A treatment benefit predictor is a function that maps patient characteristics to a putative treatment benefit for that patient. Such predictors support the optimization of individualized treatment decisions, a central idea of precision medicine. However, evaluating the predictive performance of a treatment benefit predictor is challenging, as we often cannot observe each individual's treatment benefit. This thesis theoretically underpins common predictive metrics and demonstrates conceptual and practical evaluation of prespecified treatment benefit predictors in the target population. This work focuses on a binary treatment decision made at a single time point. At a conceptual level, we define the estimands of a set of predictive performance metrics. A particular measure of discrimination is used as an illustrative example to reveal methodological concerns on multiple fronts. This example highlights specific aspects that require particular care when developing metrics to evaluate treatment benefit predictors, such as properness. We describe how to evaluate a treatment benefit predictor using observational data from the target population and explore how predictive performance metrics may change when confounding is not fully controlled. We exemplify performance assessment in the context of observational studies using another discrimination measure and a calibration measure. In the absence of full confounding control, we show that bias propagates in a more complex manner than when targeting more commonly encountered estimands. In practice, we propose and implement estimation methods for evaluating predictive performance of treatment benefit predictors, assessing their reliability through simulation studies. We then illustrate their practical use in real-world observation data, including cohort construction and modeling strategies. In general, this work helps bridge the gap between predictive modeling and causal inference, providing a framework for evaluating treatment benefit predictors using predictive performance metrics.
Yuan Xia (Thu,) studied this question.