Manajemen | Fakultas Ekonomi Universitas Maritim Raja Ali Haji 739.full
The Impact of Socioeconomic Status
on Health over the Life-Course
James P. Smith
abstract
Using data from the PSID, across the life course SES impacts future health
outcomes, although the primary influence is education and not an
individual’s financial resources in whatever form they are received. That
conclusion appears to be robust whether the financial resources are income
or wealth or whether the financial resources represent new information such
as the largely unanticipated wealth that was a consequence of the recent
stock market boom. Finally, this conclusion is robust across new health
outcomes that take place across the short and intermediate time frames of up
to 15 years in the future.
I. Introduction
It is well documented that individuals with lower socioeconomic status (SES) have much worse health outcomes (Marmot 1999; Smith 1999). But why
this is so continues to be debated (Adams et al. 2003; Deaton 2003). The central
question is whether these large differences in health by such SES indicators as income, wealth, or education largely reflect causation from SES to health or instead
the cumulative consequences of poor health outcomes throughout one’s life. Even
if SES mainly affects health, what dimensions of SES actually matter—financial
aspects such as income or wealth or nonfinancial dimensions like education?
This paper begins with the perspective that there is a distinct life-course component to the health-SES gradient so that we may be misled in trying to answer these
questions by only looking at people of a certain age—say those past age 50 as required by the widely used Health and Retirement data (HRS). To see this, Figure 1,
which is based on the PSID, displays the main contours of the SES-health gradient
by plotting at each age the fraction of people who report themselves in fair or poor
James P. Smith holds the RAND chair in Labor Market and Demographic Studies. He would like to
thank the helpful advice and assistance from Bob Schoeni. Programming assistance of Patty St. Clair
and David Rumpel is gratefully appreciated. This paper benefited from comments by participants at
seminars at Princeton University and the Population Council. This research was supported by grants
from NIA. The data used in this article can be obtained beginning May 2008 through April 2011 from
James P. Smith, 1776 Main Street, PO Box 2138, Santa Monica, CA 90407-2138. smith@rand.org.
[Submitted August 2005; accepted November 2007]
ISSN 022-166X E-ISSN 1548-8004 Ó 2007 by the Board of Regents of the University of Wisconsin System
T H E JO U R NAL O F H U M A N R E S O U R C E S
d
XLII
d
4
740
The Journal of Human Resources
Figure 1
Percent Reporting Fair or Poor Health Status by Age-Specific Income Quartiles
To construct this figure, individuals in each age-year cell between the years 1984
and 1999 were separated into four age-specific household income quartiles, and
the self-reported health status of each group was computed. The data were merged
across all years.
health by age-specific household income quartiles.1 At least until the end of life, for
the most part at each age every movement down in income is associated with being
in poorer health. Moreover, the health differences by income class are quite large.
The fraction in fair or poor health in the top income quartile is often 30 percentage
points smaller than the fraction in those health groups in the lowest income quartile.
Finally, there is a quite distinct age pattern to the SES-health gradient with health
disparities by income class expanding up to around age 50, after which the health
gradient begins slowly to fade away.
Figure 2 illustrates that a quite similar pattern emerges when our SES measure
shifts to an important SES market-household wealth—an expanding health gradient
into middle age followed by a gradual contraction.2 Given the dramatic age patterns
to the SES health gradient, the value of panels such as the PSID that span the entire
life course in addressing these difficult questions about the meaning of the SEShealth gradient would seem to be potentially quite high.
This paper is divided into four sections and an appendix. Section I is the introduction; Section II describes health and other relevant data available in the Panel Study
of Income Dynamics (PSID). In recent years, the PSID has added significant health
content to its core module. The appendix to the paper provides an evaluation of the
1. The data in Figure 1 are calculations by the author from the PSID. To construct this figure, individuals in
each age-year cell between the years 1984 and 1999 were separated into four age-specific household income quartiles and the self-reported health status of each group was computed. The data were merged
across all years. For similar graphs, see Case and Deaton (2005).
2. Figure 2 is calculated by the author from the PSID wealth modules. They are computed similarly to the income quartile graphs in Figure 1 except that only the years when the wealth modules were fielded were used.
Smith
Figure 2
Percent Reporting Excellent or Very Good Health Status by Age-Specific Wealth
Quartiles
Figure 2 is calculated similarly to the income quartile graph in Figure 1, except
that only the years when the wealth modules were fielded were used.
quality of this new mostly retrospectively collected health data for the type of research conducted in this paper. Section III examines the effect of SES on health
by looking at whether the onset of new chronic conditions is related to household
income, wealth, and education—once one conditions of a set of preexisting demographic and health conditions. This analysis attempts to take advantage of three
unique features of the PSID for this type of analysis—tracking health onsets well into
the future, assessing whether the predictive effects of the main SES measures vary
with age, and reconstructing an individual’s prior SES and health history with its
long-term panel. The final section of the paper contains the conclusions.
II. Data
The Panel Study of Income Dynamics (PSID) has gathered almost 30
years of extensive economic and demographic data on a nationally representative
sample of approximately 5,000 (original) families and 35,000 individuals who live
in those families. There are several potential advantages to using the PSID to study
issues surrounding the SES-health gradient. In contrast to the other panel surveys
such as the HRS and AHEAD that have been widely used to investigate the SEShealth gradient, the PSID spans all age groups, allowing one to examine behavior
over the complete life cycle. Figures 1 and 2 have already indicated that the nature
of the SES-health gradient may differ significantly across age groups.
The PSID is recognized as one of the premiere general-purpose panel surveys
measuring several key dimensions of SES. Details on family income and its components have been gathered in each wave since the inception of PSID in 1968. Starting
in 1984 and in five-year intervals until 1999, PSID asked a set of questions to
741
742
The Journal of Human Resources
measure household wealth. Starting in 1997, the PSID switched to a two-year periodicity, and wealth modules are now included as part of the core interview.
The PSID has been collecting information on self-reported general health status
(the standard five-point scale from excellent to poor) since 1984. Starting in 1999
and for all subsequent waves, PSID has collected information on the prevalence
and incidence of a list of chronic conditions—heart disease, stroke, heart attack, hypertension, cancer, diabetes, chronic lung disease, asthma, arthritis, and emotional,
nervous, or psychiatric problems. In addition to the prevalence in 1999, individuals
were asked the date of onset of the condition and whether it limited their normal
daily activities. Thus keeping in mind issues related to recall bias, the implications
of which I discuss in the appendix, the timing of a health onset can potentially be
identified and the impact of these new health events on labor supply, income, and
wealth can be estimated.3
The PSID offers several key potential additions to the research agenda. First, as
Figures 1 and 2 suggest, the nature of the SES-health gradient may vary considerably
over the life cycle. For example, during the age span covered by HRS, the incomehealth gradient was actually becoming considerably smaller. Labor supply effects induced by new health events may be very sensitive to life-cycle stage as for onsets that
take place in the late mid-50s or earlier-60s individuals can select an option that they
would have chosen in a few years anyway—retirement.
Second, the long-term nature of the PSID allows one to estimate the impact of health
and SES innovations over relatively long periods—at this point measured in decades.
It may well be that health responds to changes in financial measures of SES but only
after a considerable lag. Third, the PSID also permits a unique long-term perspective
in the other direction. Given that many panel members have been members of the PSID
since 1967, the entire past sequence of financial situation can be exploited in the
research. Current or last year’s income alone may be an inadequate proxy for a
household’s command over financial resources either due to standard classical measurement error issues or the familiar distinction between current and permanent income (Friedman 1957). While the British cohort samples are an important research
tool for studying the prospective evolution of both SES and health, no other American
survey can match the long history of financial SES information now available in the
PSID.
The two elements of the expanded PSID health content that potentially are analytically useful are the prospectively collected general health status available yearly
since 1984 and the 1999 retrospective health conditions information. In my view,
there is too much noise in year-to-year changes in general health status for it to
be used as an index of new health onset (Crossley and Kennedy 2002). But this past
information on general health status remains useful in that it captures at least part of
the past health histories of individuals, information that is necessary to separate out
the predictable and unpredictable components of new health onsets. Following
Smith (1999), onsets of chronic health conditions can be used to measure new health
onsets.
3. Most of these improvements in the health content of the PSID postdate the summary presented in Evans,
Level, and Simon (2000).
Smith
III. Predicting the Onset of Future Health Conditions
A. Conceptual Issues
In this section, I explore the pathway from SES to health by examining whether the
future onset of new chronic conditions is related to levels of household income,
wealth, and education, once one conditions of a set of preexisting set of demographic, economic, and health conditions. Before summarizing those results, it is
useful first to outline the essential issues in estimating effects of SES on health. Following the treatment in Smith (1999) and Adams et al. (2003), current realizations of
both economic status and health reflect a dynamic history in which both health (Ht)
and SES (Yt) are mutually affected by each other as well as by other relevant forces.
Most of the relevant ideas can be summarized by the following equation:
ð1Þ Ht ¼ a0 + a1 Ht21 + a2 Yt21 + a3 DYˆt + a4 Xt21 + u1t
where Xt-1 represents a vector of other possibly nonoverlapping time and nontime
varying factors influencing health and SES and u1t are stochastic shocks to health.
In this framework, we can estimate whether past values of SES predict health (a26¼
0).4 In the tradition of Granger causality, one view is that this provides a test of no
direct causal path (conditional on Ht21) between Yt21 and Ht.5 More precisely, this
reduces to a joint test that there is not a direct causal link between Yt21 and Ht and no
unobserved common factors that could create an ecological correlation between Yt21
and Ht. Since it is in general impossible to rule out the possibility of all such unobserved factors, some caution is warranted in going beyond the language of prediction
with these parameters on lagged values. a2 may be better viewed, as it will be here,
simply as the ability of past values of SES to predict future health onsets. The reason
is that there may well be unobserved factors correlated both with past SES and
health, even after we have conditioned on the measured components.
But even with this caution, the estimated values of a2 may be suggestive in the causality debate. Many unobserved factors would tend to produce a positive correlation between past SES and health biasing a2 upward. Health is a central part of human capital
and is positively correlated with other components of human capital that lead one to
have higher earning capacity. Even in a data source as rich as PSID, not all the components of health or other dimensions of human capital that produce higher SES can be
measured adequately. Thus, some of them surely appear as a positive correlation in past
SES and unobserved dimensions of health. In this case, the coefficients on past SES variables would be overstating the true effect of SES on health.
Of course, it is not possible to strictly sign the bias in a2 since there are other factors that would lead to a downward bias in a2. One obvious example would be classical measurement error in past SES, which would bias a2 toward zero. However, this
4. For an insightful debate about the conditions under whether coefficients are zero or stationary also
reveals something about causality, see the paper by Adams et al. (2003) and the comments on that paper
in the same volume.
5. These tests first test for parameter invariance across time to assess whether the model is ‘‘valid.’’ If invariance is accepted, the second step involves testing for noncausality using standard exclusion tests. See
Adams et al. (2003).
