Robustness of Structural Equation Modeling to Distributional Misspecification: Empirical Evidence & Research Guidelines
53 Pages Posted: 9 Apr 2009 Last revised: 3 May 2009
Date Written: April 8, 2009
A growing number of theories in information systems (IS) research are developed and tested using structural equation modeling (SEM). Use of statistical techniques for measurement and structural model assessment, reliability, and validity are facilitated by such programs as LISREL and widely reported in the literature. In contrast, identification and correction of distributional misspecification (DM) - non-normality, multicollinearity, heteroscedasticity, and combinations thereof - is rarely reported in SEM analyses, despite its potential to bias statistical estimation and inference. Four principal findings of our literature review and Monte Carlo simulations are: 1) studies using SEM rarely report tests of DM, while studies using OLS typically report such tests; 2) reduced statistical power in the presence of DM for SEM that is correctable for OLS using weighted least squares; 3) negative synergy when different types of DM occur jointly; and 4) worsening statistical power as sample size and variance explained are reduced. We provide practical guidelines for assessing, reporting, and overcoming distributional misspecification.
Keywords: structural equation models, nonnormality, multicollinearity, heteroscedasticity, distributional misspecification, statistical power
Suggested Citation: Suggested Citation