DIF Statistical Inference Without Knowing Anchoring Items.

Area of Science:

• Psychometrics
• Statistical modeling
• Measurement validity

Background:

• Measurement invariance is crucial for instrument validity, often assessed using differential item functioning (DIF) analysis.
• Traditional DIF analysis relies on anchor items, which can be problematic if misspecified or unknown.
• Existing methods like item purification and regularization lack guaranteed valid statistical inference.

Purpose of the Study:

• To propose a new DIF analysis method for multiple indicators and multiple causes (MIMIC) models.
• To enable accurate estimation of DIF effects and valid statistical inferences without prior anchor item knowledge.
• To control the type-I error rate in DIF detection.

Main Methods:

• A novel DIF analysis method is proposed using a minimal L0 norm condition for latent trait distribution identification within a MIMIC model.
• The method estimates individual item DIF effects and quantifies uncertainty without relying on a pre-specified anchor set.
• Performance is evaluated through simulation studies comparing it to anchor-set-based likelihood ratio tests and LASSO approaches.

Main Results:

• The proposed method accurately estimates DIF effects and provides valid statistical inferences, including controlled type-I error rates.
• Simulation studies demonstrate the method's effectiveness compared to existing approaches.
• The method was successfully applied to analyze personality scales in the Eysenck Personality Questionnaire-Revised (EPQ-R).

Conclusions:

• The new DIF analysis method offers a robust alternative when anchor items are unavailable or misspecified.
• It provides valid statistical inferences, enhancing the reliability of measurement invariance assessment.
• This approach contributes to more accurate and dependable psychometric evaluations.