Title:A Nonparametric Bayesian Methodology for Regression Discontinuity Designs

Abstract:Regression discontinuity designs (RDDs) are natural experiments where treatment assignment is determined by a covariate value (or "running variable") being above or below a predetermined threshold. Because the treatment effect will be confounded by the running variable, RDD analyses focus on the local average treatment effect (LATE) at the threshold. The most popular methodology for estimating the LATE in an RDD is local linear regression (LLR), which is a weighted linear regression that places larger weight on units closer to the threshold. While LLR exhibits promising bias-reducing properties, LLR tends to yield confidence intervals that undercover, in part because LLR assumes the weighting scheme is fixed, when really there is uncertainty in the weighting scheme choice. We propose an alternative nonparametric methodology utilizing Gaussian process regression that, unlike LLR, (1) does not require specifying a functional form for the expected treatment and control response, and (2) automatically incorporates the uncertainty in how units are weighted when estimating the LATE. We prove our methodology is consistent for the LATE, and we replicate previous simulation studies in the literature to show that our method exhibits better coverage and mean squared error properties than current methodologies.