Investigating a Weakly Informative Prior for Item Scale Hyperparameters in Hierarchical 3PNO IRT Models.

Area of Science:

• Statistics
• Psychometrics
• Bayesian Inference

Background:

• Bayesian hierarchical modeling often utilizes the half-t distribution for scale hyperparameters.
• The half-t distribution is defined by scale (s) and degrees-of-freedom (ν) parameters.
• Vaguely informative priors are crucial for robust Bayesian analyses, especially in complex models.

Purpose of the Study:

• To evaluate the performance of half-t prior densities in hierarchical item response theory (IRT) models.
• To compare the half-t prior with commonly used uniform and inverse-gamma priors.
• To assess the impact of varying item parameter variability on model performance.

Main Methods:

• Application of half-t prior densities with a finite scale (s) slightly larger than the standard deviation.
• Monte Carlo simulations to investigate parameter recovery and model comparison.
• Analysis of the hierarchical three-parameter item response theory (IRT) model.

Main Results:

• The half-t prior family demonstrated advantages over uniform and inverse-gamma priors.
• Effective handling of both very small and very large item parameter variability.
• Improved parameter recovery and model comparison accuracy were observed.

Conclusions:

• The half-t distribution provides a flexible and advantageous prior for scale hyperparameters in Bayesian hierarchical IRT models.
• Its performance is robust across different levels of item parameter variability.
• The findings support the use of the half-t family for enhanced precision and reliability in IRT analyses.