This article was downloaded by: [DUT Library]

On: 06 October 2014, At: 13:12

Publisher: Routledge
Informa Ltd Registered in England and Wales Registered Number:
1072954 Registered office: Mortimer House, 37-41 Mortimer Street,
London W1T 3JH, UK

Education Economics
Publication details, including instructions for
authors and subscription information:
[Link]

Educational Attainment and

Health: Evidence from a
Sample of Older Adults
Scott J. Adams
Published online: 01 Jul 2010.

To cite this article: Scott J. Adams (2002) Educational Attainment and Health:
Evidence from a Sample of Older Adults, Education Economics, 10:1, 97-109, DOI:
10.1080/09645290110110227

To link to this article: [Link]

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all
the information (the “Content”) contained in the publications on our
platform. However, Taylor & Francis, our agents, and our licensors
make no representations or warranties whatsoever as to the accuracy,
completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of
the authors, and are not the views of or endorsed by Taylor & Francis.
The accuracy of the Content should not be relied upon and should be
independently verified with primary sources of information. Taylor and
Francis shall not be liable for any losses, actions, claims, proceedings,
demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in
relation to or arising out of the use of the Content.

This article may be used for research, teaching, and private study
purposes. Any substantial or systematic reproduction, redistribution,
reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access
 and use can be found at [Link]
conditions
Downloaded by [DUT Library] at 13:12 06 October 2014
 Education Economics, Vol. 10, No. 1, 2002

Educational Attainment and Health: Evidence from

a Sample of Older Adults

SCOTT J. ADAMS

ABSTRACT This paper provides evidence of an education ± health relationship among

older adults, even after controlling for individual and family background characteristics. It
Downloaded by [DUT Library] at 13:12 06 October 2014

also explores to what extent this relationship represents an independent effect of education
on health. An identification strategy using the quarter of birth and a set of parental and
sibling characteristics as instrumental variables for education is used to purge estimates of
biases resulting from error components correlated with education. The results suggest that
the previously observed correlation between educational attainment and adult health has a
causal component.

Introduction
The positive relationship between education and health has been well-reported
(House et al. (1990) and Ross and Wu (1995) are among the recent examples). Most
often, however, the focus has not been on health outcomes measured when old. The
work of Freedman and Martin (1999) is one exception, as they show that
educational attainment is important in terms of accounting for trends in functional
limitations among older Americans.
Policy implications of such work abound, as financing of programs like US
Medicare depends on accurate assessments of population health in the future. The
effect of education becomes more critical once one recognizes the rising levels of
educational attainment among those approaching old age. Using data from the
March 1990 Current Population Survey, the mean level of education for various age
groups was calculated. For individuals in their sixties (or, roughly, born in the
1920s), the mean level of education was 11.57 years. For those in their fifties, the
level of education was 12.23 years. For individuals in their forties, it was 13.08
years. If educational attainment affects health, then there should be changes in
population health in the near future.
While more attention needs to be given to the effect of education on the health of
the aged, the literature is also limited in attempts to identify the independent effect of
education. Frequently, important factors that may jointly determine education and
health, such as childhood environment and the rate at which individuals discount

Scott J. Adams, Robert Wood Johnson Foundation Scholar in Health Policy Research, School of Public
Health, University of Michigan, Ann Arbor, MI 48109, USA. Tel: + 1 734 936± 1321; Fax: + 1 734 936
9813; E-mail: sjadams@[Link]

ISSN 0964-5292 Print; 1469-5782 Online/02/010097 ± 13 € 2002 Taylor & Francis Ltd
DOI: 10.1080/09645290110110227
 98 S.J. Adams

time, are overlooked. Failure to account for all of these factors may lead to positive
health outcomes being unduly attributed to more years of schooling.
In light of this, the aim of this paper is twofold. First, I offer evidence of a
relationship between education and health among older people using a US sample
drawn from the Health and Retirement Study (HRS). I focus on functional ability as
a measure of health. Second, I devise and implement an empirical strategy that allows
for investigation into the extent to which the education± health relationship reflects an
independent effect of education on health. I offer new evidence that exploits the
exogenous source of variation in years of schooling that stems from compulsory
school laws in the United States (first noted by Angrist & Kueger, 1991).

