Learning Latent and Hierarchical Structures in Cognitive Diagnosis Models.

Area of Science:

• Psychometrics
• Educational Measurement
• Latent Variable Modeling

Background:

• Cognitive Diagnosis Models (CDMs) are essential for educational and psychological measurement, relying on Q-matrices for item-attribute relationships.
• Current CDM applications often require subjective pre-specification of attribute hierarchies and Q-matrices, leading to potential misspecification.
• Recent studies highlight the need for data-driven approaches to learn these structures.

Purpose of the Study:

• To develop a method for jointly learning latent attribute structures and their hierarchies in CDMs with minimal assumptions.
• To address the limitations of subjective and potentially misspecified pre-specified structures in existing CDM applications.
• To provide a robust framework for attribute structure discovery in diagnostic assessments.

Main Methods:

• A penalized likelihood approach is proposed for simultaneous estimation of latent structures and attribute hierarchies.
• The method automatically selects the number of latent attributes.
• An expectation-maximization (EM) algorithm is developed for computational efficiency.

Main Results:

• The proposed penalized likelihood method effectively learns latent and hierarchical attribute structures from observed data.
• Statistical consistency is established under mild conditions, ensuring theoretical soundness.
• Simulation studies and real-data applications demonstrate the method's strong performance.

Conclusions:

• The developed approach offers a data-driven solution for uncovering complex attribute structures in CDMs.
• This method reduces subjectivity and improves the accuracy of cognitive diagnosis.
• It has significant implications for enhancing the validity and reliability of educational and psychological assessments.