Nonparametric causal inference: A preliminary bake-off
Jennifer Hill, Ph.D.
New York University
Advances in the ability to flexibly model response surfaces have led to exploration of robust approaches to causal inference that do not require matching, weighting or subclassification on the propensity score. Tradeoffs exist between these methods however with regard to their performance and flexibility, particularly when the data are high-dimensional. This work examines the strengths and weaknesses of competing approaches to nonparametric modeling of the causal response surface including Bayesian Additive Regression Trees and Gaussian Processes. We also explore the ability of such methods to exploit information in the propensity score to achieve double robust (or approximate double robust) properties. The efficacy of these approaches to causal inference will be compared with respect to bias, statistical efficiency, computational efficiency, and the ability to identify neighborhoods of common causal support.