the
two countries, and the differences are consistent with what our conceptual model would predict.
Clearly our analysis invites several improvements. On the conceptual side a more
complex model relating health, wealth, and income in a three-equation system of differential
equations appears to be natural extension. On the data side, one would want to consider more
countries and a much wider array of health measures. Although at this moment micro-panel
datasets measuring health, income and wealth are only available in a very limited set of
countries, the movement in various countries to emulate the U.S. Health and Retirement Study
Hurd and Kapteyn 27
provides an exciting perspective on making progress in quantifying the relation of health and
SES and the role of institutions in amending this relationship.
Hurd and Kapteyn 28
References
Adams, Peter, Michael D. Hurd, Daniel McFadden, Angela Merrill and Tiago Ribeiro.
2003. “Healthy, Wealthy and Wise? Tests for Direct Causal Paths between Health and
Socioeconomic Status.” Journal of Econometrics. 112 (1), pp. 3-56
Alessie, Robert, Annamaria. Lusardi, and Trea Aldershof. 1997. "Income and Wealth over the
Life Cycle: Evidence from Panel Data.” The Review of Income and Wealth 43: 1-32.
Alessie, Robert., and Arie Kapteyn. 1999a. “Savings, Pensions and Portfolio Choice in The
Netherlands.” Free University Amsterdam.
Alessie, Robert, Annamaria. Lusardi, Arie Kapteyn. 1999b. “Saving after Retirement: Evidence
from Three Different Surveys.” Labour Economics 6: 277-310.
Avery, Robert. B., Gregory E. Elliehausen, and Arthur. B. Kennickell. 1988. "Measuring Wealth
with Survey Data: An Evaluation of the 1983 Survey of Consumer Finances” The
Review of Income and Wealth 34: 339-369.
. 1988. "Measuring Wealth with Survey Data: An Evaluation of the 1983 Survey of
Consumer Finances." The Review of Income and Wealth 34: 339-369.
Avery, Robert and Arthur Kennickell. 1991. "Household Saving in the US." Review of Income
and Wealth 37: 409-432.
Hurd and Kapteyn 29
Berkman, Lisa F. 1988. “The Changing and Heterogeneous Nature of Aging and Longevity: A
Social and Biomedical Perspective.” In Annual Review of Gerontology and Geriatrics, 8,
eds. George L. Maddox and M. Powell. Lawton, 37-68.
Boffetta, Paolo, and Lawrence Garfinkel. 1990. “Alcohol Drinking and Mortality among Men
Enrolled in an America Cancer Society Prospective Study.” Epidemiology 1(5):343-48.
Camphuis, Herman. 1993. "Checking, Editing and Imputation of Wealth Data of the
Netherlands Socio-Economic Panel for the period 87-89." VSB Progress Report n. 10,
CentER, The Netherlands: Tilburg University.
Davies, James. 1979. "On the Size Distribution of Wealth in Canada." Review of Income and
Wealth 25: 237-259.
Feinstein, Jonathan. 1993. “The Relationship between Socioeconomic Status and Health: A
Review of the Literature.” The Milbank Quarterly 71(2):279-322.
Finch, Brian K., Robert A. Hummer, Maureen Reindl, William A. Vega. 2002. “The Validity of
Self-Rated Health among Latino(a)s.” Mimeo, University of California, Berkeley, School
of Public Health.
Fuchs, Victor R. “Time Preference and Health: an Exploratory Study. In Economic Aspects of
Health. 1982. Chicago: University of Chicago Press.
Goldman, Noreen. 2001. “Social Inequalities in Health: Disentangling the Underlying
Hurd and Kapteyn 30
Mechanisms,” in Population Health and Aging - Strengthening the Dialogue Between
Epidemiology and Demography. Maxine Weinstein, Albert I. Hermalin, and Michael A.
Stoto, eds. Annals of the New York Academy of Sciences. 954: 118-139.
Grossman, Michael. 1972. The Demand for Health-Atheoretical and Empirical Investigation.