Conceptual Framework
Among the several ways education can lead to better health, the most obvious is
through the enhancement of an individual’s health care utilization skills. This is the
idea of productive efficiency (Grossman, 1972). For example, a more educated
Downloaded by [DUT Library] at 13:12 06 October 2014

person might have a better understanding of his symptoms and, thus, be better able
to explain to a doctor what they are. The result will be more effective treatment and
better health outcomes later in life.
Education can also improve health through a better choice of individual health
inputs, which is called allocative efficiency. The empirical support for this
hypothesis is comprehensively surveyed by Grossman and Kaestner (1997).
Examples of research in this area include Shea et al. (1991), who find that education
reduces smoking, improves eating habits, and increases exercise, and Sander
(1999), who finds that educational attainment is inversely related to heavy drinking.
It is Kenkel (1991) who perhaps offers the most direct evidence, as he uses
measures of health knowledge to show that healthy behaviors are influenced by the
amount of information individuals have about the consequences of the behaviors.
Finally, education improves health through increased social standing. A prime
example is that more educated individuals obtain better jobs with greater earnings
and benefits (Willis, 1986) and, if unemployed, experience shorter unemployment
spells (Moen, 1999). This means that not only do more educated individuals have
better access to health care, but they are less likely to be without health insurance
for extended periods of time.
Instead of improving health independently, however, educational attainment
may merely reflect other factors that are health enhancing. For example, both
educational attainment and health are largely influenced by parental input. The
decisions made by parents concerning where and how to raise a family, self-image,
and attention from others are all family background characteristics that affect early
health and education investments. At the very least, ignoring these factors overstates
the effect of education (Elo & Preston, 1996). Clearly related to this are
intergenerational health transmissions, which have been shown to have a strong
effect on health (Smith & Kington, 1997). Thus, how healthy one’s parents were will
reflect how healthy he/she is.
In addition to unobserved childhood environment and family background
factors, many other unobserved factors affect both education and adult health. For
example, individuals with longer life expectancies will have more time during which
to earn returns to education. Thus, the human capital model suggests that these
individuals invest more in their schooling (Cawley, 1998). If one makes the logical
assumption that life expectancy early in life is formed by family health history or
 Educational Attainment and Health 99

one’s general level of health early in life, then such perception of the length of life
is an important unobserved factor. Another variable that has been mentioned as a
potential cause of both higher educational attainment and better health is the rate
at which individuals discount time. The basic idea is that some people place a higher
value on the future than do others. These individuals are more likely to make
investments in both their education and health (Fuchs, 1982).

Empirical Framework
The econometric model used to identify the independent effect of education is
summarized by the following two equations:

Ei = Xi d + Zi w + mi (1)

and
Downloaded by [DUT Library] at 13:12 06 October 2014

Hi = Xi b + gEi + «i (2)

where Ei is the educational attainment of individual i measured as years of schooling

completed. Its coefficient g is the parameter for which this paper seeks an estimate.
Hi is a health outcome for individual i. For the purposes of this paper, these are
binary indicators of whether an individual rates his own health good or better, very
good or better, and excellent or better, as well as whether someone reports relative
ease in performing a wide variety of everyday functions. Xi denotes the observed
variables that are thought to affect educational attainment and health outcomes.
These include basic individual characteristics, such as race and region of birth. Zi
consists of observed factors affecting education but not health. The error term mi
represents the factors causing variation in education that are not included in Xi or
Zi . Likewise, «i is the health error term. These error terms are allowed to be
correlated with each other. Thus, omitted factors, like the rate of time discount, can
affect both education and health through both of these terms.1
Estimation of the model can proceed in a variety of ways. First, one can estimate
equation (2) only by ordinary least squares (OLS), using proxy variables to control
for those factors causing correlation between the error terms. Alternatively, one can
estimate the two-equation model via a two-stage least squares (2SLS) approach.
This requires finding a suitable Zi to identify the model. Both approaches are
outlined in the remainder of the section.

