Peer Influence on Obesity: Evidence from a Natural Experiment of a Gene-Environment Interaction

Abstract

The role of peers in explaining the obesity epidemic is difficult to evaluate, largely due to selection (the tendency of similar individuals to make friends with each other). Our study addresses this selection issue by using data from a natural experiment of randomly assigned college roommates. We investigate whether and how peers, gender, and the FTO gene interactively influenced BMI. We find that women with a weight-prone version of the gene were about three pounds lighter if assigned frequently-exercising roommates than if assigned non-frequently-exercising roommates. However, living with frequently-exercising roommates had little impact for women without the weight-prone version of the gene or for men regardless of genotype. We find that individuals with the weight-prone version of the gene exercised more often when assigned frequently-exercising roommates. This might be a mechanism through which the effect of frequently-exercising roommates worked.

Introduction

In the United States the prevalence of obesity has increased dramatically, from 12% in 1991 to 42% in 2018 (Hales et al. 2020, Mokdad et al. 1999). This obesity epidemic is a growing threat to population health because obesity is associated with diabetes, cardiovascular disease (Mokdad et al. 2003), cancers (Lauby-Secretan et al. 2016), and increased mortality (Berrington de Gonzalez et al. 2010). One particularly interesting explanation for the obesity epidemic focuses on peer influence. Using observational data, Christakis and Fowler (2007) find that an individual is more likely to be obese if his or her friend is obese. The authors argue that obesity may spread through social networks. However, credible evidence of causal peer influence is difficult to obtain with observational data. The major difficulty stems from selection or homophily (individuals tend to choose friends who are similar to themselves) (Manski 1993, Moffitt 2001, Mouw 2006). Observational data can hardly differentiate the case in which peers cause body weight to change from the case in which the impact of peers is spurious due to selection.

Why are some individuals obese while others are not even when they share the same friends? Growing evidence from gene-environment (G × E) interaction research suggests that part of heterogeneity in individuals’ responses to similar environments is due to genetic characteristics (e.g., Belsky et al. 2016, Boardman et al. 2014, Conley and Fletcher 2018, Domingue et al. 2016, Freese 2018, Guo et al. 2015, Herd et al. 2019, Li, Liu and Guo 2015, Liu and Guo 2015, Mitchell et al. 2015, Perry 2016, Simons et al. 2011). Environmental influences may depend on an individual’s genetic susceptibility or vice versa. The incorporation of genetics can help reveal the relationship between social environment and health that might otherwise be missed. Consider a hypothetical case in which the peer effect on a group of individuals with genotype A is 1, while the effect on a group with genotype B is −1. Without considering individual genetic makeup, one might conclude that peer effect is zero by calculating the average effect across the two groups.

Gender is a critical factor in understanding G × E interaction effects on health outcomes (Short, Yang and Jenkins 2013). The powerful role of gender-differentiated social environments may lead to modification or even creation of biological differences in health outcomes (Fausto-Sterling 2005, Perry 2016). In the case of body weight, meanings and consequences attached to overweight and obesity are highly gendered (Fikkan and Rothblum 2012). The ideal female body image emphasizes thinness, whereas the ideal male body image emphasizes muscularity (Harrison 2003, Ridgeway and Tylka 2005, Thompson et al. 1999b). As a result, women may respond to peer influence differently than men because they are under more social and cultural pressure to control their weight.

In this article, we investigate whether and how peers, gender, and genetic susceptibility to obesity work interactively to influence BMI (i.e., a three-way G × E interaction, a gene × peer × gender interaction). We overcome the selection issue by using data from a natural experiment of randomly assigned college roommates. To capture genetic susceptibility to obesity, we focus on the fat mass and obesity associated (FTO) gene. Of all obesity-susceptibility genes identified by large-scale genome-wide association studies (GWAS) so far, the FTO gene stands out because it has the largest effect size (Loos and Yeo 2014) and demonstrates biological mechanistic bases for the genetic association with obesity (Claussnitzer et al. 2015).

Gene-environment correlation, or rGE (genotypes are nonrandomly associated with environments), can create a false interaction relationship between environment and genotype (Jaffee and Price 2007). Our experimental design eliminates potential bias caused by rGE. Our study combines a natural experiment design and genetic data to investigate peer influence on obesity. This is a unique combination that represents a valuable opportunity to simultaneously address the selection and rGE issues and examine G × E interaction effects.

