Structural Equation Modeling: Foundations, Recent Developments, and Applications

Structural equation modeling (SEM) is a very general and flexible multivariate technique that allows relations among variables to be examined. In practice, SEM is often used to model and test hypothesized causal relationships among observed and latent (not directly observed) variables, thereby answering a wide range of research questions. This talk will provide an overview of the SEM framework, including introducing its theoretical foundations, defining terminology and notation, explaining its differences and similarities with other modeling approaches, and reviewing the classic estimation process. Finally, examples of research questions for which SEM is a useful method, including those related to measurement, complex patterns of effects, and testing competing hypotheses, will be introduced.  

Structural equation models (SEMs) are typically fit using the maximum likelihood (ML) estimator. However, some ML assumptions might be untenable in particular applications. Of relevance to this presentation, samples might be small leading to issues with the asymptotic properties of ML, and the specified model is quite likely not the true data-generating model, which infuses bias into estimates. These issues can also lead to non-convergence, thus limiting researchers' ability to capitalize on the benefits of SEM. In this presentation, we cover modern developments in parameter estimation. Specifically, we describe the model-implied instrumental variables with two-stage least squares (MIIV-2SLS; Bollen, 1996, 2021) estimator, which affords benefits in the presence of model misspecification and violations of the traditional assumptions required in SEM. We review simulation studies using the MIIV-2SLS estimator, software accessibility, and implications for applied research. 

Methods for modeling causal relations differ in their underlying assumptions. Structural Equation Modeling (SEM) is a multivariate statistical technique for analyzing relationships among observed and latent variables, while System Dynamics (SD) considers the whole systemic structure to model multilevel feedback systems. SD models nonlinear feedback and dynamic changes, while in incorporating such characteristics, SEM requires special considerations to deal with the issue of model identifiability. SEM is a powerful tool in psychometrics, an application that can greatly enhance SD models. Previous attempts to link SD and SEM have been limited to specific questions and illustrative examples instead of a general framework. Our generative framework bridges SD and SEM. We illustrate how SD and SEM can be articulated within this framework and complement each other by evaluating the relationship between systems thinking and ecological worldview among undergraduate psychology students (N = 1058).