OLS Approach
For this subsection, equation (1) is ignored except for the fact that it is understood
that there exists a set of factors correlated with educational attainment that are also
correlated with health. In this context, equation (2) can be rewritten as

Hi = Xi b + gEi + ci + ni (3)

The vector of unobservable variables that may affect health outcomes is denoted ci .
These variables may also be correlated with the observed demographic characteristics
and education. It is in ci that one would find family background information, the rate
of time discount, and other unobserved factors affecting both education and health.
 100 S.J. Adams

The first step to estimating the education± health relationship is to find suitable
proxy variables for the types of factors that exist in ci . Several good candidates are
included in the HRS. Including in the specification a variable indicating whether or
not one’s parents are currently living can capture the effect of the intergenerational
health transmission component of an individual’s family background. Other
important proxies for childhood environment are parental education (Schultz,
1984) and height. Height has also been shown to be a good indicator of early
investments in one’s health (Thomas & Strauss, 1997). This is important here
because these investments most likely took place before one’s ultimate educational
attainment was reached. Thus, height is a good control for an individual’s initial
health condition. It is also a good additional proxy variable for the life expectancy
correlates that may affect education.
An additional concern is how income fits into the model. After all, more
educated individuals typically have higher incomes and can afford better health
care. One should not control for income, however, because, as already suggested,
one effect of education is that it affords greater health consumption possibilities
Downloaded by [DUT Library] at 13:12 06 October 2014

through increased social standing. Variables like own current earnings, spousal
earnings, and household assets are all in some way the result of educational
attainment. What is needed is a measure of the component of a person’s income that
is not the result of educational attainment (Berger & Leigh, 1989). Ideally, one
would like to know family income at the time the educational attainment decision
was made, but conventional data sets normally do not contain such information. In
the OLS estimates for this paper, the aforementioned family background proxy
variables are assumed to capture this component of income.

Instrumental Variables Approach Using the Quarter of Birth

Still, however, using proxy variables may not capture all factors that are correlated
with education and health. Therefore, equations (1) and (2) should be estimated as
a system. The first step is to find a valid Zi (i.e. an instrumental variable or
variables). To be valid, instruments must explain educational attainment but be free
of correlation with the omitted factors that are causing the correlation between the
errors in equations (1) and (2). Berger and Leigh (1989) study health outcomes for
a younger sample using family background characteristics as instruments. In one
sense, these instruments are very good in that they explain educational attainment
quite well. Arguably, however, they may be correlated with the error term in
equation (2).
The work of Angrist and Krueger (1991) appears to offer an additional
instrument. They argue that the quarter of birth affects one’s educational attainment
due to the way in which compulsory school laws operate in the United States and offer
empirical support of the claim. At the time the individuals in my sample were in
school, most states required that an individual attend school until he/she reached the
age of 16. Others required attendance until the age of 17 or 18. Also, individuals were
permitted to start the first grade in most states if they turned 6 by some specified
month, frequently by the end of the year. Thus, the older members of a class tended to
be born earlier in the year. Moreover, those born earlier in the year reached the age at
which they were allowed to drop out of school first, presumably before the current
grade was completed. Thus, one should expect that among older individuals (for
whom dropping out of high school was still fairly common) there should be less
educational attainment for those born earlier in the year.
 Educational Attainment and Health 101

Verifying the effect of the quarter of birth on educational attainment is

important because the impact of a weak first stage in 2SLS is potentially spurious
inference. This is due to finite-sample bias. It has long been known that 2SLS
estimates are biased toward OLS estimates (Nagar, 1959) in small samples. Still,
popular belief held that this finite-sample problem would be unimportant if one
obtained a large enough sample, given the consistency of instrumental variables
estimates. Bound et al. (1995) and Staiger and Stock (1997) show that, even in
studies using large data sets, finite-sample bias may still be important.
Less controversy surrounds the quarter of birth in terms of the second criteria
for a good instrument. That is, the quarter of birth is not expected to have an effect
on unobserved factors related to health. Bound and Jaeger (1996), however, do
suggest that the quarter of birth may be related to some factors other than just
educational attainment, such as schizophrenia.
With the potential problems stated up front, this paper uses the quarter of birth
to identify variation in education. First, I use it by itself, estimating a model
consisting of
Downloaded by [DUT Library] at 13:12 06 October 2014

Ei = Xi d + Qi w + mi (4)

and equation (2). Equation (4) is just equation (1) with Zi replaced by a series
of quarter of birth dummy variables interacted with year of birth dummy variables
(Qi ). Note that a series of year of birth dummies must be added to Xi , the vector
of observable characteristics. Then, the variation in education that will be
identified as exogenous is the within-year differences in educational attainment by
the quarter of birth.
Ultimately, however, I wish to combine the quarter of birth with family
background information in my vector of instruments. This will ensure first-stage
precision. Additionally, with the quarter of birth included in the specification, I
can also test the overidentifying restrictions imposed when family background
characteristics are added to the vector of instruments. I therefore can exclude
from the vector those instruments that are potentially invalid, instead using them
as control variables.

