Multilevel latent class models with dirichlet mixing distribution.

Area of Science:

• Social Science
• Biomedical Studies
• Statistical Modeling

Background:

• Latent class analysis (LCA) and latent class regression (LCR) are standard for categorical outcomes.
• Traditional methods assume data independence, limiting use in clustered designs (e.g., familial studies).

Purpose of the Study:

• To develop and evaluate methods for multilevel latent class models.
• To address computational intensity of existing maximum likelihood (ML) approaches for large clusters or classes.

Main Methods:

• Proposed a maximum pairwise likelihood (MPL) approach using a modified expectation-maximization (EM) algorithm.
• Introduced a simpler latent class analysis with robust standard errors as an alternative.
• Compared ML, MPL, and robust standard error methods via simulation studies.

Main Results:

• Multilevel latent class models with random effects provide interpretable structures.
• The MPL approach is computationally efficient for large datasets.
• MPL estimates show comparable precision to ML estimates.
• The robust standard error method is consistent and robust, though less efficient.

Conclusions:

• The proposed methods effectively analyze multilevel categorical data.
• MPL offers a computationally feasible alternative to ML for large-scale latent class analysis.
• These models enhance tools for analyzing clustered data in social and biomedical research.