New York: National Bureau of Economic Research.
Hill, Daniel H. and Nancy A. Mathiowetz. 1998. "The Empirical Validity of the HRS Medical
Expenditure Data: A Model to Account for Different Reference Periods." Ann Arbor:
HRS Working Paper Series.
Hurd, Michael D. 1987. “Saving of the elderly and desired bequests.” American Economic
Review 77: 289-312.
Hurd, Michael D. 1989. “Mortality risks and bequests”, Econometrica 57: 779-813.
Hurd, Michael D. 1998. “Mortality risk and consumption by couples.” Santa Monica: RAND
Working Paper DRU-2061.
Hurd, Michael D. and Kathleen McGarry. 1995. "Evaluation of the Subjective Probabilities of
Survival in the HRS." Journal of Human Resources 30: S268-S292.
Hurst, Erik., Ming Ching. Luoh, and Frank.P. Stafford. 1998.“Wealth Dynamics of American
Hurd and Kapteyn 31
Families: 1984-1994.” Brookings Papers on Economic Activity 1: 267-337.
Juster, F. Thomas and Richard M. Suzman. 1995, “An Overview of the Health and Retirement
Study.” Journal of Human Resources 30 (supplement): S7-S56.
Kitagawa, Evelyn M., and Philip M. Hauser. 1973. Differential Mortality in the United States: A
Study in Socioeconomic Epidemiology. Cambridge, Mass.: Harvard University Press.
Marmot, Michael G., George Davey Smith, Stephen Stansfeld, Chandra Patel, Fiona North, J.
Head, Ian White, Eric Brunner, and Amanda Feeny. 1991. “Health Inequalities among
British Civil Servants: The Whitehall II Study.” Lancet, June 8, pp. 1387-93.
McClellan, Mark. 1998. “Health Events, Health Insurance and Labor Supply: Evidence from
the Health and Retirement Survey.” In Frontiers in the Economics of Aging, ed. David
Wise, 301-346. Chicago: University of Chicago Press.
Robert, Stephanie and James House. 1994. “Socioeconomic Status and Health Over the Life
Course.” In Social Structures, Quality of Life, and Aging, Ronald P. Abeles, Helen C.
Gift and Marcia G Ory, eds., 253-274. New York: Springer Publishing Company.
. 2000. “Socioeconomic Inequalities in Health: Integrating Individual-, Community-, and
Societal-Level Theory and Research.” In Handbook of Social Studies in Health and
Medicine, Gary Albrecht, Ray Fitzpatrick and Susan Scrimshaw, eds. 115-135. London:
Sage Publications.
Hurd and Kapteyn 32
Shaper, A. G. 1990. “Alcohol and Mortality: A Review of Prospective Studies.” British
Journal of Addiction 85:837-47.
Smith, James P. 1999. “Healthy Bodies and Thick Wallets: The Dual Relationship between
Health and Economic Status.” Journal of Economic Perspectives 13: 145-166.
Van Doorslaer, Eddy, Adam Wagstaff, Han Bleichrodt, Samuel Calonge, Ulf-G. G. Gerdtham,
Michael Gerfin, José Geurts, Lorna Gross, Unto Häkkinen, Robert E. Leu, Owen
O’Donnell, Carol Propper, Frank Puffer, Marisol Rodriguez, Gun Sundberg, Olaf
Winkelhake. 1997. “Income-related inequalities in health: some international comparisons.”
Journal of Health Economics 16: 93-112.
