The Nonparametric Behrens-Fisher Project

Child Health And Human Development

Investigators

Abstract

Tests of the Nonparametric Behrens-Fisher Hypothesis are considered. In the case without censoring the Likelihood Ratio Test is obtained using Empirical Likelihood and seen to be more powerful than competitors. The distribution under the null hypothesis is approximated by simulation from the constrained EMLE. In the case with random right-censoring the Likelihood Ratio Test is again obtained using Empirical Likelihood and seen to be more powerful than the logrank. The distribution under the identity null hypothesis is approximated by special imputed permutations. In the case without censoring a semi-parametric model is used to obtain tests that are generally more powerful than the ELRT. The model is fit to the observations tranformed by the family of Box-Cox.