Exclusion bias in empirical social interaction models: causes, consequences and solutions

This paper formalises an unproven source of ordinary least squares estimation bias in standard linear-in-means peer effects models. I derive a formula for the magnitude of the bias and discuss its underlying parameters. I show the conditions under which the bias is aggravated in models adding cluster fixed effects and demonstrate how it affects inference and interpretation of estimation results. Further, I reveal that two-stage least squares (2SLS) estimation strategies eliminate the bias and provide illustrative simulations. The results may explain some counter-intuitive findings in the social interaction literature, such as the observation of OLS estimates of endogenous peer effects that are larger than their 2SLS counterparts.