Wilkinson, Richard G. 1999. “Health, Hierarchy, and Social Anxiety.” In Socioeconomic Status and
Health in Industrial Nations, Nancy E. Adler, Michael Marmot, Bruce S. McEwen, and
Judith Stewart, eds. Annals of the New York Academy of Sciences, 896: 48-63
Hurd and Kapteyn 33
Appendix 1. Tables
Table 1
Descriptive Statistics For Three Datasets
HRS CSS SEP
Mean (sd) Median
(N)
Mean (sd) Median
(N)
Mean (sd) Median
(N)
Age 58.0 58 57.0 57 56.9 57
(3.6) (41998) (4.1) (6464) (4.1) (6175)
HH-size 2.7 2 2.2 2 2.0 2
(1.6) (29020) (1.04) (3659) (.89) (2884)
Assets 302.2 126.7 161.6 125.3 103.3 85.3
(993.0) (29020) (206.6) (3099) (128.9) (2884)
Liabilities 30.0 3.3 40.3 10.1 22.5 2.6
(75.6) (29020) (68.1) (3099) (40.2) (2884)
Net worth 272.2 99.5 121.3 81.0 80.7 49.9
(991.1) (29020) (175.6) (3099) (108.4) (2884)
HH income 56.7 38.6 25.8 22.2 24.2 21.5
(94.7) (29020) (26.8) (3753) (13.6) (2791)
.22 0 .48 0 .47 0 Education l.t.
High School (41939) (6444) (5632)
.55 1 .22 0 .37 0 High School
(41939) (6444) (5632)
.23 0 .30 0 .16 More than
High School (41939) (6444) (5632)
Employed .62 1 .43 0 .40 0
(41923) (5454) (6175)
Explanation: For each variable the first row shows sample mean and median; the second row
shows standard deviation and number of observations in parentheses. Money amounts are in
thousands of dollars for HRS and thousands of euros for SEP and CSS. Household incomes in
CSS and SEP are after tax; for HRS incomes are before tax. Education definitions vary across the
three samples, so that distributions of education level are only approximately comparable across
samples. Money amounts are all reported at the household level. Fractions are reported at the
respondent level. Numbers of observations vary as a result of this, but also as a result of varying
numbers of missing observations. Entries for education and employment status are fractions. All
sample statistics are weighted.
Hurd and Kapteyn 34
Table 2
Distribution of Self-Assessed Health
HRS CSS SEP
Frequency % Frequency % Frequency %
Excellent(hrs,css)/very good(sep) 7845 18.8 841 19.7 858 13.8
Very good(hrs)/good(css,sep) 12963 31.1 2389 56.2 3239 52.1
Good(hrs)/fair(css, sep) 11876 28.5 820 19.3 1814 29.2
Fair(hrs)/not so good(css)/bad(sep) 6029 14.5 180 4.2 268 4.3
Poor(hrs,css)/very bad(sep) 2926 7.0 23 .5 38 .6
Total 41640 100 4254 100 6218 100
Frequencies are weighted; the definitions of the health categories vary by sample and are
indicated in the category names in the table.
Hurd and Kapteyn 35
Table 3
Education and Self-Reported Health
HRS CSS SE
P
<HS HS >HS <HS HS >HS <HS HS >HS
Excellent(hrs,css)/
very good(sep)
8.7 18.7 28.9 16.6 21.8 22.6 11.4 13.3 21.1
Very good(hrs)/
good(css,sep)
18.7 33.0 38.4 53.8 58.9 57.5 49.6 55.6 53.2
Good(hrs)/
fair(css, sep)
30.2 29.7 24.1 23.4 15.4 16.3 33.3 26.6 22.7
Fair(hrs)/
not so good(css)/ bad(sep)
25.7 13.4 6.4 5.3 3.7 3.1 4.9 4.1 2.3
Poor(hrs,css)/
very bad(sep)
16.6 5.2 2.3 .8 .1 .5 .7 .4 .6
Total 100 100 100 100 100 100 100 100 100
Frequencies are weighted; the definitions of the health categories vary by sample and are
indicated in the category names in the table. HS means high school; the definitions of education
are only weakly comparable across datasets.
Hurd and Kapteyn 36
Table 4
BMI and Self-Reported Health
HRS CSS
Mean Median Mean Median
Excellent(hrs,css) 25.6 25.1 24.9 24.7
Very good(hrs)/good(css) 26.7 26.2 25.2 25.0
Good(hrs)/fair(css) 27.9 27.3 26.0 25.5
Fair(hrs)/not so good(css) 28.7 27.9 24.8 24.2
Poor(hrs,css) 28.4 27.4 29.7 26.0
Means and medians are weighted.