Background

Causal Peer Influence and Roommate Studies

Interpretations as to how peers influence each other are often guided by social learning theory (Akers 1985, Bandura 1977, Sutherland 1947). Social learning theory argues that behavior is learned in social interactions. Definitions and meanings of behavior are acquired and reinforced in the process of observing peers. When a behavior is accepted by an individual and reinforced by interactions with others, this behavior likely occurs. While positive consequences often encourage the behavior, negative consequences often discourage the behavior. Christakis and Fowler (2007) examine social contagion of obesity by analyzing observational data from the Framingham Heart Study. They find that an individual’s probability of becoming obese increases if a friend is obese. The authors interpret these findings from the social learning perspective—peers may influence body weight through mechanisms including changes in weight-related behaviors such as food consumption.

Other researchers argue that it is not peer influence that makes individuals similar to one another (e.g., Cohen-Cole and Fletcher 2008). Selection may cause a similarity in weight between peers. Selection refers to the preference to associate with others like oneself. Individuals do not make friends randomly. Rather, individuals tend to make friends with those who share similar traits such as weight. Therefore, a correlation of weight between friends may be the result of friend selection, rather than social learning. Christakis and Fowler (2007) address selection by controlling for lagged values of both the individual’s and peer’s weight, assuming that peer selection is conditional on weight. However, peer selection is also conditional on a large number of other observed and unobserved factors. When using observational data, establishing causal peer influence is difficult due to selection (Manski 1993, Moffitt 2001).

Studies that use randomly assigned college roommates as a natural experiment provide a solution to the problem of peer selection (Mouw 2006). The random assignment of roommates creates an exogenous, intimate peer environment that is not self-selected. As a result, estimates of peer influence are free of selection bias in roommate studies. Researchers have used randomly assigned roommates to investigate peer effects on academic performance (Kremer and Levy 2008, McEwan and Soderberg 2006, Sacerdote 2001, Stinebrickner and Stinebrickner 2006, Zimmerman 2003), and risky behaviors such as drug use, sexual behavior, and smoking (Duncan et al. 2005, Eisenberg, Golberstein and Whitlock 2014, Li and Guo 2016, Li and Guo 2020).

So far, two roommate studies have examined peer effects on body weight, but the findings are mixed. Using a sample that consists of 144 females, Yakusheva, Kapinos and Weiss (2011) find that weight gain is negatively associated with roommate’s weight. Yakusheva, Kapinos and Eisenberg (2013) use a larger sample of 1,569 male and female college students, and find a positive association between weight gain and roommate’s weight for females. The authors do not find significant roommate influence for males.

One possible reason for the mixed findings is that these two roommate studies measure peer influence using peer’s weight. When an individual’s friend is obese, two opposite scenarios may occur. In one scenario, the individual may try not to gain weight due to the stigma linked to obesity (Brewis 2014). The opposite scenario is also possible. Attitudes towards obesity may change (Chang and Christakis 2002, Sobal 1999). When an individual does not care about the stigma attached to obesity, his or her BMI may increase by adopting a peer’s diet or behaviors leading to weight gain

Our study measures peer influence with physical activity. Unlike peer’s weight, peer’s physical activity potentially has a less ambiguous impact on body weight. Participation in physical activity is rarely associated with negative meanings. It is unlikely that an increase in peer’s physical activity leads to a decrease in physical activity. If an individual’s peer exercises often, this individual may exercise more frequently as well. In young adults, physical activity can lead to weight change (Allender, Cowburn and Foster 2006, Dishman, Sallis and Orenstein 1985).

Gender and Obesity

Gender shapes health through various gender-differentiated pathways (Short, Yang and Jenkins 2013). In the context of obesity, gender-differentiated pathways include stigma and negative social consequences. Stigma concerns social, cultural, structural and moral discrediting that individuals experience due to the negative connotations of a trait or behavior (Link and Phelan 2001, Yang et al. 2007). Common negative meanings attached to obesity include laziness, non-compliance, lack of intelligence, a weak-will, dishonesty, and lack of self-control (Brewis 2010).

Women tend to report greater levels of fat stigma than men (Puhl, Andreyeva and Brownell 2008) and are more vulnerable to processes that generate stigma (Brewis 2014). The consequences of obesity are also highly gendered. Women who are overweight or obese may face a number of negative consequences including lower wages (Baum and Ford 2004, Cawley 2004), poor educational outcomes (Crosnoe 2007), and a disadvantage in other social domains (for a review see Fikkan and Rothblum 2012). Additionally, the negative effects associated with obesity are more serious for young adults (Frisco, Houle and Martin 2010, Puhl, Andreyeva and Brownell 2008, Van Hook and Baker 2010).

