Single and Multiple Ability Estimation in the SEM Framework: A Non-Informative Bayesian Estimation Approach.

Area of Science:

• Psychometrics
• Statistical modeling
• Structural equation modeling

Background:

• Complex latent variable models with many parameters often challenge traditional frequentist methods like maximum likelihood (ML) estimation.
• High-dimensional integration in these models can lead to estimation failures.
• Bayesian estimation offers a robust alternative, particularly with diffuse priors.

Purpose of the Study:

• To compare the performance of Bayesian estimation against ML estimation.
• To evaluate these methods in the context of multidimensional item response theory (MIRT) and multiple-indicator multiple-cause (MIMIC) models.
• To assess their utility in single- and multi-level structural equation modeling frameworks.

Main Methods:

• A Monte Carlo simulation study was conducted.
• Bayesian estimation with diffuse priors and ML estimation were compared.
• Two measurement models (MIRT and MIMIC) were utilized.
• An empirical example using the Multistate Bar Examination data was analyzed.

Main Results:

• Bayesian estimation with diffuse priors yielded results comparable to ML estimation across various conditions.
• Both single- and multi-level MIRT and MIMIC models showed consistent performance.
• The empirical example demonstrated the practical utility of both MIRT and MIMIC models.

Conclusions:

• Bayesian estimation is a viable and effective alternative to ML estimation for complex latent variable models.
• It offers advantages in situations where ML estimation encounters convergence issues.
• The study provides guidance on implementing Bayesian methods in MIRT and MIMIC analyses.