Achieved educational attainment, inherited genetic endowment for education, and obesity

Abstract

This study investigates two sources of education effects on obesity—achieved educational attainment and inherited genetic endowment for education. In doing so, we accomplish two goals. First, we assess the role of genetic confounding in the association between education and health. Second, we consider the heterogeneity in the extent to which genetic potential for education is realized and examine its impact on obesity. Data come from the National Longitudinal Study of Adolescent to Adult Health. Using a polygenic score approach, we find that, net of genetic confounding, holding a college degree is associated with a lower likelihood of obesity. Moreover, we find that among individuals who hold a college degree, those with a high education polygenic score (a greater genetic propensity to succeed in education) are less likely to be obese than those with a relatively low education polygenic score. However, when individuals with a high education polygenic score do not have a college degree, their risk of obesity is similar to that of non-college-educated individuals with a low education polygenic score, suggesting that the effect of genetic endowment for education on obesity is conditional on college education.

Introduction

The association between educational attainment and health has been well-documented (Mirowsky and Ross 2003). A large body of research shows that individuals with higher levels of education report fewer health risk factors (e.g., Baker et al. 2017) and lower incidence of diseases (e.g., Hayward et al. 2000), and have lower mortality (e.g., Hummer and Lariscy 2011). Education is therefore often referred to as one of the “fundamental causes” of health (Link and Phelan 1995).

Obesity is one of the most significant contributors to ill health (Kopelman 2000). In the United States, the prevalence of adult obesity has more than tripled from 12% in 1991 to 42% in 2018 (Mokdad et al. 1999, Hales et al. 2020). This growing epidemic poses a grave threat to population health. Education has been shown to be associated with obesity (Cohen et al. 2013, Newton, Braithwaite, and Akinyemiju 2017). In this article, we distinguish two sources of education effects on obesity—achieved educational attainment and inherited genetic endowment for education. Separating the two distinct, yet related, education effects enables us to achieve two goals.

First, we obtain a more accurate estimate of the direct effect of educational attainment by addressing bias caused by genetic confounding. This education-obesity link can be explained by causal theories that higher educational attainment leads to better health (Cutler and Lleras-Muney 2006, Masters, Hummer, and Powers 2012, Smith 2004). This link, however, can also be explained by the theory that common confounders are responsible for causation as well as reverse causation (i.e., education is the result of health) (Fuchs 1982). Researchers argue that genetic influences are a critical confounder that can bias the estimate of the effects of achieved educational attainment on health (Lundborg, Lyttkens, and Nystedt 2016, Amin, Behrman, and Kohler 2015, Behrman et al. 2011).

Drawing on genetic data from the National Longitudinal Study of Adolescent to Adult Health, we evaluate the impact of genetic confounding with a polygenic score (PGS) approach. Using findings from genome-wide association studies (GWASs) of educational attainment (Lee et al. 2018), we construct the education PGS. The education PGS combines genome-wide information summarized by the GWASs to indicate genetic propensity to succeed in education, a potentially important source of genetic confounding.

Our second goal is to examine the interactive effect between achieved educational attainment and the education PGS on obesity. A rapidly growing body of research provides evidence that health and many other outcomes may depend on gene-environment interaction—environmental influences on an outcome can be contingent on genetic effects and vice versa (e.g., Gaydosh et al. 2018, Perry 2016, Conley et al. 2015, Simons et al. 2011, Mitchell et al. 2015, Boardman et al. 2014, Herd et al. 2019, Belsky et al. 2018, Domingue et al. 2018, Liu 2018, Wedow et al. 2018, Shanahan et al. 2008, Guo et al. 2015). By separating achieved educational attainment and the education PGS, our study brings to light the heterogeneity in educational level attained and genetic potential, while the existing literature typically compares health between different educational levels and ignores this heterogeneity.

1.1. Genetic confounding and the education polygenic score

The importance of identifying genetic confounding for causal inference has long been recognized (McGue, Osler, and Christensen 2010, Pingault et al. 2018). However, genotypes are indirectly observed in past studies. For example, twin studies exploit the fact that monozygotic twins share an identical genome and face similar conditions, such as parents and food (e.g., Lundborg 2013, Böckerman and Maczulskij 2016, Webbink, Martin, and Visscher 2010, Amin et al. 2015), but they do not use direct measures of genotype.