Hurd and Kapteyn 37
Table 5
Distribution of Self-Reported Health by Smoking or Non-Smoking
HRS CSS
No Yes No Yes
Excellent(hrs,css) 20.6 13.3 21.8 15.2
Very good(hrs)/good(css) 32.3 27.5 56.2 56.0
Good(hrs)/fair(css) 27.8 30.7 17.6 23.2
Fair(hrs)/not so good(css) 13.3 18.3 4.0 4.9
Poor(hrs,css) 6.0 10.3 0.4 0.8
Total 76.7 23.3 69.2 30.8
“Yes” means the respondent smokes now; “No” means the respondent does not smoke now. The
bottom row gives the percentages of smokers and non-smokers in both samples. All percentages
are based on weighted data.
Hurd and Kapteyn 38
Table 6
Distribution of Self-Reported Health by Alcohol Consumption
HRS CSS
No Yes No Yes
Excellent(hrs,css) 16.8 11.2 19.5 21.7
Very good(hrs)/good(css) 32.1 24.8 56.1 56.5
Good(hrs)/fair(css) 28.9 36.0 19.5 17.6
Fair(hrs)/not so good(css) 15.2 20.4 4.3 4.0
Poor(hrs,css) 6.9 7.5 0.6 0.3
Total 97.8 2.2 89.6 10.4
“Yes” means the respondent drinks more than four glasses of alcohol per day; “No” means the
respondent drinks less or not at all. The bottom row gives the percentages of both groups in both
samples. All percentages are based on weighted data.
Hurd and Kapteyn 39
Table 7
Health and Wealth
HRS CSS SEP
Mean Median Mean Median Mean Median
Excellent(hrs,css)/very good(sep) 477.2 233.5 160.1 121.0 96.7 82.0
Very good(hrs)/good(css,sep) 408.5 192.2 131.4 93.2 93.7 70.2
Good(hrs)/fair(css, sep) 272.3 126.1 119.6 73.9 76.8 47.3
Fair(hrs)/not so good(css)/bad(sep) 184.6 66.2 99.4 60.5 63.9 15.3
Poor(hrs,css)/very bad(sep) 98.8 30.0 62.8 .9 46.3 11.2
Money amounts are in thousands of dollars for HRS and thousands of euros for SEP and CSS.
Wealth is household net worth.
Hurd and Kapteyn 40
Table 8
Health and Income
HRS CSS SEP
Mean Median Mean Median Mean Median
Excellent(hrs,css)/very good(sep) 85.6 60.2 29.1 27.0 29.5 27.2
Very good(hrs)/good(css,sep) 70.1 52.3 26.8 23.8 28.1 26.2
Good(hrs)/fair(css, sep) 54.8 40.8 25.0 20.6 23.6 20.9
Fair(hrs)/not so good(css)/bad(sep) 38.5 26.7 22.8 19.6 19.8 18.1
Poor(hrs,css)/very bad(sep) 24.9 15.6 20.9 13.0 17.1 17.5
Money amounts are in thousands of dollars for HRS and thousands of euros for SEP and CSS.
Household incomes in CSS and SEP are after tax; for HRS incomes are before tax.
