Bias-Aware Inference in Fuzzy Regression Discontinuity Designs

Abstract

We consider the problem of constructing honest, or uniformly valid, confidence sets for treatment effects in fuzzy regression discontinuity designs based on local linear regression. We show that confidence sets based on conventional t-statistics cannot be honest under certain conditions commonly encountered in practice, such as weak identification or a discrete running variable. We therefore propose confidence sets based on an Anderson-Rubin-type approach. The confidence sets explicitly takes into account the finite-sample bias of the estimators from which they are constructed, and are honest under both strong and weak identification, as well as with both a discrete and a continuously distributed running variable. We illustrate our method through simulations and an empirical application.(joint with Claudia Noack.)