Bayesian evidence synthesis for surrogate endpoints in the era of precision medicine

This thesis considers a range of methodological challenges related to the trial-level validation of surrogate endpoints in disease areas where precision medicine have played an important role, and aims to addressed them by proposing novel statistical methodology.

Firstly, the thesis introduces two hierarchical meta-analytic methods which allow for modeling differences in trial-level surrogacy patterns. Trial-level surrogacy patterns may vary across treatment classes due to, for example, the diversity of the mechanisms of action of targeted therapies. A simple way to examine potential differences in surrogacy patterns across treatment classes is by performing subgroup analysis using a bivariate meta-analytic method. However, this approach fails to estimate trial-level association patterns effectively when data are limited in terms of the number of studies. The two hierarchical meta-analytic methods aim to improve the inference about the parameters describing the surrogacy patterns within a treatment class as they borrow information for these parameters across classes.

Secondly, the thesis proposes a new method which is appropriate for modeling correlated binomial aggregate data with very rare or frequent events. Targeted treatments are usually much more successful compared to standard of care resulting in very high numbers of treatment responses and reduced numbers of events. When standard approaches for trial-level validation of surrogate endpoints are applied to such binomial data, they may lead to poor inferences about surrogacy patterns due to inappropriate assumptions. They transform the binomial data on the log odds ratio scale and model the within-study variability using a bivariate normal distribution as data measured on this scale are assumed to be approximately normally distributed. However, this assumption is inappropriate when events occur rarely or very frequently. The proposed hierarchical method allows for modeling the within-study variability on the original binomial scale and accounts for the within-study associations leading to more precise inferences about the trial-level surrogacy patterns. Finally, this thesis develops a hierarchical method for combining data from randomised control trials and data from single-arm observational studies in a single bivariate meta-analysis. Very often data measured on a short-term final endpoint are not sufficiently mature, or there is limited number of trials published. In these situations trial-level surrogacy patterns cannot be estimated accurately as randomised control trials do not provide sufficient information. The proposed methodology aims to improve inferences for trial-level surrogacy patterns, when randomised control trials offer limited information and evidence from different study designs need to be included for such validation. The performance of the proposed methodologies were extensively assessed and compared against the standard approaches in various simulated data scenarios. They were also illustrated in data examples where targeted treatments were used.