Nonparametric Tests for Umbrella Alternatives in Stratified Datasets

Abstract

This thesis considers the problem of hypothesis testing for umbrella alternatives when there are two groups, or strata, of observations. The proposed methods extend a previously established general framework of hypothesis testing based on rankings to stratified datasets by first aligning the strata. The tests based on the Spearman and Kendall distances between ranking vectors lead to the traditional aligned-rank tests and new methods which account for “misalignment” under the alternative hypothesis. Asymptotic null distributions and simulation studies are given for the Spearman distance. Diagnostic tools for the misalignment issue are illustrated alongside the proposed tests on a dataset of IQ scores of coma patients. Extensions to three or more strata and ”adaptive” tests are provided as future research directions.