Primary Submission Category: Machine Learning and Causal Inference
Tree Priors for Feature Selection in Average Treatment Effect Estimation
Authors: Andrew Herren, Richard Hahn,
Presenting Author: Andrew Herren*
This paper builds on previous theoretical work justifying feature selection for average treatment effects in the context of discrete covariates. We extend to continuous covariates using the adaptive discretization provided by decision trees and propose a novel decision tree prior that incorporates the estimated propensity score. The methods are justified using finite sample comparisons to existing Bayesian tree methods for treatment effect estimation as well as the broader class of machine learning estimators.