Convergence and Stability of Alternative Estimators in Misspecified Covariance Structures

Abstract

This paper presents results on how Maximum Likelihood (ML) and Generalized Least Squares (GLS) estimators in covariance structure models converge to a constant difference, and how they behave differently when the degree of misspecification increases. Model fit is assessed using the RMSEA fit index. The results indicate that the ML estimator is more stable across sample sizes than the GLS estimator, is more powerful in detecting model misspecification, and produces less biased parameter estimates with lower MSE’s in substantially misspecified models. Researchers are advised to use the ML estimates in most practical modeling situations.