On the Performance of Some Multinomial Classification Rules

Abstract

This article presents and discusses a new multinomial classification procedure based on a discrete distributional distance. Its performance along with other commonly used classification procedures is assessed through Monte Carlo sampling experiments under different population structures. In addition to reporting results consistent with the work of Gilbert (1968) and Moore (1973), the article describes sampling experiments which show that the new distance procedure is generally superior, in terms of both the mean actual and mean apparent errors, to the usual full multinomial rule in situations of disproportionate sample sizes.

Key Words:

• Multinomial classification
• Full multinomial rule
• Matusita distance
• Bahadur representation
• Martin and Bradley probability model
• Monte Carlo study