Ideal body image is another important gender-differentiated pathway that relates to obesity. Body image is a dynamic and multi-faceted construct that consists of body-related attitudes, feelings, and behaviors (Cash 2004, Grogan 2016, Thompson et al. 1999b). In contemporary society, the media portrays thinness as the ideal body image for women (e.g., Fallon and Rozin 1985, Tiggemann and Miller 2010) and muscularity for men (e.g., Daniel and Bridges 2010, Pope Jr et al. 1999, Thompson et al. 1999a). Men and women, especially in their youth, may internalize media representations of ideal bodies (Arbour and Ginis 2006, Grabe, Ward and Hyde 2008), even though these ideal body types are extremely unattainable.

Hence, young women tend to be under pressure to follow this standard of thinness and may identify it as a source of motivation for physical activity. Research shows that young women are most likely to exercise in order to lose weight by taking up aerobic exercise, whereas young men are most likely to exercise to build muscle by taking up weight lifting (Egli et al. 2011, Keating et al. 2005, Kilpatrick, Hebert and Bartholomew 2005, Mueller et al. 2010, Olivardia et al. 2004). Although there are many reasons for engaging in physical activity such as maintaining health, pursuing the ideal body image remains an important reason with gendered responses.

G × E Interaction Framework

G × E interaction research often relies on the classic diathesis-stress model (Ellis et al. 2011). A key element of this model is that individuals with some “risky” genes are more susceptible to environmental influences than individuals without “risky” genes. Shanahan and Hofer (2005) propose that G × E interactions may not necessarily mean the interaction of genotype and a single environmental factor. Some G × E interactions may only apply to a particular circumstance or a subgroup of the population, suggesting that individuals with “risky” genes are susceptible to an environmental influence only when another environmental stimulus is present.

Perry (2016) examines the interaction effects by genes and two environmental variables—gender and marriage—on substance dependence (i.e., a three-way gene × marriage × gender interaction). Perry finds that among individuals with high genetic risk for substance dependence, men who are formerly married are more likely to be nicotine dependent than women who are formerly married. Among individuals with low genetic risk, married women are less likely to be nicotine dependent than married men. That is, the protective effect of marriage exists for men with the high-risk genotype and for women with the low-risk genotype, but not for other groups.

Perry’s study shows that complex G × E interaction effects can be successfully identified if researchers thoroughly consider and measure environmental factors. Perry finds no G × E interaction effect when only one environmental variable is considered. This suggests that if the “E” of the G × E framework is inadequately defined, G × E interactions could be missed.

The FTO Gene

Due to data limitations, prior studies are unable to identify sources of genetic heterogeneity in response to peer effect. For example, Christakis and Fowler (2007) use lagged obesity status as a proxy for genetic endowment to obesity.

Rapid advances in genomics have significantly improved our understanding of the genetics of obesity. The FTO gene is one of the first and most important genes identified by genome-wide associated studies (GWAS) of obesity (Dina et al. 2007, Frayling et al. 2007). The association between the FTO gene and obesity has been widely replicated in individuals across the life course and diverse ancestries. Of all BMI-associated genes identified by GWAS, the FTO gene has, by far, the largest effect size (Loos and Yeo 2014). The variance in BMI explained by the FTO gene is larger than the variances explained by the second-, third- and fourth-largest effect genes (MC4R, TMEM18, and GNPDA2) combined (Locke et al. 2015).

More importantly, the FTO gene has clearer biological pathways through which obesity status is affected. A variant in the FTO gene can repress mitochondrial thermogenesis in adipocyte precursor cells (Claussnitzer et al. 2015). As a result, there could be a cell-autonomous developmental shift from energy-dissipating adipocytes to energy-storing adipocytes, resulting in increased lipid storage and weight gain. Given its prominent role, we use the FTO gene in our model of the G × E interaction on obesity.

Gene-Environment Correlation

G × E interaction may be due to rGE because genes can play an important role in shaping one’s environments (Jaffee and Price 2007). Suppose a researcher finds that a gene interacts with marital status to influence delinquent behavior. If this gene also affects the probability of getting married, then the interaction between this gene and marital status is most likely due to the correlation between the gene and marital status. It is difficult to address rGE with observational data. In our study, bias caused by rGE is removed because the environment, peer physical activity, is uncorrelated with respondent’s genotype due to the random assignment of roommates.

Hypotheses

Insights from social learning theory, gender theory, and the diathesis-stress model suggest an interaction effect of peers, gender, and the FTO gene on obesity. When peers exercise often, an individual’s weight could change because he or she is influenced by the peers and participates in physical activity more frequently. For young women, peer influence on body weight may be stronger because they are under more social and cultural pressure to be thin. They may exercise to manage their weight. However, for young men the goal of physical activity is more likely to build muscle.

