Mutual Information Reliability for Latent Class Analysis.

Area of Science:

• Psychometrics
• Educational Measurement
• Information Theory

Background:

• Latent class models (LCMs) are crucial for classifying individuals into subgroups based on observed variables, aiding diagnostic systems.
• Existing reliability measures may not fully capture the nuances of categorical latent variables within LCMs.

Purpose of the Study:

• To propose a novel Mutual Information Reliability (MIR) coefficient for assessing the measurement quality of latent class models.
• To provide a reliability metric analogous to those in item response theory and McFadden's pseudo R-square, specifically for categorical latent variables.

Main Methods:

• The study leverages information theory, specifically mutual information, to derive the MIR coefficient.
• The coefficient is developed based on information-theoretic inequalities, ensuring desirable properties like being unitless and bounded between 0 and 1.

Main Results:

• The proposed MIR coefficient offers a theoretically sound and interpretable measure of reliability for latent class models.
• The coefficient is shown to be analogous to established reliability concepts and extends goodness-of-fit measures like McFadden's pseudo R-square.
• Simulated examples demonstrate the practical application and performance of the MIR coefficient in both fixed and adaptive testing scenarios.

Conclusions:

• The Mutual Information Reliability coefficient provides a valuable new tool for evaluating the quality of latent class models in measurement.
• This metric enhances the diagnostic capabilities of LCMs by offering a robust assessment of measurement precision for categorical latent variables.