A comparison of two methods for adaptive interim analyses in clinical trials.

Recently, two methods for planning and conducting two-stage procedures were proposed (Bauer and Köhne, 1994, Biometrics 50, 1029-1041; Proschan and Hunsberger, 1995, Biometrics 51, 1315-1324). Both procedures allow the termination of the trial with the early acceptance of H0 in the absence of a treatment effect after performing the first stage of the study. Furthermore, the observed treatment effect at stage I can be used for planning and redesigning the second stage of the study in a way that protects the Type I error rate. The exact Type I error rate of the Proschan and Hunsberger approach is derived. It is shown that the two methods lead to similar decision rules with negligibly small differences in power and expected sample size. In terms of providing design tools and practical applicability, however, they differ. The practical performance of the procedures is discussed and recommendations for their use are given.