A criterion for variable selection in multiple discriminant analysis (Q788433)

In this paper the problem of selection of variables in linear discriminant analysis is considered. A criterion is proposed for determining the ''best'' subset of variables, the aim of which is to interpret the separation among the population in terms of a few discriminant variables. This problem is solved by searching among a set of models with a criterion based on \textit{H. Akaike's} concept of information [IEEE Trans. Autom. Control AC-19, 716-723 (1974; Zbl 0314.62039)]. The selected model yields the best predictions for future observations with the same structure as the original observations. In a second part the paper tackles the more general problem of finding the best subset of variables in a multivariate linear model. Asymptotic distributions of the criterion for variable selection are obtained. The asymptotic distribution in the case when the variables are ordered a priori are shown to be reducible to a simple form.