Comparing the performance of two linear classifiers

Abstract

When multiple predictors are observed on an object in the population, it is possible to combine the information obtained from these predictors into a score, which is a scalar-valued function. Then this score can be used for classification purposes. The aim of this work is to compare the classification properties of two classical types of scores obtained from linear discriminant analysis and logistic regression. These are two of the most widely used statistical methods for classification problems and they were used in this work to model the association of several predictors with the prevalence of stroke using data from a national health interview survey. Moreover, we wanted to evaluate the accuracy of these two classification techniques. For this purpose, we evaluated and compared The results obtained from the two classification techniques by means of a receiver operating characteristic (ROC) curve. By plotting the ROC curves for the two models on the same axes, we were able to determine which classification technique is better for classification, in particular, that classification technique whose ROC curve encloses the larger area beneath it.