An adaptive permutation test for detecting treatment effect in multicenter clinical trials

An adaptive permutation test is developed for detecting treatment effect in the context of balanced, complete multicenter clinical trials. The proposed adaptive test computes a statistic for each center that measures the tail length of the response data in a center. Based on this statistic, the adaptive test will either winsorize the response data in a center or leave it unchanged. The traditional F test for performing a test of significance on a subset of regression coefficients in a linear model is then performed on the winsorized data, and a permutation method is used to compute the p-value of the adaptive test. Simulation studies are employed to compare the size and power of the adaptive test and the traditional F test. The results indicate that the adaptive test maintains an actual significance level close to its nominal significance level. Moreover, the adaptive test is seen to have significantly greater power than the traditional F test for long-tailed and highly skewed error distributions, with only slight loss in power for normal and short to moderate-tailed error distributions.