The performance of small sample correction methods for controlling type I error when analyzing parallel cluster...

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Statistical Inference

Background:

• Cluster randomized trials (CRTs) often have fewer than 50 clusters.
• Analytical methods for CRTs commonly assume a large sample size.
• This assumption may not hold, potentially affecting the reliability of treatment effect estimates.

Purpose of the Study:

• To review the simulation study literature on small sample corrections for parallel CRTs.
• To evaluate the performance of various analytical approaches under small sample conditions.
• To identify corrections that preserve nominal type I error rates.

Main Methods:

• Systematic search of Ovid Medline and Web of Science for simulation studies up to August 30, 2024.
• Inclusion of studies evaluating binary and continuous outcomes using generalized linear mixed models, generalized estimating equations, or cluster-level analyses.
• Independent duplicate full-text screening and data abstraction.

Main Results:

• For continuous outcomes, cluster-level analyses, linear mixed models with Satterthwaite correction, and GEE with Fay and Graubard correction generally preserve type I error with as few as six clusters.
• For binary outcomes, unweighted/inverse-variance weighted cluster-level analyses and GLMM with between-within correction can achieve nominal type I error with 10 clusters, but may be anticonservative or conservative in certain settings.
• More than 40 clusters are required to guarantee nominal type I error across all evaluated settings.

Conclusions:

• The performance of small sample corrections in parallel CRTs is complex and context-dependent.
• While some corrections work with very few clusters, robust control of type I error requires a larger number of clusters.
• Further research may be needed to refine corrections for specific scenarios in small-sample CRTs.