Evaluation of the Use of Standardized Weights for Interpreting Results from a Descriptive Discriminant Analysis

Abstract

When conducting descriptive discriminant analysis, many researchers make use of structure coefficients, the correlation between individual predictor variables and a discriminant function. However, previous research has demonstrated that these statistics may lead to an over-identification of variables important for group separation. An alternative to structure coefficients is the standardized discriminant function weights for the individual variables, which can be used to order variables in importance. Relatively little empirical research has been done examining how well they work in this regard. This study examined the utility of standardized weights for interpreting a discriminant function. Results suggest that the standardized weights may be a useful tool for ordering predictor variables and characterizing significant discriminant functions when the assumptions of normality and homogeneity of covariance matrices are met. When these assumptions are violated, the ability of the standardized weights to correctly order predictor variables was somewhat degraded.