You are on page 1of 14

ARTICLE IN PRESS

Social Science Research xxx (2008) xxx–xxx

Contents lists available at ScienceDirect

Social Science Research


journal homepage: www.elsevier.com/locate/ssresearch

Educational differences in health risks and illness over the life course:
A test of cumulative disadvantage theory q
Matthew E. Dupre *
Department of Sociology and Center for the Study of Aging and Human Development, Duke University, Box 3003, Duke University Medical Center,
Durham, NC 27710, USA

a r t i c l e i n f o a b s t r a c t

This study tests the cumulative disadvantage hypothesis by examining the age-varying
Keywords:
relationships between education, health risks, and disease onset and survival duration.
Cumulative disadvantage
Results based on 20 years of longitudinal data suggest that education is related to both
Disease
Education the individual and accumulated number of behavioral, social, and economic health-risks,
Health which in turn, are related to increasing educational differences in rates of disease incidence
Life course and survival. For hypertension, behavioral risks fully account for education’s negative effect
Mortality on disease onset whereas educational differences in survival are best explained by the
Hypertension accumulation of social and economic risks. For heart attack, a combination of behavioral,
Heart attack social, and economic risks mediate the association between education and incidence, but
neither the individual nor accumulated health-risks could account for education’s positive
effect on surviving after a heart attack. Similar findings for diabetes and stroke are also
discussed.
Ó 2008 Elsevier Inc. All rights reserved.

1. Introduction

Over the past half century the link between education and health has been one of the most widely documented findings in
sociological research. The enormous body of evidence accumulated to date shows a robust positive association between edu-
cational attainment and a variety of health outcomes (Antonovsky, 1967; Crimmins and Saito, 2001; Feldman et al., 1989;
Kitigawa and Hauser, 1973; Lynch, 2003). Although the causal connection between education and health appears to be well
established, its explanation is still not entirely clear. Even less understood is how the relationship varies with age.
The increasing emphasis on life-course theory and the use of longitudinal data have provided intriguing new insights for
understanding the dynamic relationship between education and health (see George, 2005). However, these advances have
also led to greater complexity and inconsistent findings. One of the current debates is whether the effect of education on
health increases with age or whether health disparities decline after middle adulthood. Cumulative disadvantage theory
emphasizes the role of accumulated risks that produce educational inequalities in health that increase over the life course
(Ross and Wu, 1996). The age-as-leveler hypothesis counters that health differentials diminish at older ages rather than in-
crease (House et al., 1994; Beckett, 2000). However, most of the evidence under debate is based on aggregated health out-
comes (prevalence) and does not consider the intervening mechanisms producing these patterns.

q
This research was partially supported by the Carolina Population Center at the University of North Carolina at Chapel Hill, NICHD Postdoctoral Training
Grant T32 HD07168-27. I thank Linda K. George, Glen H. Elder, Jr., and the anonymous reviewers for their thoughtful comments and advice.
* Fax: +1 919 668 0453.
E-mail address: med11@geri.duke.edu

0049-089X/$ - see front matter Ó 2008 Elsevier Inc. All rights reserved.
doi:10.1016/j.ssresearch.2008.05.007

Please cite this article in press as: Dupre, M.E., Educational differences in health risks and illness over the life course: ...,
Social Sci. Res. (2008), doi:10.1016/j.ssresearch.2008.05.007
ARTICLE IN PRESS
2 M.E. Dupre / Social Science Research xxx (2008) xxx–xxx

In this article, I draw upon epidemiologic methods to examine the individual-level effects of education on health from a
cumulative disadvantage framework. According to the theory, the relationship between education and health develops over
time through a process of accumulated risk (or lack of resources) that increasingly disadvantages low educated individuals
as they age (see Ross and Wu, 1995). The health consequences of these disadvantages are ultimately manifested in an acceler-
ated rate of illness and mortality for the poorly educated. Therefore, I argue that the onset and duration of illness are appropriate
outcomes for studying education’s effect across age because they quantify the individual-level rate of health decline—and ulti-
mately shape population-level disparities (prevalence = incidence  duration). To my knowledge, this research is the first to
examine education’s cumulative effect on disease onset and survival while testing the potential mediators of the relationships.
This study utilizes 20 years of nationally representative longitudinal data to analyze the age-varying relationships be-
tween education, its mechanisms, and disease onset and survival duration. I first examine how education is associated with
long-term behavioral, social, and economic health-risks, and second, how education is associated with accumulated health-
risks across age. Next, I examine the temporal effects of education on disease onset and survival. Finally, I test whether edu-
cational differences in disease onset and survival are explained by the independent or accumulated effects of behavioral, so-
cial, and economic health-risks.

2. Background

Countless studies show that education is inversely related to morbidity, disability, and mortality (e.g., Kaplan and Keil,
1993; Manton et al., 1997; Preston and Elo, 1995), and despite decades of inquiry, the explanation for the relationship is
not entirely understood. The literature currently identifies three interwoven groups of mechanisms linking education and
health (refer to Ross and Wu, 1995). The social–psychological explanation suggests that education provides a greater sense
of personal control (Mirowsky and Ross, 1998), social support (Eckenrode, 1983; Lin and Ensel, 1989), coping skills (Pearlin
and Schooler, 1978; Wheaton, 1983), and problem-solving abilities (Ross and Mirowsky, 1989; Wheaton, 1983), which pro-
mote health. Persons with low levels of education are more vulnerable to illness because they are less equipped psycholog-
ically and socially to minimize the harmful effects of stressors compared to their more educated counterparts (Ross and
Huber, 1985; Rodin, 1986; Rowe and Kahn, 1987; Turner et al., 1995).
Employment and economic resources are a second pathway by which the well educated are more likely to acquire and
maintain employment and produce higher levels of income and accumulated wealth (Crystal and Shea, 1990; Sewell and
Hauser, 1975); whereas the poorly educated generally occupy lower-status occupations and have lower earnings (Grusky
and DiPrete, 1990; Jencks et al., 1979). Lacking employment and economic resources harm health by limiting access to qual-
ity health care (Aday and Andersen, 1984), exposing workers to noxious chemicals and carcinogens (Leigh, 1983), and hin-
dering the ability to purchase healthy foods and other commodities to prevent and treat illness (Mirowsky et al., 2000; Ross
and Wu, 1995). Health lifestyles are the third and perhaps most researched explanatory variables (Lantz et al., 2001). An edu-
cation gradient has been shown—directly or indirectly—with regard to physical exercise (Grzywacz and Marks, 2001), diet
(Roos et al., 1998), body mass index (Luepker et al., 1993), smoking (Pierce et al., 1989), preventative medical care (Ross
and Wu, 1995; Williams, 1990), and immoderate alcohol consumption (Millar and Wigle, 1986). The effects of these behav-
iors on health are fairly intuitive and well documented.
Despite the vast literature identifying education’s numerous mediators, much of our knowledge is based on cross-sectional
research that is largely under-theorized and often ignores the temporal dimensions of the relationships. This study provides a
theoretical framework and analytic strategy for investigating how education is related to health across the life course.

2.1. Cumulative disadvantage theory

First proposed by Merton (1968) as the Matthew Effect, cumulative advantage theory was originally applied to the sci-
entific community to explain how inequality increases among scholars as productivity and recognition increasingly favor
the most advantaged academics. Since then, researchers have steadily adopted and expanded the theory to explain the accu-
mulation of various other social and economic inequalities (Crystal and Shea, 1990; Elman and O’Rand, 2004; O’Rand, 1996;
Rosenbaum, 1984). DiPrete and Eirich (2006) review the theoretical breadth and major developments of this literature and
underscore the empirical strengths and challenges of studying cumulative disadvantage processes.
Education is considered central to a cumulative disadvantage process because it is the nexus of numerous social pathways
and a determinant of various outcomes throughout the life course (Pallas, 2003). Although it is well known that individuals who
are poorly educated have lower economic returns than their more educated counterparts (Elman and O’Rand, 2004), the cumu-
lative effect of low education on health is less clear. The major tenet of a cumulative disadvantage in health is that individuals
who lack protective resources (e.g., knowledge of health risks, economic assets, and psychosocial buffers) or exhibit excess
adversity (e.g., poverty and obesity) are at an increasing risk of illness as they age. Low educated persons have limited employ-
ment and economic opportunities that, in turn, position them in social contexts marked by fewer resources, higher rates of
physical and psychological stress, and inadequate methods to cope with these adversities. Ross and Wu (1996) were among
the first to test the theory on health and found that education had an increasing effect on physical functioning and well-being
in two nationally representative samples. More recent findings utilizing survey and census data also show that educational
differences in survival and self-rated health increase with age within and across birth cohorts (Lauderdale, 2001; Lynch, 2003).

