Validating Surrogate Endpoints of Clinical Trials

Abstract

Clinical trials are often constructed with surrogate endpoints for practical or cost considerations, for example, lipid levels as a surrogate for arteriosclerosis, arrhythmias for coronary artery disease, and cervical smears for tubal infections (Pratt and Moye 1995; Canner et al. 1986; Riggs et al. 1990; Fleming and DeMets 1996; Boissel and Hc 1992). Such trials make inferences from surrogate observations about the effect of treatments on the supposed true endpoints without accounting the strength of association between the surrogate and true endpoints. The main problem with this practice is that the surrogate endpoint may lack sufficient validity to predict the true endpoint, giving rise to misleading trial results. The International Conference of Harmonisation (ICH) Guideline E9 Statistics Principles for Clinical Trials (Philips and Haudiquet 2003) recommends that, for the approval of a surrogate marker, (1) a statistical relationship with the true endpoint in observational studies be demonstrated, (2) evidence be given from clinical trials that treatment effects on the surrogate correspond to those on the true clinical endpoint, and (3) the surrogate marker like a diagnostic test be tested for sensitivity and specificity to predict the true endpoint. There is, thus, considerable consensus to routinely assess the accuracy of surrogate markers, but not specifically how to do so. Problems with the current sensitivity-specificity approach to validity is, that it is dual and that an overall level of validity is, therefore, hard to give (Cleophas 2005). Also, it can be used for binary (yes/no) endpoints only. As an alternative, regression-models have been proposed. (Philips and Haudiquet 2003; Chen et al. 2003) However, a correlation of borderline statistical significance between the surrogate and the true endpoint is not enough to indicate that the surrogate is an accurate predictor. The current chapter underscores the need for accuracy assessment of surrogate endpoints by comparing the required sample sizes of trials with and without surrogate endpoints, and describes two novel procedures for assessment. The first makes use of an overall level of accuracy with confidence intervals and a prespecified boundary of accuracy. The second uses a regression model that accounts both the association between the surrogate and the true endpoint, and the association between either of these variables and the treatments to be tested.