Obesity has reached epidemic proportions in children and adolescents in the United States, increasing from 5% in 1980 to over 19% in 2018 (Skinner et al., 2019, Fryar et al., 2020). This is aligned with the results of the latest cross-country large-scale study showing that since 1990, obesity among children and adolescents has quadrupled worldwide (Phelps et al., 2024). Mounting evidence suggests that the extra pounds often start children on the path to health problems such as cardiovascular diseases, diabetes, and cancer (Bendor et al., 2020) . To explain such an alarming phenomenon, a large number of studies have focused on socioeconomic factors such as growing unhealthy eating habits and the decline in time spent doing physical exercise (Papoutsi et al., 2013). Complementary to these views, health economists have also attempted to investigate the obesity epidemic from the perspective of social interactions (Christakis and Fowler, 2007, Halliday and Kwak, 2009, Trogdon et al., 2008, Yakusheva et al., 2014, Cohen-Cole and Fletcher, 2008, Fortin and Yazbeck, 2015, Lim and Cornwell, 2023). Most of these studies document the presence of positive and significant peer effects which could increase the prevalence of obesity by shaping body image and/or by boosting the social transmission of unhealthy habits related to diet and physical activity. Our paper follows the second strand of the literature by exploring the role of gender heterogeneity in the social diffusion of Body Mass Index (BMI) outcomes among teenagers, and its consequences in terms of anti-obesity interventions.1

Most studies on peer effects assume that social interactions are homogeneous (Manski, 1993, Bramoullé et al., 2009, Boucher et al., 2024). This means that the effects of all peers are equal regardless of the particular type, such as race or gender. However, this assumption is restrictive and may not accurately reflect reality, particularly when considering adolescent students’ weight. This paper proposes an econometric model allowing for heterogeneous peer effects along gender lines and estimates it using detailed network data on teenagers’ friendships from the Add Health dataset. Simulations based on our results show that ignoring gender-based heterogeneity of peer effects may lead to inefficient health interventions to curb obesity. The present study contributes novel methodology, results, and policy insights to the existing literature, which we discuss in detail below.

While the literature on dietary choices and weight outcomes of adolescents is sizable (Kapinos and Yakusheva, 2011, Mora and Gil, 2013, Corrado et al., 2019, Fortin and Yazbeck, 2015, Angelucci et al., 2019), studies focusing on the heterogeneity of peer effects in this context are rare (Arduini et al., 2019, Renna et al., 2008, Yakusheva et al., 2014). However, to our knowledge, we are the first to model heterogeneity in between-gender peer effects. In our model, two types of individuals (i.e., male vs. female students) interact within the same network (i.e., a school). This defines a ‘heterogeneous’ model with two within-gender and two between-gender peer effects, with respect to the ‘homogeneous’ setting with one peer effect term.

We characterize our model econometrically and theoretically. Our econometric approach is closely related to the ones developed by Hsieh and Lin (2017) and Arduini et al. (2020), but with important differences. Hsieh and Lin (2017) model peer effects via Bayesian methods, and estimate them through Markov Chain Monte Carlo sampling techniques. Similarly to us, Arduini et al. (2020) derive a set of identification conditions that generalize the standard linear model of Bramoullé et al. (2009) to allow for heterogeneous peer effects. Differently from them, our paper puts emphasis on the micro-foundation of the econometric model. In particular, we show that our empirical approach is consistent with the best response functions of a non-cooperative model where social interactions stem from the channel of pure spillover or pure conformity and that all its parameters are identified under some plausible assumptions.

We illustrate our econometric model using the 1996 saturation sample of the National Longitudinal Study of Adolescent Health (Add Health) which provides census data on 16 selected schools. Respondents from the sample reported their height and weight (which we use to compute the BMI), and they were also asked to name up to five male friends and up to five female friends within their school, which allows us to map the friendship networks.

When we assume that peer effects are homogeneous within and across gender lines, our findings compare well with the previous literature.2 When we relax the homogeneity assumption, we find that peers’ outcomes affect BMI in a way that is gender-specific. In particular, we find that the ‘male–female’ endogenous peer effect (that is, the effect on male students’ BMI of the BMI of their female friends) is significantly larger than the other estimated peer effects (for male–male, female–male, female–female interactions, respectively). This result adds to the growing evidence of peer-effect heterogeneity along gender lines. Previous studies on weight-related outcomes suggest that female adolescents are more responsive than males to their peers’ behavior.3 By considering both the within- and between-gender dimensions, we provide evidence that male students are particularly responsive to the weight of their female friends, which is in line with the findings by Kooreman (2007) and Hsieh and Lin (2017) for a number of documented adolescent behaviors other than BMI. This result is compatible with different explanations. For instance, it could be due to the fact that girls are more mature and presumably more influential than boys at the same age during childhood and adolescence. This hypothesis is consistent with recent studies in neurosciences (e.g., Gong et al., 2009, Lenroot and Giedd, 2010, Lim et al., 2015, Goyal et al., 2019) suggesting that girls tend to optimize brain connections earlier than boys. Also, the stronger influence of girls on boys could be imputed to the dynamics of between-gender relationships and romances (see Hill, 2015).4

One limitation of our benchmark model is that it implicitly assumes that the formation of links between students is exogenous once we account for observable attributes and school choice. However, as long as students self-select their peers partly based on unobserved factors that also appear in the equation of interest (i.e., the BMI equation), this will create an endogeneity problem. For instance, under homophily, that is, when individuals tend to bond with peers with similar preferences, a spurious correlation will arise between the individual’s BMI and his/her peers’ BMI. Thus, it is important to provide a robustness check of network exogeneity. While many approaches have been developed in recent years to test for network exogeneity (see the recent survey by Bramoullé et al., 2020), we focus on the one proposed by Jochmans (2023), which provides a natural extension of our estimation framework.

Finally, we conduct a simulation exercise to study the impact of an intervention proposing one balanced meal per week in replacement of one fast-food type serving. On the basis of our most conservative findings, we conclude that the spillovers of offering meal replacements to female students are 33% higher than the spillovers of males. This suggests that returns from (resources spent on treating) females are 8% larger than the ones from males in terms of overall BMI decrease in the student population. If we further assume that females are more responsive to the intervention, we conclude that the spillovers from females are twice the spillovers from males, which translates into a 54% gain in terms of aggregate BMI decrease from reaching out to female students. Overall, our analysis indicates that acknowledging gender-based heterogeneity of peer effects may increase dramatically the efficiency of anti-obesity policies. More generally, while ex-ante evaluations based on structural models are common in other fields of economics (e.g., Wolpin, 2007), they are novel in the context of social interactions. By providing the infrastructure to evaluate how interventions interplay with heterogeneous social diffusion, our paper may be of interest in many contexts where peer effects differ along individual dimensions (e.g., race, education).

The rest of the paper is organized as follows. In Section 2 we characterize our econometric model, and in Section 3 we discuss its microfoundation. Section 4 introduces the data, Section 5 presents our results, and Section 6 describes the simulation exercise. Section 7 concludes. Appendix A provides the mathematical derivation of the theoretical model. Appendix B formalizes the identification conditions and presents the estimation techniques in use.