Hurd and Kapteyn 41
Table 9
The Cross Section Association Between Health and SES In the Two Countries
HRS Odds
Ratios
CSS/SEP Odds
Ratios
Income low, w medium 0.797 2.22 0.098 1.10
(14.55)** (0.80)
Income low, w high 1.681 5.37 0.059 1.06
(18.55)** (0.33)
Income med, w low 0.971 2.64 0.277 1.32
(18.36)** (2.40)*
Income med, w medium 1.389 4.01 0.603 1.83
(28.38)** (5.74)**
Income. med, w high 1.814 6.13 0.615 1.85
(30.99)** (4.96)**
Income high, w low 1.406 4.08 0.451 1.57
(12.71)** (1.86)
Income. high, w medium 1.829 6.22 0.742 2.10
(30.97)** (6.07)**
Income high, w high 2.153 8.61 0.902 2.46
(36.61)** (7.18)**
High School 0.681 1.98
(17.85)**
More than High School 1.058 2.88
(21.98)**
High School (SEP) 0.051 1.05
(0.61)
More than High School (SEP) 0.375 1.45
(3.06)**
High School (CSS) 0.315 1.37
(2.52)*
More than High School (CSS) 0.168 1.18
(1.42)
Observations 42193 9423
pseudo-R2 0.07 0.03
chi2 Income/wealth 1520.32 84.38
p-value 0.00 0.00
Robust z statistics in parentheses
* - Significant at 5%
** - Significant at 1%
Explanation: The table presents ordered logits of the self reported health categories on a number
of characteristics. Controls not reported here include year dummies, a cubic in age, marital status,
household size, gender. The last column presents results for pooled data of SEP and CSS, where
education level is not pooled, since education definitions differ across the two datasets.
Hurd and Kapteyn 42
Table 10
Health Levels And Wealth Changes
(1) (2) (3) (4)
HRS CSS SEP CSS/SEP
Health=very good(hrs)/ good(css,sep) 0.380 3.125 1.577 1.901
(0.34) (1.74) (1.25) (2.12)*
Health=good(hrs)/fair(css, sep) -2.004 -1.058 0.905 -0.052
(1.76) (0.47) (0.65) (0.05)
Health=fair(hrs)/not so good(css)/bad(sep) -6.747 -2.008 -3.941 -3.598
(4.90)** (0.52) (1.68) (2.02)*
Health=poor(hrs,css)/very bad(sep) -16.029 -8.973 -0.239 -2.177
(8.87)** (0.85) (0.04) (0.47)
High School 9.924 1.483 1.823
(10.01)** (0.78) (1.91)
More than HighSchool 14.111 6.387 3.839
(11.60)** (3.80)** (3.18)**
Divorced/Separated -15.140 -9.740 -2.008 -4.062
(11.95)** (2.72)** (1.10) (2.76)**
Widowed -12.804 0.975 -0.131 -0.504
(8.34)** (0.24) (0.07) (0.31)
Not Married -9.500 -5.380 1.144 -0.182
(5.37)** (1.71) (0.95) (0.18)
High School (SEP) 1.662
(1.94)
More than HighSchool (SEP) 2.916
(2.59)**
High School (CSS) 3.404
(2.82)**
More than High School (CSS) 7.922
(7.83)**
Observations 30625 2509 3918 6518
F-test health 27.76 2.00 1.92 4.19
p-value 0.00 0.09 0.00 0.00
F-test education 73.90 7.72 5.34 15.85
p-value 0.00 0.00 0.10 0.00
Absolute value of t statistics in parentheses
* - significant at 5%
** - significant at 1%
Explanation: The table presents median regressions of percentage wealth changes from one
period to the next. Controls not reported here include a cubic in age, gender and household size.
The last column presents results for pooled data of SEP and CSS, where education level is not
pooled, since education definitions differ across the two datasets. A chi-squared test for equality
of the health dummies across CSS and SEP yields a p-value of .02.