Please cite this article in press as: Dupre, M.E., Educational differences in health risks and illness over the life course: ...,
Social Sci. Res. (2008), doi:10.1016/j.ssresearch.2008.05.007
ARTICLE IN PRESS
M.E. Dupre / Social Science Research xxx (2008) xxx–xxx 3

Two recent studies also demonstrate the cumulative impact of several risk factors on health. Ferraro and Kelley-Moore
(2003) show that obesity exerts both immediate and long-term effects on physical disability and that obesity in early life
has enduring consequences for lower-body impairments well into adulthood. They also find that compensatory exercise re-
duces the negative toll of obesity on disability over time. The cumulative effect of smoking on health is similar in that the
harmful consequences of smoking are exacerbated among persons who are socioeconomically disadvantaged (Pampel and
Rogers, 2004). The negative health effects of excess drinking, inactivity, stress, and overweight are also shown to be more
pronounced among those with low social resources.

2.2. Limitations of previous research

Until recently, most evidence linking education and health across age contradicted the cumulative disadvantage argu-
ment by showing that the association decreased at older ages rather than increased (Feldman et al., 1989; House et al.,
1990, 1994; Kitigawa and Hauser, 1973). Several explanations for the age-as-leveler argument have been posited in previous
research, although some of these are not well supported. For example, it has been suggested that social programs favoring
the elderly, such as Medicare, have an equalizing effect among heterogeneous older persons (Preston and Elo, 1995). Yet
studies often show that the diminishing effect of education on health does not occur until 10 years after eligibility for Medi-
care begins (e.g., House et al., 1994 report age 75 and older). Furthermore, inequalities should have declined among the el-
derly since 1965—when Medicare was first introduced—but inequalities appear to be growing (Lynch, 2003; Preston and Elo,
1995).
The prevailing explanation for decreasing educational inequalities in health is selective mortality. Selection occurs when poorly
educated individuals exhibit higher rates of mortality at younger ages compared to the well educated, leaving a more robust group
of low educated persons among the better-educated survivors, and thus reducing heterogeneity in health. However, this pattern
does not mean that the effect of education declines with age, but rather, it signals when the relative difference in health between
high and low educated groups becomes less observable as sick persons are removed from the population. From a cumulative dis-
advantage standpoint, failing to account for selective mortality among the low educated inevitably produces misleading evidence
for converging inequalities as selection biases become increasingly pronounced with age (Dupre, 2007).
Several limitations of previous research are addressed in the current study. First, much of the existing research is based on
cross-sectional surveys or census data. Consequently, analyses of educational differences across age are limited to static age
interactions or age-group comparisons of education and health. Although informative, such approaches are particularly sen-
sitive to the selective composition of the sample resulting from differential illness and mortality among population mem-
bers. Moreover, age effects are confounded with and indistinguishable from cohort effects. This research draws upon
20 years of longitudinal data to reduce potential cohort and selection biases when examining the individual-level patterns
of health decline across age.
A second and related limitation is that most studies use aggregated health measures (e.g., self-rated health, disability le-
vel, number of illnesses) at one or more time points. However, such prevalence measures are cross-sectional and simply indi-
cate the frequency (or existence) of a given health outcome at a given interview. I argue that disease incidence (onset) and
duration (length of illness) are better suited to quantify the rate of individual health-decline by incorporating the age-related
timing of sickness and death (prevalence = incidence  duration). From a cumulative disadvantage perspective, disease onset
and mortality occur over the life course at a faster rate for low educated persons than for well educated persons (see Dupre,
2007, for a detailed discussion).
Third, the mechanisms by which education exerts a cumulative effect on health are not well developed in literature. It is
argued that accumulated health-risks among the poorly educated leads to an acceleration of aging that produces educational
disparities in health that increase with age. Many of the frequently identified behavioral and structural risk factors require
years to exact a toll on human health. Consider the effect of physical inactivity: a sedentary lifestyle is not especially harmful
to health during young adulthood, but over time the deleterious effect of no exercise leads to higher rates of hypertension,
diabetes, and heart disease. Similar arguments can be made regarding cigarette smoking, poor social support or coping strat-
egies, economic hardship, and workplace pollutants. Prospectively, research also shows that the associations between risk
factors and disease diminish with increasing age because of disease detection, risk measurement, mortality selection, and
ceiling effects (Kaplan et al., 1999). The current analysis incorporates time-varying measures of risk to examine how multiple
risks exert separate and cumulative effects on health.

2.3. Research questions

The principal hypothesis of this study is that the effect of education on health develops through a pathway of accumu-
lating disadvantage over the life course. The following research questions guide the analysis.

(1) How is education associated with behavioral, social, and economic health-risks over the life course? I expect that education
is negatively related to poor health practices and other social and economic disadvantages. Based on cumulative dis-
advantage theory, I further predict that education’s negative association with separate and accumulated health-risks
will decrease across age as adults with greater risks exit the population due to accelerated health decline and
mortality.

Please cite this article in press as: Dupre, M.E., Educational differences in health risks and illness over the life course: ...,
Social Sci. Res. (2008), doi:10.1016/j.ssresearch.2008.05.007
ARTICLE IN PRESS
4 M.E. Dupre / Social Science Research xxx (2008) xxx–xxx

(2) How is education associated with the rate of disease onset and the duration of survival across age? Cumulative
disadvantage theory suggests that educational disparities in health increase with age because illness and mor-
tality develop at a faster rate for low educated persons than for well educated persons. Therefore, I anticipate
that education is positively related to the age of disease onset and/or the length of survival following
diagnosis.
(3) Are the effects of education on disease onset and survival explained by the individual or accumulated effects of long-
term behavioral, social, and economic health-risks? Despite extensive evidence of education’s association with
numerous risk factors and their health consequences, almost no attention is given to whether the accumulation
of disadvantages differentiates educational groups to produce health inequalities. I predict that accumulated risk
factors better explain the relationship between education and disease onset and survival than the separate risk
factors.

3. Data and measurement

Data for this study come from the National Health and Nutrition Examination Survey-I (NHANES I) Epidemiologic Follow-up
Study (NHEFS). The NHANES I collected data on a national probability sample of noninstitutionalized U.S. civilians between
the ages of 1 and 74 during 1971–75. The baseline sample for the NHEFS consists of a multistage probability sample of
14,407 participants of NHANES I ages 25–74 at the time of the initial survey (Cohen et al., 1987). Follow-up interviews with
subjects (or proxies) were conducted in 1982–1984 (n = 12,220), 1987 (n = 9998) and 1992 (n = 9281), with response rates of
84.8%, 85.1%, and 82.9%, respectively (Cox et al., 1997). This study utilizes all four interviews and provides 20 years of panel
data.
The main disease outcomes for analysis are the incidence and case-fatality rates (survival duration) of two health condi-
tions: hypertension and heart attack. These conditions are chosen because they are highly prevalent, they are among the
leading causes of death in the U.S., and they share many of the same socio-behavioral risk factors (Kaplan and Keil, 1993;
Murray and Lopez, 1996; National Center for Health Statistics, 2001). Additional analyses were conducted for diabetes
and stroke and are discussed. Age of onset (incidence) was ascertained for each condition by using the respondents’ date
of birth and the year they were first told by a doctor they had a health problem.1 Missing data on the timing of diagnosis
was minimal and efforts were made to identify onset from the prospective data. In all but one case, the panel data allowed
the age of onset to be determined from subsequent waves (the remaining case was retained as a censored observation). The
duration of illness is computed from NDI-linked death certificates as the number of years alive after the onset of a condition.2
For estimation, the age of onset and years of survival serve as the duration indicators in the hazard analyses and are used to
create the multi-record data files that include time-varying covariates.
Several behavioral, social, and economic time-varying dichotomous measures of health risk are examined. Suitable
measures of social–psychological mechanisms were not available in the survey and precluded their analysis. For the
behavioral risks, I include two variables to measure smoking: one indicating persons who ever smoked at least 100 cig-
arettes (coded 1) and another indicating persons who are current smokers (coded 1).3 Heavy drinking is coded 1 for indi-
viduals who consume three or more alcoholic beverages per day. Body mass index (BMI) is computed from the
respondents’ height and weight reported at each wave using the formula (weight in pounds/(height in inches)  (height
in inches))  703. Separate variables are used to categorize BMI: underweight (BMI < 18.5), overweight (25.0–29.9), and
obese (P30.0)—each coded 1. Medical research indicates that overweight and obesity (BMI > =25) are strong risk factors
for disease onset and that obesity is a more appropriate risk factor for survival duration because of an apparent BMI
paradox that suggests overweight and mildly obese individuals face mortality risks comparable to those with a normal
BMI (Romero-Corral et al., 2006).
For social and economic risks, I include time-varying measures of employment status, marital status, and house-
hold income. Employment status is coded using two variables: one indicating persons who are not employed (coded
1) and another to control for persons who are retired (coded 1). Marital status is coded 1 for individuals who are
divorced, separated, widowed, or never married (married = 0). Because of measurement inconsistencies across inter-
views, income is grouped into wave-specific quartiles and the lowest 25th income quartile is dichotomized as a rel-
ative risk-factor.
Accumulated health-risks are modeled with two time-varying count measures that add the respective number of behav-
ioral and socio-economic risks outlined above. Cumulative risk variables are well established in developmental research and