A contribution of this study is that we directly address genetic confounding using molecular genetic data to construct the education PGS. Complex behaviors such as education and health are affected by a very large number of social and biological factors. The effect of a single genetic variant is small (Chabris et al. 2015). Genome-wide association studies (GWASs) identify associations between typically millions of genetic variants and a behavior. PGSs summarize the effects of genetic variants discovered by GWASs to estimate individual genetic propensities (Krapohl et al. 2018). The education PGS captures a genetic predisposition for educational attainment. The education PGS may reflect the genetics that influence the development of traits and behaviors (e.g., cognitive abilities) that lead to different achievement in school (Plomin and von Stumm 2018, Belsky et al. 2018). The education PGS is also predictive of college completion in entirely independent samples (Lee et al. 2018).

The education PGS represents a potential major source of genetic confounding in the context of educational attainment and health outcomes like obesity. Here, a confounder is a variable that affects both education and obesity. Pleiotropy can explain how the education PGS may influence obesity and education. Pleiotropy occurs when a single gene creates effects for multiple phenotypic traits. Pleiotropy is a common feature of the human genome (Sivakumaran et al. 2011). There are two forms of pleiotropy—biological and mediated (Solovieff et al. 2013). Biological pleiotropy occurs when the effects of a gene are due to biological influences. Educational attainment and obesity could be affected by common genetic influences related to traits such as cognitive performance (Lee et al. 2018, Okbay et al. 2016). Mediated pleiotropy occurs when a gene influences a second phenotype by influencing a first phenotype that affects the second phenotype. Genetic correlation between education and obesity could also result from the possibility that educational attainment affects obesity (Mirowsky and Ross 2003).

1.2. Previous gene-environment interaction research

Null or weak average effects of education may conceal large differential effects across social environments. In other words, ignoring potential interaction effects could underestimate the education effects. Gene-environment interaction research is often guided by two models. The diathesis-stress model posits that unfavorable conditions trigger or magnify genetic propensities for risky behaviors, and favorable conditions protect against genetic risk (Belsky et al. 2009). The differential susceptibility model contends that some genotypes are more sensitive to environmental influences; individuals with more sensitive genotypes have more positive outcomes in favorable conditions and in unfavorable conditions they have more negative outcomes (Ellis et al. 2011). Both models suggest interaction effects that individuals with “risky” genes and individuals without “risky” genes act differently, depending on the environment in which they live.

Recently, Barcellos et al. (2018) investigate whether the effect of education on body weight depends on the education PGS and BMI PGS. The authors use a compulsory schooling age reform that raises the minimum school-leaving age from 15 to 16 as an instrumental variable (IV). The authors find no evidence that the education PGS interacts with the IV to affect body weight. They find that for individuals with higher BMI PGSs, staying in school until age 16 leads to a reduction in body weight.

The advantage of using compulsory schooling laws is that it is an exogenous variable, but a limitation of this IV is that its impact usually concerns the effect of one or two additional years of education in secondary school on a small, select group of students (Hamad et al. 2018, Galama, Lleras-Muney, and van Kippersluis 2018). For example, if a reform requires students to stay one more year at 9^th grade, only students who intend to leave at 9^th grade is affected. Students who intend to, for example, finish high school or go to college are not likely to be affected because these students would stay after 9^th grade regardless of the reform. Moreover, the effect of staying one additional year in secondary school for health outcomes many years after graduation might be limited.

In this article, we focus on college education, an arguably more significant achievement than one or two additional years of education in secondary school. College education is viewed as a salient indicator of achieved status over the life course (Hout 2012). Past research has documented the health and social advantages among college-educated individuals (Dupre 2007, Schafer, Wilkinson, and Ferraro 2013, Backlund, Sorlie, and Johnson 1999, Zheng 2017, Montez, Hummer, and Hayward 2012, Schnittker 2004, Hummer and Lariscy 2011), and found that having a college degree is a stronger predictor for body weight than other degrees such as high school diploma (Benson, von Hippel, and Lynch 2018, von Hippel and Lynch 2014, Amin, Behrman, and Kohler 2015, Baum 2017). By controlling for genetic confounding with genetic data, we aim to come closer to the goal of estimating the causal effect of college education on obesity.