743
744
The Journal of Human Resources
type of bias can be mitigated by standard practices such as averaging past values of
SES or including long lag structures in SES, which is what is done in this paper.
Another key parameter a3 measures the effect of new innovations of SES (DYˆt) on
health. The term DYˆt represents that part of the between period change in SES (DYt)
that is an innovation. To estimate a3, we require exogenous variation in SES (DYˆt)
not induced by health. In particular, this implies that it is not appropriate to use
the full between-period changes SES (DYt) to estimate these effects since such variation hopelessly confounds feedback effects. For example, a new health problem
could reduce labor supply and hence income so that regressing the change in health
on the change in income would have it all backward.
One opportunity for estimating the effects of SES not caused by health lies in the large
wealth increases that were accumulated during the large stock market runup during the
late 1980s and 1990s. Given the unusually large runup in the stock market during these
decades, it is reasonable to posit that a good deal of this surge was unanticipated and thus
captures unanticipated exogenous wealth increases that were not caused by a person’s
health. If financial measures of SES do improve health, such increases in stock market
wealth should be associated with better subsequent health outcomes at least with a lag.
While this empirical approach is followed in this paper, it is certainly not without
its drawbacks. One legitimate criticism in my view is that—given the distribution of
stock market capital gains in the population—we essentially end up studying the average treatment effect of ‘‘exogenous’’ financial wealth changes on health for mostly
the very well-to-do. Not only may we be missing the policy relevant part of the population, but the average health change induced by these innovations in the financial
dimensions of SES are measured across a population whose health is very good to
begin with and whose health is probably less likely to change.
Another relevant issue looks deeper at the source of the variation in stock wealth
used in estimating the impact. There is a key conceptual difference between withinindividual variation in stock wealth over t, on which the surge in the stock market
argument relies, and point in time variation across i, or across individuals. Variation
across i may not yield unbiased estimates if those who are have greater current exposure to stock market gains by investing more in the past in the stock market
are healthier. Some of this possible contamination is dealt with by the extensive controls for past health that are included in the models estimated below, but there still
remains the possibility of unobserved components of health that are correlated with
the size of the stock market exposure people have.6 While it undoubtedly helps to be
estimating within a time period in which the variation across t is so large, one must
acknowledge that the variation on which these estimates rely combine variation
across i and t.
A third issue is that there may be good reason to distinguish between effects of positive wealth gains and negative wealth losses on health onsets since they may have
asymmetric effects on health. There may not be a plausible mechanism (for those already with health insurance) that explains why making a good deal of extra money
makes one healthier or happier, especially if one quickly adjusts to the new state. In
6. The possibility of a downward bias also exists. Individuals who are risk averse may not only invest more
in bonds relative to stocks but also may take better care of their health. Such individuals would experience
less of a stock market gain in a boom but experience better health outcomes.
Smith
contrast, unexpectedly large poor financial outcomes may lead to distress, worry, and
depression as one tries to make ends meet in the new financial reality. Many
epidemiological studies show that prolonged periods of psychological stress may
have negative consequences for one’s health (Steptoe and Marmot 2004). Fortunately,
the problem of asymmetric impacts is more easily dealt with empirically since we can
distinguish capital gains and capital losses in the empirical specification. I report
results from several tests of this distinction between capital gains and capital losses
below.
The fundamental difference between estimates of the relation between SES and
health that are obtained from estimates of models like Equation 1 compared to the
strong relation between health and SES that exists in Figures 1 and 2 do not simply
flow from these distinctions involved in isolating new information in the measures
of SES. A more critical issue in fact involves how one measures health. A good deal
of the evidence that claims that there is a strong relationship between all measures of
SES and health uses some stock measure of health as the outcome variable. These studies are largely cross-sectional and typically use measures such as current general
health status, disease prevalence rather than incidence, or mortality as the health outcome. Such stock measures of health hopelessly confound all past interactions between health and SES, no matter which direction of causation between the two was
operative or how important some third factor was in influencing both health and SES.
In this paper, I follow instead the suggestion first made in Smith (1999) of using
the onset of new health conditions as the health outcome. The advantage of using
new health onsets is that one is at least attempting to isolate new health events that
are distinguishable from past health, especially when one simultaneously controls for
comprehensive measures of past health status, past health behaviors, and SES. Even
with a data set as rich as the PSID, one still must always acknowledge that there may
be information about past health available to the individual but not to the statistical
analyst. An alternative view is that this may be pushing individual rationality a bit
too far if people do not fully incorporate the risks inherent in their health behaviors
in formulating their views about their future health trajectories.
It is possible to estimate models based on Equation 1 with health outcomes besides
new onsets as long as one still conditions on a comprehensive array of past values of
health and SES as implied by Equation 1. A good case in point would be mortality.
Moreover, there would be no inherent contradiction between finding no influence
of financial measures of SES on disease onset but in finding an influence on subsequent mortality. It may well turn out that economic resources have significant
impacts on mortality but not on disease onset suggesting that the main influence
of financial resources is not in preventing disease onset but in dealing more effectively with treatment of an illness after it occurs thereby mitigating its impact on subsequent mortality.7
7. The PSID mortality cannot currently be used for this purpose since it is by no means a complete listing
of the death of previous PSID panel members. Once one has attrited from the PSID, there was no attempt to
find the subsequent death status of attritors. For many people who were in the PSID but are not now in the
PSID, one doesn’t know their mortality status. A project is currently underway by the PSID to do a complete accounting of the death of prior participants but the results of that effort are a few years away.
745
746
The Journal of Human Resources
B. Empirical Estimates
The key implication of Equation 1 is that one should condition on the full array of baseline health and SES status to estimate the relation between SES and health. The models
I estimate include as covariates a vector of baseline health conditions of the respondent—self-reported general health status (excellent, very good, good, fair with poor
the left-out group) and the presence of chronic conditions (arthritis, cancer, diabetes,
heart disease, hypertension, lung disease, and stroke) all measured at baseline. The
models include a standard set of demographic controls—age (dummies for whether
one is 51–61 years old or age 62 and older), race (dummy for African-American), ethnicity (dummy for Hispanic), sex (dummy for female), marital status and marital transition between waves, employment, home ownership, and state of residence.8
The PSID only allows one to back-cast one health behavior so that one can condition on it at each baseline year in the analysis. Fortunately, it is the most important
health behavior—smoking—well known to be highly correlated with both education
and income. Three smoking-related variables are included in these models—whether
one ever smoked before baseline, whether one currently smoked at baseline, and the
number of cigarettes normally smoked.
The principal interest lies in evaluating the importance of various SES measures
that include household income, baseline levels of and exogenous changes in household wealth, and respondent’s education. SES is not a one-dimensional concept and
knowing which aspect of SES affects health is key to the policy debate surrounding
the SES-health gradient. If the only pathway that operates from SES to health is education, policies directed at income redistribution could not be justified in terms of a
beneficial impact on health. In addition to years of schooling, three financial measures are used—baseline levels of household income and household wealth, and the
increase in stock wealth observed over the period covered by the health onset.
Tables 1, 2, and 3 summarize my main results using the PSID to predict the future
onset of major and minor chronic conditions.9 Because we want to condition on prior
wave values of household wealth and PSID wealth modules were collected at fiveyear intervals starting in 1984, three time periods are used with alternative baseline
years—1984, 1989, and 1994. The occurrence of new health events is then measured
over five-year intervals so that for the 1984 and 1989 baseline analyses we can look
forward for three and two five-year intervals respectively.
Because the 1999 chronic condition questions were only asked of those present in
that wave, the analysis must be limited to such people. Thus, we cumulatively start
losing respondents the further back in time because additional respondents would not
yet be members of the PSID panel. In this analysis the sample size for changes 1994–
99 is 7,205; 1989–94 is 5,783 and 1984–89 is 4,820.
8. To preserve space and since they are not the primary variables of interest, the estimated coefficients of
the marital and region variables are not included in Tables 1, 2, and 3.
9. These models and those in the next section are restricted to survivors—those who neither attrited nor
died across the waves so this analysis ignores the relationship of SES with attrition and mortality. Given
the age range of PSID respondents, mortality selection but not attrition is unlikely to be that critical. That
is clearly not the case in the AHEAD sample. For a model that incorporates mortality selection, see Adams
et al. (2003).
Smith
Table 1
Probits Predicting Onsets of New Health Conditions—Baseline 1994
Major Onset 1995–99
Age 51–61
Age 62-plus
Excellent health
Very good health
Good health
Arthritis
Cancer
Diabetes
Heart disease
High blood pressure
Lung disease
Stroke
Family income
Wealth
Education
Ever smoke
Currently smoke
Number of cigarettes
Employment
Own house
Female
Black
Hispanic
Change in stock wealth
Constant
Minor Onset 1995–99
Coefficient
z
Coefficient
z
0.420
0.809
20.315
20.266
20.226
0.164
0.048
0.240
20.296
0.104
0.093
0.130
20.0083
0.0005
20.016
0.038
0.229
0.005
20.074
20.033
0.028
20.020
20.424
0.0004
21.011
(5.67)
(10.62)
(3.37)
(3.29)
(2.98)
(2.22)
(0.36)
(2.22)
(2.36)
(1.47)
(0.68)
(0.67)
(1.78)
(1.06)
(1.40)
(0.48)
(3.21)
(2.24)
(1.21)
(0.52)
(0.52)
(0.27)
(1.85)
(1.26)
(1.49)
0.355
0.331
20.424
20.303
20.084
20.487
20.341
20.192
0.238
20.258
0.179
20.255
20.0054
0.0007
20.017
0.013
20.100
0.007
20.021
0.087
0.062
0.153
0.141
20.0010
20.056
(5.83)
(4.82)
(5.52)
(4.42)
(1.31)
(6.50)
(2.44)
(1.81)
(2.42)
(3.98)
(1.52)
(1.34)
(1.36)
(1.58)
(1.91)
(0.21)
(1.74)
(3.39)
(0.43)
(1.81)
(1.53)
(2.90)
(1.19)
(1.54)
(0.10)
Note: Models also control for Region of Residence and marital status and change in marital status between
the waves. Z values are based on robust standard errors. Financial variables measured in $10,000 units.
In all specifications, not surprisingly onsets of minor and major disease are
strongly positively related to age and negatively related to baseline levels of selfassessed health status. Collectively, prior chronic conditions are also strong predictors of future new health onsets.10 Smoking, and especially its intensity as proxied
by the number of cigarettes, continues to produce new incremental bad health outcomes, even after conditioning on an extensive set of measures of current health
10. Negative coefficients on some prior chronic conditions are not an anomaly since having some of these
conditions in the past precludes new onset of that condition.
