Accuracies of Logistic Regression, Linear Discriminant Analysis, and Classification and Regression Trees Under Controlled Conditions

• In Event: 77.056 - Fairmont Royal York Roundtable Session Seven
  In Roundtable Session: 77.056-2 - Data Mining and Unconventional Data Sources

Abstract

Logistic Regression (LR), Linear Discriminant Analysis (LDA), and Classification and Regression Trees (CART) are common classification techniques for prediction of group membership. To evaluate the performance of these methods under different controlled conditions, data were simulated with Monte Carlo procedures and a factorial ANOVA with follow-ups was employed to evaluate the effect of conditions on the performance of each technique as measured by proportions of correctly predicted observations for all groups. In most of the conditions for the two outcome measures, CART performed better than LDA and LR. However, under some conditions where there were a higher number of predictor variables and groups with low correlation, the superiority of LR to CART was observed.

Authors

• Cahit Polat, University of Denver

• Kathy E. Green, University of Denver