Furthermore, individuals who carry more risk alleles for obesity (i.e., polymorphisms associated with a BMI increase) in the FTO gene have a biological predisposition to gain weight (Claussnitzer et al. 2015). One possibility is that they may be more aware of or sensitive to their weight than those who carry fewer risk alleles in the FTO gene. Therefore, when peers exercise often, individuals who carry risk alleles may be more susceptible to peer influence. In addition, at a similar level of genetic susceptibility to obesity, young women may be more susceptible to peer influence than young men because of the gendered stigma and negative consequences associated with obesity.

Based on the preceding discussion, we hypothesize that living with a frequently-exercising roommate would lead to a lower BMI than living with a non-frequently-exercising roommate. The effect of a frequently-exercising roommate on BMI would be stronger among women than among men, and it would be stronger when individuals possessed more risk alleles for obesity in the FTO gene. In particular, this effect would be stronger among women who possessed more risk alleles. Further, a mechanism by which the roommate effect on BMI worked would be that having a frequently-exercising roommate led to an increase in the frequency of physical activity.

Data and Measurement

Data

We conducted the natural experiment, the Roommate Study, at a large racially-and-economically diverse public university in the spring semester of the 2007–2008 academic year. We collected data on pre-college and college health behaviors, demographics, and family backgrounds via a web survey. The overall objective of the study was to examine peer effects on risky behaviors.

The random assignment of roommates is key to eliminate biases caused by selection and rGE. We worked with the university’s housing department to ensure that our data come from roommates who were paired randomly. In a typical academic year, a large majority of incoming freshmen apply for campus housing. About 40% of incoming freshmen do not request a specific roommate nor do they request a themed housing program (e.g., foreign languages, health sciences, substance free, and global business). These students are randomly assigned a roommate within gender and requested type of room, making them eligible for our study (see the Appendix for the university housing assignment process).

To further ensure the randomization of roommate assignments in the Roommate Study, we use data from two independent studies, our Roommate Study and the Freshman Survey. The Freshman Survey, also known as CIRP, is designed by the Cooperative Institutional Research Program at the Higher Educational Research Institute, the University of California, Los Angeles. Results confirm the randomization of roommate assignments (Table S1). We also perform tests to verify that Roommate Study participants were representative of the university’s undergraduate population. Results validate this supposition (see the Appendix).

The Roommate Study collected DNA from saliva samples using the Oragene DNA self-collection kit. The Oragene kit has been used successfully in previous studies (Ahituv et al. 2006, McCready et al. 2005). According to the manufacturer’s instructions, DNA was extracted from 2mls of saliva, which contained buccal epithelial and white blood cells. The median DNA yield was 27.3 ug, with a minimum of 0 ug for six individuals and a maximum of 71.3 ug. DNA was plated for Illumina genotyping at 30 ul at >50 ng/ul. For quality control purposes, the DNA samples included 92 parent-parent-child trios and 46 duplicated samples. In the latter case, 46 of our study participants were randomly selected to be genotyped twice. For each pair of duplicates, one sample was labeled with the original ID and the other was labeled as blind with all related information wiped out. The trios allowed a check on Mendelian inheritance, and the duplicates tested whether the genotyping could be reproduced. Our genotyping passed both tests.

In the Roommate Study, a total of 2,664 individuals completed the web survey, and 2,080 provided saliva samples. After excluding 679 individuals whose roommates did not report frequency of pre-college physical activity and 69 with missing data for BMI and physical activity, our analytical sample consists of 1,332 individuals.

Variable Measurement

BMI is calculated as weight divided by height squared. The Roommate Study asked respondents the following questions: “How tall are you without shoes?” “How much do you weigh without shoes?” And, “When you first entered college, how much did you weigh without shoes?” The first two questions are used to calculate college BMI. The first and last questions are used to calculate pre-college BMI. Because BMI data are skewed, we follow the established convention (e.g., Frayling et al. 2007) by applying log₁₀ transformation to normalize BMI. The dependent variable is log₁₀ BMI.

Peer influence is measured by the frequency of pre-college physical activity reported by the roommate, rather than by the respondent. We dichotomize it as “exercised often” (two to four times a week or more) and “did not exercise often” (once a week or less). Biases due to simultaneity and shared environment (Manski 1993, Moffitt 2001) are removed because the roommate’s pre-college physical activity is prior to exposure to the respondent and the shared college environmental influences.