747
748
The Journal of Human Resources
Table 2
Probits Predicting Onsets of New Health Conditions—Baseline 1989
Major Onset 1990–94
Age 51–61
Age 62
Excellent health
Very good health
Good health
Arthritis
Cancer
Diabetes
Heart disease
High blood pressure
Lung disease
Stroke
Family income
Wealth
Education
Ever smoke
Currently smoke
Number of cigarettes
Employment
Own home
Female
Black
Hispanic
Change in stock wealth
Constant
Coefficient
z
Coefficient
z
0.448
0.439
20.330
20.220
20.042
0.122
0.176
0.247
20.150
0.127
0.395
20.216
0.0049
20.0007
20.029
20.029
0.167
0.010
20.189
0.041
20.046
20.378
20.288
0.0009
20.576
(5.15)
(4.13)
(2.71)
(2.01)
(0.41)
(1.07)
(0.87)
(1.61)
(0.82)
(1.26)
(1.98)
(0.62)
(0.85)
(0.80)
(2.25)
(0.28)
(1.93)
(3.80)
(2.49)
(0.53)
(0.69)
(4.05)
(1.11)
(0.47)
(0.92)
0.572
0.419
20.531
20.333
20.218
20.434
0.324
20.069
20.015
20.232
0.275
0.166
0.0101
20.0028
20.014
0.176
20.208
0.001
20.001
0.207
20.047
0.125
0.098
20.0006
20.316
(8.02)
(4.72)
(5.45)
(3.81)
(2.59)
(3.98)
(1.77)
(0.45)
(0.10)
(2.56)
(1.57)
(0.65)
(1.79)
(2.73)
(1.31)
(2.27)
(2.90)
(0.37)
(0.02)
(3.21)
(0.89)
(1.86)
(0.56)
(0.33)
(0.46)
Major Onset 1995–99
Age 51–61
Age 62
Excellent health
Very good health
Good health
Arthritis
Cancer
Minor Onset 1990–94
Minor Onset 1995–99
Coefficient
z
Coefficient
z
0.546
0.775
20.272
20.283
20.069
0.136
20.002
(7.29)
(8.77)
(2.69)
(3.02)
(0.79)
(1.43)
(0.01)
0.194
0.323
20.249
20.189
20.022
20.581
20.285
(2.95)
(4.03)
(2.91)
(2.33)
(0.28)
(5.64)
(1.42)
(continued )
Smith
Table 2 (continued)
Major Onset 1995–99
Diabetes
Heart disease
High blood pressure
Lung disease
Stroke
Family income
Wealth
Education
Ever smoke
Currently smoke
Number of cigarettes
Employment
Own house
Female
Black
Hispanic
Change in stock wealth
Constant
Minor Onset 1995–99
Coefficient
z
Coefficient
z
0.184
20.243
0.206
20.171
20.186
20.0029
0.0004
20.027
0.012
0.160
0.006
20.100
0.083
20.023
20.022
20.409
0.0005
20.376
(1.34)
(1.45)
(2.47)
(0.83)
(0.65)
(0.65)
(0.99)
(2.35)
(0.14)
(2.08)
(2.35)
(1.50)
(1.16)
(0.40)
(0.28)
(1.62)
(2.13)
(0.53)
20.145
0.035
20.199
0.282
20.189
0.0025
0.0007
20.027
20.077
20.031
0.008
0.059
0.141
0.054
0.187
0.269
20.0008
20.543
(1.04)
(0.26)
(2.53)
(1.78)
(0.74)
(0.62)
(0.19)
(2.86)
(1.10)
(0.48)
(3.76)
(1.08)
(2.66)
(1.17)
(3.25)
(2.00)
(1.01)
(0.81)
Note: Models also control for Region of Residence and marital status and change in marital status between
the waves. Z values are based on robust standard errors. Financial variables measured in $10,000 units.
status. Because smoking has a strong negative gradient with education, its inclusion
in these models reduces the predictive effects of education on new health onsets.
Whether one looks at the relatively short horizon of the next five years or a decade
or more ahead, all three financial measures of SES—income, household wealth, and
the stock market gains—are very poor predictors of future health outcomes whether
they are major or minor onsets. The estimated coefficients are rarely statistically significant, and they even vacillate in sign being more often positive than negative. To
illustrate, of the 36 coefficients estimated on the three financial variables (income,
wealth, and new stock wealth) in Tables 1–3 in only four cases is the estimated effect
statistically significant at the 5 percent level and in two of those four the effect is
actually positive.
These longer horizon PSID results on financial measures of SES are quite powerful in that they partly respond to the objection that one may have controlled for most
of the indirect effects of SES by conditioning on baseline attributes including income. In this case, the conditioning variables are sometimes measured more than
a decade before the disease onset. In sum, SES variables that directly measure or
proxy for financial resources of a family are either not related or at best only weakly
749
750
The Journal of Human Resources
Table 3
Probits Predicting Onsets of New Health Conditions—Baseline 1984
Major Onset 1985–89
Age 51–61
Age 62
Excellent health
Very good health
Good health
Arthritis
Cancer
Diabetes
Heart disease
High blood pressure
Lung disease
Stroke
Family income
Wealth
Education
Ever smoke
Currently smoke
Number of cigarettes
Employment
Own house
Female
Black
Hispanic
Change in stock wealth
Constant
Coefficient
z
Coefficient
z
0.480
0.561
20.564
20.378
20.198
0.402
0.067
0.318
20.257
0.192
20.156
(4.54)
(3.68)
(3.57)
(2.88)
(1.64)
(3.03)
(0.20)
(1.45)
(1.07)
(1.46)
(0.51)
0.0030
0.0005
20.000
0.206
0.042
0.009
0.092
0.056
20.080
20.185
(0.26)
(1.04)
(0.02)
(1.67)
(0.39)
(2.93)
(0.87)
(0.56)
(0.94)
(1.55)
0.0034
21.585
(1.18)
(3.55)
0.566
0.595
20.504
20.187
20.211
20.427
20.279
20.183
0.122
20.191
20.209
0.221
0.0050
0.0003
20.025
0.010
20.131
0.012
0.026
0.273
0.111
0.234
20.368
20.0017
20.515
(7.49)
(5.75)
(4.72)
(1.99)
(2.36)
(3.34)
(0.93)
(0.96)
(0.72)
(1.83)
(0.89)
(0.73)
(0.57)
(1.02)
(2.08)
(0.11)
(1.60)
(4.72)
(0.39)
(3.66)
(1.78)
(2.98)
(1.57)
(0.08)
(0.45)
Major Onset 1990–94
Age 51–61
Age 62
Excellent health
Very good health
Good health
Arthritis
Cancer
Minor Onset 1985–89
Minor Onset 1990–94
Coefficient
z
Coefficient.
0.519
0.345
20.555
20.263
20.087
0.087
0.245
(6.03)
(2.79)
(4.46)
(2.41)
(0.88)
(0.68)
(0.99)
0.404
0.248
20.450
20.227
20.186
20.549
0.231
z
(5.38)
(2.37)
(4.54)
(2.54)
(2.17)
(4.27)
(0.95)
(continued )
Smith
Table 3 (continued)
Major Onset 1990–94
Diabetes
Heart disease
High blood pressure
Lung disease
Stroke
Family income
Wealth
Education
Ever smoke
Currently smoke
Number of cigarettes
Employment
Own house
Female
Black
Hispanic
Change in stock wealth
Constant
Coefficient
z
Coefficient.
z
0.398
0.028
0.034
20.059
0.111
0.0046
0.0002
20.023
20.055
0.191
0.010
20.149
0.005
20.120
20.347
20.177
0.0002
20.051
(2.24)
(0.15)
(0.29)
(0.24)
(0.27)
(0.48)
(0.60)
(1.65)
(0.49)
(2.02)
(3.73)
(1.91)
(0.05)
(1.70)
(3.62)
(0.67)
(0.13)
(0.07)
0.070
0.202
20.273
0.331
0.151
0.0023
0.0003
20.023
0.170
20.150
0.002
20.018
0.231
20.028
0.181
20.039
20.0011
20.524
(0.38)
(1.20)
(2.66)
(1.70)
(0.42)
(0.32)
(1.12)
(1.98)
(2.02)
(1.95)
(0.98)
(0.28)
(3.42)
(0.49)
(2.49)
(0.20)
(1.05)
(0.82)
Major Onset 1995–99
Age 51–61
Age 62
Excellent health
Very good health
Good health
Arthritis
Cancer
Diabetes
Heart disease
High blood pressure
Lung disease
Stroke
Family income
Wealth
Education
Minor Onset 1990–94
Minor Onset 1995–99
Coefficient
z
Coefficient
z
0.558
0.829
20.424
20.334
20.124
20.034
20.310
0.277
20.178
0.202
20.282
20.162
20.0091
0.0003
20.032
(7.21)
(8.02)
(3.98)
(3.44)
(1.37)
(0.31)
(1.04)
(1.64)
(0.94)
(2.18)
(1.17)
(0.43)
(1.27)
(1.04)
(2.66)
0.170
0.143
20.206
20.158
20.001
20.581
20.280
20.259
0.062
20.254
0.289
20.538
0.0071
0.0004
20.026
(2.49)
(1.52)
(2.37)
(1.93)
(0.01)
(4.95)
(1.07)
(1.43)
(0.38)
(2.83)
(1.63)
(1.38)
(1.24)
(1.22)
(2.61)
(continued )
751
752
The Journal of Human Resources
Table 3 (continued)
Major Onset 1995–99
Ever smoke
Currently smoke
Number of cigarettes
Female
Black
Hispanic
Employment
Own house
Change in stock wealth
Constant
Minor Onset 1995–99
Coefficient
z
Coefficient
z
20.045
0.238
0.005
20.055
20.023
20.450
20.010
0.169
0.0006
20.197
(0.46)
(2.75)
(2.12)
(0.85)
(0.27)
(1.69)
(0.14)
(2.18)
(2.51)
(0.28)
20.149
20.010
0.008
0.031
0.206
0.308
20.099
0.051
0.0003
20.427
(1.92)
(0.14)
(3.68)
(0.63)
(3.26)
(2.21)
(1.79)
(0.89)
(0.90)
(0.58)
Note: Models also control for Region of Residence and marital status and change in marital status between
the waves. Z values are based on robust standard errors. Financial variables measured in $10,000 units. NA
means no coefficient estimated for that variable.
related to the future onset of disease over the time span of 15 years.11 These results
imply that whether interpreted as predictions or more boldly as causally, financial
measures of SES appear to have no discernable impact on future onsets of disease.
As mentioned above, one explanation for the absence of any significant effects of
the stock market variable on health is that there may be asymmetric effects during
stock market upturns and downturns. To test this possibility, I estimated identical
onsets models to the ones listed in Tables 1–3 for 1999–2001 and 1999–2003. The
first years during this time period were years when the stock market experienced a
sharp collapse, especially in tech stocks. The estimated effect of stock market gains
and losses vacillated in sign and in no case was an estimated z value on a stock market gain or loss variable higher than one. However, there is no reason to use only the
2001 or 2003 PSID to test this idea, a time horizon that may be too short to adequately capture any subsequent health effects. Even during periods characterized
by large and pervasive capital gains, there are individuals who selected their portfolios badly and lost considerable money. A parallel argument applies during downturns. I reestimated the models in Tables 1–3 separating out capital gains from
capital losses and tested for impact of losses and gains separately. Once again, there
were no statistically significant effects of either capital gains or capital losses on any
future health onsets.