1
The reported age at diagnosis is assumed to be a reliable indicator for incidence. The estimates of the beneficial effect of education are likely conservative
given the information-seeking behavior of better-educated individuals (leading to early diagnosis). In terms of survival, diagnosis also marks the age that
individuals learn of an illness and respond with improved health management and/or treatment to prolong life.
2
All-cause mortality is used for analysis because of the high degree of comorbidity among the conditions and many other illnesses.
3
For the initial interview, the NHEFS only collected information on smoking for half of the sample. Fortunately, smoking status could be identified for all but
5% of persons at baseline by using the detailed information gathered from all respondents at subsequent interviews. Those who report never smoking at any
follow-up wave are coded as having never smoked in the initial interview. Persons who are current smokers at baseline were determined by using multiple
questions asking about ‘‘starting” and ‘‘ending” dates/ages for smokers, which are available in later surveys.

Please cite this article in press as: Dupre, M.E., Educational differences in health risks and illness over the life course: ...,
Social Sci. Res. (2008), doi:10.1016/j.ssresearch.2008.05.007
ARTICLE IN PRESS
M.E. Dupre / Social Science Research xxx (2008) xxx–xxx 5

life-course epidemiology (Burchinal et al., 2000; Kuh and Ben Shlomo, 2004; Sameroff et al., 1993) yet are untested in most
sociological studies.4 For this analysis, a cumulative measure of behavioral risks includes a count of risks for persons who are
current smokers, heavy drinkers, and have a body mass index in the high-risk category. A cumulative measure of social and
economic risks includes a count of persons who are low income, unemployed, and not married. Both measures range from 0
(no risks) to 3, if persons report being disadvantaged in each behavioral or social and economic risk category.
Education is coded as the years of education completed at baseline (range = 0–17). A total of 105 respondents (0.7%) did
not report education level and were coded to the sample mean. Analyses using educational categories yielded similar results
to the continuous specification. A time-varying measure of education is not possible with these data, although I suspect that
the frequency of adults returning to formal schooling is low given the initial interview period (1971–75) and relatively older
birth cohorts (1890–1940). Additional variables for analysis include age, race, gender, household size, geographic region, ur-
ban/rural residence, comorbidity, and birth cohort. Age is a time-varying measure that increases monotonically according to
model specification (described below). Race and gender are dichotomized so that Whites and males equal 1, respectively. Ra-
cial categories in the NHEFS are essentially binomial (98% White and Black) and prohibit further classification. Household
size is included as a proxy for living conditions and indicates whether persons may be living alone or residing in highly occu-
pied dwellings (range = 1–20). Geographic region identifies persons living in the Northeast, Midwest, South, or West and is
measured at baseline. A variable for South is coded 1 to control for the greater rates of hypertension, heart disease, diabetes,
and stroke in the south. Rural status is coded 1 for persons residing outside urban areas. A dichotomous time-varying mea-
sure of disease comorbidity is used to indicate (1) the preexistence of hypertension, diabetes, stroke, or heart attack prior to
onset and (2) the presence of additional conditions in the survival analysis following diagnosis. Finally, three cohorts are con-
structed by grouping participants ages 25–44, 45–64, and 65 and older at baseline and are referred to as the 1930–1940s,
1910–1920s, and 1890–1900s birth cohorts, respectively.

4. Analysis

Baseline and prospective data from the NHEFS panel is used to construct an individual age-specific dataset for analysis. All
analyses use Huber–White sandwich estimators to account for the lack of independence across observations in the multi-
record data and to reduce unobserved heterogeneity bias. The first set of analyses uses binary regression models to examine
how education relates to cigarette smoking, heavy drinking, unhealthy weight, low income, unemployment, and being
unmarried. Poisson regression models are used to examine the relationships between education and the two count variables
indicating the cumulative number of behavioral and socio-economic risks. Supplementary results from negative binomial
models were consistent with the Poisson estimates and suggested adequate mode fit (i.e., minimal overdispersion). Both sets
of analyses include interactions between education and time-varying age to parameterize the changing relationships across
age in the panel. Both sets of risk models also account for differential exposure by including time as an ancillary factor in the
log-link function (Long and Freese, 2006).
The next set of analyses examines the relationships between education, disease incidence, and survival duration and test
the mechanisms thought to explain the associations. Analyses are restricted to persons who have not reported onset at base-
line.5 To estimate incidence, several age-based hazard functions were evaluated (e.g., piecewise exponential, linear, and cur-
vilinear) and a discrete-time approximation of a Weibull distribution, where age is a log-linear function of the log odds, was
selected based on model fit. Age-specific exposure to disease incidence begins at the age-at-baseline and increases until the
age at onset or censoring. Next I use a discrete-time approximation of a Gompertz distribution (log odds of time is linear) to
estimate survival differences for persons with a diagnosed condition. Mortality analyses are based on the time since onset
and model estimates account for differential exposure by including the age of onset (logged) as a product in the log-link
function. For model comparisons, I report BIC statistics for overall fit (Raftery, 1996)6 and use fully standardized coefficients
for education [xy* standardized b(educ)] and the percentage change in odds for a standard deviation (SD) increase in education
[D exp(b*SD of X)] (see Long and Freese, 2006) to assess mediating effects. The standardized coefficients for education and BIC
statistics are presented in the last three rows of the tables for disease incidence and survival duration, along with the unstan-
dardized coefficients for all independent variables.

4
This approach is not unlike many studies that use a count of chronic conditions or physical symptoms to measure health status (Greenfield et al., 1988;
House et al., 1994; Ross and Wu, 1996). Adding the individual items into a single measure necessarily discards some detail and assumes equal weight for each
risk. Nevertheless, the relative contribution of each risk factor can be ascertained from the models that include the mediators separately. However, estimates of
the individual variables only summarize an overall effect for the population and they do not indicate whether a person reports one or more risks. Simply adding
the coefficients to determine the effects of multiple disadvantages for individuals can be misleading (ecological fallacy). Therefore, the count variable is used to
further indicate level of risk; for example, a person who reports heavy drinking (or smoking) and is also obese.
5
Supplementary analyses were performed on the entire sample, which included all incident cases for each respondent with exposure beginning at age 25
(retrospective reports). The results showed that the effects of education on disease incidence and mortality were consistent with those from the sample-
restricted models, although the effects were slightly weaker in the reduced sample.
6
Both Bayesian information criterion (BIC) and Akaike Information Criterion (Akaike, 1973) were used to indicate model fit and to compare nested and non-
nested mediating models. The two methods were generally in agreement and only BIC values are reported to avoid redundancy and because BIC penalizes the
number of parameters more stringently than AIC. Smaller BIC values indicate better model fit and the difference in values from two models indicates which
model is preferred. Raftery, 1996) suggests that the strength of evidence when comparing two models can be ascertained by the value of the absolute
difference, where: 0–2 is weak, 2–6 is positive, 6–10 is strong, and 10+ is very strong.

