Handling the Label Switching Problem in Latent Class Models Via the ECR Algorithm

Abstract

Latent class models (LCMs) are specific cases of mixture models. Under a Bayesian setup, the symmetric posterior distribution of these models leads Markov chain Monte Carlo (MCMC) methods to suffer from the so-called label switching problem. In this article, we treat the corresponding MCMC outputs using a recent approach, namely, the Equivalence Classes Representative (ECR) algorithm and conclude that it can effectively solve the label switching problem by considering several examples of LCMs, such as mixtures of regressions, hidden Markov models, and Markov random fields. Moreover, the superiority of this method over other approaches becomes apparent.

Keywords:

• Bayesian analysis
• Bayesian image segmentation
• ECR Algorithm
• Hidden Markov models
• Label switching phenomenon
• Latent variables
• Mixtures of regressions

Mathematics Subject Classification:

• G2F15
• G5C40
• G5C60

Acknowledgments

[Supplementary materials are available for this article. Go to the publisher's online edition of Communications in Statistics: Stimulation and Computation for the following free supplemental files: Additional figures.]