A Note on Improving Variational Estimation for Multidimensional Item Response Theory

Area of Science:

• Psychometrics
• Statistical modeling
• Social science research

Background:

• Multidimensional item response theory (MIRT) is crucial for analyzing complex constructs in social science.
• Estimating MIRT models is computationally intensive, limiting their widespread application.
• Existing variational estimation methods, like GVEM, offer speed but can introduce bias in parameter estimates.

Purpose of the Study:

• To address the bias in discrimination parameters observed in variational estimation methods for MIRT.
• To propose an enhanced variational estimation algorithm for improved accuracy in MIRT model fitting.
• To investigate the computational efficiency and bias-correction capabilities of the proposed method.

Main Methods:

• Development of an importance-weighted version of the Gaussian variational expectation-maximization (IW-GVEM) algorithm.
• Integration of adaptive moment estimation to optimize learning rates in gradient descent.
• Simulation studies to compare IW-GVEM with existing methods like GVEM.

Main Results:

• The proposed IW-GVEM method effectively corrects bias in discrimination parameters for MIRT models.
• IW-GVEM demonstrates comparable accuracy to traditional methods with only a modest increase in computation time.
• The adaptive moment estimation enhances the stability and efficiency of the optimization process.

Conclusions:

• IW-GVEM provides a faster and more accurate approach to estimating MIRT models, overcoming limitations of previous variational methods.
• This advancement can facilitate the broader application of MIRT in large-scale social science assessments.
• The proposed techniques may offer improvements for variational estimation in other psychometric models.