View by Day

Benjamin Elsner, Ph.D.

University College Dublin, Ireland, and IZA, Bonn, Germany

Authors: Benjamin Elsner and Ingo Isphording

Abstract: This paper points to a fundamental identification problem in the estimation of ability peer effects. Even when peer groups are randomly assigned, obtaining a consistent, unbiased peer effect estimate is difficult because every person in the sample is "treated'' with a distribution of peer characteristics rather than a single binary or continuous treatment. Most peer effects regressions include one or several statistics of this distribution --- often the mean or variance of peer ability --- but inevitably omit many other statistics that are mechanically correlated with those included, such as higher moments or statistics describing a person's position within the group. In Monte Carlo simulations, we show that the direction of the omitted variable bias is ambiguous, and that commonly found non-linear effects can be the result of fundamentally different data-generating processes. Moreover, using data from five widely cited experimental studies, we demonstrate a strong sensitivity of peer effects estimates to omitting distributional features from the regression. These results raise doubts whether research designs that exploit natural variation in peer composition are suitable for obtaining unbiased causal peer effects estimates.

Full Text:  WILL BE AVAILABLE LATER