Subspace Learning for Conditional Average Treatment Effect Estimation with Unmeasured Confounding

Han Qiao

Keywords: Causal Effect Estimation, Data Fusion, Unmeasured Confounding

Abstract: Conditional average treatment effect (CATE) is the average causal effect of a treatment or an intervention (e.g., medication) on the outcome of interest, conditional on subjects' covariates. A key challenge in estimating causal effects from observational (OBS) data is to address unmeasured confounding. Mainstream methods that only rely on OBS data including sensitivity analysis, front door adjustment methods, and instrumental variables methods, may depend on strong assumptions. Recent studies suggest using randomized controlled trial (RCT) data to correct CATE estimates from biased OBS data, but existing methods may fail to efficiently utilize both data. In this paper, we present an end-to-end CATE estimation framework that addresses unmeasured confounding bias from OBS data using insights from limited unbiased RCT data. By learning representations from RCT data accounting for unmeasured confounding, our approach achieves unbiased CATE estimation. Our adaptive model structure mitigates overfitting and ensures performance across different RCT sample sizes. Extensive experiments on different datasets validate the effectiveness of the framework.

Submission Number: 26