Variance estimation for clustered recurrent event data with a small number of clusters

Area of Science:

• Biostatistics
• Epidemiology
• Clinical Trials

Background:

• Recurrent events in biomedical studies often exhibit within-subject correlation, violating independence assumptions.
• Semi-parametric proportional rates models commonly use a robust (sandwich) variance estimator, extended for clustered subjects.
• Existing variance estimators are inaccurate when dealing with a small number of moderate-to-large-sized clusters.

Purpose of the Study:

• To evaluate the accuracy of variance estimators in clustered settings with few, large clusters.
• To propose and assess improved variance estimation methods for recurrent event data in such settings.
• To compare hospitalization rates between Canada and the U.S. using the developed methods.

Main Methods:

• Simulation studies were conducted to compare the performance of different variance estimators.
• A corrected robust variance estimator was developed and evaluated.
• Jackknife and bootstrap variance estimators were also proposed and simulated.
• The methods were applied to a multi-centre dialysis study comparing hospitalization rates.

Main Results:

• The standard robust variance estimator demonstrated significant inaccuracy in the studied setting.
• The proposed corrected robust variance estimator showed considerably improved accuracy.
• The corrected robust estimator outperformed jackknife and bootstrap estimators in accuracy.
• Hospitalization rates were compared between Canada and the U.S. using the validated methods.

Conclusions:

• The corrected robust variance estimator is a more accurate method for analyzing recurrent events in clustered biomedical studies with few, large clusters.
• The proposed methods provide reliable variance estimation, enhancing the analysis of complex health data.
• The study successfully applied these methods to compare international healthcare outcomes in dialysis patients.