Primary Submission Category: Machine Learning and Causal Inference
Augmented Balancing Estimators of the Average Treatment Effect on the Treated in cross-sectional and panel data
Authors: Apoorva Lal,
Presenting Author: Apoorva Lal*
Recent developments in the use of machine learning methods for causal inference typically target the average treatment effect (ATE) and frequently rely on estimating a propensity score using nonparametric regression learners and inverting it to plug into the doubly-robust IPW score. In observational studies, however, the ATE is frequently difficult to target because of the failure of overlap, which is compounded by the inversion step; researchers often target the average treatment effect on the treated (ATT) in such cases. We propose a unified framework for augmented balancing estimators for the ATT in a wide variety of research designs used by applied researchers, including cross-sectional, two-period difference in differences, and longitudinal data settings. We propose set of estimators that combine doubly-robust estimators for the ATT with balancing weights that directly targets in-sample covariate balance. In simulation studies, we find that balancing weights outperform conventional estimators that involve inverting a propensity score, and conclude with empirical applications.