Please cite this article in press as: Dupre, M.E., Educational differences in health risks and illness over the life course: ...,
Social Sci. Res. (2008), doi:10.1016/j.ssresearch.2008.05.007
ARTICLE IN PRESS
6 M.E. Dupre / Social Science Research xxx (2008) xxx–xxx

Preliminary analyses included gender, race, and age interactions with education and the risk variables for all models.
Nearly all of the interactions were not significant and are excluded from the final analyses; however, the significant inter-
actions are discussed.

5. Results

Table 2 presents the results from the series of logistic regression models predicting the effects of education on the separate
health-risks. All models adjust for race, gender, household size, geographic region, rural residency, comorbidity, and birth co-
hort. The upper panel of the table focuses on behavioral risks and shows that education level significantly decreases the odds of
smoking and having an elevated BMI by about .93 times [=exp( .07)] for every year of schooling. Education shows no direct effect
on heavy drinking. However, as expected, education is strongly associated with all three behavioral risks when considering the
modifying effects of increasing age. For persons aged 35, the odds of smoking is 1.15 [=exp(ageb *35)+(educb*8)+(age*educb*35*8)+(bx*)] for
those with an elementary education and .33 [=exp(ageb*35)+(educb*16)+(age*educb*35*16)+(bx*)] for those with a college education. By
age 65, the difference in the odds of smoking declines substantially between elementary and college-educated individuals (odds
ratios (ORs) = .03 and .02, respectively). Similar age patterns occur for heavy drinking and unhealthy BMI by educational
attainment.
The lower portion of Table 2 presents the estimates for three social and economic risks and shows that education
is negatively associated with low income, unemployment, and nonmarried status. Comparisons of unstandardized
and standardized (not shown) coefficients indicate that the magnitude of the associations is much greater for the

Table 1
Descriptive statistics for the NHEFS study variables at baseline

Variable Total (n = 14,407) Hypertension analysis (n = 10, 578) Heart attack analysis (n = 13,634)
Mean (SD) Mean (SD) Mean (SD)
Age 48.9 (15.6) 46.3 (15.3) 48.1 (15.5)
Years of education 10.8 (3.6) 11.1 (3.4) 10.8 (3.6)
Non-White race .16 .14 .16
Male .40 .41 .39
Household size 3.3 (1.9) 3.4 (1.9) 3.4 (1.9)
Geographic region
Northeast .22 .23 .23
South .24 .25 .24
Midwest .27 .25 .26
West .28 .27 .27
Rural residence .36 .35 .35
1930–1940s birth cohort .45 .52 .47
1910–1920s birth cohort .28 .27 .28
1890–1900s birth cohort .27 .21 .25
Smoking status
Ever smoked .50 .53 .50
Current smoker .39 .42 .40
Heavy drinker (3+ drinks/day) .05 .05 .05
Body mass index
Under-weight .04 .04 .04
Over-weight .33 .31 .32
Obese .17 .12 .16
Income quartile
25th percentile .22 .18 .21
50th percentile .36 .36 .36
75th percentile .21 .23 .21
100th percentile .20 .22 .21
Employment status
Not employeda .52 .48 .51
Retired .10 .09 .09
Nonmarried .25 .23 .25
No. of behavioral risks .94 (.72) .91 (.73) .94 (.72)
No. of social and economic risks 1.3 (.95) 1.2 (.91) 1.2 (.95)
Disease incidenceb
Hypertension .46 .46 —
Heart attack .14 — .14
Comorbidity (2+ conditions) .07 .01 .04

Note: Hypertension analysis n = 166,974 person years and heart attack analysis n = 239,815 person years.
a
Includes retired and those keeping house.
b
Summarizes incident cases across all waves.

Please cite this article in press as: Dupre, M.E., Educational differences in health risks and illness over the life course: ...,
Social Sci. Res. (2008), doi:10.1016/j.ssresearch.2008.05.007
ARTICLE IN PRESS
M.E. Dupre / Social Science Research xxx (2008) xxx–xxx 7

Table 2
Logistic regression coefficients for the associations between education and health risks across age

Behavioral risks
Smoking Heavy drinking Body mass (overweight/obese)
Education 0.070*** 0.292*** 0.001 0.254*** 0.066*** 0.239***
Age (time-varying) 0.113*** 0.159*** 0.087*** 0.140*** 0.055*** 0.089***
Education * age — 0.004*** — 0.004*** — 0.003***
Social and economic risks
Lowest income quartile Unemployed Nonmarried
Education 0.197*** 0.355*** 0.133*** 0.332*** 0.088*** 0.181***
Age (time-varying) 0.074*** 0.102*** 0.106*** 0.151*** 0.062*** 0.079***
Education * age — 0.003*** — 0.004*** — 0.001***

Models adjust for race, gender, household size, geographic region, rural residency, comorbidity, and birth cohort. Estimation includes a log-link function
that factors ln (time) to incorporate differential exposure and uses Huber–White sandwich estimators to account for intra-individual correlation in the
panel data. *p 6 .05, **p 6 .01, ***p 6 .001 (two-tailed tests).

socio-economic risks compared to the behavioral factors. Again, the relationships between education and the risks
diminish as age increases. For the lowest income group, I find that the difference in odds narrows significantly between
persons with an elementary and college education from ages 35 to 65 (ORs from .27 to .02 and from .03 to .01, respec-
tively). The educational gaps in unemployment and nonmarried status decrease similarly across the ages of the NHEFS
participants. These findings concur with the cumulative disadvantage argument that excess risks among the low edu-
cated gradually approach the risk levels of their more educated counterparts as a consequence of mortality selection
across age.
Fig. 1 illustrates the results from the Poisson estimates of the age-related associations between education and accumu-
lated risk factors using the same covariates and model adjustments as Table 1. The figure shows large educational differences
in the cumulative number of behavioral and socio-economic risks at most ages. Consistent with the binary models, I also find
that the educational-risk gaps decrease exponentially with advancing age (education, age, and education–age interactions,

Behavioral Risks
1.2
elementary
college
0.8
Exp (b)

0.4

0
25 35 45 55 65 75 85
Age

Social and Economic Risks


1.2
elementary
college
0.8
Exp (b)

0.4

0
25 35 45 55 65 75 85
Age

Fig. 1. Poisson regression estimates of the age-related association between education and accumulated behavioral and social and economic risk factors.
Note: Models adjust for race, gender, household size, geographic region, rural residency, comorbidity, and birth cohort. Poisson estimation includes a log-
link function that factors in time to incorporate differential exposure and uses Huber–White sandwich estimators to account for intra-individual correlation
in the panel data. Results are significant at the .001 level.

Please cite this article in press as: Dupre, M.E., Educational differences in health risks and illness over the life course: ...,
Social Sci. Res. (2008), doi:10.1016/j.ssresearch.2008.05.007
ARTICLE IN PRESS
8 M.E. Dupre / Social Science Research xxx (2008) xxx–xxx

p 6 .001). In terms of level of risk, the predicted probability (PP) for college-educated persons exhibiting zero behavioral risks
at age 35 is .74 compared to .56 for persons with 8 years of education. The probability that persons with a post-secondary
level of education exhibit two behavioral risks is .22 at the same age, whereas the probability increases by 45% for those with
an elementary level of education (PP = .32). By age 65, these differences are nearly eliminated. The difference in accumulated
social and economic risks is comparable with a slightly larger initial educational disparity (PPs = .83 and .61) and a less pre-
cipitous decline across age. Interestingly, the predicted differentials in the Poisson incident-rate ratios [=exp(bx)] by age indi-
cate that the educational advantage in cumulative risks reverses by age 80 for behavioral risks and age 100 for social and
economic risks. For the separate risks, education remains negatively related to each risk until an apparent reversal by
approximately age 60 for heavy drinking, age 80 for smoking and high BMI, age 90 for unemployment, and over age 100
for low income and nonmarried status.