Data and Results

Initially, I verify whether there exists an education± health relationship for older
individuals using the first wave of the HRS, which was conducted in 1992. The
longitudinal feature of the HRS is not helpful here because older people tend not to
obtain additional years of schooling. The sample is restricted to US born individuals
between the ages of 51 and 61 who report valid health information. Sample means
for selected demographic variables are listed in Table 1. The average female in the
sample is 56.08 years of age and has 12.37 years of completed schooling. The
average male is 56.05 years old and has completed 12.65 years of schooling. There
are 4577 men in the sample and 4059 women.
The most promising measures of health in the HRS from an empirical
standpoint are those that look at functional limitation. For instance, individuals are
asked to rate how easy they find walking a block or picking up a small coin. Since
almost everyone does these things each day, or at least attempts them every once in
a while, most people can give a dependable account of their ability to perform these
functions. There have been efforts to determine the reliability and validity of such
 102 S.J. Adams

Table 1. Sample description

Selected variable means Women Men

Age 56.08 56.05

Educational attainment (years completed) 12.37 12.65
White (%) 86.13 87.55
Non-white 13.87 12.45
Married 68.73 82.51
Separated or divorced 16.93 11.76
Widowed 10.66 1.59
Never married 3.68 4.14
Sample size 4577 4059

measures of functional ability as an indicator of health status. Typical reliability tests

Downloaded by [DUT Library] at 13:12 06 October 2014

involve asking a set of health questions of individuals at one point in time and
returning a short time later to ask the same questions, presumably before health
status has changed. In such tests, measures of functional ability are quite reliable
(Strauss et al., 1993). Validity tests can be carried out by way of internal consistency
checks. That is, if ability to perform certain functions is a good measure of health,
then it should be the case that those reporting high degrees of difficulty in physical
functioning should have a low opinion of their own health. Such tests of internal
consistency are normally passed by measures of functional limitation.
Functional ability measures are also a good measure of health because they
reflect an individual’s level of independence. This is why measures of functional
limitation have been used extensively in studying the health of the aged. Smith and
Kington’s (1997) work is a recent example.
The upper half of Table 2 presents means of the health variables of this paper
by educational attainment for females. The differences in the health status
between these education groups are striking and exist for both self-reported
general health measures as well as measures of functional ability. The lower half
of Table 2 presents the proportions of the same health measures by educational
attainment for men. The differences among education groups appear to be
smaller, but they are still striking.

OLS Results
For women, estimates of g in equation (3) when no controls are included are
reported in column 1 of Table 3. Each row contains estimates from separate
linear probability models with heteroscedasticity-corrected standard errors, where
the dependent variable is the binary health indicator appearing in the left-hand
column. The estimates certainly reflect a positive relationship between education
and health. All of the effects on general health and functional ability are
significant at the 0.05 level of significance. The estimates for men appear in
column 4 of Table 3. The coefficient estimates also indicate that education is
positively correlated with good health.
In columns 2 and 5 of Table 3, the vector of individual demographic controls,
Xi , is added for women and men, respectively. When these controls are included, the
coefficient estimates fall. Next, a proxy variable approach is employed in an attempt
 Educational Attainment and Health 103

to control for ci in equation (3). Columns 3 and 6 present estimates of the effect of
education on each health measure when the proxy variables for intergenerational
transmissions (whether one’s mother and father are still living) and for childhood
environment (mother’s education, father’s education, and height) are added. For
women, the coefficient estimates tend to fall slightly, but a significant, positive effect
on health still remains. For men, the coefficient estimates fall as well. Actually,
education’s positive effect on the ease in bathing and showering is no longer
significant, and its effect on the ease in picking up a small coin is significant at the
0.10 level of significance. 2

Instrumental Variables Results

The 2SLS results using only the quarter of birth as an instrument appear in Table
4 and are quite close to the OLS estimates. Education has a positive effect on all but
one of the health measures, and the p values of Hausman tests indicate that
instrumenting may not be necessary. There are a few reasons to be skeptical of these
Downloaded by [DUT Library] at 13:12 06 October 2014

results, however. First, the standard errors of the 2SLS estimates are quite high.
Second, the F-statistic for a test of the joint significance of the first-stage excluded

Table 2. Descriptive statistics for health outcomes by educational attainment

Never Graduated Attended

Entire graduated high school some Graduated
population high school only college college

