A Bayesian framework for the logistic positive exponent and its reflection IRT models

Area of Science:

• Psychometrics
• Educational Measurement
• Statistical Modeling

Background:

• Traditional item response theory (IRT) models often assume symmetric item characteristic curves, limiting their ability to capture complex response patterns.
• Asymmetric IRT models, such as the logistic positive exponent (LPE) and its reflection (RLPE), offer greater flexibility but require robust estimation methods.
• Bayesian estimation for LPE and RLPE models presents challenges, including optimal prior specification and model selection, which have been underexplored.

Purpose of the Study:

• To comprehensively investigate Bayesian estimation for LPE and RLPE models.
• To evaluate parameter recovery, computational efficiency, and sensitivity to prior specifications for the asymmetry parameter.
• To assess model selection performance in distinguishing asymmetric from symmetric IRT models and demonstrate practical application.

Main Methods:

• Extensive simulation studies were conducted to evaluate parameter recovery, computational efficiency, and sensitivity to various prior distributions for the asymmetry parameter.
• Model selection performance was systematically evaluated using criteria to differentiate between asymmetric (LPE/RLPE) and symmetric IRT models.
• Theoretical comparisons were made with alternative asymmetric IRT models to delineate the advantages of LPE/RLPE.
• An empirical analysis was performed on mathematics assessment data to showcase the practical utility of the LPE/RLPE framework.

Main Results:

• Simulation studies demonstrated good parameter recovery and computational efficiency for LPE/RLPE models under various prior specifications.
• The RLPE model showed superior performance in discriminating between asymmetric and symmetric models, particularly in identifying meaningful item asymmetry.
• Theoretical comparisons highlighted specific conditions where LPE/RLPE models offer advantages over other asymmetric approaches.
• Empirical application revealed that the RLPE model significantly outperformed symmetric alternatives, uncovering important item-specific asymmetry patterns in mathematics assessment.

Conclusions:

• Bayesian estimation provides a viable and effective framework for LPE and RLPE models, offering flexibility in prior specification for asymmetry.
• The RLPE model is a valuable tool for detecting and modeling item characteristic curve asymmetry, outperforming symmetric counterparts in empirical settings.
• The findings support the practical utility of LPE/RLPE models in educational measurement for a more nuanced understanding of item response behavior.