SA-Learner: Surrogate-Assisted Meta-Learner with Missing Outcomes

Yan-Yu Chen, Qixian Zhong, Jane-Ling Wang

Keywords: causal inference, heterogeneous treatment effect, missing mechanism

Abstract: Estimating heterogeneous treatment effects is essential for personalized decision-making across various applications. While existing methods primarily focus on the conditional average treatment effect (CATE) for fully observed outcomes, real-world data often suffer from missingness. Direct CATE estimation using only complete cases can introduce bias and reduce efficiency. To address these challenges, we propose the Surrogate-Assisted Learner (SA-learner), which leverages surrogate outcomes—auxiliary variables expected to predict the effect of a treatment on the primary outcome and is more readily observed—to improve CATE estimation. The SA-learner enjoys double robustness, ensuring consistent CATE estimates even under misspecification of certain nuisance functions. We also establish its convergence rate, requiring only slower-rate convergence of nuisance function estimators without restrictive model assumptions. This property enables flexible implementation using off-the-shelf machine learning algorithms. Extensive experiments on synthetic data further demonstrate effectiveness of the proposed method.

Primary Area: causal reasoning

Submission Number: 23140