Women
Rate own health as
Good or better 0.7948 0.6027 0.8198 0.8653 0.9384
Very good or better 0.5397 0.3059 0.5553 0.6372 0.7403
Excellent or better 0.2275 0.0948 0.2279 0.2878 0.3589
Can do the following with ease
Climb flights of stairs 0.6823 0.5369 0.6860 0.7481 0.8170
Stoop, kneel, or crouch 0.7432 0.6211 0.7452 0.8148 0.8379
Walk a block 0.9351 0.8735 0.9431 0.9645 0.9721
Bathe or shower 0.9777 0.9571 0.9776 0.9899 0.9949
Pick up a dime 0.9726 0.9464 0.9752 0.9840 0.9918
Number 4577 1241 1846 860 630

Men
Rate own health as
Good or better 0.8093 0.6271 0.8176 0.8663 0.9247
Very good or better 0.5303 0.3060 0.5156 0.5892 0.7188
Excellent or better 0.2373 0.1274 0.2065 0.2500 0.3781
Can do the following with ease
Climb flights of stairs 0.7881 0.6462 0.7790 0.8158 0.9152
Stoop, kneel, or crouch 0.8259 0.7396 0.8210 0.8357 0.9085
Walk a block 0.9482 0.8914 0.9531 0.9641 0.9827
Bathe or shower 0.9811 0.9655 0.9845 0.9804 0.9918
Pick up a dime 0.9725 0.9493 0.9802 0.9658 0.9894
Number 4059 1050 1377 769 863
 104 S.J. Adams

Table 3. OLS estimates of the effect of educational attainment on health outcomes

Women Men

Column 1 Column 2 Column 3 Column 4 Column 5 Column 6

Rate own health as

Good or better 0.0440 0.0379 0.0336 0.0371 0.0334 0.0296
(0.0024) (0.0024) (0.0027) (0.0021) (0.0023) (0.0026)

Very good or better 0.0544 0.0473 0.0410 0.0479 0.0436 0.0376

(0.0028) (0.0030) (0.0033) (0.0023) (0.0025) (0.0028)

Excellent or better 0.0323 0.0285 0.0244 0.0275 0.0255 0.0228

(0.0025) (0.0026) (0.0029) (0.0022) (0.0024) (0.0026)

Can do the following with ease

Climb flights of stairs 0.0385 0.0358 0.0336 0.0337 0.0319 0.0293
(0.0027) (0.0028) (0.0033) (0.0021) (0.0023) (0.0026)
Downloaded by [DUT Library] at 13:12 06 October 2014

Stoop, kneel, or crouch 0.0311 0.0300 0.0264 0.0194 0.0182 0.0159

(0.0027) (0.0028) (0.0031) (0.0021) (0.0022) (0.0025)

Walk a block 0.0158 0.0134 0.0108 0.0120 0.0114 0.0100

(0.0018) (0.0018) (0.0021) (0.0015) (0.0016) (0.0019)

Bathe or shower 0.0058 0.0051 0.0041 0.0031 0.0024 0.0011

(0.0010) (0.0010) (0.0011) (0.0009) (0.0009) (0.0010)

Pick up a dime 0.0058 0.0053 0.0048 0.0041 0.0036 0.0024

(0.0010) (0.0010) (0.0011) (0.0011) (0.0012) (0.0013)

Includes
Demographic controls No Yes Yes No Yes Yes
Proxy variables No No Yes No No Yes

Note: Reported are estimated coefficients of the education variable in separate linear probability models where
the dependent variable is whether one fits in the health category in the left-hand column. Heteroscedasticity-
consistent standard errors are reported in parentheses. Demographic controls include a quadratic in age, race,
marital status, region of birth, and whether one answered the survey questions himself/herself. Proxy variables
include whether or not one’s mother is living, whether or not one’s father is living, the education levels attained
by each parent, respondent height, and dummy variables indicating that values were set to zero due to missing
information.

instruments is not that large. Thus, the 2SLS estimates themselves may be biased
toward the OLS estimates.
We can allow for additional instruments to be added to the first stage to improve
its ability to explain variation in education. Specifically, this paper adds some family
background characteristics to the basic two-equation system summarized in
equations (1) and (2). The new system can be written as:

