Bayesian imputation of time-varying covariates in linear mixed models

Area of Science:

• Statistics
• Biostatistics
• Data Science

Background:

• Large observational datasets frequently contain multiple missing values, posing a significant analytical challenge.
• Multiple imputation using chained equations (MICE) is common but can be suboptimal for complex longitudinal data.
• Existing methods struggle with time-varying covariates, their endogeneity, and functional relationships with outcomes.

Purpose of the Study:

• To extend and evaluate a flexible Bayesian alternative for handling complex incomplete longitudinal data.
• To address limitations of standard imputation methods, particularly concerning time-varying covariates and their assumptions.
• To compare Bayesian approaches regarding covariate endogeneity and functional form assumptions.

Main Methods:

• Developed and studied a flexible Bayesian approach beyond the standard multivariate normal distribution.
• Investigated assumptions about covariate endogeneity and functional form in Bayesian imputation models.
• Utilized simulation studies and two real-world data examples for evaluation.

Main Results:

• The proposed flexible Bayesian method demonstrates improved handling of complex longitudinal data with missing values.
• Violations of standard assumptions regarding covariate endogeneity and functional form can lead to bias.
• The study quantifies the consequences of these assumption violations in both simulated and real data.

Conclusions:

• Flexible Bayesian methods offer a robust alternative to standard approaches for complex incomplete longitudinal data.
• Careful consideration of covariate endogeneity and functional form assumptions is crucial for valid inference.
• The proposed approach provides a more reliable tool for analyzing challenging observational datasets.