Estimating the Structure of Social Interactions Using Panel Data

Abstract

I consider settings where outcomes depend on own characteristics and on the characteristics of other individuals in the sample. I propose a method to identify individuals generating spillovers and their strength using panel data on outcomes and characteristics. This is in contrast to existing approaches, which require a priori knowledge of the structure of interactions. The method is suitable when the structure of interactions is stable over time and few individuals generate spillovers distinct in magnitude from the rest. To estimate the model, I introduce the Pooled Lasso estimator, a panel- data counterpart of the Lasso estimator and develop an iterative algorithm for computation that alternates between Lasso estimation and pooled panel regression. Average marginal spillover effects across individuals enjoy gains in the rate of convergence under independence of model selection noise. I apply this methodology to study technological spillovers in productivity in a panel of US firms. I find evidence that spillovers are asymmetric across firms, arising mostly from small, highly productive firms.