Assessing the Performance of Bayesian and Traditional Estimators in Confirmatory Factor Analysis with Sparse Likert-Scale Data

• In Event: Extending the General Linear Model: From Classical Regression to Bayesian Estimation and Machine Learning

Abstract

Applied researchers often encounter sparse data when few respondents endorse certain item categories. Collapsing categories is a common strategy to address this issue. This study used simulations of a confirmatory factor analysis (CFA) model with sparse data, followed by recoding to collapse sparse responses. Estimators evaluated included MLR, MLMV, WLSMV, and Bayesian methods. Results showed that collapsing categories generally improved model fit under correctly specified models across all estimators. However, in misspecified models, collapsing increased the risk of selecting incorrect models. Bayes showed the highest rejection rates under misspecification, while WLSMV had the lowest rejection rates in correctly specified conditions. These findings offer practical guidance for handling sparse data in CFA. 

Authors

• Jin Liu, University of Texas - Arlington

• Liu Liu, University of Georgia

• Wei Jiang, University of Texas - Arlington