How Many Clusters Do You Need in a Cluster-Randomized Trial?

Is There a Minimum Number of Schools, Clinics, or Other Groups You Must Randomize?

You want to evaluate a new instructional intervention across several schools. Each school has hundreds of eligible students, and the intervention must be implemented at the institutional level. You can recruit eight schools, four for the intervention and four for the comparison condition.

Would eight schools be enough? What about twelve, twenty, or forty?

Researchers sometimes look for a universal minimum number of clusters, hoping to find a threshold that automatically makes a trial acceptable. Unfortunately, the adequacy of a cluster-randomized trial cannot be established through a single number.

The required number depends on the intervention effect the study aims to detect or estimate, the similarity of outcomes within clusters, the number of participants per cluster, and the planned statistical analysis. A trial with few clusters may be feasible for one purpose yet inadequate for another.

No Universal Minimum Applies to Every Cluster-Randomized Trial

How to Determine a Defensible Number of Randomized Clusters

Why the number of clusters matters independently of participant count

In a cluster-randomized trial, intact groups are assigned to intervention conditions. These groups are the independent units of random allocation, even when outcomes are measured for many individuals within each group.

If 12 schools are randomized and each contributes 200 students, the trial has 12 randomized clusters and 2,400 participating students. The student observations provide information, but their treatment assignments are determined by only 12 school-level allocations.

Participants within schools may share educational resources, teachers, administrative policies, and other influences. Their outcomes can therefore be correlated.

This dependence is one reason cluster randomization usually requires larger samples than comparable individual randomization.

The number of clusters also affects how reliably treatment effects and uncertainty can be estimated. A large participant count does not create additional independent randomization units.

What determines the required number of clusters?

For a conventional parallel, two-arm cluster-randomized trial, the main planning inputs include the primary outcome, target intervention effect, outcome variability, ICC, cluster size, allocation ratio, desired power or precision, and significance level.

The planned statistical analysis also matters, especially when the number of clusters is small.

Why a fixed minimum such as 10, 20, or 30 clusters is misleading

Rules of thumb can be useful for recognizing potential problems, but they should not be treated as universal design standards.

A trial with 20 clusters may have inadequate power for a small intervention effect and substantial within-cluster correlation. Another trial with fewer clusters may provide useful estimation for a large effect under favorable assumptions, although its inferential limitations still require careful treatment.

The relevant distinction is between the minimum number needed to meet a specified statistical objective and an arbitrary threshold considered acceptable without reference to that objective.

Researchers should therefore avoid claiming that a trial is adequately powered simply because it exceeds a commonly cited cluster count.

A simplified calculation illustrates how cluster requirements arise

For a continuous outcome in a balanced two-arm parallel trial with equal cluster sizes, a simplified normal-approximation formula can illustrate the relationship between the number of clusters and the design parameters.

The calculation is deliberately simplified. It assumes independent clusters, approximately equal cluster sizes, a continuous outcome, a common variance, and a basic unadjusted comparison using normal approximations.

Five clusters per arm should not be interpreted as a general recommendation. With so few clusters, finite-sample corrections and appropriate critical values may materially change the required number. The illustrative formula can underestimate requirements when conventional normal approximations are inappropriate.

A defensible trial calculation should incorporate the planned inferential method, especially when only a small number of clusters is available.

Why increasing cluster size cannot always solve a shortage of clusters

Increasing the number of participants within each cluster can improve precision, but the gain diminishes when outcomes are positively correlated.

For a fixed number of clusters and positive ICC, the variance contribution associated with between-cluster differences cannot be eliminated simply by recruiting more individuals.

Hemming and colleagues describe how power and precision curves can reveal the point at which increasing cluster size provides little additional benefit. They recommend determining cluster number and cluster size jointly rather than independently.

For example, recruiting 100 additional students from each existing school may produce less improvement than recruiting several additional schools with smaller student samples.

