Variable Importance in Multivariate Group Comparisons

Three views of relative variable contribution in a multivariate analysis of variance context are discussed: contribution to (a) latent construct definition, (b) linear discriminant function scores, and (c) grouping variable effects. It is argued that the third view is most meaningful. The problem of assessing the relative importance of a set of outcome variables is reviewed. Two numerical ranking methods are proposed. The bootstrap approach is used to compare the methods utilizing two real data sets. No definitive preference is advanced. The variable-ordering problem in two other multivariate analysis contexts is briefly discussed.