5.1. Incidence

I turn next to education’s association with hypertension and heart attack incidence using discrete-time logit models. The
mediating effects of the individual and accumulated risks are tested in a series of models with the covariate adjustments de-
scribed above. The findings for hypertension incidence are reported in the left panel of Table 3. The first model establishes a sig-
nificant education gradient in incidence showing that low educational attainment accelerates the age-related onset of
hypertension (OR = .98), as predicted by cumulative disadvantage theory. Model 2 tests the behavioral mechanisms by includ-
ing time-varying variables for smoking, excess alcohol consumption, and unhealthy BMI. The odds of hypertension onset are
significantly greater (p 6 .001) for current smokers (OR = 1.19), heavy drinkers (OR = 1.27), and for persons who are overweight
or obese (OR = 2.08). I also find that education becomes non-significant once behavioral differences are taken into account.
Model 3 includes the second group of explanatory variables and there are no significant effects related to low income,
unemployment, and nonmarried status. In contrast to the separate behavioral risks, the inclusion of social and economic
risks yields little change in the education coefficient and significance level. Model 4 includes both the behavioral and so-
cio-economic indicators and shows minimal change in the coefficients—with the exception of low income, which becomes
marginally significant (b = .097, p 6 .05).
Model 5 includes the cumulative risks measures to test whether multiple disadvantages underlie the educational dif-
ferences in incidence. Results show that the rate of onset increases by nearly 70% [=100  exp(.51) 1] for every addi-
tional negative behavior; however there is no evidence that the number of social and economic risks is directly
related to hypertension incidence or indirectly through education. In terms of overall model fit, the BIC values indicate
that Model 2 (behavioral mediators) is the preferred model for hypertension incidence over the social and economic
mediators (Model 3), a combination of both behavioral and socio-economic risks (Model 4), and the cumulative risks
(Model 5). However, the cumulative-risk models exhibit the strongest mediating factors indicated by the lowest stan-
dardized coefficient for education ( .017) and the smallest percent change in odds for a SD increase in education
( 3.1). Thus, I find some evidence that accumulated disadvantages exert greater effects on hypertension vis-a-vis edu-
cation than the independent effects of health risks.
The results for heart attack are presented in the right panel of Table 3. Education decreases the odds of a heart attack by about
3% [=100  exp( .03) 1] for every added year of schooling. Consistent with hypertension, this finding demonstrates that low
educational attainment accelerates the age-related onset of heart attacks in accordance with cumulative disadvantage theory.
The behavioral measures are included in Model 2, and though highly significant for smoking and BMI, they have little influence
on the effect of education. The social and economic measures are included in Model 3 and partially reduce the education coef-
ficient from .031 (p 6 .001) in Model 1 to .023 (p 6 .01). The odds of a heart attack are much greater for the unemployed
(OR = 1.43) and unmarried (OR = 1.17), but not significantly different for persons categorized in the lowest 25th income
quartile.
Model 4 includes each of the separate risks and shows minor changes in their direct effects (compared to Models 2 and 3),
but collectively, they produce a greater reduction in the size and significance of the education effect (b = .020, p 6 .05). The
results from Model 5 suggest that accumulated risks are robust predictors of differences in heart attack rates. For every
added behavioral risk the odds of a heart attack increase by 31%, and by 20% for every added social and economic disadvan-
tage. Comparisons of BIC values concur that Model 5 is the best-fitting model, although there is modest evidence that the
individual-risk model (Model 4) is more salient in mediating education’s effect on heart attack incidence than cumulative
risks. The independent-risk models exhibit the lowest standardized coefficient for education ( .033) and the smallest change
in odds for a SD increase in education ( 6.8).
Race interactions were tested in the onset models and yielded no significant findings. Gender interactions showed
that unmarried women and those with multiple socio-economic risks had slightly greater risks of a heart attack and
hypertension than men, respectively. Additional analyses on diabetes and stroke incidence yielded results similar to
those of hypertension and heart attack. A negative educational gradient in diabetes onset indicated that low educated
persons exhibited onset at younger ages on average than the well educated and that diabetes was best explained by
behavioral risks (overweight and obesity) and education was mediated most by their cumulative effect. I also found that
educational attainment delayed the onset of a stroke and that social and economic factors best captured disease differ-
entials (employment and marital status) and that the corresponding accumulated risks largely accounted for the educa-
tional differences in stroke.

Please cite this article in press as: Dupre, M.E., Educational differences in health risks and illness over the life course: ...,
Social Sci. Res. (2008), doi:10.1016/j.ssresearch.2008.05.007
Social Sci. Res. (2008), doi:10.1016/j.ssresearch.2008.05.007
Please cite this article in press as: Dupre, M.E., Educational differences in health risks and illness over the life course: ...,

Table 3
Discrete-time parameter estimates for the effects of education and health risks on hypertension and heart attack incidence

Variable Hypertension incidence Heart attack incidence


a b a,b a,b
Model 1 Model 2 Model 3 Model 4 Model 5 Model 1 Model 2a Model 3b Model 4a,b Model 5a,b
** ** *** *** ** *
Education 0.019 0.011 0.017 0.010 0.009 0.031 0.027 0.023 0.020 0.021*

M.E. Dupre / Social Science Research xxx (2008) xxx–xxx


Behavioral risks
Current smoker 0.172** 0.169** 0.400*** 0.386***

ARTICLE IN PRESS
Heavy drinker 0.239** 0.237** 0.051 0.057
Overweight/obese 0.732*** 0.734*** 0.275*** 0.285***
Social and economic risks
Income (25th quartile) 0.023 0.001 0.109 0.087
Not employed 0.077 0.097* 0.360*** 0.351***
Nonmarried 0.003 0.004 0.160* 0.153*
No. of behavioral risks 0.509*** 0.267***
No. of social and economic risks 0.025 0.180***
Constant 8.339*** 8.758*** 8.536*** 9.01*** 9.406*** 17.179*** 19.120*** 18.190*** 20.094*** 19.578***
Log likelihood 14,215.52 14,024.48 14,210.54 14,021.63 14,064.07 7494.25 7452.33 7481.05 7440.30 7446.87
xy* standardized b(educ) 0.034 0.020 0.032 0.018 0.017 0.052 0.045 0.039 0.033 0.034
Dexp(b*SD of X) (%) 6.1 3.7 5.6 3.3 3.1 10.2 9.1 7.8 6.8 7.0
BIC 28,557 28,241 28,601 28,284 28,321 15,125 15,103 15,148 15,128 15,092

Note: All models adjust for log-linear age, race, gender, household size, geographic region, rural residency, comorbidity, and birth cohort. Huber–White sandwich estimators are used to account for intra-individual
correlation in the panel data.
*
p 6 .05, ** p 6 .01; *** p 6 .001 (two-tailed tests).
a
Model also adjusts for ever smoking and low BMI (<18.5).
b
Model also adjusts for retirement status.

