Email Bookmark and Share


Advances in Structural Equation Modeling (SEM)

Friday, May 22, 2009, 1:00 PM - 2:20 PM
Nob Hill C - D

Chair: Victoria Savalei
University of British Columbia

Structural equation modeling (SEM) is a popular modeling tool in psychology, which involves proposing a theoretically motivated set of relationships among observed and latent variables. The filed of SEM has experienced rapid growth over the last several decades. This symposium will provide an overview of several recent SEM developments.

Conducting Specification Searches in SEM Using a Ruin and Recreate Principle
George A. Marcoulides
University of California, Riverside
An SEM specification search approach using a new optimization algorithm based on a ruin and recreate (R & R) principle is introduced. The principle ruins a large fraction of an obtained modeling solution and then tries to obtain a new solution that is better in terms of model fit than the previous one. Using data with known structure, the performance of the new algorithm is illustrated. The results demonstrate the capabilities of the algorithm for conducting specification searches in SEM.

Expected Versus Observed Information in SEM With Incomplete Normal and Nonnormal Data
Victoria Savalei
University of British Columbia
Standard errors for ML estimates in SEM are obtained from the associated information matrix, which can be "expected" or "observed." With complete data, both types of estimates are consistent. However, with incomplete MAR data, standard errors based on expected information will not be valid. Further, type of information affects the computation of the robust standard errors and of the corresponding scaled test statistic often computed in the presence of nonnormality. This talk summarizes a simulation study that investigated the different types of standard errors for incomplete nonnormal data.

Specification Testing in Structural Equation Models Based on Heywood Cases
Stanislav Kolenikov
University of Missouri, Columbia
This talk reviews issues related to estimates of variance below or near zero in SEMs. We provide examples of models where the population values of the variance parameters are below zero. In all of those examples, the structural model is severely misspecified. We present a variety of tests for zero variance, along with the simulation studies of their performance. The sandwich standard errors were found to be the only reliable way to estimate variability in parameter estimates, while the signed root of Satorra-Bentler chi-square difference was the best overall specification test.

Ken Bollen, University of North Carolina at Chapel Hill

Bayesian SEM: Current Developments and Future Directions
Johnny Zhang
University of Notre Dame
In the past decade, there has been a rapid growth in Bayesian SEM partly due to the development of MCMC techniques and data augmentation algorithms. Besides their unique ability to incorporate prior information, Bayesian methods are especially useful for estimating complex SEM models with mixture of categorical and continuous data, with longitudinal and multilevel data, and with incomplete or missing data. In this presentation, I will review the current Bayesian SEM techniques, including both methods and software, as well as discuss future research directions.

Go back

U.S. Public Health Service Commissioned Corps.