The Model Size Effect in Confirmatory Factor Analysis With Ordinal Missing Data: A Comparison of Imputation Methods

• In Event: Estimation and Interpretation of Educational Data

Abstract

This study aims to compare three multiple imputation (MI) methods: ordinal-logistic-regression-based MI (MI-LOGIT), latent-variable-model-based MI (MI-LV), and random-forest-based MI (MI-RF) in confirmatory factor analysis (CFA) with ordinal missing data, with a focus on the model size effect. These MI methods were compared for handling ignorable missingness: missing completely at random (MCAR) and missing at random (MAR), under various simulation conditions. The WLSMV method was used to compute parameter estimates and fit indices after the imputation. Preliminary results revealed that model size, missingness, and sample size affected the performance of parameter estimates and model fit. MI-LOGIT outperformed MI-RF in parameter estimates when sample size was small and model size was small to moderate, while MI-RF yielded better model fit.

Authors

• Ji Li, University of Arkansas for Medical Sciences

• Xinya Liang, University of Arkansas