Robust Bayesian hierarchical model using normal/independent distributions.

Area of Science:

• Biostatistics
• Longitudinal Data Analysis
• Clinical Trials

Background:

• Multilevel item response theory (MLIRT) models are valuable for longitudinal clinical studies with multiple mixed-type outcomes.
• Standard MLIRT models often assume normal distributions for continuous outcomes and random effects, which can be unrealistic.
• Normality assumptions may lead to misleading results, particularly with outliers or heavy-tailed data.

Purpose of the Study:

• To develop a robust Bayesian approach for MLIRT models using normal/independent (NI) distributions.
• To address limitations of normality assumptions in continuous outcomes and random effects within MLIRT.
• To provide more reliable inference for longitudinal clinical data.

Main Methods:

• Implemented a Bayesian framework incorporating NI distributions for continuous outcomes and random effects in MLIRT.
• Explored various strategies for implementing NI distributions within the Bayesian MLIRT models.
• Conducted extensive simulation studies to evaluate model performance.

Main Results:

• The proposed NI-based MLIRT models demonstrated improved parameter estimation with reduced bias.
• Coverage probabilities for parameter estimates were more reasonable compared to standard models.
• The models effectively handled potential outliers and heavy-tailed distributions in longitudinal data.

Conclusions:

• The Bayesian MLIRT models with NI distributions offer a robust alternative when normality assumptions are questionable.
• These models enhance the validity of inference in longitudinal clinical studies analyzing mixed outcomes.
• Application to the DATATOP study highlights the utility in Parkinson's disease research for treatment effect evaluation.