Susceptibility to obesity in the FTO gene is measured by a single nucleotide polymorphism (SNP) rs9939609. This SNP shows a highly significant association with BMI in a GWAS of 38,759 individuals of European ancestry (Frayling et al. 2007). Subsequent studies independently replicate this association in individuals of European ancestry (e.g., Cornes et al. 2009, González-Sánchez et al. 2009), and in samples of African, East Asian, and Hispanic individuals (Graff et al. 2013, Liu et al. 2010, Monda et al. 2013, Villalobos-Comparán et al. 2008, Wen et al. 2012, Wing et al. 2009). SNP rs9939609 is in Hardy-Weinberg Equilibrium (P > 0.01) in our data. Although some studies do not find an association between the SNP rs9939609 and BMI in individuals of African ancestry (see Grant 2013), large-scale GWAS find that the FTO gene is among the most significant genes associated with BMI in African ancestry populations (Monda et al. 2013). Moreover, the effect of the FTO gene in African ancestry populations is similar to that observed in European ancestry populations (Loos and Yeo 2014). As a robustness check, we repeat our analyses without respondents of African ancestry. Results are similar.

The association between BMI and SNP rs9939609 is replicated in our data. Table 2 shows that each additional copy of the risk allele, the A allele, is associated with a BMI increase of approximately 0.4 kg/m². The effect size is similar to the GWAS that identifies SNP rs9939609 (Frayling et al. 2007), which reports that each additional copy of the risk allele is associated with a BMI increase of 0.4 kg/m².

Table 2.

Effect of the FTO gene on BMI

Independent variable
Genotype at rs9939609 in the FTO gene	.007**
Female (ref: male)	−.017***
N	2,039

Note: The model controls for the first ten principal components to address population stratification.

^*p < .05

^**p < .01

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

To further ensure that the estimate of the roommate effect is based solely on variability within the dorm room, we control for two characteristics of residence halls. The first is gender composition of the residential hall or dormitory floor. As we contend in this article, male and female college students often have different motives to engage in physical activity, therefore gender composition might affect respondent’s physical activity. The second is location on campus. Campus location indicates the residence hall cluster on campus. Students living in the same vicinity share a set of characteristics such as the same distance to gym.

Covariates are pre-college and college physical activity, pre-college BMI, race/ethnicity, grade point average, school year, and socioeconomic status. Table 1 reports coding definitions and descriptive statistics for all variables in our analysis. Missing values in covariates are imputed using multiple imputation (Rubin 1987). However, we do not impute missing values for BMI or physical activity, key variables in our analysis. The Roommate Study asked college students about pre-college behaviors retrospectively. To verify the validity of these pre-college measures, we conduct analyses using the two independent sources of data, the Freshman Survey and the Roommate Study (see the Appendix). Findings in Table S2 verify the validity of the pre-college measures in the Roommate Study.

Table 1.

Descriptive statistics

Variable	Definition/coding	Mean or percentage
College BMI	Body mass index in the spring semester of the 2007–2008 academic year	23.22 (S.D. = 3.88)
Roommate’s pre-college physical activity	Exercised often: 2–4 times a week or more Did not exercise often: once a week or less Question: How often did you exercise or participate in physical activity for at least 20 minutes that made you sweat and breathe hard during the 12 months before entering college? Answers: never; less than once a month; once or twice a month; about once a week; 2–4 times a week; and every day or almost every day.	67% 33%
Gender	Female	60%
	Male	40%
Genotype at rs9939609 in the FTO gene	AA—2 risk alleles	15%
	AT—1 risk allele	46%
	TT—0 risk alleles	39%
Pre-college physical activity	Exercised often: 2–4 times a week or more Did not exercise often: once a week or less	68% 32%
College physical activity	Exercised often: 2–4 times a week or more	56%
	Did not exercise often: once a week or less	44%
Pre-college BMI	Body mass index when first entered college	22.78 (S.D. = 3.92)
Race/ethnicitya	Asian	6%
	Black	11%
	Hispanic	6%
	Other	9%
	White	67%
GPAa	Grade point average	3.25  (SD = .53)
School year	Freshman	31%
	Sophomore or junior	69%
Family incomea	<$75,000	30%
	$75,000–$150,000	39%
	>$150,000	32%
Mother’s educationa	No college	25%
	College or more	75%
Hall gender	Co-ed	84%
	All female, all male, or segregated co-ed	16%
Campus location	South campus	62%
	North and mid campuses	38%
Change in the monthly number of physical activities	Questions are the same as roommate’s college and pre-college physical activity.	−2.94 (SD = 9.30)
	Coding: never = 0, less than once a month = .5, once or twice a month = 1.5, about once a week = 4.3, 2–4 times a week = 12.9, and every day or almost every day = 25.
	Calculation: number of college physical activities minus number of pre-college physical activities
N		1,317–1,332

^aDescriptive statistics for race/ethnicity, GPA, family income, and mother’s education are similar between respondents and their roommates, therefore roommate statistics are not shown for these variables.

