Learning Conditional Average Treatment Effects in Regression Discontinuity Designs using Bayesian Additive Regression Trees

BART (Bayesian additive regression trees) has been established as a leading supervised learning method, particularly in the field of causal inference. This paper explores the use of BART models for learning conditional average treatment effects (CATE) from regression discontinuity designs, where treatment assignment is based on whether an observed covariate (called the running variable) exceeds a pre-specified threshold. A purpose-built version of BART that uses linear regression leaf models (of the running variable and treatment assignment dummy) is shown to out-perform off-the-shelf BART implementations as well as a local polynomial regression approach and a CART-based approach. The new method is evaluated in thorough simulation studies as well as an empirical application looking at the effect of academic probation on student performance.