Designing Learning Intervention Studies: Identifiability of Heterogeneous Hidden Markov Models

Area of Science:

• Statistics
• Psychometrics
• Educational Research

Background:

• Hidden Markov Models (HMMs) are widely used for longitudinal data but often assume constant emission probabilities and irreducible Markov chains.
• These assumptions may not hold in educational and psychological research, necessitating more flexible models.
• Restricted HMMs (RHMMs) integrate HMMs with Restricted Latent Class Models (RLCMs) for detailed analysis of attribute profile changes over time.

Purpose of the Study:

• To generalize identifiability conditions for conventional HMMs to heterogeneous HMMs and RHMMs.
• To establish critical identifiability conditions for RHMMs, ensuring accurate statistical inference.
• To provide insights for designing interventions and analyzing attribute profiles in longitudinal studies.

Main Methods:

• Generalizing identifiability theory for homogeneous HMMs to accommodate time-varying emission probabilities and absorbing states in heterogeneous HMMs.
• Establishing identifiability conditions for RHMMs by combining HMM and RLCM frameworks.
• Applying a heterogeneous HMM to longitudinal data on daily positive and negative affect.

Main Results:

• New identifiability conditions were established for heterogeneous HMMs and RHMMs.
• These conditions extend existing theory by relaxing assumptions of constant emission probabilities and irreducible transition matrices.
• The study demonstrated the practical application of a heterogeneous HMM to affect data.

Conclusions:

• The established identifiability conditions for heterogeneous HMMs and RHMMs are vital for researchers using these models.
• These findings offer guidance for designing more effective interventions and assessment strategies in psychological and educational research.
• The application highlights the utility of advanced HMMs for analyzing complex, time-varying phenomena.