Transfer Learning for Meta-analysis Under Covariate Shift

A team of researchers led by Zilong Wang has introduced a new statistical machine learning framework designed to solve a critical flaw in medical research: the fact that randomized controlled trials (RCTs) often don't represent the real-world populations where treatments are ultimately used. This mismatch, known as covariate shift, can invalidate standard meta-analysis methods. Their proposed 'placebo-anchored transport' framework innovatively treats outcomes from source trials as abundant but imperfect proxy signals, while using placebo outcomes from the target population as scarce, high-fidelity 'gold labels' to calibrate a model's understanding of baseline risk.

This calibration anchors the model to the target population with a sparse correction. The anchored model is then integrated into a cross-fitted, doubly robust learner, creating a Neyman-orthogonal estimator for patient-level treatment effects. The method operates in two key regimes: for 'connected' targets with a treated arm, it provides fully identified effect estimates; for 'disconnected' targets with only placebo data, it becomes a principled screening and transport procedure. In experiments on synthetic data and the semi-synthetic IHDP benchmark, the method demonstrated superior or near-best performance across metrics like CATE accuracy, ATE error, and policy regret, showing substantial improvements over proxy-only, target-only, and standard transport baselines, particularly when target sample sizes were small.