Empirical comparison of error rate-estimators in logistic discriminant analysis

Abstract

A simulation study related to the comparison of 15 estimators of the actual-error rate in two-group logistic discriminant analysis is achieved in 528 cases taking into account some characteristics of the groups such as the common distribution of the variables, the overlap of the populations, the number of variables, etc. For each combination of these parameters, the relative errors of the estimators with regard to the theoretical actual-error rate (empirically determined) are computed and help to compare the performance of the estimators. The results show an overall best performance of the estimators e632 and eOS. On the contrary, the estimators eA, epp and eppcv record the lowest performance in terms of the mean relative error and the mean relative bias. Moreover, the performance of most of the parametric estimators depends on the normality and heteroscedasticity degrees of the populations.

Keywords:

• Simulation
• Actual-error rate
• Logistic rule
• Estimators
• Comparison

2000 Mathematics Subject Classification :

• 62H30
• 62G09
• 62H12
• 65C05