Title: Bayesian tree models with targeted smoothing for causal inference
Authors: Jared Murray - University of Texas at Austin (United States) [presenting]
Abstract: Bayesian tree models like Bayesian additive regression trees (BART) and Bayesian causal forests (BCF) are popular and effective methods for inferring heterogeneous causal effects. However, their function estimates are necessarily discontinuous and ``rough'' in their arguments, a significant disadvantage in applications involving continuous treatments or effect moderators thought to have smoothly evolving relationships with treatment efficacy. We extend Bayesian tree models with ``targeted smoothing'' to allow for (possibly) irregularly spaced continuous treatment variables or moderators while maintaining computational efficiency through the use of carefully constructed basis expansions.