Extracting Post-Treatment Covariates for Heterogeneous Treatment Effect Estimation

Qiang Huang, Defu Cao, Yi Chang, Yan Liu

Primary Area: causal reasoning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Causality, Deep Learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: The exploration of causal relationships between treatments and outcomes, and the estimation of causal effects from observational data, have garnered considerable interest in the scientific community in recent years. However, traditional causal inference methods implicitly assume that all covariates are measured prior to treatment assignment, while in many real-world scenarios, some covariates are affected by the treatment and collected post-treatment. In this paper, we demonstrate how ignoring or mishandling post-treatment covariates can lead to biased estimates of heterogeneous treatment effects, referred to as the "post-treatment bias" problem. We discuss the possible cases in which post-treatment bias may appear and the negative impact it can have on causal effect estimation. Methodologically, we propose a novel variable decomposition approach to account for post-treatment covariates and eliminate post-treatment bias, based on a newly proposed causal graph for post-treatment causal inference analyses. Extensive experiments on synthetic, semi-synthetic, and real-world data demonstrate the superiority of our proposed method over state-of-the-art models for heterogeneous treatment effect estimation.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6641