Testing for Selection Bias

Abstract

This paper uses the control function to develop a framework for testing for selection bias. The idea behind our framework is if the usual assumptions hold for matching or IV estimators, the control function identifies the presence and magnitude of potential selection bias. Averaging this correction term with respect to appropriate weights yields the degree of selection bias for alternative treatment effects of interest. One advantage of our framework is that it motivates when is appropriate to use more efficient estimators of treatment effects, such as those based on least squares or matching. Another advantage of our approach is that it provides an estimate of the magnitude of the selection bias. We also show how this estimate can help when trying to infer program impacts for program participants not covered by LATE estimates.

Keywords: selection bias, program evaluation, average treatment effects

JEL Classification: C21, C26, D04