Use a Cluster Design Effect as an Assumption Check

Describe the cluster before counting people

Patients may share a clinic, school, household, or other unit that creates correlated outcomes. Record the number of clusters, their sizes, the allocation level, and the outcome being considered. Two studies with the same participant count can contain very different information when their cluster structures differ.

The Cochrane guidance on clustered trials describes an approximate design effect using cluster size and intracluster correlation, or ICC. Its purpose is to account for dependence under specified assumptions. It is not a complete design or power calculation.

Document a plausible ICC

The ICC describes similarity within clusters for a particular outcome and setting. Record its source, measurement time, population, and uncertainty. An estimate for one outcome should not silently become the assumption for every endpoint. Keep an assumed value explicitly labeled as assumed.

The cluster assumptions worksheet gives the team a place to preserve the cluster counts and the rationale. Include a summary of the smallest and largest cluster sizes. Substantial variation in size requires more consideration than replacing all clusters with a convenient average.

Work through a fictional calculation

For equal cluster size m, the simple design effect is 1 + (m − 1) × ICC. Suppose 12 clusters contain 20 participants each, for 240 participants, and ICC is assumed to be 0.04. The design effect is 1 + 19 × 0.04 = 1.76. Dividing 240 by 1.76 gives an approximate effective sample size of 136.36.

The cluster design-effect tool explores this calculation. With the same cluster size and ICC 0.01, the design effect becomes 1.19; with ICC 0.08 it becomes 2.52. All numbers in this example are fictional planning assumptions.

Keep the interpretation within its limits

An effective sample size is an information approximation, not a count of actual participants or clusters. Do not round it into a recruitment promise. The number of clusters, unequal cluster sizes, small-sample inference, covariate adjustment, and analysis model can all matter beyond this shortcut.

If a published effect already comes from an appropriate cluster-adjusted analysis, applying the design effect again can double-adjust its uncertainty. Record which estimate you plan to use before modifying any input.

Finish with a short assumption table and unresolved questions for the analysis plan. A useful conclusion is that precision is sensitive to the ICC range, followed by the evidence needed to narrow that range. It is not that the study now has adequate power.

Source checked on 31 October 2026; fictional arithmetic independently recalculated.