Recommendations for choosing an analysis method that controls Type I error for unbalanced cluster sample designs with...

Area of Science:

• Biostatistics
• Clinical Trials Methodology

Background:

• Cluster-randomized designs are frequently used in various research fields.
• Accurate Type I error rate control is crucial for valid statistical inference in these designs.
• Analytic methods must account for within-cluster correlation, especially with unbalanced cluster sizes.

Purpose of the Study:

• To evaluate Type I error rates of one-stage and two-stage analytic methods for cluster-randomized designs.
• To compare the performance of different hypothesis testing approaches under unbalanced data conditions.
• To identify optimal methods for maintaining accurate Type I error control in practice.

Main Methods:

• Theoretical derivations and simulation-based approaches were employed.
• One-stage methods used observed data with general linear mixed models.
• Two-stage methods utilized cluster-specific means in general linear univariate models.

Main Results:

• Both one-stage and two-stage models achieve exact Type I error rates with equal cluster sizes.
• With unbalanced data, Type I error inflation can occur as exact tests do not exist.
• A weighted two-stage model performed best with ≥6 clusters; Kenward-Roger one-stage method performed well with ≥14 clusters.

Conclusions:

• The choice of analytic method significantly impacts Type I error control in cluster-randomized trials with unbalanced data.
• A weighted two-stage approach is recommended for studies with at least six clusters per arm.
• The Kenward-Roger method is the preferred one-stage approach for unbalanced data, particularly when small sample sizes and low intracluster correlation are present.