ASSESSING THE RELATIVE CONTRIBUTION OF VARIABLES IN CANONICAL DISCRIMINANT-ANALYSIS

A procedure by which subsets of variables can be selected and their re lative importance assessed in canonical discriminant analysis (CDA) is discussed. It is apparent from a review of the literature that only t he standardized coefficients can be used for this purpose. However, an example relative to morphological variation of chives (Allium schoeno prasum L.) in eastern North America indicates that it would be hazardo us, at least with small sample size, to select variables by the magnit ude of their standardized coefficients alone, considering their exceed ing instability when different sets of observations are included in th e analysis. We argue that selection of variables, to be conducted with confidence, should be based upon t-values associated with jackknife s tandardized coefficients.