.jpg){kind=link}
- Published
- Updated
When sites, schools or care units are randomised, participants are not independently assigned. Shared processes within a cluster can correlate outcomes. Selection of the analytical unit should explicitly respect that structure instead of relying only on the level at which measurements are stored.
Define the question and assumptions
Define allocation, intervention delivery and measurement levels separately. Availability of participant-level records does not justify treating every observation as independent. Align sample size and analysis with clustering, and plan rules for changing cluster boundaries or people moving between units during follow-up. These operational details can affect what comparison the analysis actually represents.
Hypothetical study example
In a hypothetical study, hospital units receive the intervention while outcomes are measured in individuals. Thousands of people are observed, but relatively few units participate. The team uses an approach accounting for clustering and unit differences rather than presenting the total person count as an equal amount of independent statistical information.
Design review checks
- Map allocation, intervention delivery and outcome measurement levels explicitly.
- Record the source of cluster membership and any changes over time.
- Check how clustering enters both sample size planning and final analysis.
Record the decision and its limits
Accepted reporting shows the flow of clusters and individuals. Assess contamination and different recruitment processes separately. A cluster design does not automatically settle consent or ethical questions. Do not change methods simply for a smaller p-value; explain the analytical target and the uncertainty remaining under the method used.
Frequently asked questions
- Do individual records justify independent analysis?
- No. Shared cluster membership can correlate observations.
- Is an individual participant-flow diagram enough?
- No. Cluster counts and cluster losses should also be visible.