Semiparametric marginal regression for clustered competing risks data with missing cause of failure.

Area of Science:

• Statistics
• Biostatistics
• Epidemiology

Background:

• Multicenter studies often generate clustered competing risks data, where outcomes are influenced by cluster size (informative cluster size or ICS).
• Incomplete observation of failure causes is common in real-world health data.
• Existing statistical methods struggle to handle both ICS and missing failure causes simultaneously in population-averaged analyses.

Purpose of the Study:

• To develop a novel statistical methodology for analyzing clustered competing risks data with informative cluster size (ICS) and missing causes of failure.
• To provide a robust framework for population-averaged analysis that accommodates complex data structures common in multicenter health research.

Main Methods:

• Proposes a semiparametric marginal proportional cause-specific hazards model.
• Introduces a maximum partial pseudolikelihood estimator under a missing at random assumption, utilizing auxiliary variables to strengthen the assumption.
• The method is designed to be robust, not requiring assumptions on within-cluster dependence and explicitly handling ICS.

Main Results:

• Asymptotic properties of estimators for regression coefficients and marginal cumulative incidence functions are rigorously established.
• Simulation studies demonstrate the proposed method's effectiveness and highlight the invalidity of inferences from methods ignoring ICS or within-cluster dependence.
• The method was successfully applied to HIV data from a sub-Saharan African multicenter study with substantial missing cause-of-failure information.

Conclusions:

• The proposed statistical method offers a valid approach for analyzing clustered competing risks data with informative cluster size and missing failure causes.
• This methodology is crucial for accurate population-averaged inference in complex observational studies, particularly in public health research.
• Ignoring informative cluster size and missing data mechanisms can lead to significantly biased results in competing risks analyses.