Under the instruction of Prof. Yajuan Si at UW-Madison, I conducted my undergraduate thesis on application of BART on controlled trail data analysis.
In the context of causal inference, researcher may be interested in estimating test effect. Instead of traditional approaches like propensity score matching, Jennifer L. Hill proposed a new idea of fitting a highly flexible non-parametric method (like BART) on the observed data. One of the goals of my research was to investigate whether BART fits the interaction between covariates and treatment label, from the perspective of a prediction problem. Another goal was to explore the situation where BART is suitable for extrapolation, i.e. predicting the non-overlapping area.
The simulation code is uploaded on my Github repo, and a brief summary slides on this study is available as below: