Individual participant data meta-analyses should not ignore clustering

Area of Science:

• Biostatistics
• Epidemiology
• Medical Research Methodology

Background:

• Individual participant data (IPD) meta-analyses are crucial for synthesizing evidence.
• A common practice is to analyze IPD as if from a single study, potentially overlooking data structure.
• Patient data within studies exhibit clustering, which may influence statistical analyses.

Purpose of the Study:

• To compare effect estimates from IPD meta-analyses that ignore versus account for patient clustering.
• To evaluate the impact of clustering on prognostic and treatment effect estimates.
• To provide guidance on robust statistical approaches in IPD meta-analyses.

Main Methods:

• Comparison of logistic regression models using real-world and simulated data.
• Analysis of prognostic factors (e.g., age in traumatic brain injury) and treatment effects (e.g., nicotine gum for smoking cessation).
• Assessment of model performance, including effect estimate bias and confidence interval coverage.

Main Results:

• Prognostic effect of age in traumatic brain injury was similar whether clustering was accounted for or not.
• A family history of thrombophilia was identified as a significant diagnostic marker for deep vein thrombosis only when clustering was considered.
• The treatment effect of nicotine gum for smoking cessation was significantly attenuated when clustering was ignored compared to when it was accounted for.
• Simulations demonstrated that models accounting for clustering performed well, while ignoring clustering led to biased estimates and low coverage.

Conclusions:

• Clustering of patients within studies must be routinely accounted for in IPD meta-analyses.
• Failure to account for clustering can result in misleading effect estimates and erroneous conclusions.
• Adopting methods that incorporate clustering is essential for accurate and reliable meta-analysis findings.