Rasch Model Parameter Estimation via the Elastic Net

Area of Science:

• Psychometrics
• Statistical modeling
• Educational measurement

Background:

• The Rasch model is a fundamental tool in psychometrics for analyzing item response data.
• Traditional estimation methods like conditional maximum likelihood estimation (CMLE), marginal maximum likelihood estimation (MMLE), and marginal Bayes modal estimation (MBME) have limitations, particularly when the number of items exceeds the number of examinees.

Purpose of the Study:

• To introduce and evaluate a novel estimation method: penalized joint maximum likelihood estimation (PJMLE).
• To assess the performance of PJMLE in estimating Rasch model parameters under various conditions.
• To address the limitations of traditional methods in scenarios with more items than examinees.

Main Methods:

• Implemented penalized joint maximum likelihood estimation (PJMLE) using joint maximum likelihood estimation (JMLE) with elastic net penalization via the R 'glmnet' package.
• Conducted simulation studies to compare PJMLE with CMLE, MMLE, and MBME.
• Evaluated parameter estimation accuracy for item difficulties and examinee abilities.

Main Results:

• PJMLE successfully estimates Rasch model parameters when the number of items is greater than the number of examinees, overcoming a key limitation of traditional techniques.
• PJMLE demonstrates comparable performance to traditional estimation methods when the number of examinees exceeds the number of items.
• PJMLE achieves this performance without requiring the specification of a mixing distribution or a prior distribution.

Conclusions:

• Penalized joint maximum likelihood estimation (PJMLE) offers a robust alternative for Rasch model parameter estimation, especially in data-rich item settings.
• PJMLE provides a valuable tool for psychometricians and researchers dealing with complex item response datasets.
• The method's flexibility and comparable accuracy make it a significant advancement in Rasch model estimation.