Blinded sample size reassessment in non-inferiority and equivalence trials.

Friede, Tim and Kieser, Meinhard (2003) Blinded sample size reassessment in non-inferiority and equivalence trials. Statistics in Medicine, 22 (6). pp. 995-1007. ISSN 1097-0258

Abstract

Even in situations where the design and conduct of clinical trials is highly standardized, there may be a considerable between-study variation in the observed variability of the primary outcome variable. As a consequence, performing a study in a fixed sample size design implies a considerable risk of resulting in a too high or too low sample size. This difficulty can be alleviated by applying a design with internal pilot study. After a provisional sample size calculation in the planning stage, a portion of the planned sample is recruited and the sample size is recalculated on the basis of the observed variability. To comply with the requirement of some regulatory guidelines only blinded data should be used for the reassessment procedure. Furthermore, the effect on the type I error rate should be quantified. The current literature presents analytical results on the actual level in the t-test situation only for superiority trials. In these situations, blinded sample size recalculation does not lead to an inflation of the type I error rate. We extended the methodology to non-inferiority and equivalence trials with normally distributed outcome variable and hypotheses formulated in terms of the ratio and difference of means. Surprisingly, in contrast to the case of testing superiority, we observed actual type I error rates above the nominal level. The extent of inflation depends on the required sample size, the sample size of the internal pilot study, and the standardized equivalence or non-inferiority margin. It turned out that the elevation of the significance level is negligible for most practical situations. Nevertheless, the consequences of sample size reassessment have to be discussed case by case and regulatory concerns with respect to the actual size of the procedure cannot generally be refuted by referring to the fact that only blinded data were used.