Solving the problem of latent class selection

Abstract

The correct selection of the number of classes represents a difficult step in classification methods because it can significantly affect interpretations and may lead to an incorrect analysis of the studied phenomenon. This problem arises in latent class analysis, which is a method of classifying individuals based on categorical data. Determining the number of classes constituting the profiles of a population is generally done by using a likelihood ratio test, but the use of this test is not correct theoretically. To solve this problem, we use the information criteria to select the number of classes. Several information criteria have been proposed for this purpose previously, but the difficulty of selection remains because these criteria are influenced by other parameters such as the number of classes and the sample size. In this paper, we conduct a comparative study of these information criteria and use a panorama of the best-adapted criteria according to the data for an exact selection of models.