Meta-analytic evaluation of surrogate endpoints at multiple time points in randomized controlled trials with...

Valid surrogate endpoints are of great interest for efficient evaluation of novel therapies. With surrogate and true time-to-event endpoints, meta-analytic approaches for surrogacy validation commonly rely on the hazard ratio, ignore that randomized trials possibly contribute to the meta-analysis for different follow-up durations, overlook the importance of the time lag between surrogate and true endpoints in determining surrogate utility, and assume that treatment effects and the strength of surrogacy remain constant over time. In this context, we introduce a novel two-stage meta-analytic model to evaluate trial-level surrogacy.

Our model employs restricted mean survival time (RMST) differences to quantify treatment effects at the first stage. At the second stage, the model is based on the between-study covariance matrix of RMSTs and differences in RMST to assess surrogacy through coefficients of determination at multiple timepoints. This framework integrates estimates from each component RCT without extrapolation beyond the trial-specific time support, can explicitly model a time lag between endpoints, and remains valid under non-proportional hazards.

Simulation studies indicate that our model yields unbiased and precise estimates of the coefficient of determination. In an individual patient data meta-analysis in gastric cancer, estimates of coefficients of determination from our model reflect the temporal lag between endpoints and reveal dynamic changes in surrogacy strength over time compared to the Clayton survival copula model, a widely used reference method in surrogate endpoint validation for time-to-event outcomes.

Our new meta-analytic model to evaluate trial-level surrogacy using the difference in RMST as the measure of treatment effect does not require the proportional hazard assumption, captures the strength of surrogacy at multiple time points, and can evaluate surrogacy with a time lag between surrogate and true endpoints. The proposed method enhances the rigor and practicality of surrogate endpoint validation in time-to-event settings.