Structural Equation Modeling With Misspecification and Nonnormal Data: A Comparison of Bayesian and Maximum Likelihood Estimation

• In Event: Estimation Accuracy and Conditions for Complex Latent Variable Models

Abstract

This simulation study evaluated the performance of robust maximum likelihood (MLR), Bayesian with non-informative priors (BN), and Bayesian with weakly informative priors (BI) in confirmatory factor analysis under conditions with model misspecification and nonnormal data. The design factors included factor structures, levels of nonnormality, sample sizes, and different priors. We evaluated model fit and parameters at both population and sample levels. Preliminary results indicate that Bayesian methods did not over reject models with small sample sizes and symmetric distributions. Generally, BW provided reasonable loadings estimates, and MLR provided accurate factor correlation estimates.

Authors

• Xinya Liang, University of Arkansas at Fayetteville

• Yanyun Yang, Florida State University