Estimating three- and four-parameter MIRT models with importance-weighted sampling enhanced variational auto-encoder.

Area of Science:

• Psychometrics
• Educational Measurement
• Machine Learning

Background:

• Multidimensional Item Response Theory (MIRT) is crucial for educational and psychological assessments.
• Existing estimation methods for MIRT, particularly the multidimensional three-parameter (M3PL) and four-parameter logistic (M4PL) models, struggle with computational demands on large datasets.
• Scalability is a significant challenge for current MIRT estimation techniques with big data.

Purpose of the Study:

• To develop a computationally efficient and scalable algorithm for estimating MIRT models, specifically M3PL and M4PL.
• To address the limitations of existing methods in handling large-scale assessment data.
• To leverage machine learning techniques for improved MIRT parameter estimation.

Main Methods:

• Proposed an importance-weighted sampling enhanced Variational Autoencoder (VAE) approach.
• Utilized variational inference to approximate intractable marginal likelihoods.
• Employed importance-weighted samples to enhance the VAE's log-likelihood approximation.

Main Results:

• Simulation studies demonstrated superior computational efficiency and scalability compared to Monte Carlo EM and Metropolis-Hastings Robbins-Monro methods.
• The proposed algorithm showed good performance on a real-world NAEP multistage testing dataset.
• The VAE-based approach provides a robust alternative for estimating complex MIRT models.

Conclusions:

• The importance-weighted sampling enhanced VAE offers a computationally efficient and scalable solution for MIRT estimation, particularly for M3PL and M4PL models.
• This method effectively handles large-scale assessment data, overcoming limitations of traditional algorithms.
• The approach shows promise for modern educational and psychological measurement applications.