Multiple imputation of longitudinal categorical data through bayesian mixture latent Markov models

Area of Science:

• Statistics
• Biostatistics
• Longitudinal Data Analysis

Background:

• Multiple imputation (MI) is crucial for handling missing data in longitudinal studies.
• Existing methods like latent class models are effective for cross-sectional data but limited for longitudinal applications.
• There is a need for advanced imputation techniques that capture complex dependencies in longitudinal categorical data.

Purpose of the Study:

• To introduce and evaluate Bayesian mixture Latent Markov (BMLM) models for multiple imputation (MI) of missing categorical covariates in longitudinal studies.
• To demonstrate the capability of BMLM models in capturing both within-occasion and between-occasion dependencies.
• To compare the performance of BMLM for MI against complete case analysis and MICE.

Main Methods:

• Development of Bayesian mixture Latent Markov (BMLM) models.
• Application of BMLM for multiple imputation (MI) of categorical longitudinal data.
• Performance evaluation through simulation studies and empirical data analysis.
• Comparison with complete case analysis and Multiple Imputation by Chained Equations (MICE).

Main Results:

• The proposed BMLM model demonstrates strong performance in accurately retrieving analysis model parameters.
• BMLM effectively captures complex relationships, including lagged dependencies and associations between time-varying and time-constant variables.
• Complete case analysis and MICE showed limitations, providing correct estimates only for certain data aspects.

Conclusions:

• BMLM models offer a flexible and powerful tool for multiple imputation (MI) in longitudinal studies with categorical data.
• The BMLM approach preserves variable measurement scales and complex relationships within and across time points.
• BMLM significantly outperforms competing methods in accurately estimating parameters from longitudinal datasets with missing categorical covariates.