9
ARTICLE IN PRESS
10 M.E. Dupre / Social Science Research xxx (2008) xxx–xxx

5.2. Survival duration

Table 4 presents the results from discrete-time models estimating the effect of education on mortality following hyper-
tension and heart attack incidence. The findings for hypertension survival in Model 1 show that educational attainment is
significantly related to lower rates of mortality after the diagnosis of hypertension (OR = .94, p 6 .001). Similar to disease
incidence, cumulative disadvantage is evidenced by the shorter duration of survival after onset for the poorly educated com-
pared to their well educated counterparts. Model 2 shows that the odds of dying from hypertension are not related to heavy
drinking or smoking; however, the latter effect is suppressed by the excess risk of persons who quit smoking (OR = 1.44, not
shown). The counterintuitive finding that the odds of dying are significantly lower for persons who are obese (OR = .60) is
somewhat expected given the known obesity paradox in hypertension survival.
Model 3 includes the second group of explanatory variables and shows persons who are not married or unemployed are
significantly more likely to die (OR = 1.63 and 2.43, respectively). The size of the education coefficients decrease from .058
(Model 1) to .040, rather than increase with the inclusion of behavioral risks ( .061 in Model 2). Model 4 adds both sets of
risk and each parameter estimate slightly weakens and the education effect remains strong despite the inclusion of both
groups of mediators. Model 5 includes the cumulative risk measures and shows that the rate of mortality increases by over
47% for every additional social and economic disadvantage. The 19% decline related to accumulated behavioral-risks is lar-
gely attributed to the resilience of hypertensive obese individuals. Considering this (and Model 2), there is no evidence that
behavioral risks reduce survival following hypertension onset.
Unlike the incidence findings, BIC statistics favor the individual social and economic risks as the best-fitting model for
hypertension survival. However, the cumulative-risk indicator is the strongest mediating factor that is shown by the lowest
standardized coefficient for education ( .056) and the smallest percent change in odds for a SD increase in education
( 11.2). Indeed, BIC values favor Model 5 when excluding the poor performance of the cumulative behavioral risks that
are comprised of a counteractive factor (BMI). Taken together, the number of social and economic disadvantages exerts a
strong direct effect on mortality and mediates the relationship between education and survival more than the independent
effects of health risks.
The results for heart attack survival in Model 1 (right panel) show that education decreases the risk of dying after a heart
attack by about 5% for every added year of schooling. Low educated individuals are not only more likely to have a heart at-
tack at younger ages but also die sooner thereafter. The behavioral measures included in Model 2 do not mediate the rela-
tionship between education and heart attack survival, although the effects for obesity are marginally significant. The
inclusion of social and economic indicators in Model 3 are similarly muted with a moderate increase in dying following a
heart attack for only the unemployed (OR = 1.49, p 6 .01). Including both sets of explanatory variables in Model 4 produces
little change in effect sizes of the covariates and is the poorest fitting model. Finally, unlike hypertension survival, the accu-
mulated number of disadvantages does not affect mortality nor explain educational differences in surviving after a heart at-
tack (Model 5).
In terms of overall model fit, the BIC values indicate that the initial model with covariates (Model 1) is the preferred mod-
el for heart attack incidence over all other parameterizations. Interestingly, the cumulative-risk models are considered the
strongest mediating factors because of the lowest standardized coefficient for education ( .084) and the smallest percent
change in odds for a SD increase in education ( 15.7). However, I am cautious drawing conclusions from these results until
more evidence is permits.
Race interactions indicated that unemployed Whites and those with multiple behavioral and socio-economic risks were
slightly more likely to die following a heart attack than Black individuals. Men with multiple behavioral risks and those who
were obese were slightly more likely to die sooner after the diagnosis of hypertension and a heart attack than women,
respectively. The results for diabetes and stroke survival were mixed. For diabetes, the low number of incident cases
(12%)—coupled with low case-fatality—yielded few significant effects that could be interpreted with confidence. However,
further analyses with retrospective data confirmed that educational attainment significantly prolonged survival following
diabetes onset when considering incidence prior to baseline. For stroke, education was negatively associated with mortality
(p 6 .05) and the effects were best explained by the accumulation of social and economic risk factors.

6. Discussion

This research examines the age-varying relationships between education, health risks, and chronic disease and provides
new evidence to support cumulative disadvantage theory. To my knowledge, this study is the first to demonstrate increasing
educational disparities in health while testing the behavioral, social, and economic mediators thought to account for the
association. Results based on 20 years of panel data show that education is related to both the individual and accumulated
number of health risks, which in turn, influence educational differences in the rates of onset and/or survival for hypertension,
heart attack, diabetes, and stroke.
Consistent with a well-established body of literature I find that education has a strong negative relationship with smok-
ing, unhealthy BMI, low income, unemployment status, and nonmarried status. In addition, I find that education is negatively
related to the cumulative number of risky behaviors and social and economic disadvantages. The strength of the association
appears to be stronger for the social and economic health-risks compared to the behavioral risks and there is evidence that

Please cite this article in press as: Dupre, M.E., Educational differences in health risks and illness over the life course: ...,
Social Sci. Res. (2008), doi:10.1016/j.ssresearch.2008.05.007
Social Sci. Res. (2008), doi:10.1016/j.ssresearch.2008.05.007
Please cite this article in press as: Dupre, M.E., Educational differences in health risks and illness over the life course: ...,

Table 4
Discrete-time parameter estimates for the effects of education and health risks on mortality by disease

Variable Hypertension mortality Heart attack mortality


a b a,b a,b
Model 1 Model 2 Model 3 Model 4 Model 5 Model 1 Model 2a Model 3b Model 4a,b Model 5a,b
*** *** ** ** * *** *** *** ***
Education 0.058 0.061 0.040 0.043 0.037 0.054 0.056 0.052 0.054 0.048***

M.E. Dupre / Social Science Research xxx (2008) xxx–xxx


Behavioral risks
Current smoker 0.013 0.021 0.087 0.131

ARTICLE IN PRESS
Heavy drinker 0.223 0.182 0.184 0.172*
Obese 0.517*** 0.478*** 0.328* 0.314
Social and economic risks
Income (25th quartile) 0.164 0.154 0.158 0.156
Not employed 0.887*** 0.852*** 0.402** 0.405**
Nonmarried 0.490*** 0.471*** 0.146 0.139
No. of behavioral risks 0.206* 0.071
No. of social and economic risks 0.386*** 0.116
*** *** *** ***
Constant 9.386 9.429 10.298 10.308 10.260*** 5.963 ***
5.894 ***
6.290 ***
6.222 ***
6.233***
Log likelihood 2241.90 2224.44 2207.01 2192.28 2213.29 1962.55 1956.54 1954.29 1948.46 1958.60
xy* standardized b(educ) 0.090 0.093 0.060 0.064 0.056 0.093 0.098 0.090 0.094 0.084
Dexp(b*SD of X) (%) 17.1 17.7 12.0 12.8 11.2 17.2 18.0 16.7 17.5 15.7
BIC 4596 4612 4567 4588 4590 4022 4054 4041 4073 4058

Note: All models adjust for age of onset, race, gender, household size, geographic region, rural residency, comorbidity, and birth cohort. Huber–White sandwich estimators are used to account for intra-individual
correlation in the panel data.
*
p 6 .05, ** p 6 .01; *** p 6 .001 (two-tailed tests).
a
Model also adjusts for ever smoking and low BMI (<18.5).
b
Model also adjusts for retirement status.

11
ARTICLE IN PRESS
12 M.E. Dupre / Social Science Research xxx (2008) xxx–xxx

the educational differences in individual and accumulated risks diminish across age. These findings suggest that low edu-
cated individuals are not only more likely to exhibit one of many possible health risks, but they are also more likely to accu-
mulate multiple risks. Moreover, the reduction in risks across age is congruent with the cumulative disadvantage tenet that
accelerated onset and mortality disproportionately removes high-risk individuals from the population.
The theory suggests that educational differences in health increase with age because of differences in risk and their accu-
mulation. I find strong evidence that education’s effect on chronic disease increases with age, and unlike many existing stud-
ies, I find no evidence that these effects decline at older ages. The results indicate that education is positively associated with
the age of onset and/or length of survival for hypertension, heart attack, diabetes, and stroke. Support for the cumulative
disadvantage hypothesis is manifested as an accelerated rate of health decline for the poorly educated compared to the well
educated. In results not presented here, I also calculated the prevalence of the four conditions from the onset and duration
data (prevalence = incidence  duration) and examined the effects of education across age. As anticipated, educational dis-
parities in disease prevalences declined in late adulthood (age-as-leveler), presumably because of differences in the rates of
aging and selective mortality. This paradoxical finding should alert researchers to be cautious when modeling and interpret-
ing health differentials at the individual and aggregate level.
The effects of education on incidence and survival are similar across illnesses, yet the explanation for their age-based pat-
terning differs. For hypertension and diabetes, education’s effect on disease onset is better explained by behavioral risks than
the social and economic factors. Alternatively, it is the social and economic risks that explain a significant amount of the edu-
cational difference in hypertension survival (diabetes survival-models lacked statistical power). Moreover, the accumulation
of behavioral risks better explained educational differences in hypertension and diabetes onset, whereas educational differ-
ences in hypertension survival are best explained by the cumulative number of social and economic risks. For heart attack
and stroke, the individual and accumulated social and economic risks exert a greater mediating effect on the relationship
between education and incidence, respectively, but only the cumulative socio-economic risks could account for education’s
effect on surviving after a stroke.
These findings have potential policy implications. For example, the behavioral factors appear to have a greater influence
on hypertension and diabetes onset than surviving after the diagnosis. Each of the behavioral risks—smoking, heavy drinking,
and high BMI—is associated with onset and their accumulation mediates the effect of education. The physical health of indi-
viduals is generally better at younger ages, and because of this, individuals with less education may underestimate the ben-
efits of health maintenance and overlook their own mortality. Therefore, it is reasonable to expect that engaging in risky
behaviors has a greater impact on disease onset rather than living with the illness. Among persons who are hypertensive,
it appears that social and economic factors play a greater role in survival than behavioral factors. After disease recognition,
individuals may be more inclined to moderate their lifestyle when faced with a crucial turning point in health. Alternatively,
social and economic factors become more important for surviving with hypertension because they determine access to qual-
ity care, support, and enable individuals to purchase costly treatments. In addition, it appears that the accumulated number
of social and economic risks is largely responsible for the educational differences in survival.
In contrast to hypertension and diabetes onset, social and economic risks seem to have a greater effect on heart attack and
stroke incidence than behavioral risks. Two of the social and economic risks—unemployment and unmarried status—are re-
lated to onset, and together, they mediate a large portion of the education effect. In addition, the larger the number of dis-
advantages, the greater its consequences for having a stroke. The reasons vary for why social and economic factors may be
more important than behaviors for predicting heart attack or stroke. For example, in some cases a heart attack or stroke oc-
curs among people who have been diagnosed with cardiovascular disease and were taking measures to treat their illness.
Unlike behavioral risks, which are often modifiable, social and economic risks have deeper roots in the social structure
and will influence health regardless of changes in lifestyle. Moreover, persons who are less educated are less likely to possess
the social and economic resources needed to identify symptoms or receive medical check-ups to diagnose heart disease (or a
related illness), and consequently, are more likely experience a heart attack or stroke at younger ages than the well educated.
A somewhat unexpected finding is that income has no effect on the onset or duration of illness for any of the four con-
ditions. It is possible that the crude measurement of income in the NHEFS plays a role in its non-effect. However, it is impor-
tant to note that the majority of existing evidence is based on income’s relationship with overall mortality and not survival
following the onset of a particular disease, which is far less documented. Instead, the results here suggest that being unem-
ployed or unmarried are the primary detriments to survival. It may be because these individuals lack access to health care or
a supportive care-giving spouse. The implication of this finding is that financial resources may not be as important to sur-
viving with a chronic illness compared to the structural factors that determine the availability to basic care and support. Cer-
tainly, more research is needed to evaluate this claim.
The finding that obesity prolongs survival for persons with hypertension is counterintuitive. Several reasons may account
for this anomaly. First, the NHEFS likely introduces some degree of selective survivorship due to sampling, which I address
later as a study limitation. Second, to put the findings in context, the results from Table 3 indicated that overweight and ob-
ese individuals were over twice as likely to develop hypertension at younger ages as persons with lower BMI values. In other
words, the mortality analyses were largely comprised of overweight and obese persons (60%) and the obese were diagnosed
with hypertension 5 years earlier than persons with a lower BMI. Third, in life-course terms, the diagnosis of a serious illness
may have precipitated a turning point in health marked by improved diet, exercise, and lower cholesterol, which may have
gone undetected with modest changes in overall weight. Lastly, the findings support recent medical studies that find an