Analytical Strategy

To determine whether the effect of frequently-exercising roommates on BMI depended on gender and the FTO gene, we first compare the observed mean BMI and observed median BMI by gender and genotype. Next, we examine the gene × peer × gender interaction in a single regression model. In the model, we include the main effects of roommate’s pre-college physical activity, gender, and SNP rs9939609, all the two-way interactions, and the three-way interaction (i.e., all constitutive terms). To adjust for clustering at the room level, we use a generalized estimating equation (GEE) model (Liang and Zeger 1986) and specify exchangeable working correlation structure. Equation (1) describes the model.

$$\begin{gathered} {\text{CollegeBMI}}_{\text{ij}}={\mathrm{\beta }}_0+{\mathrm{\beta }}_1\text{Gende}{\text{r}}_{\text{ij}}+{\mathrm{\beta }}_{2}FTO\text{Genotyp}{\text{e}}_{\text{ij}}+ \\ {\mathrm{\beta }}_3{\text{RoommatePreCollegePhysicalActivity}}_{\text{ij}}+{\mathrm{\beta }}_4{\text{Gender}}_{\text{ij}}\times FTO{\text{Genotype}}_{\text{ij}}+{\mathrm{\beta }}_5{\text{Gender}}_{\text{ij}} \\ \times {\text{RoommatePreCollegePhysicalActivity}}_{\text{ij}}+{\mathrm{\beta }}_6{\text{RoommatePreCollegePhysicalActivity}}_{\text{ij}}\times \\ FTO{\text{Genotype}}_{\text{ij}}+{\mathrm{\beta }}_7{\text{Gender}}_{\text{ij}}\times {\text{RoommatePreCollegePhysicalActivity}}_{\text{ij}}\times FTO{\text{Genotype}}_{\text{ij}}+ \\ {\mathrm{\beta }}_8{\text{Covariates}}_{\text{ij}} \end{gathered}$$ (1)

where i denotes individual and j denotes dorm room.

To address population stratification, we calculate principal components (PCs) (Price et al. 2006) and control for the first 10 PCs in the models. In Equation (1), we try including the 10 PCs as covariates, replacing race/ethnicity with the PCs, and entering race/ethnicity and the PCs simultaneously in the model. Results are similar.

Results

Figures 1A and 1B compare observed mean and median BMI by gender and genotype, respectively (note that the Y axis starts at the BMI value of 20). Both figures show that women with two risk alleles (i.e., two A alleles) exhibited a lower BMI when they lived with frequently-exercising roommates than women with two risk alleles who lived with non-frequently-exercising roommates. The mean BMI was 1.75 kg/m² lower, and the median BMI was 0.77 kg/m² lower. We find no significant roommate effects among women with one or zero risk alleles (i.e., genotype AT or TT) or among men.

Figure 1. Effect of pre-college physical activity of roommates on college BMI depended on gender and the FTO gene.

Note: The numbers of women inheriting 2 risk alleles, 1 risk allele and 0 risk alleles are 112, 380, and 309, respectively. The corresponding numbers for men are 83, 239, and 209.

(A) Observed mean BMI by rs9939609 genotype and by gender. Error bars represent standard errors. P-values are obtained from the t-tests.

(B) Observed median BMI by rs9939609 genotype and by gender. P-values are obtained from the Mann-Whitney-Wilcoxon tests.

(C) Predicted BMI by rs9939609 genotype and by gender. P-values are derived from the interaction model in Table 3.

We now turn to results of the GEE models. The main effect model in Table 3 shows that gender is associated with BMI, but the main effects of roommate and SNP rs9939609 are statistically insignificant. The interaction model shows findings that are not discovered by the main effect model. The three-way interaction is statistically significant, suggesting that peer influence could be contingent on gender and genotype. It should be noted that in the interaction model as well as the main effect model, coding SNP rs9939609 as two risk alleles, one risk allele, and zero risk alleles and coding it as a dichotomous variable (two risk alleles versus one or zero risk alleles) yield similar results. We present results based on the dichotomous variable to simplify our interpretation. This is consistent with findings in Figures 1A and 1B that the roommate effect was more pronounced for individuals with two risk alleles than for individuals with one or zero risk alleles.

Table 3.

