arXiv · 2611.18970
A Distributionally Robust Framework for Safe and Generalizable Utilization of Surrogate Markers
Abstract
Surrogate endpoints can accelerate treatment evaluation, but their validity is challenging to establish because the joint distribution of counterfactual outcome surrogate pairs is not identifiable. We propose a distributionally robust framework for safe and generalizable surrogate evaluation that explicitly accounts for this uncertainty. By treating cross arm dependence as partially identified, we formulate a maximin optimization problem that learns an optimal surrogate transformation by maximizing the worst case explained variance of individual treatment effects over admissible counterfactual dependence structures. The resulting robust surrogacy measure provides a conservative lower bound on explained treatment effect variation. Moreover, our maximin construction guarantees nonnegative alignment between the treatment effects on the transformed surrogate and the primary outcome over the admissible counterfactual couplings. To generalize surrogacy evaluation across populations under covariate shift, we integrate transfer learning with debiased machine learning to estimate the required target population moments using source trial data and target baseline covariates. The computational tractability and asymptotic properties of our method are established. Simulations demonstrate good finite sample performance and show that the proposed method guards against counterfactual misalignment by shrinking the learned transformation toward zero, revealing a robustness prediction tradeoff. We illustrate the framework by evaluating 18 month glomerular filtration rate change as a surrogate for two year change using multiple randomized clinical trial data sets.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Molei Liu, Tian Jiang, Keyao Zhan, Zirui Zhang, Xuan Wang. 2026-11-13. A Distributionally Robust Framework for Safe and Generalizable Utilization of Surrogate Markers. https://searcharxiv.com/abs/2611.18970
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.