Classification Methods—Part 3. Inferential Considerations in the MANOVA

Abstract

I am going to conclude the general topic of classification and discrimination with a consideration of null hypothesis testing. Much of this chapter deals with the multivariate analysis of variance (MANOVA) and related themes. I have mentioned earlier at several points of the text that testing a multivariate hypothesis of centroid (vector, profile) differences is more complex than testing a univariate hypothesis of mean (location) difference. The basic point to remember is that an inferential test that is the most powerful for detecting a difference when centroids are concentrated is not necessarily the most powerful test for detecting a difference when centroids were diffuse. The general strategy is to treat all unknown differences in structure as if they are diffuse.