Ei = Xi d + Wi v + wZi + mi (5)

and

Hi = Xi b + gEi + Wi r + «i (6)
 Educational Attainment and Health 105

where Wi is a vector containing some childhood environment controls, specifically

the intergenerational health transmission proxies (whether or not one’s mother and
father are still alive) and height. Once these controls are added, one can conduct
specification tests to determine whether the remaining family background variables
belong in Wi or whether these variables can serve as instrumental variables (i.e.
belong in Zi ). The family background variables that are tested are the education of
each parent, the number of siblings, the percentage of siblings that are female, a
dummy variable denoting that one is the youngest child, and a dummy variable
denoting whether one is the middle child.
This approach utilizes the advantages of first-stage precision enjoyed by Berger
and Leigh (1989), but it does so with particular care as to not use instrumental
variables that are correlated with the equation (2) error term. This is because the
parental and sibling information is tested before being added to the vector of
instruments. More importantly, instruments are used that have a higher degree of
Downloaded by [DUT Library] at 13:12 06 October 2014

Table 4. 2SLS estimates of the effect of educational attainment on health outcomes

using the quarter of birth as an instrument

Women Men

p value for p value for

Hausman Hausman
2SLS exogeneity test 2SLS exogeneity test

Rate own health as

Good or better 0.0579 0.40 0.0358 0.91
(0.0241) (0.0212)

Very good or better 0.0409 0.83 ± 0.0008 0.12

(0.0306) (0.0283)

Excellent or better 0.0462 0.52 0.0507 0.30

(0.0272) (0.0248)

Can do the following with ease

Climb flights of stairs 0.0150 0.48 0.0153 0.47
(0.0295) (0.0230)

Stoop, kneel, or crouch 0.0047 0.36 0.0172 0.97

(0.0280) (0.0221)

Walk a block 0.0125 0.94 0.0131 0.89

(0.0148) (0.0132)

Bathe or shower 0.0104 0.54 0.0055 0.68

(0.0086) (0.0076)

Pick up a dime 0.0047 0.95 0.0022 0.88

(0.0100) (0.0097)

F-statistic of a joint test of the first-stage

significance of the excluded instruments 1.24 ± 1.14 ±

Note: Specifications contain controls for year of birth, race, marital status, region of birth, a quadratic in age,
and a dummy variable denoting whether one answered the survey himself or herself. The set of instrumental
variables includes three quarter of birth dummies interacted with 11 year dummies. Thus, there are 33 quarter
of birth variables in all.
 106 S.J. Adams

explanatory power than just the quarter of birth alone. Quarter of birth along with
the exogenous family characteristics explain much of the variation in education.
The results appear for women in Table 5. The first column reports the 2SLS
estimates using the quarter of birth and the family background characteristics that are
deemed valid as instruments. These estimates indicate that education has a positive
and significant effect on all of the health measures except the ability to stoop, kneel, or
crouch. For several of the measures, the p value of the Hausman test in the second
column suggests OLS estimates are, if anything, biased downward. Table 6 reports
the results for men. As with the results for women, these indicate that the effect of
education becomes stronger after one instruments for education.

Table 5. Alternative 2SLS estimates using the quarter of birth and family
characteristics as instruments, for women
Downloaded by [DUT Library] at 13:12 06 October 2014

p value of a test of
p value of misspecification
Estimated Hausman when additional
effect of exogeneity Additional instruments are
education test instruments a added

Rate own health as

Good or better 0.0484 0.02 P,S 0.91
(0.0054)

Very good or better 0.0634 0.00 P,S 0.19

(0.0068)

Excellent or better 0.0422 0.01 P,S 0.74

(0.0061)

Can do the following with ease

Climb flights of stairs 0.0405 0.36 P,S 0.17
(0.0063)

Stoop, kneel, or crouch ± 0.0136 0.10 S 0.64

(0.0246)

Walk a block 0.0235 0.00 P,S 0.26

(0.0034)

Bathe or shower 0.0075 0.13 P,S 0.93

(0.0020)

Pick up a dime 0.0063 0.56 P,S 0.60

(0.0021)

F-statistic of a joint test of the first-stage

significance of the excluded instruments 20.71 ± ± ±

Note: Each specification includes as regressors year of birth dummy variables, race controls, marital controls,
region of birth controls, a quadratic in age, a dummy variable indicating whether or not one answered the survey
herself, height, dummy variables indicating whether each parent is alive, and dummy variables indicating
missing information. The set of instrumental variables always includes quarter of birth. Additional instruments
include P (mother’s and father’s education) and/or S (number of siblings, percentage of siblings that are female,
a dummy variable denoting that one is the youngest child, and a dummy variable denoting whether one is the
middle child).
a
P, Parents’ education; S, sibling information.
 Educational Attainment and Health 107

