New Method for Policy Learning with Missing Treatment Data

Policy learning is crucial for allocating treatments under budget constraints, but real-world applications often face missing treatment data, which can lead to biased estimates and suboptimal decisions. This paper introduces an extension of efficient estimators for average treatment effect (ATE) to policy value and conditional average treatment effect (CATE) when treatment data is missing at random (MAR) or missing completely conditionally at random (MCCAR). The research demonstrates that the MAR estimator, which utilizes partially observed units, is both valid and more efficient than the MCCAR estimator, even when MCCAR assumptions are met. This finding provides a formal justification for prioritizing MAR-based estimation in policy learning across various missing data scenarios. Extensive experiments using synthetic and semi-synthetic datasets confirm that accurate specification of the missingness mechanism is vital. Misspecified estimators consistently produce biased results, whereas the proposed estimators achieve near-optimal performance when their underlying assumptions are satisfied, offering practitioners reliable tools for robust policy learning.