Primary Submission Category: Generalizability/Transportability
Beyond Covariate Shift: Fusing Trial Data for Treatment Comparisons” to: “Trial Data Fusion: Indirect Comparisons Beyond Covariate Shift
Authors: Kuan-Hung Yeh, Ronghui Xu, Siddharth Singh,
Presenting Author: Kuan-Hung Yeh*
Data fusion methods enable indirect treatment comparisons by combining evidence from randomized trials without head-to-head comparisons. Most existing approaches assume covariate shift, meaning that cross-trial differences are fully explained by observed covariates and that covariate–outcome relationships remain invariant across populations. This assumption may be oversimplified, as it ignores conditional shift, where these relationships differ between trials, potentially leading to biased treatment effect estimates in target populations.
We develop a unified data fusion framework that accommodates both covariate and conditional shift when two randomized trials share a common arm. In this setting, we establish identifiability of average treatment effects in a target trial population under a relaxed set of assumptions that allow for changes in both covariate and conditional outcome distributions. We introduce two estimators, a weighting-based and a resampling-based estimator recently proposed in literature, that leverage discriminative learning to estimate joint density ratios for both shifts. We evaluate their finite-sample performance through simulation studies. Results show that existing estimators can exhibit substantial bias in the presence of conditional shift, whereas proposed methods remain robust. An application to inflammatory bowel disease trials illustrates the practice of this framework in comparative effectiveness research when direct comparisons are unavailable.