Table 6. Alternative 2SLS estimates using the quarter of birth and family
characteristics as instruments, for men

p value of a test of
p value of misspecification
Estimated Hausman when additional
effect of exogeneity Additional instruments are
education test instruments a added

Rate own health as

Good or better 0.0441 0.01 P,S 0.44
(0.0052)

Very good or better 0.0225 0.42 S 0.15

(0.0184)

Excellent or better 0.0320 0.13 P,S 0.50

(0.0057)
Downloaded by [DUT Library] at 13:12 06 October 2014

Can do the following with ease

Climb flights of stairs 0.0385 0.13 P,S 0.20
(0.0053)

Stoop, kneel, or crouch 0.0258 0.09 P,S 0.58

(0.0051)

Walk a block 0.0175 0.01 P,S 0.57

(0.0029)

Bathe or shower 0.0076 0.00 P,S 0.64

(0.0019)

Pick up a dime 0.0084 0.02 P,S 0.48

(0.0023)

F-statistic of a joint test of the first-stage

significance of the excluded instruments 16.24 ± ± ±

Note: See Table 5 note for further details.

a
P, Parents’ education; S, sibling information.

Conclusion
This paper provides new evidence that there exists an education± health relation-
ship that is quite pronounced for older individuals. Moreover, this paper provides
a framework for empirical investigation into the independent effect of education
on health and applies it. Specifically, one’s quarter of birth is used in an
instrumental variables approach aimed at purging estimates of biases resulting
from individual-specific error components correlated with education. Such evi-
dence provides greater confidence that education has an independent effect on
health, even at advanced ages.
The OLS results indicate that with higher levels of educational attainment
come healthier outcomes for older people. This holds even after one controls for
the observable demographic characteristics of the individual and his/her family
background. The effect is especially strong among women. For men, the effects
of education on the health measures of the paper are also significant but not quite
as strong.
 108 S.J. Adams

The 2SLS results of the paper provide evidence that education’s effect on
health for both older women and men is independent of the influence of omitted
variables. Most of the estimates in the paper indicate a positive and significant (at
least at the 0.10 level of significance) effect of education on health. Moreover,
almost all of the estimates that are not significant are positive. Altogether, the
results indicate that even after correcting for biases in the estimates resulting from
omitted variables, education’s effect on health remains positive and significant for
most reliable health measures.

Acknowledgements
The author thanks David Neumark, John Strauss, and an anonymous referee for
helpful comments and suggestions, and also thanks the Robert Wood Johnson
Foundation for financial support.
Downloaded by [DUT Library] at 13:12 06 October 2014

Notes
1. The model can be amended to account for the potential of reverse causality. For instance, Hi
could be added to the right side of equation (1). This would not be appropriate here, however,
because the health outcomes of this paper are measured when old. These are not known when
the education decision is being made. The more sensible approach is to allow factors related to
life expectancy to be components of the error terms in both equations (1) and (2). Thus,
incorporating the potential of reverse causality (as it pertains to the analysis of this paper)
requires no explicit change to the model.
2. The results presented thus far do not change markedly if a probit model is used to estimate
equation (3). Moreover, when the measures of general health and functional ability are re-
defined to be on a scale of ill health and an ordered probit model is estimated, education
continues to have a positive effect on health. Specifically, general health is given a value of 1± 5.
A value of one indicates that an individual rates his/her own health as excellent. Values of 2± 5
indicate ratings of very good, good, fair, and poor, respectively. Likewise, ability to perform a
daily function is given a value of 1± 5, based on one’s own rating of relative difficulty in
performing the task. A value of 1 represents no difficulty. The effect of education on ill health
is negative and highly significant for all measures. Even the effects of education on ease in
bathing or showering and ease in picking up a dime, which became weaker when the proxy
variables were added in the OLS specifications, are positive and significant when ordered
probits are used.
Although outside of the scope of this paper, the effect of education controlling for current
income is an interesting question and one that the HRS can help answer. For a subsample of
respondents, individual earnings are reported. When I include earnings as a control variable, the
effect of education weakens slightly but the results remain qualitatively similar for both men and
women. The coefficient estimates on earnings tend to be small. Moreover, when I include a set
of dummy variables denoting levels of household equity, the education coefficient remains
strong and significant. Thus, education significantly increases health status even after
controlling for earnings and household equity. This provides support for the allocative and
productive efficiency hypotheses described in the second section while providing less support
for the increased social standing hypothesis.