Materials and methods

2.1. Data

Data come from the National Longitudinal Study of Adolescent to Adult Health (Add Health). Add Health is a nationally representative survey of U.S. adolescents enrolled in grades 7 through 12 during the 1994 to 1995 school year (Harris et al. 2019). More than 20,000 respondents participated in the in-home phase of the survey at Wave I (1994–95). Respondents were then followed. Wave II, III, and IV data were collected in 1996, 2001–02, and 2008–09, respectively.

Add Health collected saliva samples for genotyping via Oragene saliva collection at Wave IV. A total of 15,072 respondents provided saliva samples and 12,058 provided consent for genotyping. After quality control, 9,975 respondents retained. We use the genotype dataset posted to dbGaP (Study Accession phs001367.v1.p1). After genotyping quality control and excluding respondents with missing data, our final sample consists of 5,105 non-Hispanic whites. We restrict the final sample to non-Hispanic whites because the GWAS results we use focus on individuals of European ancestry (Lee et al. 2018, Locke et al. 2015), and because there are known problems associated with applying PGSs in diverse ancestry groups (Martin et al. 2017). This approach is consistent with other PGS studies that limit analysis to non-Hispanic whites (e.g., Herd et al. 2019).

2.2. Measures

During Wave IV in-home survey, interviewers measured participant’s weight to the nearest 1/2 pound and height to the nearest 1/8 inch. BMI is calculated as measured weight in kilograms divided by measured height in meters squared. A respondent is classified as obese if this respondent’s BMI is 30 or greater (CDC 2017). If a participant held a bachelor’s degree or higher by Wave IV, this participant’s educational level is coded as with a college degree. Otherwise, we code it as without a college degree.

The most recent education GWAS of 1.1 million individuals (Lee et al. 2018) identifies single nucleotide polymorphisms (SNPs) associated with educational attainment. The GWAS estimates an effect, β, for every SNP. A PGS is calculated as the weighted sum of risk alleles (the allele positively associated with educational attainment) of SNPs:

$$\mathrm{P}\mathrm{G}{\mathrm{S}}_i\mathrm{=}{\mathrm{\Sigma }}_{\mathit{j}\mathrm{=}\mathrm{1}}^{\mathit{J}}{\mathit{\beta }}_{\mathit{j}}{\mathit{G}}_{\mathit{i}\mathit{j}}$$

where i denotes individual and j denotes SNP, β is the coefficient for SNP j estimated by the GWAS, and G_ij is the number of risk alleles j for individual i.

To control for genetic confounding, we also construct the BMI PGS, derived from a large GWAS of BMI (Locke et al. 2015). The BMI PGS represents another potential major source of genetic confounding for the education-obesity link. Recall that a confounder in the present case is a variable that influences both obesity and educational attainment. The BMI PGS is associated with some causal genetic variants for body weight (Locke et al. 2015). As a result, the BMI PGS could influence obesity. We discuss above how the education PGS could influence educational attainment and obesity. For similar reasons, the BMI PGS may influence education and obesity through biological effects (e.g., genetic cognitive potential) and mediated effects (i.e., pleiotropic effects).

The education PGS and BMI PGS are constructed with PRSice software (Euesden, Lewis, and O’Reilly 2015), after removing highly correlated SNPs (SNPs having r² > 0.1 within 250k base pairs of the index SNP). The GWAS of education (Lee et al. 2018) includes the Add Health data. We use the effect, β, derived from data that exclude the Add Health data to construct the education PGS. The PGSs are standardized to have a mean of 0 and a standard deviation (SD) of 1. Table 2 shows that a 1 SD higher education PGS increases the odds of holding a college degree by 81%; a 1 SD higher BMI PGS increases the odds of being obese by 46%.

