Leveraging non-concurrent controls for time-to-event endpoints in platform trials for infectious diseases

Platform trials offer an efficient framework for evaluating multiple therapies against a common control, with the flexibility to add or remove investigational arms. For newly added arms, leveraging non-concurrent controls can further enhance efficiency. However, naively pooling non-concurrent controls with concurrent controls can introduce bias due to time trends. In trials involving time-to-event outcomes, methods for addressing this problem are limited. This study addresses this gap by developing and evaluating frequentist and Bayesian approaches to incorporate non-concurrent controls data into the analysis of time-to-event endpoints, with a particular focus on drug development for infectious diseases. We propose extensions to both frequentist and Bayesian methods, employing a Cox proportional-hazards model and a Bayesian piecewise exponential model, respectively. We assess the practical performance of these methods through comprehensive simulations, including a resampling-based simulation of the COVID-19 platform trial, the ACTIV-1 IM study. Our simulations demonstrate that no single method is uniformly superior; the optimal approach is contingent on the specific characteristics of the underlying time trend. This finding highlights that the analytical method must be carefully selected based on the anticipated time trend and reinforces that naive pooling should be avoided. Therefore, we recommend pre-specifying sensitivity analyses using a range of methods to ensure the robustness of the trial conclusions.