Estimation of a Simple Structure in a Multidimensional IRT Model Using Structure Regularization.

Area of Science:

• Psychometrics
• Statistical modeling
• Educational measurement

Background:

• Multidimensional item response theory (MIRT) models are complex.
• Existing methods can be difficult to interpret and may struggle with small sample sizes.
• A need exists for simpler, more interpretable MIRT estimation techniques.

Purpose of the Study:

• To develop a novel method for estimating a simple structure matrix in MIRT.
• To enhance the interpretability of latent traits in MIRT.
• To enable effective clustering of test items based on latent traits.

Main Methods:

• Utilizing the product-based elastic net (prenet) penalty, adapted from factor analysis.
• Employing a combination of stochastic EM algorithms, proximal gradient methods, and coordinate descent for optimization.
• Developing a method where each test item corresponds to a single latent trait.

Main Results:

• The proposed method yields easily interpretable results by assigning single latent traits to test items.
• Effective clustering of test items based on their corresponding latent traits is achieved.
• Numerical experiments show superior performance compared to existing L1 penalty methods, particularly with small numbers of test subjects.

Conclusions:

• The prenet penalty offers an effective approach for estimating simple structure matrices in MIRT.
• The method provides enhanced interpretability and item clustering capabilities.
• This technique is particularly advantageous in scenarios with limited sample sizes, offering a robust alternative to existing L1 penalty methods.