Using Lasso and Adaptive Lasso to Identify DIF in Multidimensional 2PL Models.

Area of Science:

• Psychometrics
• Educational Measurement
• Item Response Theory

Background:

• Differential item functioning (DIF) analysis is crucial for ensuring assessment item fairness across different groups.
• Current DIF research predominantly focuses on unidimensional item response theory (IRT) models.
• Multidimensional IRT (MIRT) models offer enhanced insights but require advanced DIF detection methods.

Purpose of the Study:

• To explore and compare regularization methods for DIF detection within MIRT models.
• To evaluate the performance of regularization techniques against the traditional likelihood ratio test.
• To investigate the efficacy of lasso (EM, EMM) and adaptive lasso (EM) algorithms for DIF detection in MIRT.

Main Methods:

• Implementation of regularization methods: lasso with expectation-maximization (EM), lasso with expectation-maximization-maximization (EMM), and adaptive lasso with EM.
• Comparison of regularization methods against the classic likelihood ratio test for DIF detection.
• Simulation studies to assess performance under various conditions, including sample size.

Main Results:

• Regularization methods offer advantages over traditional approaches, including bypassing iterative purification and handling multiple covariates.
• Lasso EMM and adaptive lasso EM demonstrated superior performance compared to lasso EM, particularly with large sample sizes.
• The study identified promising regularization techniques for robust DIF detection in MIRT.

Conclusions:

• Regularization methods, specifically lasso EMM and adaptive lasso EM, are effective for DIF detection in MIRT models.
• These methods provide a valuable alternative to traditional approaches, especially in complex assessment scenarios.
• The findings support the use of advanced regularization techniques for improved psychometric analysis and fairness in assessments.