Classification in the Presence of Heterogeneous Groups: A Monte Carlo Comparison of Several Statistical Methods

• In Event: Alternative Methods for Data Analysis

Abstract

Abstract

Prediction of group membership is common in social science research. Many methods exist for such applications, and have been studied extensively in terms of their accuracy for group prediction. Nearly all of this prior research has assumed that the known groups are homogeneous. However, there is evidence that in practice actual groups may be much more heterogeneous than this previous work would imply. Thus, the goal of this simulation study was to examine the performance of several categorical prediction methods when known groups were heterogeneous with respect to the predictor variables. Results show that classification and regression trees and mixture discriminant analysis were particularly accurate in terms of classification. Complete results and their implications are discussed.

Authors

• William Holmes Finch, Ball State University

• Jocelyn E. Holden, Ball State University

• Ken Kelley, University of Notre Dame

• John Michael Starling, Ball State University