A Bayesian Model for Inferring Overall Survival Benefit from Progression-Free Survival in Oncology Trials

In oncology trials for indolent cancers, demonstrating a statistically significant overall survival (OS) benefit is often infeasible due to long survival times and limited follow-up. Current regulatory approaches focus on monitoring OS to rule out unacceptable harm, borrowing from the cardiovascular safety framework used in diabetes trials. However, unlike HbA1c reduction in diabetes, in many cases progression-free survival (PFS) has no independent clinical value apart from its relationship to survival. Therefore, we argue that the more relevant question is not whether OS harm can be excluded, but whether there is sufficient reason to believe that the treatment improves survival.We propose a Bayesian model that formalizes the belief that a PFS benefit translates into an OS benefit and combines this belief with observed trial data to produce a posterior probability of OS benefit. The model has three parameters: r, the proportion of the PFS benefit (on the log-hazard ratio scale) expected to be preserved as an OS benefit; v, a variance capturing uncertainty in this relationship; and a threshold posterior probability for concluding benefit. We introduce a triangulation approach for calibrating these parameters based on hypothetical boundary cases and illustrate the model with analysis and design examples.