All this does not imply that at a more general level SES cannot predict future
health events. Even within the contours of the models estimated in Tables 1–3, education is mostly a statistically significant predictor of major health events across
11. For supporting evidence on these statements using the HRS, see Smith (2003).
Smith
both a short- and long-term horizon. Even after one controls for an extensive array of
baseline health and financial economic conditions, those with less schooling appear
to be more likely to experience a major negative health onset. Of the 12 education
coefficients estimated in these three tables, all are negative and eight are statistically
significant at the 10 percent level of significance. This is true both over the short horizon
of the next five years as well as over the longest horizon represented in the data—11–15
years out. The effect of education appears to become more statistically significant the
longer the horizon allowed for the detection of the onset. Even after conditioning on all
prior health events, themselves related to education, years of schooling still has predictive effects on diminishing the likelihood of a new health onset.
A legitimate issue about the models listed in Tables 1–3 is their reliance on singleperiod Markovs in both income and health. For several reasons, of which measurement error in either income or health would be but one good example, one-period
Markovs are quite limited. Even without measurement error, current income may
not be the appropriate concept for predicting future disease trajectories. Future onset
of disease may be more influenced by longer-term measures of financial resources,
which may well be one reason why education matters so much. Similarly, measures
relying only on last period’s health ignore the duration dimension of any health problems that might exist, a dimension quite likely to be associated with the onset of new
disease. With this in mind, the models were reestimated using up to four-period income lags and four-period lags in self-reported general health status. The baseline
1984 model only would allow a Year 1 lag in general health status and a baseline
1994 would only permit a short window for a new onset to occur. Therefore, models
with these longer lag structures in income and health were estimated for the 1989
baseline using a ten-year future horizon for the onset of a disease. All other covariates in these models remain the same as those in Tables 1–3.
Because the principal issue is whether these lags matter, Table 4 summarizes the
results with a series of F-tests and the associated probabilities for differences from
zero for a one-period lag, the three additional lags beyond the last year, and the
Table 4
Does SES Predict Future Major Onset with Lag Values of Income and Health?
(F-tests and p values; 1989 Baseline—10 Year Forecast)
Major Onset
Income lag 1
Income lag 2, 3, 4
Income lag 1,2,3,4
Health lag 1
Health lag 2,3,4
Health lag 1,2,3,4
Wealth
Change in stock wealth
Education
0.32
2.99
3.01
5.60
25.32
48.70
0.01
3.92
4.46
(0.57)
(0.39)
(0.56)
(0.13)
(0.01)
(0.00)
(0.89)
(0.05)
(0.02)
Minor Onset
0.05
4.13
5.12
17.67
34.71
69.22
1.53
1.42
5.43
(0.83)
(0.25)
(0.27)
(0.01)
(0.00)
(0.00)
(0.22)
(0.23)
(0.02)
753
754
The Journal of Human Resources
full four periods. The results obtained with lag structures on income are easy to
summarize—in no case do any of the variants of the income lags matter. Since a
four-period year should serve as a reasonable approximation to long-run measures
of income, this suggests that the reason current income doesn’t matter does not lie
in the familiar permanent-transitory income distinction. Moreover, the prior summary that they also do not predict future onset for the other financial variables included in the model—wealth and the change in wealth—remains intact even with
the full set of income and health lags in place. Finally, education still predicts future
major and minor disease onset indicating that it is not simply serving as a proxy for
permanent income, which should be adequately captured by the lag structure in past
incomes.
The story for lagged health, however, is quite different. A one-period lag for health
is clearly rejected and the full four-period lag structure is needed to predict future
health onset. As mentioned above, this probably reflects both measurement error
in health, especially in a noisy self-reported general health status (Crossley and
Kennedy 2002), as well as the fact that these health lags pick up the impact of duration of illness. However, even that four-period structure does not alter the conclusion that even after conditioning on these measures of health over the last four years,
education still predicts future onset and financial measures of SES do not.
One advantage of the PSID is that it allows us to investigate whether predictive
effects of SES on health onset vary with age. Much current research on this topic
is based either on the original HRS (Smith 1999) or the AHEAD cohort (Adams
et al. 2003), which were originally restricted to those between the ages of 51–61
and 70-plus, respectively. To investigate this possibility, separate models were estimated over three age groups—those younger than 40, those older than 60, and those
between those ages. While certainly less common than for people in the HRS and
AHEAD age ranges, health episodes for PSID respondents who are younger than
50 years old are not negligible. For example, among those in their 40s in the 1999
wave, one in seven had previously experienced a major disease onset at some time
in their lives and 40 percent have a minor chronic health condition. In the five years
before 1999, 7 percent of these 40-year-olds experienced a major disease onset while
one in four reported a new minor onset.
Sample sizes are necessarily much smaller when the models are estimated separately into these three age groups. This reduction in sample size becomes especially
problematic when onsets are rare—that is the younger the sample (especially for the
major onsets) and the shorter the duration over which one allows an onset to occur.
Because of these considerations, I estimated onset models for these age group models over ten-year horizons for the1984 and 1989 baseline year’s specifications and
over the full 15-year horizon for the 1984 baseline model. With the exception of
age, which is the stratifying variable, these models include the same set of covariates
as in Tables 1–3.
The principal results are summarized for all SES variables in Table 5. The column
labeled ‘‘financials’’ presents p values for tests that the three financial variables—
family income, wealth, and changes in stock market wealth—are jointly statistically
significant. In these age-specific models, once again, financial SES measures did not
matter in any of the age groups. In all but one of the four cases in which the associated p value was 0.10 or smaller, this was due to an ÔincorrectÕ positive sign on the
Smith
Table 5
Does SES Predict Future Onsets by Age?
Financials*
Education Coefficient
Education Z Score
1984 Baseline Models
Age group < 40
Major 1985–94
Minor 1985–94
Major 1985–99
Minor 1985–99
Age group 40–60
Major 1985–94
Minor 1985–94
Major 1985–99
Minor 1985–99
Age group > 60
Major 1985–94
Minor 1985–94
Major 1985–99
Minor 1985–99
0.29
0.31
0.94
0.65
20.045
20.044
20.030
20.035
21.79
22.57
21.49
22.50
0.21
0.09
0.32
0.05
20.019
20.009
20.050
20.032
21.04
20.56
22.97
22.08
0.40
0.44
0.24
0.21
0.026
20.038
0.015
20.029
0.81
21.39
0.57
21.05
1989 Baseline Models
Age group < 40
Major 1990–99
Minor 1990–99
Age group 40–60
Major 1990–99
Minor 1990–99
Age group > 60
Major 1990–99
Minor 1990–99
0.10
0.43
20.029
20.042
21.32
22.69
0.88
0.82
20.050
20.017
22.89
21.20
0.20
0.04
20.026
20.011
21.37
20.57
Note: This financial column reports the p value associated with the joint statistical significance for the coefficients for the three financial variables—income, wealth, and changes in stock market wealth. Models include all variables included in models in Tables 1–3. Using the minor onset models, sample sizes for the
1984 baseline were 2,788 (40 and 60). The corresponding sample sizes
for the 1989 baseline models were 2,972 (40 and 60).
wealth variable.12 Apparently, the ability of financial variables to predict future
onsets is not sensitive to the stage of the adult life cycle one happens to be in.13
12. The sole exception is the 60-plus age group for the 1989 baseline where the stock market variable is
statistically negative.
13. In contrast, family income as a child appears to matter a great deal. See Case, Lubotsky, and Paxson
(2002), Case, Fertig, and Paxson (2005), and Currie and Hansen (1999).
755
756
The Journal of Human Resources
The remaining columns in Table 5 list estimated coefficients and associated z values for the education variable in each model. Data necessarily gets thinner with these
age stratifications (especially in the oldest age group) and conclusions must be
tempered. But for those in the two age groups ages less than 60 and younger, the conclusion that education remains a strong predictor of future onset of disease apparently remains intact. This is far less apparent for the 60–plus age groups where
the estimated effects of education are statistically significant. This is consistent with
the findings of Adams et al. (2003) who used the AHEAD sample (aged 70-plus) and
reported virtually no effects of any SES variables including education on subsequent
health outcomes.
It may also reflect an inherent limitation of the PSID for research in this age
group. The 1984 baseline estimates are especially problematic with this age stratification. The 60-plus population in 1984 would be aged 75-plus and living in 1999
when the questions on past disease onset were first asked. Very few people of that
age who suffered a severe disease onset would still be alive in 1999, and they would
be quite unrepresentative of the actual onsets that occurred. While somewhat less severe, as they would have to be only aged 70-plus in 1999, the same type of problem
exists with the 1989 baseline models. This raises a more general point that the earlier
years of the PSID for this type of analysis are probably best restricted to younger age
groups.
Additional research on why education matters so much should receive high priority. It is not simply due to education’s association with poor health behaviors. Smoking, perhaps the most important health behavior, is controlled for in this analysis, and
other work with the HRS using more comprehensive health measures including
smoking, drinking, and obesity show similar net impacts of education (Smith 2005).
The basic source of the large estimated impact of schooling on success in many
life domains, of which health is only one example, may be the most important unanswered question in the social sciences. One possibility is that the education experience itself has little to do with it, but is simply a marker for personal traits (reasoning
ability, rates of time preference, etc.) that lead people to acquire more education and
to be healthier. But education may not be that passive. It may help train people in
decision-making, problem-solving, perseverance, and adaptive skills, all of which
have direct applications to a healthier life. Education may have biological effects
on the brain, which result in improved cognitive function and problem-solving ability, some of which may impart benefits to choices made regarding one’s health. This
is similar to the argument that more active brain functioning when younger pushes
off the onset of dementia.
One approach to isolating such a causal path of schooling follows the lead of
Lleras-Muney (2005) who provided evidence that laws affecting the age of school
entry and the age at which a child could obtain a work permit significantly affected
years of schooling completed during the 1915–39 period.14 Lleras-Muney found statistically significant effects of the component of schooling induced by compulsory
schooling laws on subsequent mortality.
14. Lleras-Muney presents evidence that the compulsory schooling laws are good instruments in the sense
that the F-test of joint significance of the laws yields a large value (around 5).
Smith
One difficulty with relying on this approach with the PSID is that those affected by
such laws would be relatively older cohorts in the PSID, the very group identified
above as not being PSID’s strong suit. A more promising approach may lie in applications in other countries where the impact of such laws were larger and more recent.
For example, England raised the statutory minimum from 14 to 15 shortly after the
Second World War and again from 15 to 16 in 1974.
