Nonparametric tests for treatment effect heterogeneity in observational studies

TL;DR

This paper introduces a nonparametric test for treatment effect heterogeneity in observational studies, leveraging reweighted data and multisample U-statistics, without relying on parametric outcome models.

Contribution

It develops a novel nonparametric testing method that accounts for confounders and does not require modeling treatment effects for subgroups.

Findings

The test is asymptotically normal.

It outperforms competing methods in simulations.

Demonstrated on real-world employment and mental health data.

Abstract

We consider the problem of testing for treatment effect heterogeneity in observational studies, and propose a nonparametric test based on multisample U-statistics. To account for potential confounders, we use reweighted data where the weights are determined by estimated propensity scores. The proposed method does not require any parametric assumptions on the outcomes and bypasses the need for modeling the treatment effect for each study subgroup. We establish the asymptotic normality for the test statistic, and demonstrate its superior numerical performance over several competing approaches via simulation studies. Two real data applications including an employment program evaluation study and a mental health study of China's one-child policy are also discussed.