Comparison of Regularized Full- and Limited-Information Structural Equation Modeling

• In Event: Innovative Methods in Structural Equation Modeling

Abstract

This study compared performance of regularized structural equation modeling (SEM) via maximum likelihood (MLE) and 2-stage least squares (2SLS) estimators. Both frameworks have been shown to outperform standard maximum likelihood when models are misspecified. However, MLE and 2SLS regularization have not been directly compared in prior research. Furthermore, prior work has not examined the impact of factor loading magnitude, or very small sample size on all of these regularization estimators. This study describes a Monte Carlo simulation designed which compared these estimators under a variety of conditions, with special focus on small samples, weak loadings, and model misspecification, all of which are often found in applied educational research.

Authors

• Joseph E Miller, The University of Texas

• William Holmes Finch, Ball State University