Nonlinear Structural Equation Modeling With Nonnormal Data and Small Sample Sizes

• In Event: Mediation, Ordinal, and Nonlinear Modeling

Abstract

Data collected from surveys are commonly placed on a Likert scale although the metric of the targeted latent construct is interval and continuous. Nonlinear models are more appropriate for such data. With computer generated data, this study evaluated the performance of WLSMV and Bayesian estimation methods in fitting CFA models with polychoric correlation matrix. Conditions included sample size, various underlying distributions, and categorical distributions. We expected that Bayesian method perform better when sample size is small particularly when the categorical distributions vary across items.

Authors

• Xinya Liang, University of Arkansas

• Yanyun Yang, Florida State University