A Bayesian Shrinkage Model for Incomplete Longitudinal Binary Data with Application to the Breast Cancer Prevention...

Area of Science:

• Biostatistics
• Longitudinal Data Analysis
• Missing Data Methods

Background:

• Longitudinal studies frequently encounter missing data due to skipped visits and loss to follow-up.
• Standard methods often require unverifiable assumptions for valid inference when data is missing.
• Full data estimands are typically unidentified without imposing strong, unverified assumptions.

Purpose of the Study:

• To develop a robust statistical methodology for inference in randomized longitudinal studies with complex missing data patterns.
• To address the identification problem of full data estimands under non-ignorability and intermittent missingness.
• To apply the developed methodology to real-world data from the Breast Cancer Prevention Trial.

Main Methods:

• Proposed an exponential tilt model linking non-identifiable and partially identifiable distributions.
• Assumed a non-future dependence model for dropout and partial ignorability for intermittent missingness.
• Employed a Bayesian shrinkage model for the observed data distribution to mitigate dimensionality issues.

Main Results:

• Demonstrated that full data estimands can be expressed as functionals of the observed data distribution under the proposed model.
• Simulation studies compared the novel approach against fully parametric and fully saturated models.
• The methodology proved effective in handling missing data in longitudinal settings.

Conclusions:

• The proposed Bayesian exponential tilt model offers a viable solution for inference in longitudinal studies with missing data.
• This approach allows for valid estimation of full data parameters without relying on untestable assumptions.
• The method is applicable to complex longitudinal datasets, as shown by its application to breast cancer trial data.