Model Estimation for Ordered Categorical Data Under Nonnormal Latent Factor and Error Scores: A Comparison of PIV, ULSMV, and WLSMV Estimators

• In Event: AERA Poster Session 7
  In Poster Session: Division D Section 2 Poster Session 2

Abstract

Impact of underlying non-normal distribution on parameter estimation, standard errors, and reliability with different estimation methods for ordered categorical data have been studied. A recently proposed estimator PIV is likely to perform better than WLSMV and ULSMV. However, the underlying distribution comes from two sources: non-normal latent factor scores and/or non-normal error scores. Previous studies did not disentangle them. We propose to conduct a Monte Carlo study to compare these three estimators (WLSMV, ULSMV, and PIV) in the context of confirmatory factor analysis by manipulating the level of non-normality of factor scores, level of non-normality of error scores, the number of response category, heterogeneity of categorical distributions across items, model misspecification, and sample size.

Authors

• Yachen Luo, Florida State University

• Yanyun Yang, Florida State University