Related Experiment Videos
Bayesian structural equation modeling: a more flexible representation of substantive theory.
Bengt Muthén1, Tihomir Asparouhov
1Muthén & Muthén, 3463 Stoner Avenue, Los Angeles, CA 90066, USA. bmuthen@ucla.edu
Psychological Methods
|September 12, 2012
Summary
This study introduces a novel Bayesian approach for factor analysis and structural equation modeling, improving model accuracy by using approximate zeros instead of exact zeros. This method enhances the reflection of theories and aids in analyzing complex models.
Area of Science:
- Psychometrics
- Statistical Modeling
- Bayesian Analysis
Background:
- Traditional factor analysis and structural equation modeling (SEM) often rely on exact zero specifications, which may not accurately represent complex theoretical relationships.
- Maximum-likelihood estimation can fail with non-identified models, limiting the analysis of certain complex structures.
Purpose of the Study:
- To propose and evaluate a new Bayesian approach for factor analysis and SEM that utilizes informative priors for approximate zero specifications.
- To demonstrate the utility of this approach in handling complex models, including those with cross-loadings, residual correlations, and potential misspecification.
Main Methods:
- The proposed method replaces exact zero parameter constraints with small-variance priors, allowing for more flexible model specification.
- Model estimation, testing (using posterior predictive checking), and modification are integrated within the Bayesian framework.
- Analyses were conducted using Mplus software, incorporating Monte Carlo simulations and real-world datasets.
Main Results:
- The Bayesian approach effectively incorporates approximate zero specifications, leading to analyses that better align with substantive theories.
- The method successfully handles non-identified models and provides insights into measurement aspects of latent variable modeling.
- Applications in confirmatory factor analysis (cross-loadings, residual correlations) and full SEM demonstrated efficient model misspecification detection.
Conclusions:
- The proposed Bayesian methodology offers a robust alternative to traditional methods for factor analysis and SEM, particularly for complex models.
- This approach enhances the interpretability and accuracy of statistical models by allowing for nuanced parameter specifications.
- The study validates the method's effectiveness through simulations and real-data analyses across various psychological and educational domains.