Table 2.

Polygenic scores (PGSs) predicting respective phenotypes, logistic regressions

	Dependent variable	Independent variable	Odds ratio	N
Model 1	College enrollment	Education PGS	1.81***	5,105
Model 2	Obesity	BMI PGS	1 46***	5,105

Note: All models control for age, sex, and the first 10 principal components to address population stratification.

^*p < .05

^**p < .01

^***p < .001 (two-tailed tests).

In our analysis, if an individual’s education PGS is above the 50^th percentile of the PGS distribution, this individual is assigned to the high education PGS group. If the score is at or below the 50^th percentile, the individual is assigned to the low education PGS group. We also try the 40^th and 60^th percentiles as cut-off points. Results are similar. Control variables are reported in Table 1. A sense of self-control is a key benefit provided by education (Mirowsky and Ross 2003). Self-control at earlier stages of life is important for health (Moffitt et al. 2011). To measure self-control, we select items similar to previous research (Perrone et al. 2004). Physical activity and the inactivity score are constructed following previous research (Gordon-Larsen, McMurray, and Popkin 2000, Graff et al. 2016).

Table 1.

Sample characteristics

Variable	Description and coding	Percentage/ mean (S.D.)
Obesity	BMI > 30, Wave IV	34.12%
College degree	Holding a bachelor’s degree or higher by Wave IV	32.43%
Age	Age, Wave IV	28.37
		(1.75)
Sex	Female = 1, male = 0	53.42%
Parental education	Highest degree attained by parents, Wave I
	More than high school	60.80%
	High school	31.26%
	Below high school	7.93%
Biological parents	Lived with biological father and mother, Wave I	58.77%
PVT score	Verbal IQ score, Wave I	105.09 (11.82)
Self-control	Sum of responses to the following questions, Wave I. Higher scores indícate	8.25
	higher levels of self-control. Score ranges from 1 to 19. 1. had trouble getting along with teachers; 2. had trouble paying attention in school; 3. had trouble getting homework done; 4. had trouble keeping your mind on what you were doing; and 5. felt like doing everything just about right.	(2.25)
Household size	The number of individuals living with the respondent, Wave IV	3.07 (1.42)
Employment	Employed: worked for pay at least 10 hours a week, Wave IV	81.52%
Marital status	Married = 1, not married = 0, Wave IV	50.40%
Physical activitya	Sum of frequency occurrences (bouts) of the following moderate to vigorous physical activities (MVPA) in the past seven days, Wave IV bicycle, skateboard, dance, hike, hunt, yard work, roller blade, roller skate, downhill ski, snow board, racquet sports, aerobics, football, soccer, basketball, lacrosse, rugby, field hockey, ice hockey, run, wrestle, swim, cross-country ski, cycle race, martial arts, gymnastics, weight lifting, strength training, golf, fish, bowl, softball, baseball, walk for exercise
	0 bouts	13.40%
	1 to 4 bouts	32.20%
	5 bouts or more	54.40%
Inactivity scorea	In the past seven days, hours spent on watching television/videos, playing video/computer games, and using a computer, Wave IV
	Low: 10 hours or below	46.97%
	Mediumml: 11 to 24 hours	34.67%
	High: 25 hours or more	18.35%
N		5,105

2.3. Analytical strategy

Given that the dependent variable is binary (obese or not), we estimate a logistic regression. First, we estimate the main effects of educational attainment and the education PGS on obesity in Equation (1).

$$\log \mathrm{(}\frac{Probabiltyofobesity}{\mathrm{1}-probabilityofobesity}\mathrm{)}\mathrm{=}\alpha \mathrm{+}{\beta }_{\mathrm{1}}\mathrm{E}\mathrm{d}\mathrm{u}\mathrm{+}{\beta }_{\mathrm{2}}\text{EduPGS}\mathrm{+}{\beta }_{\mathrm{3}}\text{Controls}\mathrm{+}{\beta }_{\mathrm{4}}\mathrm{B}\text{MIPGS}\mathrm{+}{\beta }_{\mathrm{5}}\mathrm{P}\mathrm{C}\mathrm{s}$$ (1)

where PCs represent the top 10 principal components (PCs) of genetic ancestry. PCs are included to control for population stratification (Price et al. 2006).

