UNPACKING SOCIAL INTERACTIONS

Abstract

I. INTRODUCTION Following seminal contribution of Manski (1993) to the econometric literature on social interactions, as well as comprehensive treatment by Brock and Durlauf (2001), researchers have taken seriously the identification problems that surround the empirical detection of social influences in individual behavior. To date, virtually any empirical work in this field addresses the so-called reflection problem, using either econometric or experimental strategies. A possibly trickier problem arises from the consideration that identification of social influences says nothing about their source. Suppose that one identifies correctly the causal effect of the prevalence of a certain behavior--such as crime, work effort, labor supply, human capital investment, use, risky sexual behavior, unhealthy habits--in a reference group on individual behavior within that group. One cannot say whether the social effect is due to information sharing (people communicate and pass along information that increase the likelihood of behaving in a certain way), social learning (people observe others' behavior and infer the distribution of outcomes from taking certain actions), stigmatization (the prevalence of a certain behavior makes its adoption less embarrassing), identity (a certain behavior is the hallmark of membership and loyalty to a group characterized by a precise identity), pure desire for imitation (keeping up with the Joneses), and so on. More likely, the estimated social effect will be the compound effect of all or some of these. Separate identification of different social effects, that is, of the channel through which social interactions operate, is only apparently of second-order importance compared to identification toutcourt. For instance, the aforementioned effects do not work necessarily in the same direction and may offset each other. Therefore, in principle, it is possible that estimated social effects are negligible, while underlying social interactions do affect behavior. Albeit extreme, this possibility suggests that the issue is a relevant one. In particular, the separate identification of different social effects is important for policy evaluation because different channels of influence generally require different policies. Consider, for example, the literature on why welfare cultures emerge, that is, which social factors affect participation. Information sharing and stigma are natural candidates--and in fact have received considerable attention in both research and policy circles. Suppose a policymaker wants programs to reach all individuals they are designed for, and wants to take actions in order to increase take-up rates (which are notoriously low in the United States), the appropriate measures depend on the relative magnitude of stigma and information effects. If stigma is the main cause of nonparticipation, then one could think of ways to hide recipiency--such as the replacement of food stamps with plastic cards. If instead information sharing is the main cause, then one should think of ways to inform eligible nonparticipants. The point is that policies that work if one kind of social effect is predominant may be ineffective if instead another kind is predominant. However, even the most recent empirical works that identify social interactions in participation such as Bertrand, Luttmer, and Mullainathan (2000) do not attempt to identify them separately. An estimate of aggregate social interactions, despite being important, may be of limited use because without knowledge of the composition and/or source of the social effect, one is hard-pressed to make policy recommendations. The issue, which we label the conflation problem, (1) was brought to light by Manski (2000) and is a pressing one for empirical study of social interactions. In this article, we suggest a simple and easily usable procedure to separately identify different kinds of social interactions. …