The Subset Problem in Multivariate Analysis of Variance

Summary As with most other multivariate techniques, methods such as discriminant analysis which depend on multivariate analysis of variance are affected adversely by the inclusion of irrelevant variables. This paper discusses the problem of finding a parsimonious subset of variables that is essentially as good as the full set. As a result of the close computational similarity to multiple regression, a stepwise procedure, and a best subset method are derived from their regression counterparts. The inferential aspects of the stepwise procedure are also investigated.