Next, we examine the interactive effect between college education and the education PGS in Equation (2). Because the coefficient of the interaction term, β₃ in Equation (2), cannot be readily evaluated due to the nonlinearity of logistic regression (Ai and Norton 2003), we will plot the predicted probability of being obese to better understand the interaction effect.

$$\begin{gathered} \log \mathrm{(}\frac{Probabiltyofobesity}{\mathrm{1}-p\tau obabilityofobesity}\mathrm{)}\mathrm{=}\alpha \mathrm{+}{\beta }_{\mathrm{1}}\mathrm{E}\mathrm{d}\mathrm{u}\mathrm{+}{\beta }_{\mathrm{2}}\text{EduPGS}\mathrm{+}{\beta }_{\mathrm{3}}\mathrm{E}\mathrm{d}\mathrm{u}\times \text{EduPGS}\mathrm{+}{\beta }_{\mathrm{4}}\text{Controls}+ \\ {\beta }_5\text{BMIPGS}+{\beta }_6\text{PCs} \end{gathered}$$ (2)

Results

Models 1 and 2 in Table 3 report odds ratios from the main effects models. The difference between the two models is that Model 1 excludes the education PGS and the BMI PGS whereas Model 2 includes both. We can examine the impact of the two PGSs as confounders by comparing Models 1 and 2. In Table 3 an odds ratio greater than 1 indicates a higher likelihood of being obese.

Table 3.

College education and the education polygenic score (PGS) predicting obesity: Odds ratios from logistic regressions

	Model 1	Model 2	Model 3
Independent variable
College degree vs. no college degree	0.61***	0.62***	0.75*
High education PGS vs. low education PGS	------	0.95	1.04
College degree × High education PGS	------	------	0.71*
Age	1.03	1.03	1.03
Female vs. male	0.97	0.98	0.98
Parental education
High school vs. below high school	0.97	0.96	0.95
More than high school vs. below high school	0.77*	0.76*	0.75*
Lived with two biological parents vs. no	1.03	1.05	1.04
Verbal IQ (PVT) score	1.00	1.00	1.00
Self-control	0.94***	0.94***	0.94***
Employed vs. unemployed	1.04	1.06	1.06
Married vs. unmarried	1.13	1.13	1.12
Household size	1.07**	1.07**	1.07**
Weekly physical activitya
1 to 4 bouts vs. 0 bouts	0.88	0.88	0.88
5 bouts or more vs. 0 bouts	1.02	1.04	1.03
Composite inactivity scorea
Medium vs. low	1.40***	1.43***	1.43***
High vs. low	1.50***	1.54***	1.54***
BMI PGS	------	1.46***	1.47***
−2 Log-likelihood	6553.29	6553.29	6553.29
LR Chi-square	194.99***	378.85***	377.06***
N	5,105	5,105	5,105

Note: An odds ratio greater than 1 indicates a higher likelihood of being obese.

^*p < .05

^**p < .01

^***p < .001 (two-tailed tests)

Model 1 shows that the odds of obesity are about 39% (= 1 − 0.61) lower for individuals who hold a college degree than for individuals who do not hold a college degree. The coefficient of holding a college degree is similar in Mode 2—the odds are 37% lower. Hence, evidence suggests that adding the education PGS and BMI PGS in the model may not affect the effect of college education on obesity. In Model 2, the coefficient for the education PGS is not statistically significant. A 1 SD higher BMI PGS increases the odds by about 46% (= 1.46 − 1). The coefficients of the control variables are similar between Models 1 and 2. When parental educational level is above high school, the odds of obesity are 23% lower for their offspring than for those whose parents do not finish high school. Individuals who score higher on the self-control scale, live in a smaller household, or are in the low inactivity group are generally less likely to be obese.

