What does outcome similarity capture? Identification and inference in pairwise regressions of peer effects.

Abstract

A growing number of empirical studies analyze peer effects via pairwise regressions, assessing whether the similarity of individual outcomes within a pair is higher when the two agents are connected. We provide the first analysis of identification and inference in these pairwise regressions. We find major problems with this design. We show, first, that the pairwise coefficient picks up clustering and that this bias is not solved by network randomization. Second, there is no systematic relationship between the sign and size of peer effects and the sign and size of the pairwise coefficient. Third, the pairwise parameter generally lacks a causal interpretation because of interference. Finally, we establish a central limit theorem for the pairwise OLS estimator and propose a new dyadic-robust variance estimator that is valid in the presence of peer effects. (joint with Zheng Wang - NYU Abu Dhabi)