SELECTION OF VARIABLES, AND ASSESSMENT OF THEIR PERFORMANCE, IN MIXED-VARIABLE DISCRIMINANT-ANALYSIS

Two problems frequently encountered by the practitioner when faced with multivariate situations are how to select the ''best'' variables for the objective in hand, and how to assess accurately subsequent performance of the selected variables. A typical framework in which these problems arise is discriminant analysis, and the variables available are often mixtures of qualitative and quantitative ones. Various methods have been proposed in recent years for selecting variables in such mixed-variable discrimination situations, and a brief review is first given of the possibilities. Some of the methods are simply variations on the same basic underlying model (the location model); some simulation studies comparing these variations are then described. Finally, the practical assessment of performance of a variable selection method is addressed. Analysis of some real data highlights the problems that can arise.