Assessing non-inferiority of a new treatment in a three-arm trial in the presence of heteroscedasticity

In this paper, we describe an adjusted method to facilitate non-inferiority tests in a three-arm design. While the methodology is readily available in the situation of homogeneous group variances, the adjusted method will also maintain the alpha-level in the presence of heteroscedasticity. We propose explicit criteria for an optimal allocation. Depending on the pattern of heterogeneity, remarkably unbalanced designs are power optimal. We will apply the method to a randomized clinical trial and a toxicological experiment. Copyright (c) 2007 John Wiley & Sons, Ltd.