On model prespecification in confirmatory randomized studies.

Abstract

Typically, the primary purpose of confirmatory randomized trials, such as drug trials sponsored by the pharmaceutical industry, is to determine whether there is a treatment effect, and if there is, to estimate the size of the effect. For such studies it is accepted practice to prespecify the statistical model to be used in the primary analysis. The reason for this is a concern that if the model were to be chosen on the basis of the data, the model most favourable to the sponsor might be chosen, with consequent inflation of the type I error. The purpose of this article is to show that, in a sense, this concern is needless. It is shown that if the model is chosen in a blinded fashion and randomization-based tests for no treatment effect are used, then the type I error is controlled. A similar technique to derive unbiased estimates of treatment effect is also described. This approach may be of value when there is uncertainty as to the correct model when the study is being planned.