Accounting for complex intracluster correlations in longitudinal cluster randomized trials: a case study in malaria...

Area of Science:

• Biostatistics
• Epidemiology
• Public Health

Background:

• Cluster randomized trials (CRTs) are crucial for evaluating malaria vector control interventions.
• Longitudinal CRTs require accounting for the intra-cluster correlation coefficient (ICC) for accurate analysis.
• Standard ICC structures include exchangeable, block exchangeable, and exponential decay.

Purpose of the Study:

• To empirically explore the impact of various correlation structures on treatment effect inferences in longitudinal CRTs.
• To identify methodological literature gaps concerning ICC in malaria trials.
• To provide practical recommendations for the design and analysis of longitudinal CRTs.

Main Methods:

• Data from a parallel-arm CRT in Tanzania comparing insecticide-treated bed-nets were re-analyzed.
• Mixed-effects logistic regression was used with five different correlation structures and a robust variance estimator.
• Malaria prevalence was assessed via repeated cross-sectional surveys at multiple time points.

Main Results:

• Correlation structures varied significantly across models, with the unstructured model showing the best fit.
• While point estimates were similar, flexible correlation structures altered conclusions regarding statistical significance.
• Robust variance estimators generally resulted in wider confidence intervals; under-specification led to lower coverage probabilities.

Conclusions:

• Flexible correlation structures should be considered in longitudinal CRTs, especially for malaria trials with fluctuating outcomes.
• Researchers should assess the sensitivity of results to different ICC assumptions.
• Robust variance estimators are a valuable consideration in the absence of definitive methods for selecting the best-fitting correlation structure.