Cunha and Heckman (2006) instead emphasize the importance of cognitive and
noncognitive skills such as motivation, persistence and self-control in explaining
schooling and one’s eventual success in multiple domains of life. They see these
skills as produced mainly by the family and by repetitive cumulative personal decisions whereby
on Health over the Life-Course
James P. Smith
abstract
Using data from the PSID, across the life course SES impacts future health
outcomes, although the primary influence is education and not an
individual’s financial resources in whatever form they are received. That
conclusion appears to be robust whether the financial resources are income
or wealth or whether the financial resources represent new information such
as the largely unanticipated wealth that was a consequence of the recent
stock market boom. Finally, this conclusion is robust across new health
outcomes that take place across the short and intermediate time frames of up
to 15 years in the future.
I. Introduction
It is well documented that individuals with lower socioeconomic status (SES) have much worse health outcomes (Marmot 1999; Smith 1999). But why
this is so continues to be debated (Adams et al. 2003; Deaton 2003). The central
question is whether these large differences in health by such SES indicators as income, wealth, or education largely reflect causation from SES to health or instead
the cumulative consequences of poor health outcomes throughout one’s life. Even
if SES mainly affects health, what dimensions of SES actually matter—financial
aspects such as income or wealth or nonfinancial dimensions like education?
This paper begins with the perspective that there is a distinct life-course component to the health-SES gradient so that we may be misled in trying to answer these
questions by only looking at people of a certain age—say those past age 50 as required by the widely used Health and Retirement data (HRS). To see this, Figure 1,
which is based on the PSID, displays the main contours of the SES-health gradient
by plotting at each age the fraction of people who report themselves in fair or poor
James P. Smith holds the RAND chair in Labor Market and Demographic Studies. He would like to
thank the helpful advice and assistance from Bob Schoeni. Programming assistance of Patty St. Clair
and David Rumpel is gratefully appreciated. This paper benefited from comments by participants at
seminars at Princeton University and the Population Council. This research was supported by grants
from NIA. The data used in this article can be obtained beginning May 2008 through April 2011 from
James P. Smith, 1776 Main Street, PO Box 2138, Santa Monica, CA 90407-2138. smith@rand.org.
[Submitted August 2005; accepted November 2007]
ISSN 022-166X E-ISSN 1548-8004 Ó 2007 by the Board of Regents of the University of Wisconsin System
T H E JO U R NAL O F H U M A N R E S O U R C E S
d
XLII
d
4
740
The Journal of Human Resources
Figure 1
Percent Reporting Fair or Poor Health Status by Age-Specific Income Quartiles
To construct this figure, individuals in each age-year cell between the years 1984
and 1999 were separated into four age-specific household income quartiles, and
the self-reported health status of each group was computed. The data were merged
across all years.
health by age-specific household income quartiles.1 At least until the end of life, for
the most part at each age every movement down in income is associated with being
in poorer health. Moreover, the health differences by income class are quite large.
The fraction in fair or poor health in the top income quartile is often 30 percentage
points smaller than the fraction in those health groups in the lowest income quartile.
Finally, there is a quite distinct age pattern to the SES-health gradient with health
disparities by income class expanding up to around age 50, after which the health
gradient begins slowly to fade away.
Figure 2 illustrates that a quite similar pattern emerges when our SES measure
shifts to an important SES market-household wealth—an expanding health gradient
into middle age followed by a gradual contraction.2 Given the dramatic age patterns
to the SES health gradient, the value of panels such as the PSID that span the entire
life course in addressing these difficult questions about the meaning of the SEShealth gradient would seem to be potentially quite high.
This paper is divided into four sections and an appendix. Section I is the introduction; Section II describes health and other relevant data available in the Panel Study
of Income Dynamics (PSID). In recent years, the PSID has added significant health
content to its core module. The appendix to the paper provides an evaluation of the
1. The data in Figure 1 are calculations by the author from the PSID. To construct this figure, individuals in
each age-year cell between the years 1984 and 1999 were separated into four age-specific household income quartiles and the self-reported health status of each group was computed. The data were merged
across all years. For similar graphs, see Case and Deaton (2005).
2. Figure 2 is calculated by the author from the PSID wealth modules. They are computed similarly to the income quartile graphs in Figure 1 except that only the years when the wealth modules were fielded were used.
Smith
Figure 2
Percent Reporting Excellent or Very Good Health Status by Age-Specific Wealth
Quartiles
Figure 2 is calculated similarly to the income quartile graph in Figure 1, except
that only the years when the wealth modules were fielded were used.
quality of this new mostly retrospectively collected health data for the type of research conducted in this paper. Section III examines the effect of SES on health
by looking at whether the onset of new chronic conditions is related to household
income, wealth, and education—once one conditions of a set of preexisting demographic and health conditions. This analysis attempts to take advantage of three
unique features of the PSID for this type of analysis—tracking health onsets well into
the future, assessing whether the predictive effects of the main SES measures vary
with age, and reconstructing an individual’s prior SES and health history with its
long-term panel. The final section of the paper contains the conclusions.
II. Data
The Panel Study of Income Dynamics (PSID) has gathered almost 30
years of extensive economic and demographic data on a nationally representative
sample of approximately 5,000 (original) families and 35,000 individuals who live
in those families. There are several potential advantages to using the PSID to study
issues surrounding the SES-health gradient. In contrast to the other panel surveys
such as the HRS and AHEAD that have been widely used to investigate the SEShealth gradient, the PSID spans all age groups, allowing one to examine behavior
over the complete life cycle. Figures 1 and 2 have already indicated that the nature
of the SES-health gradient may differ significantly across age groups.
The PSID is recognized as one of the premiere general-purpose panel surveys
measuring several key dimensions of SES. Details on family income and its components have been gathered in each wave since the inception of PSID in 1968. Starting
in 1984 and in five-year intervals until 1999, PSID asked a set of questions to
741
742
The Journal of Human Resources
measure household wealth. Starting in 1997, the PSID switched to a two-year periodicity, and wealth modules are now included as part of the core interview.
The PSID has been collecting information on self-reported general health status
(the standard five-point scale from excellent to poor) since 1984. Starting in 1999
and for all subsequent waves, PSID has collected information on the prevalence
and incidence of a list of chronic conditions—heart disease, stroke, heart attack, hypertension, cancer, diabetes, chronic lung disease, asthma, arthritis, and emotional,
nervous, or psychiatric problems. In addition to the prevalence in 1999, individuals
were asked the date of onset of the condition and whether it limited their normal
daily activities. Thus keeping in mind issues related to recall bias, the implications
of which I discuss in the appendix, the timing of a health onset can potentially be
identified and the impact of these new health events on labor supply, income, and
wealth can be estimated.3
The PSID offers several key potential additions to the research agenda. First, as
Figures 1 and 2 suggest, the nature of the SES-health gradient may vary considerably
over the life cycle. For example, during the age span covered by HRS, the incomehealth gradient was actually becoming considerably smaller. Labor supply effects induced by new health events may be very sensitive to life-cycle stage as for onsets that
take place in the late mid-50s or earlier-60s individuals can select an option that they
would have chosen in a few years anyway—retirement.
Second, the long-term nature of the PSID allows one to estimate the impact of health
and SES innovations over relatively long periods—at this point measured in decades.
It may well be that health responds to changes in financial measures of SES but only
after a considerable lag. Third, the PSID also permits a unique long-term perspective
in the other direction. Given that many panel members have been members of the PSID
since 1967, the entire past sequence of financial situation can be exploited in the
research. Current or last year’s income alone may be an inadequate proxy for a
household’s command over financial resources either due to standard classical measurement error issues or the familiar distinction between current and permanent income (Friedman 1957). While the British cohort samples are an important research
tool for studying the prospective evolution of both SES and health, no other American
survey can match the long history of financial SES information now available in the
PSID.
The two elements of the expanded PSID health content that potentially are analytically useful are the prospectively collected general health status available yearly
since 1984 and the 1999 retrospective health conditions information. In my view,
there is too much noise in year-to-year changes in general health status for it to
be used as an index of new health onset (Crossley and Kennedy 2002). But this past
information on general health status remains useful in that it captures at least part of
the past health histories of individuals, information that is necessary to separate out
the predictable and unpredictable components of new health onsets. Following
Smith (1999), onsets of chronic health conditions can be used to measure new health
onsets.
3. Most of these improvements in the health content of the PSID postdate the summary presented in Evans,
Level, and Simon (2000).
Smith
III. Predicting the Onset of Future Health Conditions
A. Conceptual Issues
In this section, I explore the pathway from SES to health by examining whether the
future onset of new chronic conditions is related to levels of household income,
wealth, and education, once one conditions of a set of preexisting set of demographic, economic, and health conditions. Before summarizing those results, it is
useful first to outline the essential issues in estimating effects of SES on health. Following the treatment in Smith (1999) and Adams et al. (2003), current realizations of
both economic status and health reflect a dynamic history in which both health (Ht)
and SES (Yt) are mutually affected by each other as well as by other relevant forces.
Most of the relevant ideas can be summarized by the following equation:
ð1Þ Ht ¼ a0 + a1 Ht21 + a2 Yt21 + a3 DYˆt + a4 Xt21 + u1t
where Xt-1 represents a vector of other possibly nonoverlapping time and nontime
varying factors influencing health and SES and u1t are stochastic shocks to health.
In this framework, we can estimate whether past values of SES predict health (a26¼
0).4 In the tradition of Granger causality, one view is that this provides a test of no
direct causal path (conditional on Ht21) between Yt21 and Ht.5 More precisely, this
reduces to a joint test that there is not a direct causal link between Yt21 and Ht and no
unobserved common factors that could create an ecological correlation between Yt21
and Ht. Since it is in general impossible to rule out the possibility of all such unobserved factors, some caution is warranted in going beyond the language of prediction
with these parameters on lagged values. a2 may be better viewed, as it will be here,
simply as the ability of past values of SES to predict future health onsets. The reason
is that there may well be unobserved factors correlated both with past SES and
health, even after we have conditioned on the measured components.
But even with this caution, the estimated values of a2 may be suggestive in the causality debate. Many unobserved factors would tend to produce a positive correlation between past SES and health biasing a2 upward. Health is a central part of human capital
and is positively correlated with other components of human capital that lead one to
have higher earning capacity. Even in a data source as rich as PSID, not all the components of health or other dimensions of human capital that produce higher SES can be
measured adequately. Thus, some of them surely appear as a positive correlation in past
SES and unobserved dimensions of health. In this case, the coefficients on past SES variables would be overstating the true effect of SES on health.
Of course, it is not possible to strictly sign the bias in a2 since there are other factors that would lead to a downward bias in a2. One obvious example would be classical measurement error in past SES, which would bias a2 toward zero. However, this
4. For an insightful debate about the conditions under whether coefficients are zero or stationary also
reveals something about causality, see the paper by Adams et al. (2003) and the comments on that paper
in the same volume.
5. These tests first test for parameter invariance across time to assess whether the model is ‘‘valid.’’ If invariance is accepted, the second step involves testing for noncausality using standard exclusion tests. See
Adams et al. (2003).
