Bayesian meta-analytical methods to incorporate multiple surrogate endpoints in drug development process

摘要

A number of meta-analytical methods have been proposed that aim to evaluate surrogate endpoints. Bivariate\udmeta-analytical methods can be used to predict the treatment effect for the final outcome from the treatment\udeffect estimate measured on the surrogate endpoint while taking into account the uncertainty around the effect\udestimate for the surrogate endpoint. In this paper, extensions to multivariate models are developed aiming to include\udmultiple surrogate endpoints with the potential benefit of reducing the uncertainty when making predictions. In\udthis Bayesian multivariate meta-analytic framework, the between-study variability is modelled in a formulation\udof a product of normal univariate distributions. This formulation is particularly convenient for including multiple\udsurrogate endpoints and flexible for modelling the outcomes which can be surrogate endpoints to the final outcome\udand potentially to one another. Two models are proposed, first using an unstructured between-study covariance\udmatrix by assuming the treatment effects on all outcomes are correlated and second using a structured between-\udstudy covariance matrix by assuming treatment effects on some of the outcomes are conditionally independent.\udWhile the two models are developed for the summary data on a study level, the individual-level association is taken\udinto account by the use of the Prentice’s criteria (obtained from individual patient data) to inform the within study\udcorrelations in the models. The modelling techniques are investigated using an example in relapsing remitting\udmultiple sclerosis where the disability worsening is the final outcome, while relapse rate and MRI lesions are\udpotential surrogates to the disability progression.