Identification of Social Effects through Networks and Groups

In this paper, we propose new solutions to the well-known problem of identification of social effects. Manski (1993) showed that endogenous and contextual (or exogenous) social effects cannot, in general, be disentangled in the linear-in-means model. Our main innovation is that we allow individuals to have different reference groups. That is, social interactions are structured through a network. We have two main results. First, if the network is not partitioned into groups, the model is identified. Second, even when individuals interact in groups, as soon as two groups have different sizes the model is identified. This second result is particularly surprising since it means that endogenous and contextual effects could, in principle, be disentangled with traditional data