Primary Submission Category: Randomization Tests
Fisher Meets BART: Integrating Causal Machine Learning with Randomization Tests
Authors: JungHo Lee, Panos Toulis, David Puelz,
Presenting Author: JungHo Lee*
The Fisher randomization test provides an attractive, robust testing methodology that is finite-sample valid. However, it cannot be immediately applied for testing non-sharp, or weak, null hypotheses that are often of greater empirical interest. This paper develops an approach for testing a general class of weak null hypotheses. The key idea is to use a flexible causal machine learning model to provide plausible values of the individual treatment effects under the null, and then reject or accept these values through a standard randomization test. Hence, a contribution of this work is to integrate randomization testing and machine learning methods for causal inference. We demonstrate our methodology on weak null hypotheses involving the sample average treatment effect and treatment effect heterogeneity in simulated and real-world scenarios.