.jpg){kind=link}
- Published
- Updated
Adding more people within one cluster may not provide the same information gain as adding an independent cluster. With few randomised units, balance and uncertainty estimation need particular attention, even when the overall participant count looks substantial.
Define the question and assumptions
Show cluster numbers, size distributions and within-cluster similarity assumptions separately. Large units may contribute disproportionately, while small units can represent different populations. The analysis method needs specialist assessment for a small number of clusters; individual sample size alone cannot justify a claim of adequate power or precision for the actual allocation design.
Hypothetical study example
In a hypothetical study, one arm contains a few large centres and the other more small centres. Person counts appear similar. The team reviews cluster sizes and baseline features rather than assuming that large enrolment removes the shared influence of practice patterns within each centre or makes the design equivalent to individual randomisation.
Design review checks
- Report cluster counts and size distributions separately for every study arm.
- Inspect how large clusters contribute to and influence the overall estimate.
- Evaluate planning assumptions across different levels of within-cluster similarity.
Record the decision and its limits
An accepted design also considers the impact of losing an entire cluster. Chance differences between a small number of units may remain important and should not disappear from reporting. Select corrections using statistical rationale. Instead of promising that more people fix every structural problem, explain which independent information remains limited.
Frequently asked questions
- Does a large person count solve having few clusters?
- Not automatically. Independent cluster count also determines available information.
- Do equal person counts imply equivalent designs?
- No. Cluster distributions and within-cluster relationships can differ.