Identification and Estimation of Social Interactions through Variation in Equilibrium Influence

This paper presents a new method for estimating social interaction effects. The proposed approach is based on using network interaction structure induced variation in equilibrium influence to construct conditionally balanced interaction structures. As equilibrium influence is determined by the known interaction structure and the unknown endogenous social interaction parameter, interaction structures are constructed for different imputed values of the unknown parameter. Each constructed interaction structure is conditionally balanced in the sense that when it is combined with observations on the outcome variable to construct a new variable, the constructed variable is a valid instrumental variable for the endogenous social interaction regressor if the true and imputed parameter values are the same. Comparison of each imputed value with the associated instrumental variable estimate thus yields a confidence set estimate for the endogenous social interaction parameter as well as for other model parameters. We provide conditions for point identification and partial identification. The contrast between the proposed and existing approaches is stark. In the existing approach instruments are constructed from observations on exogenous variables, whereas in the proposed approach instruments are constructed from observations on the outcome variable. Both approaches have advantages, and the two approaches complement one another. We demonstrate the feasibility of the proposed approach with analyses of the determinants of subjective college completion and income expectations among adolescents in the Add Health data and with Monte Carlo simulations of Erdös-Rényi and small-world networks.