Are individual outcomes allowed in a cluster trial?

Yes, but individual-level data do not erase cluster-level assignment or dependence.

Does doubling participants within each cluster double useful information?

Usually not when outcomes are positively correlated within clusters; the gain depends on the design.

Must every cluster contain the same number of people?

No, although unequal sizes require appropriate planning and analysis rather than automatic use of an equal-size formula.

AI Glossary / Statistics and probability / Intermediate

Cluster-randomized controlled trial

Definition

A cluster-randomized controlled trial assigns groups, rather than individual people, at random to intervention or comparison conditions. Groups may be clinics, schools, workplaces or communities. Outcomes can still be measured for individual members, but the analysis must reflect the group-level assignment and dependence among members. For AI projects, this design can evaluate a shared workflow when individual assignment would be impractical or create substantial spillover.

What it means

The cluster is the unit randomized, while the unit measured may be a person, encounter or organization. People in one clinic share staff, routines and local conditions, so their outcomes often contain less independent information than the same number of unrelated observations. A trial with many patients but very few clinics does not automatically have strong evidence about an intervention assigned to clinics. The number of clusters and the variation between them matter alongside the total participant count.

How it works

Define eligible clusters, the intervention, comparison and intended treatment effect before randomization. Decide when and how individuals will be recruited, because recruitment after assignments are known can introduce selection bias. Randomize clusters using an appropriate procedure, potentially with stratification or matching specified in advance. Measure outcomes consistently and record both cluster and participant flow. Analysis may use cluster summaries, suitable mixed models or marginal approaches, depending on the question and design. Account for clustering, allocation restrictions and limited cluster numbers. Report effect estimates and uncertainty, along with assumptions about dependence and losses to follow-up.

Three practical examples

Clinical workflow: assign clinics to an AI-supported review process or usual practice while measuring patient-level outcomes.
Education research: assign schools to a tutoring program, recognizing that pupils share teachers and classroom conditions.
Workplace evaluation: assign service teams to a drafting assistant when staff within a team routinely exchange procedures and outputs.

Think of it this way

Think of testing a new recipe by assigning whole kitchens. Fifty meals from one kitchen share its equipment and cook; they are not equivalent to fifty meals prepared independently in fifty kitchens. Collecting more meals helps, but learning whether the recipe works broadly also requires enough separate kitchens.

Worked example

Suppose an illustrative trial includes 40 clinics with exactly 20 participants each, giving 800 participants. Assume a simple exchangeable correlation model with intracluster correlation 0.05 and equal cluster sizes. The usual design-effect approximation is 1 + (m - 1)rho, where m is cluster size. Here it is 1 + 19 x 0.05 = 1.95. Dividing 800 by 1.95 gives about 410 independent-observation equivalents for this simplified variance comparison. This does not turn the design into a trial with 410 independently randomized people, nor does it supply a complete power calculation. Increasing cluster size to 40 would produce 1,600 participants but a design effect of 2.95, approximately 542 equivalents. The example shows diminishing information from enlarging existing clusters under these assumptions.

Illustrative case study

Fictional case: Juniper Care Network plans to evaluate an AI scheduling intervention. Staff would find it difficult to use two incompatible scheduling processes within the same clinic, so the researchers consider clinic-level assignment. They first examine how many clinics can participate and how much outcomes vary between clinics. The protocol defines a consistent patient-selection procedure and protects recruitment from knowledge of assignment where feasible. Analysis planning includes the limited number of randomized units, rather than relying on the large number of appointments. This invented example illustrates design choices; it does not claim that the intervention improves access or health.

Common misconceptions and mistakes

Cluster randomization is not the same as grouping people into clusters after data collection. Ignoring within-cluster dependence commonly understates uncertainty when correlation is positive. A small intracluster correlation can still matter when clusters are large. The simple equal-size design-effect formula is not universal: unequal sizes, repeated periods, outcome models and informative cluster size can change the calculation. Two large clusters, one per condition, cannot reliably separate an intervention effect from the difference between those particular clusters. Group assignment also does not remove ethical responsibilities toward individual participants.

Business use

Businesses evaluating AI across branches or teams can use cluster trials when workflows are shared and contamination between individually assigned users is likely. The practical benefit must be balanced against greater design and analysis demands. Plan enough distinct sites, preserve the randomization record and define whether the target is an average effect across individuals or across clusters. Those targets can differ when cluster sizes vary. Track implementation and cross-cluster spillover so a result can be interpreted as evidence about the actual intervention delivered.

Check your understanding

Why does the worked example not justify treating all 800 participants as independent?

Answer: Participants within each clinic share an assumed correlation of 0.05. The resulting variance inflation is approximately 1.95 under the stated model. The original 40 randomized clinics remain central to inference and design-specific uncertainty.

Related terms

Clinical trial
Design effect
Generalized estimating equation
Allocation concealment
cluster randomization
intracluster correlation
contamination
mixed-effects model
generalized estimating equations
cluster size
unit of analysis

Learn next

distinguish the unit of randomization from the unit of observation and the target of inference. Study intracluster correlation, unequal cluster sizes and small-sample analysis. Then compare parallel cluster trials with cluster crossover and stepped-wedge designs, which add time structure and require additional assumptions about trends and correlations.

Sources

Campbell and colleagues, CONSORT extension to cluster randomised trials: https://www.bmj.com/content/345/bmj.e5661 Kerry and Bland, The intracluster correlation coefficient in cluster randomisation: https://www.bmj.com/content/316/7142/1455.1 Hemming and colleagues, How to design efficient cluster randomised trials: https://www.bmj.com/content/358/bmj.j3064

Previous: Cluster labeling Next: Cluster Sampling