Coefficients for GEE regressions of college BMI on pre-college physical activity of roommates, gender, and the FTO gene

	Main effect		Interaction effect
Estimator	Coefficient	p-value	Coefficient	p-value
Roommate’s pre-college physical activity
Exercised often	−.001	.368	.003	.226
Did not exercise often (ref.)	----	----	----	----
Gender
Female	−.007***	.000	−.008*	.009
Male (ref.)	----	----	----	----
Genotype at rs9939609 in the FTO gene
AA—2 risk alleles	.004	.075	.009	.082
AT or TT—1 or 0 risk alleles (ref.)	----	----	----	----
Two-way interaction
Female × 2 risk alleles	----	----	−.008	.222
Female × exercised often	----	----	−.003	.385
Exercised often × 2 risk alleles	----	----	−.013*	.046
Three-way interaction
Female × exercised often × 2 risk alleles	----	----	.022**	.007
N		1,332

Note: The models control for gender, school year, respondent and roommate’s race/ethnicity, GPA, family income, mother’s education, gender composition of the dorm, dorm campus location, and the 10 principal components.

^*p < .05

^**p < .01

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

To better understand the interaction effect in Table 3, we plot the findings and conduct significance tests in Figure 1C (note that the Y axis starts at the BMI value of 20). We find that for women inheriting two risk alleles, BMI was approximately 0.48 kg/m² lower when they were randomly paired with frequently-exercising roommates instead of non-frequently-exercising roommates. We find no significant roommate influence among women with one or zero risk alleles. Nor do we find significant roommate effect among men at the 0.05 level. However, Figure 1C shows suggestive evidence that living with frequently-exercising roommates would lead to weight gain among men with two risk alleles (p = 0.09). In sum, findings support the hypothesis that the roommate effect would be stronger among women who possessed more risk alleles.

Exploring Mechanism

Next, we use a GEE model to test the potential mechanism that roommate’s pre-college physical activity affected BMI by affecting physical activity. The dependent variable is the change in monthly number of physical activities (i.e., the monthly number of physical activities in college minus the monthly number of pre-college physical activities). Findings in Figure 1C show that the roommate effect on BMI only existed among individuals inheriting two risk alleles. This suggests that the roommate effect on physical activity depended on the FTO genotype. Hence, the model includes a two-way interaction between roommate’s pre-college physical activity and SNP rs9939609 (we try a three-way interaction model and find no gender difference). This analysis also benefits from our study design—the random assignment of roommates and estimating the effect of pre-college behavior on college behavior. In other words, the results are not threatened by the selection, simultaneity, shared environment, or the rGE issue.

Table 4 shows that the interaction term is significant. To better interpret this result, Figure 2 compares the predicted monthly number of college physical activities. For individuals with two risk alleles, the number of college physical activities increased by about two activities per month when they were assigned frequently-exercising roommates than when they were assigned non-frequently-exercising roommates. However, we find no significant roommate effect for individuals with one or zero risk alleles. As a robustness check, we use a logistic regression to examine whether or not the respondent would exercise more often in college. Results are similar to those reported in Table 4 and Figure 2.

Table 4.

Exploring mechanism: Coefficients for GEE regression of change in physical activity before and after college on pre-college physical activity of roommates and the FTO gene

Estimator	Coefficient	p-value
Roommate’s pre-college physical activity
Exercised often	−.043	.416
Did not exercise often (ref.)	----	----
Genotype at rs9939609 in the FTO gene
AA—2 risk alleles	−1.709	.095
AT or TT—1 or 0 risk alleles (ref.)	----	----
Two-way interaction
Exercised often × 2 risk alleles	2.811*	.021
N	1,332

Note: The model controls for gender, school year, respondent and roommate’s race/ethnicity, GPA, family income, mother’s education, gender composition of the dorm, dorm campus location, and the 10 principal components.

^*p < .05

^**p < .01

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

Figure 2. Effect of pre-college physical activity of roommates on college physical activity depended on the FTO gene.

Note: P-values are derived from Table 4.

Discussion

The prevalence of obesity has increased dramatically in the United States in recent decades. Using unique natural experiment data in combination with genetic data, we reevaluate the role of peers in explaining the obesity epidemic. Our study design overcomes methodological problems that have hindered previous studies on peer influence and obesity: selection, shared environment, simultaneity, and rGE. The randomization of roommates enables us to estimate peer influence that is not self-selected. By estimating the effect of roommate’s pre-college exercise on college BMI, our analysis is free of biases due to shared environment and simultaneity.

Evidence obtained in our study can be woven into the following account. We replicate the effect of the GWAS-identified SNP rs9939609 in the FTO gene on BMI in our data. Results suggest that peer influence on physical activity was contingent on the FTO genotype. Living with frequently-exercising roommates increased the number of physical activities by about two activities per month for individuals inheriting two risk alleles. However, the effect of roommate’s physical activity on BMI was highly gendered. Among women with two risk alleles, BMI was 0.48 kg/m² lower when the respondent was paired with a frequently-exercising roommate as opposed to a respondent who was paired with a non-frequently-exercising roommate. This peer effect amounts to a reduction of about three pounds for a woman of average weight and height in our data. This size is nontrivial. It is equivalent to about 60% of the effect of two risk alleles in the FTO gene (0.48 kg/m² /(0.4 kg/m² × 2) = 60%) (Frayling et al. 2007). When men who possessed two risk alleles lived with frequently-exercising roommates, however, they did not lose weight. Suggestive evidence in Figure 1C indicates that the effect of frequently-exercising roommates might lead to weight gain instead.

