Identification of Variables Associated With Group Separation in Descriptive Discriminant Analysis: Comparison of Methods for Interpreting Structure Coefficients

Discriminant Analysis (DA) is a tool commonly used for differentiating among 2 or more groups based on 2 or more predictor variables. DA works by finding 1 or more linear combinations of the predictors that yield maximal difference among the groups. One common goal of researchers using DA is to characterize the nature of group difference by interpreting the contributions of the individual predictors to this linear combination, often using structure coefficients (SC). The authors of this simulation study examine the utility of several methods for interpreting SCs. Results indicate that with samples greater than 100, a bootstrap confidence interval may be optimal, whereas with smaller samples, common rules of thumb may work best. Furthermore, nonnormal data and unequal covariance matrixes diminish the effectiveness of SCs as an interpretive tool.