A SEQUENTIAL CLUSTERING STRATEGY FOR CLASSIFYING GENE BANK ACCESSIONS

Statistical classification methods such as Normix, applied to groups p reviously formed (a priori groups) using hierarchical methods, ean imp rove the a priori groups and give estimation of probabilities Far memb ership. In this study, we used data from the evaluation of 29 maize (Z ea mays L.) races to (i) apply Normix to the a priori groups formed by the Ward method (Normix after Ward), (ii) use tile upper tail approac h and a likelihood ratio test to estimate the optimal number of groups , and (iii) use three geometrical and statistical criteria (Mahalanobi s distances, multivariate statistics, and trace and determinant of the within- and among-group variability matrices) to compare the groups o btained from Ward per se and the Normix after Ward method. Results ind icated that the reclassification of accessions by Normix substantially improved the a priori groups obtained by Ward in four ways: (i) the a verage Mahalanobis distance between groups was larger among Normix tha n among Ward groups; (ii) the ratio of variability between groups to v ariability within groups is increased after Normix compared with that obtained after Ward; (iii) the generalized variance within groups is d ecreased in Normix compared with Ward; and (iv) the correlation coeffi cients between variables are increased after Normix compared with thos e correlation values obtained by Ward. Results suggested that the grou ps Foamed by the Normix after Ward procedure allows (i) a good estimat ion of the optical group numbers, (ii) formation of more compact and s eparates: groups than the a priori groups formed by Ward per se, and ( iii) estimation of the probability of accession membership to the grou ps.