Estimating Heterogeneous Treatment Effects for General Responses

TL;DR

This paper introduces DINA, a new method for estimating heterogeneous treatment effects in exponential family models, enhancing practical modeling for binary and survival outcomes with robust machine learning integration.

Contribution

The paper proposes DINA, a novel measure for treatment effects in exponential families, and a meta-algorithm enabling flexible, robust estimation using existing machine learning tools.

Findings

DINA effectively models treatment effects for binary and survival data.

The meta-algorithm improves robustness to nuisance function estimation errors.

Empirical results show strong performance on simulated and real datasets.

Abstract

Heterogeneous treatment effect models allow us to compare treatments at subgroup and individual levels, and are of increasing popularity in applications like personalized medicine, advertising, and education. In this talk, we first survey different causal estimands used in practice, which focus on estimating the difference in conditional means. We then propose DINA, the difference in natural parameters, to quantify heterogeneous treatment effect in exponential families and the Cox model. For binary outcomes and survival times, DINA is both convenient and more practical for modeling the influence of covariates on the treatment effect. Second, we introduce a meta-algorithm for DINA, which allows practitioners to use powerful off-the-shelf machine learning tools for the estimation of nuisance functions, and which is also statistically robust to errors in inaccurate nuisance function estimation. We demonstrate the efficacy of our method combined with various machine learning base-learners on simulated and real datasets.