Design effect and the intra-cluster correlation (ICC)
Last reviewed 16 October 2026
When you randomise or sample groups, such as clinics, villages or schools, people within a group are more alike than people across groups. Treating them as independent overstates your information, so cluster designs need a larger sample.
Open the sample size calculatorThe design effect
The design effect is DE = 1 + (m − 1) × ICC, where m is the average cluster size and the ICC (intra-cluster correlation) is the share of the outcome's variation that lies between clusters. Multiply the sample size for an individually randomised trial by DE.
A worked example
The trial of 30% versus 20% needs 294 per arm if individuals are randomised. Run it in clinics of about 20 patients with an ICC of 0.05 and the design effect is 1 + 19 × 0.05 = 1.95, so you need 574 per arm, 1,148 in total.
How much the ICC matters
With the same clinics of 20 patients, an ICC of 0.01 gives a total of 700, 0.05 gives 1,148 and 0.10 gives 1,706. Because the ICC is rarely known precisely, show a range and take it from a similar published study.
More clusters beat bigger clusters
Adding patients to each cluster has diminishing returns, because the design effect grows with cluster size. Adding clusters is more efficient, so a trial with very few clusters is usually weak.
Analysis must match the design
Analyse cluster trials with methods that account for clustering, such as mixed-effects models or generalised estimating equations. Ignoring it gives confidence intervals that are too narrow.
Common questions
- What is a typical ICC?
- It varies by outcome and setting. Values around 0.01 to 0.05 are common for many health outcomes, but use a published value for a similar outcome and population rather than a general figure.
- What if cluster sizes vary?
- Use the average size, and add a margin when sizes differ a lot, because unequal clusters reduce the effective sample further.
- Can I ignore clustering if the ICC is small?
- Not if clusters are large. Even an ICC of 0.02 gives a design effect of 1.38 with clusters of 20, and larger with bigger ones.
Keep reading
General methods guidance, not a substitute for advice from a statistician on your specific study.