References
Angrist, J. & Krueger, A. (1991) Does compulsory school attendance affect schooling and
earnings?, Quarterly Journal of Economics, 106, pp. 979± 1014.
Berger, M. & Leigh, J. (1989) Schooling, self-selection, and health, The Journal of Human
Resources, 24, pp. 433± 455.
Bound, J., Jaeger, D. & Baker, R. (1995) Problems with instrumental variables estimation when the
correlation between the instruments and the endogenous explanatory variable is weak, Journal
of the American Statistical Association, 90, pp. 443± 450.
 Educational Attainment and Health 109
Bound, J. & Jaeger, D. (1996) On the validity of season of birth as an instrument in wage
equations: a comment on Angrist and Krueger’s `Does compulsory schooling attendance affect
schooling and earnings?’ , NBER Working Paper No. 5835 (Cambridge, MA, National Bureau of
Economic Research).
Cawley, J. (1998) The correlation between schooling and mortality, mimeo, University of Chicago,
IL.
Elo, I. & Preston, S. (1996) Educational differences in mortality: United States, 1979± 1985, Social
Science and Medicine, 42, pp. 47± 57.
Freedman, V. & Martin, L. (1999) The role of education in explaining and forecasting trends in
functional limitations among older Americans, Demography, 36, pp. 461± 473.
Fuchs, V. (Ed.) (1982) Time preference and health: an exploratory study, Economic Aspects of
Health (Chicago, IL, University of Chicago Press).
Grossman, M. (1972) On the concept of health capital and the demand for health, Journal of
Political Economy, 80, pp. 223± 255.
Grossman, M. & Kaestner, R. (1997) Effects of education on health, in: Behrman, J. R. & Stacey,
N. (Eds) The Social Benefits of Education (Ann Arbor, MI, University of Michigan Press).
House, J., Kessler, R., Herzog, A., Mero, R., Kinney, A. & Breslow, M. (1990) Age, socioeconomic
status, and health, The Milbank Quarterly, 68, pp. 383± 411.
Kenkel, D. (1991) Health behavior, health knowledge, and schooling, Journal of Political Economy,
Downloaded by [DUT Library] at 13:12 06 October 2014

99, pp. 287± 305.

Moen, E. (1999) Education, ranking, and competition for jobs, Journal of Labor Economics, 17, pp.
694± 723.
Nagar, A. (1959) The bias and moment matrix of the general k-class estimators of the parameters
in simultaneous equations, Econometrica, 27, pp. 575± 595.
Ross, C. & Wu, C. (1995) The links between education and health, American Sociological Review,
60, pp. 719± 745.
Sander, W. (1999) Cognitive ability, schooling, and the demand for alcohol by young adults,
Education Economics, 7, pp. 53± 66.
Schultz, T. (1984) Studying the impact of household economic and community variables on child
mortality, Population and Development Review, 10, pp. S215± S235.
Shea, S., Stein, A., Basch, C., Lantigua, R., Maylahn, C., Strogatz, D. & Novick, L. (1991)
Independent associations of educational attainment and ethnicity with behavioral risk factors for
cardiovascular disease, American Journal of Epidemiology, 134, pp. 567± 582.
Smith, J. & Kington, R. (1997) Demographic and economic correlates of health in old age,
Demography, 34, pp. 159± 170.
Staiger, D. & Stock, J. (1997) Instrumental variables regressions with weak instruments,
Econometrica , 65, pp. 557± 586.
Strauss, J., Gertler, P., Rahman, O. & Fox, K. (1993) Gender and life-cycle differentials in the
patterns and determinants of adult health, Journal of Human Resources, 28, pp. 791± 837.
Thomas, D. & Strauss, J. (1997) Health and wages: evidence on men and women in urban Brazil,
Journal of Econometrics, 77, pp. 159± 185.
Willis, R. (1986) Wage determinants: a survey and reinterpretation of human capital earnings
functions, in: Ashenfelter, O. & Layard, R. (Eds) Handbook of Labor Economics, Volume I
(Amsterdam, North-Holland Publishing Company).