On group sequential tests based on robust location and scale estimators in the two-sample problem

Abstract

The behaviour of group sequential tests in the two-sample problem is investigated if one replaces the classical non-robust estimators in the t-test statistic by modern robust estimators of location and scale. Hampel’s 3-part redescending M-estimator 25A used in the Princeton study and the robust scale estimators length of the shortest half proposed by Rousseeuw and Leroy and Q proposed by Rousseeuw and Croux are considered. Of special interest are level, power, average sample size number of the tests, and the bias of the estimated standardized treatment difference. It is investigated, whether commerical software can be used to apply these tests.