A comparative study on the performance of GSCA and CSA in parameter recovery for structural equation models with ordinal observed variables

Kwanghee Jung, Pavel Panko, Jaehoon Lee, Heungsun Hwang

Research output: Contribution to journalArticlepeer-review

6 Scopus citations

Abstract

A simulation based comparative study was designed to compare two alternative approaches to structural equation modeling-generalized structured component analysis (GSCA) with the alternating least squares (ALS) estimator vs. covariance structure analysis (CSA) with the maximum likelihood (ML) estimator or the weighted least squares mean and variance adjusted (WLSMV) estimator-in terms of parameter recovery with ordinal observed variables. The simulated conditions included the number of response categories in observed variables, distribution of ordinal observed variables, sample size, and model misspecification. The simulation outcomes focused on average root mean square error (RMSE) and average relative bias (RB) in parameter estimates. The results indicated that, by and large, GSCA-ALS recovered structural path coefficients more accurately than CSA-ML and CSA-WLSMV in either a correctly or incorrectly specified model, regardless of the number of response categories, observed variable distribution, and sample size. In terms of loadings, CSA-WLSMV outperformed GSCA-ALS and CSA-ML in almost all conditions. Implications and limitations of the current findings are discussed, as well as suggestions for future research.

Original languageEnglish
Article number2461
JournalFrontiers in Psychology
Volume9
Issue numberDEC
DOIs
StatePublished - Dec 5 2018

Keywords

  • Alternating least squares estimation
  • Covariance structure analysis
  • Diagonally weighted least squares estimation
  • Generalized structured component analysis
  • Maximum likelihood estimation
  • Monte carlo simulation
  • Structural equation modeling

Fingerprint Dive into the research topics of 'A comparative study on the performance of GSCA and CSA in parameter recovery for structural equation models with ordinal observed variables'. Together they form a unique fingerprint.

Cite this