Extensions of the concept of exchangeability and their applications

Abstract

Permutation tests provide exact p-values in a wide variety of practical testing situations. But permutation tests rely on the assumption of exchangeability, that is, under the hypothesis, the joint distribution of the observations is invariant under permutations of the subscripts. Observations are exchangeable if they are independent, identically distributed (i.i.d.), or if they are jointly normal with identical covariances. The range of applications of these exact, powerful, distribution-free tests can be enlarged through exchangeability-preserving transforms, asymptotic exchangeability, partial exchangeability, and weak exchangeability. Original exact tests for comparing the slopes of two regression lines and for the analysis of two-factor experimental designs are presented. Copyright © 2002 JMASM, Inc.