The Statistical Evaluation of Surrogate Endpoints in Clinical Trials

Over the past 35 years, surrogate endpoints, to be used in lieu of so-called true endpoints, have become increasingly important as potential shortcuts for the development of medicinal products. In various therapeutic areas, current-day endpoints once were surrogates that underwent favorable evaluation, such as progression-free survival in colorectal cancer, or simply replaced a former true endpoint out of necessity, such as CD4 counts and then viral load in HIV/AIDS. Of course, a surrogate endpoint will never exactly capture the same information as the endpoint it is designed to replace, and the relationship between both may change when there are drastic changes in, for example, the therapies applied. In any case, meticulous evaluation of candidate surrogates is in order. To this end and over time, a variety of approaches have been proposed to evaluate candidate surrogate endpoints, starting from the seminal work of Prentice (Stat Med 8; 431–440, 1989) and that of Freedman et al. (Stat Med 11; 167–178, 1992). The earliest work focused on a single trial, while at the turn of the century, focus shifted to the meta-analytic approach. The framework was developed for different outcome types and permits the distinction between individual-level and trial-level surrogacy. It also enables examination of the predictive power of the treatment effect on the surrogate to the treatment effect on the true endpoint. Surrogates have also entered the regulatory framework for clinical trials. More recently, causal inference methodology has entered surrogate endpoint evaluation. In this setting, a counterfactual pair of endpoints is considered for every patient, under control and experimental conditions, for the true endpoint and perhaps also for the surrogate. The lack of identification that results is addressed via sensitivity analysis. A case study in psychiatry completes the overview.