The Perils of Peer Effects

TLDR: In this article, the authors introduce pre-determined peer characteristics as covariates in a model linking individual outcomes with group averages, and the question of whether peer effects or social spillovers exist is econometrically identical to that of whether a 2SLS estimator using group dummies to instrument individual characteristics differs from OLS estimates of the effect of these characteristics.

Abstract: Individual outcomes are highly correlated with group average outcomes, a fact often interpreted as a causal peer effect. Without covariates, however, outcome-on-outcome peer effects are vacuous, either unity or, if the average is defined as leave-out, determined by a generic intraclass correlation coefficient. When pre-determined peer characteristics are introduced as covariates in a model linking individual outcomes with group averages, the question of whether peer effects or social spillovers exist is econometrically identical to that of whether a 2SLS estimator using group dummies to instrument individual characteristics differs from OLS estimates of the effect of these characteristics. The interpretation of results from models that rely solely on chance variation in peer groups is therefore complicated by bias from weak instruments. With systematic variation in group composition, the weak IV issue falls away, but the resulting 2SLS estimates can be expected to exceed the corresponding OLS estimates as a result of measurement error and other reasons unrelated to social effects. Randomized and quasi-experimental research designs that manipulate peer characteristics in a manner unrelated to individual characteristics provide the strongest evidence on the nature of social spillovers. As an empirical matter, designs of this sort have uncovered little in the way of socially significant causal effects.