Primary Submission Category: Machine Learning and Causal Inference
Efficient and Flexible Heterogeneous Treatment Effect Estimation with Random BART Features
Authors: Cory McCartan, Melody Huang,
Presenting Author: Cory McCartan*
Bayesian Additive Regression Trees (BART) models have shown promise for flexibly estimating causal response functions and heterogeneous treatment effects. However, they require custom Markov chain Monte Carlo (MCMC) sampling algorithms for computation, which limits their scalability and applicability to other model classes and data structures. We show how a recent reformulation of BART as an random-features approximation to a certain Gaussian Process can be applied to the estimation of heterogeneous treatment effects on large data in a variety of causal settings, including survival outcomes, differences-in-differences, and spatial data.