Title: On the Evaluation of Surrogate Markers in Real World Data Settings

Abstract: Shortcomings of randomized clinical trials are pronounced in urgent health crises, when rapid identification of effective treatments is critical. Leveraging short-term surrogates in real-world data (RWD) can guide policymakers evaluating new treatments. In this paper, we develop novel estimators for the proportion of treatment effect (PTE) on the true outcome explained by a surrogate in RWD settings. We propose inverse probability weighted and doubly robust (DR) estimators of an optimal transformation of the surrogate and PTE by semi-nonparametrically modeling the relationship between the true outcome and surrogate given baseline covariates. We show that our estimators are consistent and asymptotically normal, and the DR estimator is consistent when either the propensity score model or outcome regression model is correctly specified. We compare our proposed estimators to existing estimators and show a reduction in bias and gains in efficiency through simulations. We illustrate the utility of our method in obtaining an interpretable PTE by conducting a cross-trial comparison of two biologic therapies for ulcerative colitis.