Identifying Social Effects In Economic Networks Gal Oestreicher-Singer 1 and Arun Sundararajan 2 1 New York University and Tel-Aviv University. 2 New York University. Overview and Theory A central measurement challenge when estimating models using networked data is of identification. Many theories one might test using such networked data would suggest that the presence of a link (or edge) between two entities (or nodes) explains why their outcomes are correlated. In other words, such theory posits that the presence of the link causes a “social” or “peer” effect. However, in order to establish that the presence of the link is what is in fact causing this correlation, it is important to separate this social effect from two others: co-variation that would have been observed even if the link was not present (owing to some inherent similarity that the entities have that causes the link to appear), and co-variation caused by some exogenous factor that is not observed by the researcher. A seminal paper by Manski (1993) separated these effects into three categories: (1) Social or peer effects: that the propensity of an entity to choose varies with the outcomes chosen by the group the entity is connected to directly. (2) Exogenous (or contextual) effects: that the propensity of an entity to choose or demonstrate an outcome varies with the exogenous characteristics of the group the entity is connected to; the outcomes by themselves play no role in the co-variation. (3) Correlated effects: that entities choose (or have observed) outcomes that are similar to the group the entity is connected to because the entities in the group have similar individual characteristics or face a similar environment. Manski further established that, under fairly general conditions involving groups that partition the set of entities, it is impossible to identify social effects. The first contribution of our paper is to show that under fairly general conditions, when the groups in question are defined by a social network rather than a partition, social effects can in fact be identified and measured . Our result exploits the fact that researchers often have access to multiple overlapping networks that connect the entities whose choices or behaviors one is trying to explain. Since the explosion of social networking sites and other sites that make connections between entities visible (such as recommendation networks) leads to the widespread availability of overlapping networked data sets associated with the entities in question. This is not surprising, of course, given that sociologists in the past have observed that people have multiple overlapping communities they belong to. Our result requires some minimal conditions about the structure of the overlapping social network which are far less restrictive than those in prior papers (for instance, Braumoulle et al, 2007). Briefly, we require that (1) the two networks differ in some minimal way (one has an edge that the other does not) and (2) there exist a pair of nodes whose degree is the same in one network while being different in the other. In our empirical estimation of the model, we have verified that these conditions are easily met for a sample of over 200 instances of a copurchases network (more on this later). CT 3.2.3