Dr. Zhu Wang, PhD, University of Tennessee Health Science Center

Outcome weighted learning (OWL) estimates individualized treatment rules by casting treatment selection as weighted classification. We study single-stage binary OWL with known treatment-assignment probabilities and nonnegative rewards. Our main contribution is a quantitative comparison between excess treatment risk and excess surrogate risk for general nonnegative surrogate losses. A variational transform of the loss yields this comparison for measurable scoring functions, with nontrivial bounds under a positive-definiteness condition. We relate policy calibration to classification calibration and derive simplified transforms for convex and selected nonconvex losses. We then apply the risk comparison to regularized kernel estimators, obtaining learning rates under two general approximation frameworks: smoothness assumptions for Matérn kernels and geometric-noise assumptions for Gaussian kernels. For robust binomial losses, a special bounded-nonconvex class for which smoothness assumptions can fail, we derive separate rates using sign-correct reference functions under weighted small-margin conditions. As a computational application, we develop an iteratively reweighted convex optimization algorithm for robust surrogate losses in OWL and residual weighted learning. Simulations and an analysis of the AIDS Clinical Trials Group Study 175 illustrate the methods.