Comparative assessment of trial-level surrogacy measures for candidate time-to-event surrogate endpoints in clinical trials

Various meta-analytical approaches have been applied to evaluate putative surrogate endpoints (S) of primary clinical endpoints (T), however a systematic assessment of their performance is lacking. Existing methods in the meta-analytic framework can be grouped into two types-conventional and model-based trial-level surrogacy (TLS) measures. Both conventional and model-based TLS measures assess the ability to predict the treatment effect on T based on an observed treatment effect on putative S. Conventional TLS measures include correlation coefficients and R-square measures from weighted linear regression. Model-based TLS includes Copula R^2 proposed by Burzykowski et al. (2001). We examined and compared the estimation performance of these frequently used surrogacy measures in a large-scale simulation study. The impact of several key factors on the estimation performance was assessed, including the strength of the true surrogacy, the amount of effective information provided by available data, and the range of within-trial treatment effect on S and T. The TLS can be estimated accurately and precisely by both types of surrogacy measures when the true surrogacy is strong, number of trials is large, and the range of within-trial treatment effects is wide. When one or more factors deviate from the ''best'' scenarios, both types of TLS measures tend to underestimate the true surrogacy with increased variability. The estimation performance of conventional measures is similar to model-based measures, but with higher computational efficiency. The findings are applied to a large individual patient data pooled analysis in colon cancer.