PhD Student, Faculty of Economics, University of Cambridge
Bayesian Causal Forests using Bayesian Model Averaging
Eoghan O’Neill and Melvyn Weeks Tree based methods, including CART (Breiman et al. 1984) and Random forests (Breiman 2001) produce predictions through sets of simple decision rules selected by a greedy algorithm. Bayesian Additive Regression Trees (BART) (Chipman 2010) is a remarkably accurate prediction method that involves drawing sum-of-trees models in a MCMC back-fitting algorithm. Bayesian Causal Forests (Hahn et al. 2017) are an extension of BART to treatment effect estimation. In this paper, I extend an alternative BART fitting algorithm, BART-BMA (Hernandez et al. 2018) to treatment effect estimation. This method, Bayesian Causal Forests using Bayesian Model Averaging (BCF-BMA), is an alternative implementation of Bayesian Causal Forests. The motivation for BCF-BMA is that it may result in improvements on BCF similar to those observed for BART-BMA over BART. BCF-BMA is more feasible than current implementations of BCF when the number of covariates is very large. In addition to improved computational speed, it is hoped that this method exhibits improved accuracy when the number of covariates is large.
Eoghan O’Neill is an Economics PhD student at the University of Cambridge. He holds a BA in Mathematics and Economics from Trinity College Dublin and an MPhil in Economic Research from University of Cambridge. He is currently researching machine learning methods for the estimation of causal effects. Eoghan is also interested in energy economics and applications of machine learning to electricity smart meter data.