Hurd and Kapteyn 43
Table 11
Health Levels And Income Changes
(1) (2) (3) (4)
HRS CSS SEP CSS/SEP
Health=very good(hrs)/ good(css,sep) -0.331 -0.273 -1.221 -0.946
(0.57) (0.27) (2.11)* (2.19)*
Health=good(hrs)/fair(css, sep) -0.151 -0.991 -0.245 -0.201
(0.26) (0.76) (0.38) (0.40)
Health=fair(hrs)/not so good(css)/bad(sep) -1.882 4.059 -1.224 -0.503
(2.67)** (1.80) (1.14) (0.58)
Health=poor(hrs,css)/very bad(sep) -0.303 -7.701 -2.082 -5.189
(0.33) (1.20) (0.81) (2.29)*
High School 1.319 0.026 0.002
(2.63)** (0.02) (0.00)
More than HighSchool 2.266 1.022 0.457
(3.64)** (1.07) (0.82)
Divorced/Separated -2.134 1.804 1.751 1.766
(3.35)** (0.89) (2.10)* (2.48)*
Widowed -2.607 3.026 0.673 2.561
(3.39)** (1.35) (0.73) (3.23)**
Not Married 0.739 0.876 1.046 1.010
(0.84) (0.49) (1.87) (1.98)*
High School (SEP) 0.215
(0.50)
More than High School (SEP) 0.827
(1.45)
High School (CSS) -0.717
(1.31)
More than High School (CSS) -0.188
(0.40)
Observations 31226 2895 3643 6624
F-test health 2.31 1.59 1.96 2.75
p-value 0.06 0.50 0.10 0.03
F-test education 6.77 0.70 0.39 1.32
p-value 0.00 0.17 0.68 0.26
Absolute value of t statistics in parentheses
* - Significant at 5%
** - Significant at 1%
Hurd and Kapteyn 44
Table 12
Ordered Logits: Health In TheTop-Two Categories At Baseline
(1) (2) (3) (4)
HRS Odds
Ratios
CSS/SEP Odds
Ratios
Income low, w med 0.389 1.48 -0.166 .847
(3.65)** (0.98)
Income low, w high 0.911 2.49 0.088 1.09
(7.18)** (0.38)
Income. med, w low 0.405 1.50 -0.025 .975
(4.00)** (0.16)
Income med, w med 0.671 1.96 0.198 1.22
(7.58)** (1.50)
Income med, w high 0.949 2.58 0.338 1.40
(10.03)** (2.27)*
Income. high, w low 0.760 2.14 0.153 1.17
(4.36)** (0.62)
Income high, w med 0.935 2.55 0.236 1.27
(9.69)** (1.54)
Income high, w high 1.038 2.82 0.320 1.38
(10.99)** (2.07)*
High School 0.553 1.74
(9.85)**
More than High School 0.770 2.16
(12.55)**
Health==very
good(hrs)/good(css,sep)
-1.476 .228
(36.73)**
High School(SEP) -0.031 .969
(0.28)
More than High School (SEP) 0.376 1.46
(2.58)**
Health=good(SEP) -2.007 .135
(15.02)**
High School (CSS) 0.015 1.02
(0.13)
More than High School (CSS) -0.116 .890
(1.01)
Health=good(CSS) -2.522 .080
(19.50)**
Observations 16272 4408
pseudo-R2 0.08 0.13
chi2 inc/wealth 205.32 16.66
p-value 0.00 0.03
Robust z statistics in parentheses
* - Significant at 5%
Hurd and Kapteyn 45
** - Significant at 1%
Explanation: The table presents ordered logits of next period’s self reported health status.
Controls not reported here include a cubic in age, gender, year dummies, marital status, gender,
and household size. The last column presents results for pooled data of SEP and CSS, where
education level is not pooled, since education definitions differ across the two datasets. A chi-
squared test for equality of the income and wealth coefficients in SEP and CSS yields a p-value
of .60.
