Unified estimation for Cox regression model with nonmonotone missing at random covariates.

Area of Science:

• Biostatistics
• Statistical Modeling
• Survival Analysis

Background:

• Missing covariate data in Cox regression models pose significant challenges to accurate statistical inference.
• Existing methods like complete case analysis can lead to biased and inefficient estimates.
• Multiple Imputation (MI) offers a robust solution but can be complex to implement.

Purpose of the Study:

• To develop and evaluate a unified estimator for Cox regression models with missing at random covariate data.
• To extend the use of parametric working models for extracting information from incomplete observations.
• To provide a method that is consistent, efficient, and easily implementable.

Main Methods:

• Utilizing parametric working models to leverage partial information from incomplete observations.
• Developing a unified estimator that incorporates auxiliary variables to reduce bias and enhance efficiency.
• Comparing the proposed unified estimator with the substantive model compatible modification of the fully conditional specification MI (SMC-FCS) estimator through simulation studies.

Main Results:

• The unified estimator is demonstrated to be consistent and more efficient than the (weighted) complete case estimator.
• Simulation studies show the unified estimator is as efficient as the SMC-FCS MI estimator.
• The proposed method is easily implementable using standard statistical software.

Conclusions:

• The unified estimator provides a consistent and efficient approach for Cox regression with missing covariate data.
• It offers a practical alternative to multiple imputation, simplifying the analysis of incomplete datasets.
• The method is validated through simulations and illustrated with a clinical trial dataset.