Please cite this article in press as: Dupre, M.E., Educational differences in health risks and illness over the life course: ...,
Social Sci. Res. (2008), doi:10.1016/j.ssresearch.2008.05.007
ARTICLE IN PRESS
M.E. Dupre / Social Science Research xxx (2008) xxx–xxx 13

obesity paradox among the mildly obese (and overweight) with hypertension who exhibit lower mortality compared to
those with a normal BMI (Romero-Corral et al., 2006).
The results of this study are not without limitations. First, I consider only a limited number of illnesses. It is not clear
whether the cumulative disadvantage hypothesis would be supported when focusing on cancer, lung or kidney diseases, Alz-
heimer’s disease, or even infectious diseases. I suspect that if an association exists between education and any of these health
problems, as studies often find, it is plausible that the effects of education would increase with age—evidenced with earlier
disease onset and shorter survival duration for the poorly educated. Future research utilizing different health outcomes is
clearly warranted.
Second, although the NHEFS data used for this study have numerous advantages, they also have several shortcomings. An
important contribution of this analysis is that it focuses on disease onset and survival to minimize the effects of mortality
selection on observed health. However, the data used here are not unaffected. The baseline sample includes adults age 25–74
and inherently includes persons considered robust survivors. By restricting the analyses to persons experiencing disease on-
set following the baseline interview, heterogeneity in the sample is further reduced. The consequences of this bias are not
entirely known and difficult to eliminate. Presumably, the results are a conservative under-estimation of education’s asso-
ciation with the health risks, disease onset, and survival duration. However, additional studies need to assess the current
findings in other national data sets and with other health conditions.
Third, I consider only education and a limited number of explanatory variables (see DiPrete and Eirich, 2006). Other dee-
ply-rooted SES indicators (e.g., parental characteristics, childhood adversity, and occupational history) are not available in
the data and certainly warrant more investigation. Likewise, research should consider psychosocial variables and other po-
tential mechanisms that were not included in the analysis, such as social support, psychological characteristics, health care
coverage, physical activity, mobility, and occupational type. A related avenue of research also should advance the longitudi-
nal conceptualization of the mediating variables linking education to health. This research attempted to extend our knowl-
edge by examining education’s mechanisms separately and with measures that incorporate their accumulation and change
over time. Unfortunately, the gap between NHEFS survey waves is fairly lengthy (5–10 years) and may oversimplify the tem-
poral variations (and effects) of the behavioral, social, and economic risk factors. However, existing longitudinal data sets
with smaller intervals between panels—though more suitable for modeling changes in mediators—are limited in analyzing
age variations in health.
The results of this research are among a small but growing body of evidence that supports a cumulative disadvantage in
health. Continued efforts to study and ultimately reduce health disparities in populations is a tall order and one that can only
be achieved when properly understood. This study provides new clues as to how educational attainment relates to health
risks and why certain subgroups in the population experience accelerated health decline. We also should recognize the value
of studying individuals who defy their aggregated likelihood of following a ‘‘predicted” life-course trajectory (outliers). It is
not enough to know the process by which disadvantaged persons exhibit poor health outcomes, but also, why some individ-
uals prosper despite facing cumulative disadvantages across the life course.

References

Aday, L.A., Andersen, R.M., 1984. The national profile to access to medical care: where do we stand? American Journal of Public Health 74, 1331–1339.
Akaike, H., 1973. Information theory and an extension of the maximum likelihood principle. In: Petrov, B.N., Csaki, F. (Eds.), The Second International
Symposium on Information Theory. Akad’emiai Kiado, Budapest.
Antonovsky, A., 1967. Social class, life expectancy and overall mortality. Milbank Memorial Fund Quarterly 45, 31–73.
Beckett, M., 2000. Converging health inequalities in later life—an artifact of mortality selection? Journal of Health and Social Behavior 41, 106–119.
Burchinal, M.R., Roberts, J.E., Hooper, S., Zeisel, S.A., 2000. Cumulative risk and early cognitive development: a comparison of statistical risk models.
Developmental Psychology 36 (6), 793–807.
Cohen, B.B., Barbano, H.E., Cox, C.S. et al. 1987. Plan and Operation of the NHANES I Epidemiologic Follow-up Study, 1982-84. National Center for Health
Statistics. Vital and Health Statistics, Report Series 1, No. 22.
Cox, C.S., Mussolino, M.E., Rothwell, S.T. et al. 1997. Plan and Operation of the NHANES I Epidemiologic Followup Study, 1992. National Center for Health
Statistics. Vital Health Statistics, Report Series 1, No. 35.
Crimmins, E.M., Saito, Y., 2001. Trends in healthy life expectancy in the united states, 1970–1990: gender, racial, and educational differences. Social Science
and Medicine 52 (11), 1629–1641.
Crystal, S., Shea, D., 1990. Cumulative advantage, cumulative disadvantage, and inequality among elderly people. The Gerontologist 10, 437–443.
DiPrete, T., Eirich, G., 2006. Cumulative advantage as a mechanism for inequality: a review of theory and evidence. Annual Review of Sociology 32, 271–297.
Dupre, M.E., 2007. Educational differences in age-related patterns of disease reconsidering cumulative disadvantage and age-as-leveler. Journal of Health
and Social Behavior 48 (1), 1–15.
Eckenrode, J., 1983. The mobilization of social supports: some individual constraints. American Journal of Community Psychology 11, 509–528.
Elman, C.E., O’Rand, A.M., 2004. The race is to the swift: socioeconomic origins, adult education, and wage attainment. American Journal of Sociology 110
(1), 123–160.
Feldman, J.J., Makuc, D.M., Kleinman, J.C., Cornoni-Huntley, J., 1989. National trends in educational differentials in mortality. American Journal of
Epidemiology 129, 919–933.
Ferraro, K.F., Kelley-Moore, A. Jessica., 2003. Cumulative disadvantage and health: long-term consequences of obesity? American Sociological Review 68 (5),
707–729.
George, L.K., 2005. Socioeconomic status and health across the life course: progress and prospects. Journals of Gerontology: Series B 60B (Special issue II),
135–139.
Greenfield, S., Aronow, H.U., Elashoff, R.M., Watanabe, D., 1988. Flaws in mortality data. the hazards of ignoring comorbid disease. Journal of the American
Medical Association 260 (15), 2253–2255.
Grusky, D.B., DiPrete, T.A., 1990. Recent trends in the process of stratification. Demography 27, 617–637.
Grzywacz, J.G., Marks, N.F., 2001. Social inequalities and exercise during adulthood: toward an ecological perspective. Journal of Health and Social Behavior
42, 202–220.

