New Evidence on Linear Regression and Treatment Effect Heterogeneity

In this paper I provide new evidence on the implications of treatment effect heterogeneity for least squares estimation when the effects are inappropriately assumed to be homogenous. I prove that under a set of benchmark assumptions linear regression provides a consistent estimator of the population average treatment effect on the treated times the population proportion of the nontreated individuals plus the population average treatment effect on the nontreated times the population proportion of the treated individuals. Consequently, in many empirical applications the linear regression estimates might not be close to any of the standard average treatment effects of interest.