A projection-based model checking for heterogeneous treatment effect

TL;DR

This paper introduces a projection-based model checking method for assessing the significance of covariates on heterogeneous treatment effects, effectively handling high-dimensional data and avoiding slow convergence issues.

Contribution

It proposes a novel projection-based, nonparametric test that is robust to high dimensionality and sensitive to a wide range of alternative hypotheses.

Findings

Test performs well in high-dimensional settings.

Method detects broad classes of local alternatives.

Numerical studies and real data illustrate effectiveness.

Abstract

In this paper, we investigate the hypothesis testing problem that checks whether part of covariates / confounders significantly affect the heterogeneous treatment effect given all covariates. This model checking is particularly useful in the case where there are many collected covariates such that we can possibly alleviate the typical curse of dimensionality. In the test construction procedure, we use a projection-based idea and a nonparametric estimation-based test procedure to construct an aggregated version over all projection directions. The resulting test statistic is then interestingly with no effect from slow convergence rate the nonparametric estimation usually suffers from. This feature makes the test behave like a global smoothing test to have ability to detect a broad class of local alternatives converging to the null at the fastest possible rate in hypothesis testing. Also, the test can inherit the merit of lobal smoothing tests to be sensitive to oscillating alternative models. The performance of the test is examined by numerical studies and the analysis for a real data example for illustration.