Aggregate Bayesian Causal Forests

This repository provides the code for implementing the aggregate Bayesian Causal Forest (aBCF) model. It is forked from the original bcf package (https://github.com/jaredsmurray/bcf). You can find methodological details on the underlying modeling approach in the original BCF paper: Hahn, Murray, and Carvalho 2017, and information on the adapatations made for aBCF in our forthcoming paper Thal, Forrow, Lipman, Starling, and Finucane, forthcoming

Why BCF?

BCF is a cutting-edge causal inference methodology that builds on Bayesian Additive Regression Trees (BART, Chipman, George, and McCulloch 2010). Specifically, it applies BART to causal inference (Hill 2011). BART and BCF both combine Bayesian regularization with regression trees to provide a highly flexible response surface that, thanks to the Bayesian regularizing priors, does not overfit to the training data. BCF extends BART's flexibility by specifying different models for relationships between (1) covariates and the outcome and (2) covariates and the treatment effect.

BCF performs remarkably well in simulation and has led the pack at recent rigorous causal inference competitions, such as those held at the Atlantic Causal Inference Conference (see, for example, Dorie et al. 2019).

Why aBCF?

aBCF extends BCF to work with aggregate data (e.g. school- or hospital-level data aggregated up from student- or patient-level data). In an aggregate data context we must account for heteroskedasticity of errors due to varying sample sizes, and we also want to account for homoskedastic aggregate-unit effects.

Installation

This package requires compilation, so make sure you have Rtools properly installed -- see this site for details.

Install from GitHub:

if (!require("devtools")) {
  install.packages("devtools")
}
devtools::install_github("mathematica-org/abcf")