Selezione delle variabili nell’analisi discriminante quadratica tra due gruppi

The purpose of the present study is to examine the problem of the selection of variables in a two-group quadratic discriminant analysis. We are particularly interested in the conditional deletion method that will be based on three different solutions of the multivariate Behrens- Fisher problem: James’s test, Yao’s test and Bennett’s test. On the basis of different sets of real and simulated data we find that a ranking procedure, when the covariances matrices are unequal and the sample sizes are both unequal and small, should be based on the James’s Test. However, The Hotelling’s T2 test, currently used for testing equality of means when covariances matrices are equal, is a valuable alternative to the James’s Test.