Latent Class Analysis - A Variational Approach

Abstract

Choosing number of classes is a major modeling decision in latent class analysis. This is most often carried out by fitting a number of models with increasing number of classes. In the frequentist methodology, one records a number of fit criteria for comparison, while in Bayesian methodology, this exercise is more difficult, as incremental fitting is computationally burdensome, more so when number of manifest variables is large and/or dataset is big. A methodology that can provide good approximation to number of classes with little user intervention, enjoys benefits of Bayesian methodology and can circumvent incremental fitting of latent classes will be a valuable tool. In this article, we bring to the attention of latent class modelers the methodology of variational Bayes for latent class modeling and extend it to the case of polychotomous manifest variables.