Performance of Model-Implied Instrumental Variable Estimators and a Maximum Likelihood Estimator for Structural Equation Models

• In Event: 106 - Open Refereed Roundtable Session I
  In Refereed Roundtable: A. Table 14. Quantitative Methodology

Abstract

This article compares the performance of two model-implied instrument variable (MIIV) estimators (MIIV-2SLS and MIIV-GMM) with a maximum likelihood (ML) estimator for structural equation models (SEMs). As potentially limited information estimators, the MIIV estimators are less prone to spreading bias due to model misspecification than the standard full information ML estimator for SEMs. The ML estimator, however, is theoretically more efficient than the MIIV estimators. In addition, a key feature of the MIIV-GMM estimator is the ability to estimate subsets of equations up to and including the full model. A simulation study is designed to compare the performance of the MIIV-2SLS, MIIV-GMM, and ML estimators with respect to bias, efficiency, and root mean squared error (RMSE) for two SEMs with increasing degrees of misspecification. Results indicate that the MIIV-2SLS and MIIV-GMM estimators perform better with respect to bias and the RMSEs than the ML estimator. In addition, the MIIV-2SLS estimator performs comparably to the MIIV-GMM estimator, except that the MIIV-GMM estimator is preferred for one parameter with a greater degree of misspecification. The results of this simulation study contribute to our understanding of the performance of alternatives to the popular ML estimator for SEMs and suggest that researchers might consider one of the MIIV estimators for SEMs.

Author

• Shawn Bauldry, University of Alabama-Birmingham