743
744
The Journal of Human Resources
type of bias can be mitigated by standard practices such as averaging past values of
SES or including long lag structures in SES, which is what is done in this paper.
Another key parameter a3 measures the effect of new innovations of SES (DYˆt) on
health. The term DYˆt represents that part of the between period change in SES (DYt)
that is an innovation. To estimate a3, we require exogenous variation in SES (DYˆt)
not induced by health. In particular, this implies that it is not appropriate to use
the full between-period changes SES (DYt) to estimate these effects since such variation hopelessly confounds feedback effects. For example, a new health problem
could reduce labor supply and hence income so that regressing the change in health
on the change in income would have it all backward.
One opportunity for estimating the effects of SES not caused by health lies in the large
wealth increases that were accumulated during the large stock market runup during the
late 1980s and 1990s. Given the unusually large runup in the stock market during these
decades, it is reasonable to posit that a good deal of this surge was unanticipated and thus
captures unanticipated exogenous wealth increases that were not caused by a person’s
health. If financial measures of SES do improve health, such increases in stock market
wealth should be associated with better subsequent health outcomes at least with a lag.
While this empirical approach is followed in this paper, it is certainly not without
its drawbacks. One legitimate criticism in my view is that—given the distribution of
stock market capital gains in the population—we essentially end up studying the average treatment effect of ‘‘exogenous’’ financial wealth changes on health for mostly
the very well-to-do. Not only may we be missing the policy relevant part of the population, but the average health change induced by these innovations in the financial
dimensions of SES are measured across a population whose health is very good to
begin with and whose health is probably less likely to change.
Another relevant issue looks deeper at the source of the variation in stock wealth
used in estimating the impact. There is a key conceptual difference between withinindividual variation in stock wealth over t, on which the surge in the stock market
argument relies, and point in time variation across i, or across individuals. Variation
across i may not yield unbiased estimates if those who are have greater current exposure to stock market gains by investing more in the past in the stock market
are healthier. Some of this possible contamination is dealt with by the extensive controls for past health that are included in the models estimated below, but there still
remains the possibility of unobserved components of health that are correlated with
the size of the stock market exposure people have.6 While it undoubtedly helps to be
estimating within a time period in which the variation across t is so large, one must
acknowledge that the variation on which these estimates rely combine variation
across i and t.
A third issue is that there may be good reason to distinguish between effects of positive wealth gains and negative wealth losses on health onsets since they may have
asymmetric effects on health. There may not be a plausible mechanism (for those already with health insurance) that explains why making a good deal of extra money
makes one healthier or happier, especially if one quickly adjusts to the new state. In
6. The possibility of a downward bias also exists. Individuals who are risk averse may not only invest more
in bonds relative to stocks but also may take better care of their health. Such individuals would experience
less of a stock market gain in a boom but experience better health outcomes.
Smith
contrast, unexpectedly large poor financial outcomes may lead to distress, worry, and
depression as one tries to make ends meet in the new financial reality. Many
epidemiological studies show that prolonged periods of psychological stress may
have negative consequences for one’s health (Steptoe and Marmot 2004). Fortunately,
the problem of asymmetric impacts is more easily dealt with empirically since we can
distinguish capital gains and capital losses in the empirical specification. I report
results from several tests of this distinction between capital gains and capital losses
below.
The fundamental difference between estimates of the relation between SES and
health that are obtained from estimates of models like Equation 1 compared to the
strong relation between health and SES that exists in Figures 1 and 2 do not simply
flow from these distinctions involved in isolating new information in the measures
of SES. A more critical issue in fact involves how one measures health. A good deal
of the evidence that claims that there is a strong relationship between all measures of
SES and health uses some stock measure of health as the outcome variable. These studies are largely cross-sectional and typically use measures such as current general
health status, disease prevalence rather than incidence, or mortality as the health outcome. Such stock measures of health hopelessly confound all past interactions between health and SES, no matter which direction of causation between the two was
operative or how important some third factor was in influencing both health and SES.
In this paper, I follow instead the suggestion first made in Smith (1999) of using
the onset of new health conditions as the health outcome. The advantage of using
new health onsets is that one is at least attempting to isolate new health events that
are distinguishable from past health, especially when one simultaneously controls for
comprehensive measures of past health status, past health behaviors, and SES. Even
with a data set as rich as the PSID, one still must always acknowledge that there may
be information about past health available to the individual but not to the statistical
analyst. An alternative view is that this may be pushing individual rationality a bit
too far if people do not fully incorporate the risks inherent in their health behaviors
in formulating their views about their future health trajectories.
It is possible to estimate models based on Equation 1 with health outcomes besides
new onsets as long as one still conditions on a comprehensive array of past values of
health and SES as implied by Equation 1. A good case in point would be mortality.
Moreover, there would be no inherent contradiction between finding no influence
of financial measures of SES on disease onset but in finding an influence on subsequent mortality. It may well turn out that economic resources have significant
impacts on mortality but not on disease onset suggesting that the main influence
of financial resources is not in preventing disease onset but in dealing more effectively with treatment of an illness after it occurs thereby mitigating its impact on subsequent mortality.7
7. The PSID mortality cannot currently be used for this purpose since it is by no means a complete listing
of the death of previous PSID panel members. Once one has attrited from the PSID, there was no attempt to
find the subsequent death status of attritors. For many people who were in the PSID but are not now in the
PSID, one doesn’t know their mortality status. A project is currently underway by the PSID to do a complete accounting of the death of prior participants but the results of that effort are a few years away.
745
746
The Journal of Human Resources
B. Empirical Estimates
The key implication of Equation 1 is that one should condition on the full array of baseline health and SES status to estimate the relation between SES and health. The models
I estimate include as covariates a vector of baseline health conditions of the respondent—self-reported general health status (excellent, very good, good, fair with poor
the left-out group) and the presence of chronic conditions (arthritis, cancer, diabetes,
heart disease, hypertension, lung disease, and stroke) all measured at baseline. The
models include a standard set of demographic controls—age (dummies for whether
one is 51–61 years old or age 62 and older), race (dummy for African-American), ethnicity (dummy for Hispanic), sex (dummy for female), marital status and marital transition between waves, employment, home ownership, and state of residence.8
The PSID only allows one to back-cast one health behavior so that one can condition on it at each baseline year in the analysis. Fortunately, it is the most important
health behavior—smoking—well known to be highly correlated with both education
and income. Three smoking-related variables are included in these models—whether
one ever smoked before baseline, whether one currently smoked at baseline, and the
number of cigarettes normally smoked.
The principal interest lies in evaluating the importance of various SES measures
that include household income, baseline levels of and exogenous changes in household wealth, and respondent’s education. SES is not a one-dimensional concept and
knowing which aspect of SES affects health is key to the policy debate surrounding
the SES-health gradient. If the only pathway that operates from SES to health is education, policies directed at income redistribution could not be justified in terms of a
beneficial impact on health. In addition to years of schooling, three financial measures are used—baseline levels of household income and household wealth, and the
increase in stock wealth observed over the period covered by the health onset.
Tables 1, 2, and 3 summarize my main results using the PSID to predict the future
onset of major and minor chronic conditions.9 Because we want to condition on prior
wave values of household wealth and PSID wealth modules were collected at fiveyear intervals starting in 1984, three time periods are used with alternative baseline
years—1984, 1989, and 1994. The occurrence of new health events is then measured
over five-year intervals so that for the 1984 and 1989 baseline analyses we can look
forward for three and two five-year intervals respectively.
Because the 1999 chronic condition questions were only asked of those present in
that wave, the analysis must be limited to such people. Thus, we cumulatively start
losing respondents the further back in time because additional respondents would not
yet be members of the PSID panel. In this analysis the sample size for changes 1994–
99 is 7,205; 1989–94 is 5,783 and 1984–89 is 4,820.
8. To preserve space and since they are not the primary variables of interest, the estimated coefficients of
the marital and region variables are not included in Tables 1, 2, and 3.
9. These models and those in the next section are restricted to survivors—those who neither attrited nor
died across the waves so this analysis ignores the relationship of SES with attrition and mortality. Given
the age range of PSID respondents, mortality selection but not attrition is unlikely to be that critical. That
is clearly not the case in the AHEAD sample. For a model that incorporates mortality selection, see Adams
et al. (2003).
Smith
Table 1
Probits Predicting Onsets of New Health Conditions—Baseline 1994
Major Onset 1995–99
Age 51–61
Age 62-plus
Excellent health
Very good health
Good health
Arthritis
Cancer
Diabetes
Heart disease
High blood pressure
Lung disease
Stroke
Family income
Wealth
Education
Ever smoke
Currently smoke
Number of cigarettes
Employment
Own house
Female
Black
Hispanic
Change in stock wealth
Constant
Minor Onset 1995–99
Coefficient
z
Coefficient
z
0.420
0.809
20.315
20.266
20.226
0.164
0.048
0.240
20.296
0.104
0.093
0.130
20.0083
0.0005
20.016
0.038
0.229
0.005
20.074
20.033
0.028
20.020
20.424
0.0004
21.011
(5.67)
(10.62)
(3.37)
(3.29)
(2.98)
(2.22)
(0.36)
(2.22)
(2.36)
(1.47)
(0.68)
(0.67)
(1.78)
(1.06)
(1.40)
(0.48)
(3.21)
(2.24)
(1.21)
(0.52)
(0.52)
(0.27)
(1.85)
(1.26)
(1.49)
0.355
0.331
20.424
20.303
20.084
20.487
20.341
20.192
0.238
20.258
0.179
20.255
20.0054
0.0007
20.017
0.013
20.100
0.007
20.021
0.087
0.062
0.153
0.141
20.0010
20.056
(5.83)
(4.82)
(5.52)
(4.42)
(1.31)
(6.50)
(2.44)
(1.81)
(2.42)
(3.98)
(1.52)
(1.34)
(1.36)
(1.58)
(1.91)
(0.21)
(1.74)
(3.39)
(0.43)
(1.81)
(1.53)
(2.90)
(1.19)
(1.54)
(0.10)
Note: Models also control for Region of Residence and marital status and change in marital status between
the waves. Z values are based on robust standard errors. Financial variables measured in $10,000 units.
In all specifications, not surprisingly onsets of minor and major disease are
strongly positively related to age and negatively related to baseline levels of selfassessed health status. Collectively, prior chronic conditions are also strong predictors of future new health onsets.10 Smoking, and especially its intensity as proxied
by the number of cigarettes, continues to produce new incremental bad health outcomes, even after conditioning on an extensive set of measures of current health
10. Negative coefficients on some prior chronic conditions are not an anomaly since having some of these
conditions in the past precludes new onset of that condition.
