Mixed-effect hybrid models for longitudinal data with nonignorable dropout.

Area of Science:

• Biostatistics
• Longitudinal Data Analysis
• Missing Data Methods

Background:

• Nonignorable dropout is a significant challenge in longitudinal studies, potentially biasing results.
• Existing methods like selection models and pattern-mixture models address this using different distributional factorizations.
• Shared-parameter models (SPMs) are commonly used but rely on a conditional independence assumption that may not always hold.

Purpose of the Study:

• To introduce a novel class of models, mixed-effect hybrid models (MEHMs), for analyzing longitudinal data with nonignorable dropout.
• To provide a unified framework that integrates features of both selection and pattern-mixture models.
• To generalize existing shared-parameter models by relaxing the conditional independence assumption.

Main Methods:

• Developed mixed-effect hybrid models (MEHMs) by factorizing the joint distribution of outcome and dropout processes.
• MEHMs model the dropout process conditional on random effects and the outcome process conditional on dropout patterns and random effects.
• Utilized the nested structure of SPMs within MEHMs to enable likelihood ratio tests for the conditional independence assumption.

Main Results:

• MEHMs directly model the missingness process, similar to selection models.
• MEHMs offer computational advantages akin to pattern-mixture models.
• The proposed framework allows for formal testing of the conditional independence assumption inherent in SPMs.

Conclusions:

• Mixed-effect hybrid models (MEHMs) provide a flexible and powerful approach to handling nonignorable dropout in longitudinal studies.
• MEHMs generalize shared-parameter models and allow for evaluation of their underlying assumptions.
• The utility of MEHMs was demonstrated using data from a pediatric AIDS clinical trial.