Peer effects on risky behaviors: New evidence from college roommate assignments

Highlights

• We use a natural experiment with college roommates to estimate peer effects.

• Peer effects on binge drinking are robust, regardless of gender and prior drinking.

• There is tentative evidence that peer effects on smoking are stronger among men.

• There is little evidence of peer effects on illicit drug use, gambling, and sexual activity.

• Similarity in pre-existing risky behaviors predicts the closeness of friendships.

Abstract

Social scientists continue to devote considerable attention to spillover effects for risky behaviors because of the important policy implications and the persistent challenges in identifying unbiased causal effects. We use the natural experiment of assigned college roommates to estimate peer effects for several measures of health risks: binge drinking, smoking, illicit drug use, gambling, having multiple sex partners, suicidal ideation, and non-suicidal self-injury. We find significant peer effects for binge drinking but little evidence of effects for other outcomes, although there is tentative evidence that peer effects for smoking may be positive among men and negative among women. In contrast to prior research, the peer effects for binge drinking are significant for all subgroups defined by sex and prior drinking status. We also find that pre-existing risky behaviors predict the closeness of friendships, which underscores the significance of addressing selection biases in studies of peer effects.

Introduction

The spread of substance use and other risky behaviors in social networks is important to understand in order to inform health and social policy. Information about spillover effects can improve predictions about the dynamics of behaviors in populations and assist the design of interventions that mitigate harmful spillovers or leverage beneficial spillovers. Behaviors such as heavy alcohol consumption and other substance use have substantial impacts on health, functioning, and educational outcomes (Rice, 1999, Carpenter and Dobkin, 2011, Carrell et al., 2011).

Economists and other social scientists continue to devote considerable attention to measuring spillover effects for risky behaviors not only because of the important policy implications but also because of the challenges in identifying unbiased causal effects. As Manski (1993) and others have described, there are three main factors that may bias estimates of social interaction effects: (1) the reflection problem, in which the effect of others on the self cannot be disentangled from the reverse; (2) selection into social networks, which may lead to correlations in unmeasured individual characteristics and generate spurious correlations in outcomes; and, (3) unmeasured contextual factors, or “common shocks,” which may also generate spurious correlations in outcomes. In addition, from a policy perspective it is useful to distinguish between “endogenous” peer effects that imply multiplier effects (behavior A by one person is directly influenced by behavior A by another person), versus “contextual” peer effects that imply causal but not necessarily multiplier effects (behavior A by one person is influenced by being around another person engaging in behavior A, but the causal mechanism is through other peer characteristics correlated with behavior A).

In this study we use the natural experiment of assigned college roommates to estimate peer effects for substance use and other risky behaviors. This empirical approach addresses the identification issues noted above and thus yields unbiased estimates. The approach has been used in previous studies mainly to look at academic outcomes, particularly grade point average (GPA), using administrative data from colleges and universities. For this study we collected new survey data to examine a range of behaviors with important implications for health and wellbeing: binge drinking, cigarette smoking, illicit drug use, gambling, sexual activity, suicidal ideation, and non-suicidal self-injury.

We find significant peer effects for binge drinking but little evidence of effects for the other outcomes. The effects for binge drinking are robust to controlling for a range of additional peer characteristics, suggesting that these are true spillover effects (“endogenous” rather than “exogenous” effects, in Manski's terminology), although we cannot rule out the possibility that the effects are driven at least in part by unmeasured roommate characteristics. The magnitude of the effects—a 8.6 percentage point increase in the probability of binge drinking, as a result of having a binge-drinking roommate—is somewhat smaller than in most previous studies of peer effects. As compared to a previous study based on college roommate assignments, which finds significant peer effects on binge drinking only for men with prior binge drinking (Duncan et al., 2005), we find more widespread effects: for both women and men, and for both prior binge drinkers and prior non-binge drinkers. We also examine the closeness of roommates’ relationships, as reported in the follow-up survey. This analysis indicates that similarity in pre-existing behaviors predicts closeness of relationships, for the most part. Also, roommates who end up being close friends exhibit stronger apparent peer effects on binge drinking; this differential is less robust, however, when we look at predicted friendship levels based on baseline measures, rather than the endogenous actual friendship levels.