Unraveling the Relationship between Education and Health: Genetic Controls, Heterogeneity across Sociodemographic Groups, and Variation across Biomarkers of Health Risk

Despite decades of research demonstrating better health among the higher educated, the causal
effect of education on health is still debated. This is due in part to mixed evidence obtained in quasiexperimental work. These puzzling patterns could be explained by the influence of uncontrolled
confounders in observational research, by effect heterogeneity across individuals or environments, or by
variation in effects across manifestations of health. The empirical chapters of this dissertation draw
motivation from these observations to further unravel the relationship between education and health
among older adults in the United States.
First, I assess the utility of a novel control variable: a measure of genetic selection into education.
Genetic selection is operationalized using a polygenic score (PGS) that predicts years of schooling based
on many hundreds of thousands of genetic variants across the genome. Among European-ancestry
respondents to the Health and Retirement Study (HRS) and the Wisconsin Longitudinal Study (WLS), I
find that controlling for the PGS significantly attenuates the association between education and later
health. The level of attenuation I observe is comparable to that obtained when controlling instead for
measures of other known confounders, including family background and childhood health. Additional
results suggest that the education PGS reflects more proximal confounders of the education-health link
that may not be adequately controlled using survey measures alone. Crucially, however, the positive
relationship between education and health is robust to this particular measure of genetic selection into
years of schooling.
Next, I evaluate whether the association of education with health varies across sociodemographic
groups defined by socioeconomic (SES) origin, race, and gender using data from the HRS. In so doing, I
take a more complex intersectional perspective than has been used in prior work. This is important, as
exposure to discrimination, which shapes opportunities to use resources in support of health, may depend
on multiple sociodemographic characteristics simultaneously. Results underscore the importance of one
intersection in particular: that between SES origin and race. In line with prior work, I find that the
association of years of schooling with self-reported health is stronger for those from low-SES
backgrounds; however, this is only the case among whites. Seen from the other angle, the association of
education with self-reported health and mortality is weaker for blacks than for whites, but primarily
among those from low-SES origins. For both self-reported health and mortality, I find the smallest gain in
health per year of schooling among low-SES origin black men, the group with the highest risk of poor
health and mortality overall.
In the final empirical chapter, I use data from the HRS to assess whether educational disparities in
biomarkers of health risk vary across their distributions. Fundamental cause theory implies that such
disparities will be largest where related resources can most successfully be leveraged to improve
outcomes. For many biomarkers, this could be in the unhealthy tail of the distribution, where unequal
access to and efficacy of medical interventions may exacerbate disparities. Consistent with this theory, I
find that educational disparities in blood sugar and blood pressure are largest at their least healthy levels,
precisely the points where impacts on subsequent morbidity and mortality are greatest. Meanwhile, highdensity lipoprotein (HDL) or “good” cholesterol—a biomarker that is not regularly targeted by
medication—does not display such a pattern. These results are not only of theoretical and substantive
interest; they also provide methodological guidance for future work on biomarkers of health risk, which is
timely given the recent proliferation of such measures in social science datasets.