A note on MAR, identifying restrictions, model comparison, and sensitivity analysis in pattern mixture models with...

Area of Science:

• Statistics
• Biostatistics

Background:

• Pattern mixture models are common for incomplete longitudinal data.
• These models often lack inherent identifiability, requiring specific restrictions.
• Existing identification strategies, particularly for missing at random (MAR) assumptions, can be problematic with multivariate normal models and covariates.

Purpose of the Study:

• To investigate the conditions under which identifying restrictions for MAR exist in pattern mixture models with multivariate normality.
• To explore strategies for identifying sensitivity parameters for sensitivity analysis or Bayesian analysis.
• To propose alternative modeling and sensitivity analysis approaches using less restrictive distributional assumptions and address issues with covariates.

Main Methods:

• Exploration of necessary conditions for MAR identifying restrictions under multivariate normality.
• Development of strategies for identifying sensitivity parameters.
• Proposal of alternative modeling strategies with less restrictive distributional assumptions.
• Utilizing the deviance information criterion (DIC) for model comparison.
• Conducting a simulation study and applying methods to a longitudinal clinical trial.

Main Results:

• Identified conditions for MAR identifying restrictions under multivariate normality.
• Demonstrated potential issues with baseline covariates and proposed a residual-based identifying restriction as a solution.
• Evaluated the performance of different modeling approaches through simulation and real-world data.

Conclusions:

• The study clarifies conditions for identifiability in pattern mixture models, particularly concerning MAR assumptions.
• Alternative methods are proposed to overcome limitations of existing approaches, especially when dealing with complex covariate structures.
• The findings offer practical solutions for analyzing incomplete longitudinal data more robustly.