Please cite this article in press as: Dupre, M.E., Educational differences in health risks and illness over the life course: ...,
Social Sci. Res. (2008), doi:10.1016/j.ssresearch.2008.05.007
ARTICLE IN PRESS
14 M.E. Dupre / Social Science Research xxx (2008) xxx–xxx

House, J.S., Kessler, R.C., Herzog, A.R., 1990. Age, socioeconomic status, and health. Milbank Memorial Quarterly 68, 383–411.
House, J.S., Lepkowski, J.M., Kinney, A.M., Mero, R.P., Kessler, R.C., Herzog, A.R., 1994. The social stratification of aging and health. Journal of Health and Social
Behavior 35 (September), 13–34.
Jencks, C.S., Bartlett, S., Corcoran, M., Crouse, J., Eaglesfield, D., Jackson, G., McClelland, K., Mueser, P., et al, 1979. Who Gets Ahead? The Determinants of
Economic Success in America. Basic, New York.
Kaplan, G.A., Keil, J.E., 1993. Socioeconomic factors and cardiovascular disease: a review of the literature. Circulation 88 (4), 1973–1998.
Kaplan, G.A., Haan, M.N., Wallace, R.B., 1999. Understanding changing risk factor associations with increasing age in adults. Annual Review of Public Health
20, 89–108.
Kitigawa, E.M., Hauser, P.M., 1973. Differential Mortality in the United States: A Study of Socioeconomic Epidemiology. Harvard University Press,
Cambridge, MA.
Kuh, D., Ben Shlomo, Y. (Eds.), 2004. Life Course Approach to Chronic Disease Epidemiology, second ed. Oxford University Press, Oxford UK.
Lantz, P.M., Lynch, J.W., House, J.S., Lepkowski, J.M., Mero, R.P., Musick, M.A., Williams, D.R., 2001. Socioeconomic disparities in health change in a
longitudinal study of U.S. adults: the role of health-risk behaviors. Social Science and Medicine 53, 29–40.
Lauderdale, D.S., 2001. Education and survival: birth cohort, period, and age effects. Demography 38 (4), 551–561.
Leigh, J.P., 1983. Education, working conditions, and workers’ health. Social Science Journal 20, 99–107.
Lin, N., Ensel, W.M., 1989. Life stress and health: stressors and resources. American Sociological Review 54, 382–399.
Long, J.S., Freese, J., 2006. Regression Models for Categorical Dependent Variables Using Strata, second ed. Stata Press, College Station, Texas.
Luepker, R.V., Rosamond, W.D., Murphy, R., Sprafka, J.M., Folsom, A.R., McGovern, P.G., Blackburn, H., 1993. Socioeconomic-status and coronary heart-
disease risk factor trends: the Minnesota-heart-survey. Circulation 88 (5), 2172–2179.
Lynch, S.M., 2003. Cohort and life-course patterns in the relationship between education and health: a hierarchical approach. Demography 40, 309–331.
Manton, K.G., Stallard, E., Corder, L., 1997. Education-specific estimates of life expectancy and age-specific disability in the U.S. elderly population. Journal of
Aging and Health 9, 419–450.
Merton, R.K., 1968. The Matthew effect in science. Science 159, 56–63.
Millar, W.J., Wigle, D.T., 1986. Socioeconomic disparities in risk factors for cardiovascular disease. Canadian Medical Association Journal 134, 127–132.
Mirowsky, J., Ross, C.E., 1998. Education, personal control, lifestyle and health: a human capital hypothesis. Research on Aging 20 (4), 415–449.
Mirowsky, J., Ross, C., Reynolds, J., 2000. Links between social status and health. In: Bird, C.E., Conrad, P., Fremont, A.C. (Eds.), Handbook of Medical
Sociology, 5th ed. Prentice Hall, Upper saddle river.
Murray, C.J.L., Lopez, A.D., 1996. The Global Burden of Mortality and Disability from Diseases, Injuries, and Risk Factors in 1990 and Projected to 2020.
Harvard University Press, Cambridge, MA.
National Center for Health Statistics, 2001. National Vital Statistics Reports, vol. 49, issue 8. Centers for Disease Control.
O’Rand, A.M., 1996. The precious and the precocious: understanding cumulative disadvantage and cumulative advantage over the life course. The
Gerontologist 36 (2), 230–238.
Pallas, A.M., 2003. Educational transitions, trajectories, and pathways. In: Mortimer, J.T., Shanahan, M.J. (Eds.), Handbook of the Life Course. Kluwer
Academic/Plenum Publishers, New York.
Pampel, F.C., Rogers, R.G., 2004. Socioeconomic status, smoking, and health: a test of competing theories of cumulative advantage. Journal of Health and
Social Behavior 45 (3), 306–321.
Pearlin, L., Schooler, C., 1978. The structure of coping. Journal of Health and Social Behavior 19 (1), 2–21.
Pierce, J.P., Fiore, M.C., Novotny, T.E., Hatziandreu, E.J., Davis, R.M., 1989. Trends in cigarette smoking in the united states: educational differences are
increasing. Journal of the American Medical Association 261 (1), 56–60.
Preston, S.H., Elo, I.T., 1995. Are educational differentials in adult mortality increasing in the United States? Journal of Aging and Health 7, 476–496.
Raftery, A.E., 1996. Bayesian model selection in social research. In: Marsden, P.V. (Ed.), Sociological Methodology, vol. 26. Blackwell, Oxford, pp. 11–163.
Rodin, J., 1986. Aging and health: effects of the sense of control. Science 233, 1271–1276.
Romero-Corral, E., Montori, V.M., Somers, V.K., Korinek, J., Thomas, R.J.T.G., et al, 2006. Association of bodyweight with total mortality and with
cardiovascular events in coronary artery disease: a systematic review of cohort studies. Lancet 368, 666–678.
Roos, E., Lahelma, E., Virtanen, M., Prattala, R., Pietinen, P., 1998. Gender, socioeconomic status and family status as determinants of food behavior. Social
Science and Medicine 46 (12), 1519–1529.
Rosenbaum, J.E., 1984. Career Mobility in a Corporate Hierarchy. Academic Press, New York.
Ross, C.E., Huber, J., 1985. Hardship and depression. Journal of Health and Social Behavior 26, 312–327.
Ross, C.E., Mirowsky, J., 1989. Explaining the social patterns of depression: control and problem solving—or support and talking. Journal of Health and Social
Behavior 30, 206–209.
Ross, C.E., Wu, C., 1995. The link between education and health. American Sociological Review 60, 719–745.
Ross, C.E., Wu, C., 1996. Education, age, and the cumulative advantage in health. Journal of Health and Social Behavior 37 (March), 104–120.
Rowe, J.W., Kahn, R.L., 1987. Human aging: usual and successful. Science 143, 143–149.
Sameroff, A.J., Seifer, R., Baldwin, A., Baldwin, C., 1993. Stability of intelligence from preschool to adolescence: the influence of social and family risk factors.
Child Development 64, 80–97.
Sewell, W.H., Hauser, R.M., 1975. Education Occupation and Earnings. Academic Press, New York.
Turner, R.J., Wheaton, B., Lloyd, D.A., 1995. The epidemiology of social stress. American Sociological Review 60, 104–125.
Wheaton, B., 1983. Stress personal coping resources and psychiatric symptoms: an investigation of interactive models. Journal of Health and Social Behavior
24 (3), 208–229.
Williams, D.R., 1990. Socioeconomic differentials in health: a review and redirection. Social Psychology Quarterly 53 (2), 81–99.

Please cite this article in press as: Dupre, M.E., Educational differences in health risks and illness over the life course: ...,
Social Sci. Res. (2008), doi:10.1016/j.ssresearch.2008.05.007