Transfer Learning for Meta-analysis Under Covariate Shift

arXiv stat.ML / 4/6/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper addresses how covariate shift across randomized controlled trials can invalidate standard individual-patient-data (IPD) meta-analysis and transport estimators.
It proposes a placebo-anchored transport framework that uses abundant source-trial outcomes as proxy signals and scarce target-trial placebo outcomes as high-fidelity labels to calibrate baseline risk.
A sparse, low-complexity correction anchors proxy models to the target population, and the anchored models are used within a cross-fitted doubly robust learner to produce a Neyman-orthogonal, target-site doubly robust estimator for heterogeneous treatment effects.
The method distinguishes between connected targets (with a treated arm) where effects can be target-identified, and disconnected targets (placebo-only) where it becomes a screen-then-transport approach under explicit working-model assumptions.
Experiments on synthetic data and a semi-synthetic IHDP benchmark show strong performance—especially at small target sample sizes—with good ranking for targeting in both regimes, while pointwise accuracy in disconnected settings depends on how accurate the working transport condition is.

Abstract

Randomized controlled trials often do not represent the populations where decisions are made, and covariate shift across studies can invalidate standard IPD meta-analysis and transport estimators. We propose a placebo-anchored transport framework that treats source-trial outcomes as abundant proxy signals and target-trial placebo outcomes as scarce, high-fidelity gold labels to calibrate baseline risk. A low-complexity (sparse) correction anchors proxy outcome models to the target population, and the anchored models are embedded in a cross-fitted doubly robust learner, yielding a Neyman-orthogonal, target-site doubly robust estimator for patient-level heterogeneous treatment effects when target treated outcomes are available. We distinguish two regimes: in connected targets (with a treated arm), the method yields target-identified effect estimates; in disconnected targets (placebo-only), it reduces to a principled screen--then--transport procedure under explicit working-model transport assumptions. Experiments on synthetic data and a semi-synthetic IHDP benchmark evaluate pointwise CATE accuracy, ATE error, ranking quality for targeting, decision-theoretic policy regret, and calibration. Across connected settings, the proposed method is best or near-best and improves substantially over proxy-only, target-only, and transport baselines at small target sample sizes; in disconnected settings, it retains strong ranking performance for targeting while pointwise accuracy depends on the strength of the working transport condition.