747
748
The Journal of Human Resources
Table 2
Probits Predicting Onsets of New Health Conditions—Baseline 1989
Major Onset 1990–94
Age 51–61
Age 62
Excellent health
Very good health
Good health
Arthritis
Cancer
Diabetes
Heart disease
High blood pressure
Lung disease
Stroke
Family income
Wealth
Education
Ever smoke
Currently smoke
Number of cigarettes
Employment
Own home
Female
Black
Hispanic
Change in stock wealth
Constant
Coefficient
z
Coefficient
z
0.448
0.439
20.330
20.220
20.042
0.122
0.176
0.247
20.150
0.127
0.395
20.216
0.0049
20.0007
20.029
20.029
0.167
0.010
20.189
0.041
20.046
20.378
20.288
0.0009
20.576
(5.15)
(4.13)
(2.71)
(2.01)
(0.41)
(1.07)
(0.87)
(1.61)
(0.82)
(1.26)
(1.98)
(0.62)
(0.85)
(0.80)
(2.25)
(0.28)
(1.93)
(3.80)
(2.49)
(0.53)
(0.69)
(4.05)
(1.11)
(0.47)
(0.92)
0.572
0.419
20.531
20.333
20.218
20.434
0.324
20.069
20.015
20.232
0.275
0.166
0.0101
20.0028
20.014
0.176
20.208
0.001
20.001
0.207
20.047
0.125
0.098
20.0006
20.316
(8.02)
(4.72)
(5.45)
(3.81)
(2.59)
(3.98)
(1.77)
(0.45)
(0.10)
(2.56)
(1.57)
(0.65)
(1.79)
(2.73)
(1.31)
(2.27)
(2.90)
(0.37)
(0.02)
(3.21)
(0.89)
(1.86)
(0.56)
(0.33)
(0.46)
Major Onset 1995–99
Age 51–61
Age 62
Excellent health
Very good health
Good health
Arthritis
Cancer
Minor Onset 1990–94
Minor Onset 1995–99
Coefficient
z
Coefficient
z
0.546
0.775
20.272
20.283
20.069
0.136
20.002
(7.29)
(8.77)
(2.69)
(3.02)
(0.79)
(1.43)
(0.01)
0.194
0.323
20.249
20.189
20.022
20.581
20.285
(2.95)
(4.03)
(2.91)
(2.33)
(0.28)
(5.64)
(1.42)
(continued )
Smith
Table 2 (continued)
Major Onset 1995–99
Diabetes
Heart disease
High blood pressure
Lung disease
Stroke
Family income
Wealth
Education
Ever smoke
Currently smoke
Number of cigarettes
Employment
Own house
Female
Black
Hispanic
Change in stock wealth
Constant
Minor Onset 1995–99
Coefficient
z
Coefficient
z
0.184
20.243
0.206
20.171
20.186
20.0029
0.0004
20.027
0.012
0.160
0.006
20.100
0.083
20.023
20.022
20.409
0.0005
20.376
(1.34)
(1.45)
(2.47)
(0.83)
(0.65)
(0.65)
(0.99)
(2.35)
(0.14)
(2.08)
(2.35)
(1.50)
(1.16)
(0.40)
(0.28)
(1.62)
(2.13)
(0.53)
20.145
0.035
20.199
0.282
20.189
0.0025
0.0007
20.027
20.077
20.031
0.008
0.059
0.141
0.054
0.187
0.269
20.0008
20.543
(1.04)
(0.26)
(2.53)
(1.78)
(0.74)
(0.62)
(0.19)
(2.86)
(1.10)
(0.48)
(3.76)
(1.08)
(2.66)
(1.17)
(3.25)
(2.00)
(1.01)
(0.81)
Note: Models also control for Region of Residence and marital status and change in marital status between
the waves. Z values are based on robust standard errors. Financial variables measured in $10,000 units.
status. Because smoking has a strong negative gradient with education, its inclusion
in these models reduces the predictive effects of education on new health onsets.
Whether one looks at the relatively short horizon of the next five years or a decade
or more ahead, all three financial measures of SES—income, household wealth, and
the stock market gains—are very poor predictors of future health outcomes whether
they are major or minor onsets. The estimated coefficients are rarely statistically significant, and they even vacillate in sign being more often positive than negative. To
illustrate, of the 36 coefficients estimated on the three financial variables (income,
wealth, and new stock wealth) in Tables 1–3 in only four cases is the estimated effect
statistically significant at the 5 percent level and in two of those four the effect is
actually positive.
These longer horizon PSID results on financial measures of SES are quite powerful in that they partly respond to the objection that one may have controlled for most
of the indirect effects of SES by conditioning on baseline attributes including income. In this case, the conditioning variables are sometimes measured more than
a decade before the disease onset. In sum, SES variables that directly measure or
proxy for financial resources of a family are either not related or at best only weakly
749
750
The Journal of Human Resources
Table 3
Probits Predicting Onsets of New Health Conditions—Baseline 1984
Major Onset 1985–89
Age 51–61
Age 62
Excellent health
Very good health
Good health
Arthritis
Cancer
Diabetes
Heart disease
High blood pressure
Lung disease
Stroke
Family income
Wealth
Education
Ever smoke
Currently smoke
Number of cigarettes
Employment
Own house
Female
Black
Hispanic
Change in stock wealth
Constant
Coefficient
z
Coefficient
z
0.480
0.561
20.564
20.378
20.198
0.402
0.067
0.318
20.257
0.192
20.156
(4.54)
(3.68)
(3.57)
(2.88)
(1.64)
(3.03)
(0.20)
(1.45)
(1.07)
(1.46)
(0.51)
0.0030
0.0005
20.000
0.206
0.042
0.009
0.092
0.056
20.080
20.185
(0.26)
(1.04)
(0.02)
(1.67)
(0.39)
(2.93)
(0.87)
(0.56)
(0.94)
(1.55)
0.0034
21.585
(1.18)
(3.55)
0.566
0.595
20.504
20.187
20.211
20.427
20.279
20.183
0.122
20.191
20.209
0.221
0.0050
0.0003
20.025
0.010
20.131
0.012
0.026
0.273
0.111
0.234
20.368
20.0017
20.515
(7.49)
(5.75)
(4.72)
(1.99)
(2.36)
(3.34)
(0.93)
(0.96)
(0.72)
(1.83)
(0.89)
(0.73)
(0.57)
(1.02)
(2.08)
(0.11)
(1.60)
(4.72)
(0.39)
(3.66)
(1.78)
(2.98)
(1.57)
(0.08)
(0.45)
Major Onset 1990–94
Age 51–61
Age 62
Excellent health
Very good health
Good health
Arthritis
Cancer
Minor Onset 1985–89
Minor Onset 1990–94
Coefficient
z
Coefficient.
0.519
0.345
20.555
20.263
20.087
0.087
0.245
(6.03)
(2.79)
(4.46)
(2.41)
(0.88)
(0.68)
(0.99)
0.404
0.248
20.450
20.227
20.186
20.549
0.231
z
(5.38)
(2.37)
(4.54)
(2.54)
(2.17)
(4.27)
(0.95)
(continued )
Smith
Table 3 (continued)
Major Onset 1990–94
Diabetes
Heart disease
High blood pressure
Lung disease
Stroke
Family income
Wealth
Education
Ever smoke
Currently smoke
Number of cigarettes
Employment
Own house
Female
Black
Hispanic
Change in stock wealth
Constant
Coefficient
z
Coefficient.
z
0.398
0.028
0.034
20.059
0.111
0.0046
0.0002
20.023
20.055
0.191
0.010
20.149
0.005
20.120
20.347
20.177
0.0002
20.051
(2.24)
(0.15)
(0.29)
(0.24)
(0.27)
(0.48)
(0.60)
(1.65)
(0.49)
(2.02)
(3.73)
(1.91)
(0.05)
(1.70)
(3.62)
(0.67)
(0.13)
(0.07)
0.070
0.202
20.273
0.331
0.151
0.0023
0.0003
20.023
0.170
20.150
0.002
20.018
0.231
20.028
0.181
20.039
20.0011
20.524
(0.38)
(1.20)
(2.66)
(1.70)
(0.42)
(0.32)
(1.12)
(1.98)
(2.02)
(1.95)
(0.98)
(0.28)
(3.42)
(0.49)
(2.49)
(0.20)
(1.05)
(0.82)
Major Onset 1995–99
Age 51–61
Age 62
Excellent health
Very good health
Good health
Arthritis
Cancer
Diabetes
Heart disease
High blood pressure
Lung disease
Stroke
Family income
Wealth
Education
Minor Onset 1990–94
Minor Onset 1995–99
Coefficient
z
Coefficient
z
0.558
0.829
20.424
20.334
20.124
20.034
20.310
0.277
20.178
0.202
20.282
20.162
20.0091
0.0003
20.032
(7.21)
(8.02)
(3.98)
(3.44)
(1.37)
(0.31)
(1.04)
(1.64)
(0.94)
(2.18)
(1.17)
(0.43)
(1.27)
(1.04)
(2.66)
0.170
0.143
20.206
20.158
20.001
20.581
20.280
20.259
0.062
20.254
0.289
20.538
0.0071
0.0004
20.026
(2.49)
(1.52)
(2.37)
(1.93)
(0.01)
(4.95)
(1.07)
(1.43)
(0.38)
(2.83)
(1.63)
(1.38)
(1.24)
(1.22)
(2.61)
(continued )
751
752
The Journal of Human Resources
Table 3 (continued)
Major Onset 1995–99
Ever smoke
Currently smoke
Number of cigarettes
Female
Black
Hispanic
Employment
Own house
Change in stock wealth
Constant
Minor Onset 1995–99
Coefficient
z
Coefficient
z
20.045
0.238
0.005
20.055
20.023
20.450
20.010
0.169
0.0006
20.197
(0.46)
(2.75)
(2.12)
(0.85)
(0.27)
(1.69)
(0.14)
(2.18)
(2.51)
(0.28)
20.149
20.010
0.008
0.031
0.206
0.308
20.099
0.051
0.0003
20.427
(1.92)
(0.14)
(3.68)
(0.63)
(3.26)
(2.21)
(1.79)
(0.89)
(0.90)
(0.58)
Note: Models also control for Region of Residence and marital status and change in marital status between
the waves. Z values are based on robust standard errors. Financial variables measured in $10,000 units. NA
means no coefficient estimated for that variable.
related to the future onset of disease over the time span of 15 years.11 These results
imply that whether interpreted as predictions or more boldly as causally, financial
measures of SES appear to have no discernable impact on future onsets of disease.
As mentioned above, one explanation for the absence of any significant effects of
the stock market variable on health is that there may be asymmetric effects during
stock market upturns and downturns. To test this possibility, I estimated identical
onsets models to the ones listed in Tables 1–3 for 1999–2001 and 1999–2003. The
first years during this time period were years when the stock market experienced a
sharp collapse, especially in tech stocks. The estimated effect of stock market gains
and losses vacillated in sign and in no case was an estimated z value on a stock market gain or loss variable higher than one. However, there is no reason to use only the
2001 or 2003 PSID to test this idea, a time horizon that may be too short to adequately capture any subsequent health effects. Even during periods characterized
by large and pervasive capital gains, there are individuals who selected their portfolios badly and lost considerable money. A parallel argument applies during downturns. I reestimated the models in Tables 1–3 separating out capital gains from
capital losses and tested for impact of losses and gains separately. Once again, there
were no statistically significant effects of either capital gains or capital losses on any
future health onsets.
