The Specification and Impact of Prior Distributions for Categorical Latent Variable Models

ABSTRACT

Latent class models can exhibit poor parameter recovery and low convergence rates under the traditional frequentist estimation approach. Bayesian estimation may be a viable alternative for estimating latent class models–especially when categorical items are present and priors can be placed directly on the categorical item-thresholds. We present a simulation study involving Bayesian latent class analysis (LCA) with categorical items. We demonstrate that the frequentist framework and the Bayesian framework with diffuse (non-informative) priors are unable to properly recover parameters (e.g., latent class item-thresholds); a substantive interpretation of the obtained results would lead to improper conclusions under these estimation conditions. However, specifying (weakly) informative priors within the Bayesian framework generally produced accurate parameter recovery, indicating that this may be a more viable estimation approach for LCA models with categorical indicators. The paper concludes with a general discussion surrounding the advantages of Bayesian estimation for LCA models.

KEYWORDS:

• Bayesian estimation
• latent class analysis
• categorical latent variable models
• prior distributions

Disclosure statement

No potential conflict of interest was reported by the author(s).