Hurd and Kapteyn 46
Table 13
Ordered Logits: Health In The Middle Category At Baseline
(1) (2) (3) (4)
HRS Odds
Ratios
CSS/SEP Odds
Ratios
Income low, w med 0.327 1.39 0.154 1.17
(3.52)** (0.79)
Income low, w high 0.573 1.77 -0.143 .867
(3.44)** (0.51)
Income med, w low 0.407 1.50 0.181 1.20
(4.46)** (0.84)
Income med, w med 0.587 1.80 0.327 1.39
(7.20)** (1.87)
Income med, w high 0.785 2.19 0.179 1.20
(7.79)** (0.82)
Income high, w low 0.640 1.90 -0.699 .497
(3.73)** (1.28)
Income high, w med 0.720 2.05 0.537 1.71
(7.36)** (2.22)*
Income high, w high 0.769 2.16 0.818 2.27
(7.90)** (3.36)**
High School 0.277 1.32
(5.44)**
More than High School 0.459 1.58
(6.93)**
High School (SEP) 0.021 1.20
(0.14)
More than High School (SEP) 0.367 1.44
(1.70)
High School (CSS) 0.440 1.55
(1.50)
More than High School (CSS) 0.406 1.50
(1.74)
Observations 9857 1568
pseudo-R2 0.02 0.02
chi2 inc/wealth 88.09 20.24
p-value 0.00 0.01
Robust z statistics in parentheses
* - Significant at 5%
** - Significant at 1%
Explanation: The table presents ordered logits of next period’s self reported health status.
Controls not reported here include a cubic in age, gender, year dummies, marital status, gender,
and household size. The last column presents results for pooled data of SEP and CSS, where
education level is not pooled, since education definitions differ across the two datasets. A chi-
squared test for equality of the income and wealth coefficients in SEP and CSS yields a p-value
of .61.
Hurd and Kapteyn 47
Table 14
Ordered Logits: Health In The Bottom Two Categories At Baseline
(1) (2) (3) (4)
HRS Odds
Ratios
CSS/SEP Odds
Ratios
Income low, w med 0.214 1.24 -0.243 .784
(2.94)** (0.51)
Income low, w high 0.161 1.17 0.456 1.56
(0.85) (0.74)
Income med, w low 0.332 1.39 -0.195 .823
(4.29)** (0.33)
Income. med, w med 0.436 1.55 0.271 1.31
(6.27)** (0.57)
Income med, w high 0.619 1.86 -0.554 .575
(5.98)** (0.80)
Income. high, w low 0.520 1.68 -0.986 .373
(2.14)* (1.06)
Income. high, w med 0.803 2.32 0.389 1.48
(6.57)** (0.59)
Income. high, w high 0.913 2.50 0.621 1.86
(7.31)** (0.72)
High School 0.174 1.19
(3.53)**
More than High School 0.256 1.29
(2.81)**
Health==fair(hrs)/not so
good(css)/bad(sep)
1.803 6.06
(29.78)**
High School (SEP) -0.230 .795
(0.64)
More than High School(SEP) -0.405 .667
(0.55)
Health=bad(SEP) 1.704 5.50
(2.20)*
High School (CSS) -0.442 .642
(0.85)
More than High School(CSS) 0.186 1.20
(0.31)
Health=not so good(CSS) 3.181 24.1
(4.35)**
Observations 7799 296
pseudo-R2 0.09 0.11
chi2 inc/wealth 87.45 7.73
p-value 0.00 0.46
Robust z statistics in parentheses
* - Significant at 5%
Hurd and Kapteyn 48
** - Significant at 1%
Explanation: The table presents ordered logits of next period’s self reported health status. Controls
not reported here include a cubic in age, gender, year dummies, marital status, gender, and
household size. The last column presents results for pooled data of SEP and CSS, where education
level is not pooled, since education definitions differ across the two datasets. A chi-squared test for
equality of the income and wealth coefficients in SEP and CSS yields a p-value of .27.
Hurd and Kapteyn 49
1 See Kitagawa and Hauser (1973); Berkman (1988); Marmot et al. (1991); and Feinstein (1993) and Smith (1999).
2 See Davies (1979); Avery, Elliehausen, and Kennickell (1988); Avery and Kennickell (1991); and Hurst, Luoh,
and Stafford (1998).
3 In some cases, missing data on assets and liabilities could be imputed. See Camphuis (1993) for more details on
the data imputation and Alessie, Lusardi and Aldershof (1997) for a description of the criteria used to calculate total
net worth.
4 We have not combined income and wealth data of CSS and SEP in overlapping years, in view of the clear
differences in measurement properties of income and wealth in the two datasets.