All this does not imply that at a more general level SES cannot predict future
health events. Even within the contours of the models estimated in Tables 1–3, education is mostly a statistically significant predictor of major health events across
11. For supporting evidence on these statements using the HRS, see Smith (2003).
Smith
both a short- and long-term horizon. Even after one controls for an extensive array of
baseline health and financial economic conditions, those with less schooling appear
to be more likely to experience a major negative health onset. Of the 12 education
coefficients estimated in these three tables, all are negative and eight are statistically
significant at the 10 percent level of significance. This is true both over the short horizon
of the next five years as well as over the longest horizon represented in the data—11–15
years out. The effect of education appears to become more statistically significant the
longer the horizon allowed for the detection of the onset. Even after conditioning on all
prior health events, themselves related to education, years of schooling still has predictive effects on diminishing the likelihood of a new health onset.
A legitimate issue about the models listed in Tables 1–3 is their reliance on singleperiod Markovs in both income and health. For several reasons, of which measurement error in either income or health would be but one good example, one-period
Markovs are quite limited. Even without measurement error, current income may
not be the appropriate concept for predicting future disease trajectories. Future onset
of disease may be more influenced by longer-term measures of financial resources,
which may well be one reason why education matters so much. Similarly, measures
relying only on last period’s health ignore the duration dimension of any health problems that might exist, a dimension quite likely to be associated with the onset of new
disease. With this in mind, the models were reestimated using up to four-period income lags and four-period lags in self-reported general health status. The baseline
1984 model only would allow a Year 1 lag in general health status and a baseline
1994 would only permit a short window for a new onset to occur. Therefore, models
with these longer lag structures in income and health were estimated for the 1989
baseline using a ten-year future horizon for the onset of a disease. All other covariates in these models remain the same as those in Tables 1–3.
Because the principal issue is whether these lags matter, Table 4 summarizes the
results with a series of F-tests and the associated probabilities for differences from
zero for a one-period lag, the three additional lags beyond the last year, and the
Table 4
Does SES Predict Future Major Onset with Lag Values of Income and Health?
(F-tests and p values; 1989 Baseline—10 Year Forecast)
Major Onset
Income lag 1
Income lag 2, 3, 4
Income lag 1,2,3,4
Health lag 1
Health lag 2,3,4
Health lag 1,2,3,4
Wealth
Change in stock wealth
Education
0.32
2.99
3.01
5.60
25.32
48.70
0.01
3.92
4.46
(0.57)
(0.39)
(0.56)
(0.13)
(0.01)
(0.00)
(0.89)
(0.05)
(0.02)
Minor Onset
0.05
4.13
5.12
17.67
34.71
69.22
1.53
1.42
5.43
(0.83)
(0.25)
(0.27)
(0.01)
(0.00)
(0.00)
(0.22)
(0.23)
(0.02)
753
754
The Journal of Human Resources
full four periods. The results obtained with lag structures on income are easy to
summarize—in no case do any of the variants of the income lags matter. Since a
four-period year should serve as a reasonable approximation to long-run measures
of income, this suggests that the reason current income doesn’t matter does not lie
in the familiar permanent-transitory income distinction. Moreover, the prior summary that they also do not predict future onset for the other financial variables included in the model—wealth and the change in wealth—remains intact even with
the full set of income and health lags in place. Finally, education still predicts future
major and minor disease onset indicating that it is not simply serving as a proxy for
permanent income, which should be adequately captured by the lag structure in past
incomes.
The story for lagged health, however, is quite different. A one-period lag for health
is clearly rejected and the full four-period lag structure is needed to predict future
health onset. As mentioned above, this probably reflects both measurement error
in health, especially in a noisy self-reported general health status (Crossley and
Kennedy 2002), as well as the fact that these health lags pick up the impact of duration of illness. However, even that four-period structure does not alter the conclusion that even after conditioning on these measures of health over the last four years,
education still predicts future onset and financial measures of SES do not.
One advantage of the PSID is that it allows us to investigate whether predictive
effects of SES on health onset vary with age. Much current research on this topic
is based either on the original HRS (Smith 1999) or the AHEAD cohort (Adams
et al. 2003), which were originally restricted to those between the ages of 51–61
and 70-plus, respectively. To investigate this possibility, separate models were estimated over three age groups—those younger than 40, those older than 60, and those
between those ages. While certainly less common than for people in the HRS and
AHEAD age ranges, health episodes for PSID respondents who are younger than
50 years old are not negligible. For example, among those in their 40s in the 1999
wave, one in seven had previously experienced a major disease onset at some time
in their lives and 40 percent have a minor chronic health condition. In the five years
before 1999, 7 percent of these 40-year-olds experienced a major disease onset while
one in four reported a new minor onset.
Sample sizes are necessarily much smaller when the models are estimated separately into these three age groups. This reduction in sample size becomes especially
problematic when onsets are rare—that is the younger the sample (especially for the
major onsets) and the shorter the duration over which one allows an onset to occur.
Because of these considerations, I estimated onset models for these age group models over ten-year horizons for the1984 and 1989 baseline year’s specifications and
over the full 15-year horizon for the 1984 baseline model. With the exception of
age, which is the stratifying variable, these models include the same set of covariates
as in Tables 1–3.
The principal results are summarized for all SES variables in Table 5. The column
labeled ‘‘financials’’ presents p values for tests that the three financial variables—
family income, wealth, and changes in stock market wealth—are jointly statistically
significant. In these age-specific models, once again, financial SES measures did not
matter in any of the age groups. In all but one of the four cases in which the associated p value was 0.10 or smaller, this was due to an ÔincorrectÕ positive sign on the
Smith
Table 5
Does SES Predict Future Onsets by Age?
Financials*
Education Coefficient
Education Z Score
1984 Baseline Models
Age group < 40
Major 1985–94
Minor 1985–94
Major 1985–99
Minor 1985–99
Age group 40–60
Major 1985–94
Minor 1985–94
Major 1985–99
Minor 1985–99
Age group > 60
Major 1985–94
Minor 1985–94
Major 1985–99
Minor 1985–99
0.29
0.31
0.94
0.65
20.045
20.044
20.030
20.035
21.79
22.57
21.49
22.50
0.21
0.09
0.32
0.05
20.019
20.009
20.050
20.032
21.04
20.56
22.97
22.08
0.40
0.44
0.24
0.21
0.026
20.038
0.015
20.029
0.81
21.39
0.57
21.05
1989 Baseline Models
Age group < 40
Major 1990–99
Minor 1990–99
Age group 40–60
Major 1990–99
Minor 1990–99
Age group > 60
Major 1990–99
Minor 1990–99
0.10
0.43
20.029
20.042
21.32
22.69
0.88
0.82
20.050
20.017
22.89
21.20
0.20
0.04
20.026
20.011
21.37
20.57
Note: This financial column reports the p value associated with the joint statistical significance for the coefficients for the three financial variables—income, wealth, and changes in stock market wealth. Models include all variables included in models in Tables 1–3. Using the minor onset models, sample sizes for the
1984 baseline were 2,788 (40 and 60). The corresponding sample sizes
for the 1989 baseline models were 2,972 (40 and 60).
wealth variable.12 Apparently, the ability of financial variables to predict future
onsets is not sensitive to the stage of the adult life cycle one happens to be in.13
12. The sole exception is the 60-plus age group for the 1989 baseline where the stock market variable is
statistically negative.
13. In contrast, family income as a child appears to matter a great deal. See Case, Lubotsky, and Paxson
(2002), Case, Fertig, and Paxson (2005), and Currie and Hansen (1999).
755
756
The Journal of Human Resources
The remaining columns in Table 5 list estimated coefficients and associated z values for the education variable in each model. Data necessarily gets thinner with these
age stratifications (especially in the oldest age group) and conclusions must be
tempered. But for those in the two age groups ages less than 60 and younger, the conclusion that education remains a strong predictor of future onset of disease apparently remains intact. This is far less apparent for the 60–plus age groups where
the estimated effects of education are statistically significant. This is consistent with
the findings of Adams et al. (2003) who used the AHEAD sample (aged 70-plus) and
reported virtually no effects of any SES variables including education on subsequent
health outcomes.
It may also reflect an inherent limitation of the PSID for research in this age
group. The 1984 baseline estimates are especially problematic with this age stratification. The 60-plus population in 1984 would be aged 75-plus and living in 1999
when the questions on past disease onset were first asked. Very few people of that
age who suffered a severe disease onset would still be alive in 1999, and they would
be quite unrepresentative of the actual onsets that occurred. While somewhat less severe, as they would have to be only aged 70-plus in 1999, the same type of problem
exists with the 1989 baseline models. This raises a more general point that the earlier
years of the PSID for this type of analysis are probably best restricted to younger age
groups.
Additional research on why education matters so much should receive high priority. It is not simply due to education’s association with poor health behaviors. Smoking, perhaps the most important health behavior, is controlled for in this analysis, and
other work with the HRS using more comprehensive health measures including
smoking, drinking, and obesity show similar net impacts of education (Smith 2005).
The basic source of the large estimated impact of schooling on success in many
life domains, of which health is only one example, may be the most important unanswered question in the social sciences. One possibility is that the education experience itself has little to do with it, but is simply a marker for personal traits (reasoning
ability, rates of time preference, etc.) that lead people to acquire more education and
to be healthier. But education may not be that passive. It may help train people in
decision-making, problem-solving, perseverance, and adaptive skills, all of which
have direct applications to a healthier life. Education may have biological effects
on the brain, which result in improved cognitive function and problem-solving ability, some of which may impart benefits to choices made regarding one’s health. This
is similar to the argument that more active brain functioning when younger pushes
off the onset of dementia.
One approach to isolating such a causal path of schooling follows the lead of
Lleras-Muney (2005) who provided evidence that laws affecting the age of school
entry and the age at which a child could obtain a work permit significantly affected
years of schooling completed during the 1915–39 period.14 Lleras-Muney found statistically significant effects of the component of schooling induced by compulsory
schooling laws on subsequent mortality.
14. Lleras-Muney presents evidence that the compulsory schooling laws are good instruments in the sense
that the F-test of joint significance of the laws yields a large value (around 5).
Smith
One difficulty with relying on this approach with the PSID is that those affected by
such laws would be relatively older cohorts in the PSID, the very group identified
above as not being PSID’s strong suit. A more promising approach may lie in applications in other countries where the impact of such laws were larger and more recent.
For example, England raised the statutory minimum from 14 to 15 shortly after the
Second World War and again from 15 to 16 in 1974.
Cunha and Heckman (2006) instead emphasize the importance of cognitive and
noncognitive skills such as motivation, persistence and self-control in explaining
schooling and one’s eventual success in multiple domains of life. They see these
skills as produced mainly by the family and by repetitive cumulative personal decisions whereby