The gender-differential effect of roommate’s physical activity on BMI can be explained by the gendered stigma attached to obesity and the gendered ideal body type. To avoid negative consequences and to meet social expectations, young women are often motivated to manage their weight, while men participate in physical activity more often to build muscle (Egli et al. 2011, Keating et al. 2005, Kilpatrick, Hebert and Bartholomew 2005, Olivardia et al. 2004). Muscle building may increase BMI (Prentice and Jebb 2001). For example, a muscular football linebacker has a very high BMI, but he is not obese. The effect of physical activity on BMI for men might be opposite that for women. This might be a reason for the positive relationship between BMI and roommate’s physical activity among males, as reported in Figure 1C. We observe a peer effect on BMI among women with two risk alleles and a peer effect on physical activity among both men and women with two risk alleles. Individuals inheriting two risk alleles might be more self-conscious of changes in their weight and had a tendency to respond accordingly. Future research might test this explanation and investigate the relationship between predisposition to weight gain and weight-related behaviors.

In their investigation of roommate influence on body weight, Yakusheva, Kapinos and Eisenberg (2013) determine heterogeneity by self-reported weight. The authors stratify the sample into subsamples based on the weight-quartiles. Individuals from different subsamples are considered heterogeneous. Doing this captures heterogeneity, but also has the potential to introduce errors because self-reported weight represents weight along with a combination of numerous weight-related social factors (e.g., income and diet preference) and biological factors (e.g., hormones and genes). In this study, we find that no interaction effect will be revealed if pre-college BMI replaces the FTO genotype to capture heterogeneity (results available upon request). Our study suggests that genetic susceptibility to obesity detects heterogeneity better in the sense that it only represents characteristics at the gene level, and it is fixed over the life course except for rare mutations.

Our findings demonstrate that obesity may spread through social networks, but in a more complicated way than previously thought. Peer influence may depend on gender. The same behavior can carry different meanings for men and women, making social learning a gendered process. This means that gender can affect how social and genetic factors interact to produce individual health outcomes. In an auxiliary analysis we find that only when gender-specific interactions are taken into consideration in the model can peer effect be discovered. Our findings support Short and colleagues’ (2013) argument that scholarship on health and genetics should pay greater attention to gender as a social and cultural environment.

Our results from the three-way interaction model support Shanahan and Hofer’s (2005) argument that to better depict the contingencies and complexities of social environments, a more sophisticated modeling of G × E interactions is in order. We find that a model with only the two-way G × E interaction does not reveal any peer effect (see Table S5). Similarly, Perry (2016) reports that there are no significant interaction effects if the three-way G × E interaction is ignored. Taken together, these findings underscore the need to carefully define and measure complex social environmental influences in G × E interaction research.

This study has several limitations. First, given the unique nature of the Roommate Study, it is not yet possible to conduct a replication study. Second, the Roommate Study did not collect genome-wide genotype data. We are unable to construct the polygenic scores for obesity. The polygenic scores based on SNPs across the whole genome are more powerful in explaining statistical variance in BMI than a single FTO gene. However, a strength of using the FTO gene, whose association with obesity has been widely replicated in large-scale GWAS, is that the FTO gene has clearer biological roles (Claussnitzer et al. 2015) than the polygenic scores, which summarize the individual regression coefficients of SNPs. Third, our sample focuses on college-age individuals in the contemporary United States. Our findings may not generalize to other cohorts or cultures. One should use caution when interpreting our findings in another setting. Fourth, BMI has been commonly used as a measure of obesity, but our results suggest that BMI cannot distinguish unwanted weight gain among women from desirable muscle building among men. Lastly, this study uses self-reported height and weight to calculate BMI, and, as a result, may introduce reporting errors. In addition, our estimates could be biased if underreporting of weight is related to gender and peer. Despite these limitations, the Roommate Study provides a rare and valuable opportunity to investigate G × E interactions in a natural experiment.

Our randomized study design plays a crucial role in obtaining the findings of peer influence on both physical activity and BMI. These findings might be useful for designing intervention programs that promote physical activity to prevent obesity and related diseases for young people. Our work demonstrates that social scientists are well-positioned to make use of genetic findings to advance understanding of the relationship between the social environment and genes, and how they relate to health and well-being.