Bayesian Latent Class Analysis for Person-Centered Research: Simulation Study on Conditional Dependence and Prior Choice

• In Event: 18.018 - Advances in Structural Equation Modeling

Abstract

Bayesian structural equation modeling relaxes the conditional independence assumption in latent class analysis (LCA) by freely estimating all correlations among class indicators using an informative prior with zero mean but nonzero variance (i.e., approximate independence). This simulation study investigated the consequences of ignoring conditional dependence in Bayesian LCA, and using a range of different priors to specify approximate independence in a model. Simulation results favored the use of a less informative prior such that model fit was always satisfactory when nonzero correlations were neglected in any degree; and that power to detect nonzero correlations was acceptable (i.e., > .80) with N > 300 for large correlations or N > 500 for moderate correlations. Implications and analytic guidelines are provided.

Authors

• Jaehoon Lee, Texas Tech University

• Kwanghee Jung, Texas Tech University

• Jungkyu Park, Kyungpook National University

• Hyeran Park