We now turn to the interaction effect results. Model 3 shows that holding a college degree decreases the odds of obesity by 25% (odds ratio 0.75). The coefficient for the education PGS is insignificant and the interaction term is significant. To interpret this interaction effect, we plot it in Figure 1 below. Model 3 also shows that the coefficients of the control variables are similar to those of the main effects models.

Figure 1.

Predicted probability of obesity by college education and the education polygenic score (PGS), with 95% confidence interval

Note: Results are derived from Model 3 of Table 3.

We find evidence that the influence of the education PGS is contingent on college education. Figure 1 shows that individuals with no college degree are generally more likely to be obesity than individuals with college degree. Among the former group, individuals with a high education PGS and those with a low education PGS have a similar probability of obesity, as the two 95% confidence intervals overlap. This finding suggests that in an unfavorable environment (no college degree) the genes (education PGS) are not likely to make a difference in the risk of being obese. However, among those who hold a college degree, the high education PGS group is associated a lower probability of obesity than the low education PGS group (the two 95% confidence intervals do not overlap). The predicted probability of obesity for individuals with a high education PGS is about 0.25, whereas the predicted probability is 0.32 for individuals with a low education PGS. Therefore, in a favorable environment (college degree) individuals without “risky” genes (individuals with a high education PGS) tend to have a more positive outcome (lower risk of obesity) than individuals with “risky” genes (individuals with a low education PGS).

Discussion

In this article, we find that, net of the education PGS and BMI PGS, holding a college is associated with a lower probability of being obese. Genetic influences do not appear to confound the effect of college education. This finding supports the idea that educational attainment is a cause of health. We also find an interaction effect between college education and the education PGS. The education PGS may indicate intrinsic academic and cognitive abilities (Okbay et al. 2016), and be related to the development of brain and central nervous system (Lee et al. 2018). A relatively high education PGS might indicate stronger deliberative abilities, and enable individuals to adopt a healthy lifestyle through calculations about the consequences of life choices. Consequently, in the favorable environment (college education) individuals with high education PGSs might enjoy more benefits including good health provided by “learned effectiveness” (Mirowsky and Ross 2005) than those with relatively low education PGSs. However, in the unfavorable environment (no college education) it is plausible that individuals with high education PGSs face limited opportunities to gain knowledge about healthy lifestyles or develop skills and abilities that would help them be rational about life choices. Without much “learned effectiveness,” the high education PGS group might exhibit levels of knowledge and capability similar to those of the low education PGS group in the unfavorable environment. Future research might test this possibility and explore other explanations.

The PGS approach has been extensively used in the social and biological sciences, but some caveats apply. The associations between genetic variants and education discovered in the GWASs reflect direct genetic influences on education, but they also reflect indirect influences such as genetic nurturing (Kong et al. 2018). In addition, within-family association between the education PGS and education is weaker than that of general population (Lee et al. 2018). This finding suggests that the GWAS education effects may be overestimated because educational attainment is also influenced by family environments. Most GWASs analyze samples of individuals of predominantly European genetic ancestry (e.g., Locke et al. 2015, Lee et al. 2018, Okbay et al. 2016, Rietveld et al. 2013). When GWASs include more non-European ancestry groups in the future, the PGS approach can be expanded.

Our estimates are obtained after controlling for genetic confounding, but another caveat—for the present study, twin and fixed-effects models, the IV strategy, and observational studies more broadly—is that unobserved confounders may affect results. For example, our finding that the education PGS predicts a lower probability of obesity in the college education group may be caused by a confounder that makes some individuals with a high education PGS go to college, and, at the same time, less likely to be obese than those who have a low education PGS. In twin studies, monozygotic twins have different non-shared environments (e.g., different experiences in adult life). Non-shared environments can be a confounder (McGue, Osler, and Christensen 2010). For the IV strategy, there could be confounders like large secular improvements in education over the years (Galama, Lleras-Muney, and van Kippersluis 2018).

Estimating the causal effects of education on health is a challenging task. This article is an example of how genetic data can be a useful tool to improve our understanding of education’s role in population health. As we learn more about the genetics of educational attainment and obesity, we can explore new ways to unpack the dynamic and complicated processes that generate the health disparity.