The appropriate choice depends on the ICC, target effect, implementation costs, and the number of clusters already available.

Very few clusters create problems beyond low power

Small numbers of clusters can produce unstable treatment-effect estimates, substantial baseline imbalance, and unreliable uncertainty estimates under some conventional analytical methods.

Randomization remains genuine when only a few clusters are assigned, provided the allocation procedure is valid. However, the number of possible allocations is limited, and chance differences between arms may be considerable.

Standard mixed-effects models and generalized estimating equations can perform poorly when small-sample issues are ignored. Depending on the design, suitable corrections, cluster-level analyses, or randomization-based inference may be necessary.

Methodological simulation studies have shown that small-sample corrections can improve control of false-positive error rates, but they may also reduce power relative to uncorrected methods. Consequently, power calculations should incorporate the analysis actually intended.

The distinction between having many observations and having enough independently randomized clusters is particularly important in small trials.

Allocation balance matters when the cluster count is small

Suppose a trial randomizes eight schools, four to each condition. By chance, one arm may contain schools with substantially higher baseline achievement, greater resources, or more experienced teachers.

Such imbalance can occur even when randomization is conducted correctly.

Stratification, matching, or restricted randomization may improve balance on important baseline characteristics. However, the procedures must be planned before allocation and reflected appropriately in the analysis.

Researchers should not use restricted allocation merely to make observed groups appear similar without considering the inferential consequences.

The decision about matching or stratifying clusters should be based on the design's substantive and statistical requirements.

Cluster-size variability affects the required number

Clusters may differ substantially in the number of eligible or recruited participants. A school with 80 students and a school with 500 students each represent one randomized cluster when allocation occurs at the school level.

Unequal sizes can affect efficiency and the weighting of intervention effects. Calculations based only on average cluster size may be misleading when size variability is substantial.

Researchers should estimate the likely distribution of cluster sizes and use an approach compatible with the intended analysis and estimand.

When size variation is large, simulation-based planning may be particularly useful.

Cluster loss requires a separate allowance

Loss of an entire cluster is different from losing several individual participants. A school withdrawing from a trial may remove hundreds of student outcomes and reduce the number of randomized units contributing to the comparison.

Researchers should distinguish expected participant attrition from expected cluster attrition.

If cluster withdrawal is plausible, the planned number of clusters may need to be increased. The allowance should be based on credible assumptions rather than a generic percentage applied without justification.

Power-based and precision-based planning answer different questions

Power calculations determine whether a study is likely to reject a specified null hypothesis under an assumed effect and statistical model. Precision-based calculations focus on the expected uncertainty around the estimated effect, often expressed through confidence interval width.

A trial intended to test a specific effectiveness hypothesis may prioritize power. A trial intended to estimate an intervention effect for policy planning may instead emphasize precision.

Neither approach eliminates the need to account for clustering. The required number of clusters depends on the chosen inferential objective.

For exploratory or feasibility studies, a smaller number of clusters may be appropriate when the objectives are explicitly limited to implementation, recruitment, acceptability, or preliminary parameter estimation. Such studies should not be presented as adequately powered effectiveness trials without supporting evidence.

Simulation may be necessary for complex designs

Analytical formulas are useful for conventional parallel cluster-randomized trials, but more complex arrangements may require simulation.

Examples include repeated measurements, unequal cluster sizes, binary outcomes with low event rates, complex covariate adjustment, restricted randomization, or nonstandard allocation structures.

Simulation involves generating data under plausible assumptions, applying the intended analysis repeatedly, and estimating how often the procedure achieves its statistical objective.

Its validity depends on the assumed data-generating process, parameter values, and analysis method. Simulation should be accompanied by sensitivity analyses when important assumptions are uncertain.

Feasibility must be evaluated alongside statistical requirements

A calculation might indicate that 40 schools are needed, while the research team can recruit only 12. The appropriate response is not simply to increase the number of students per school until the total participant count appears large.

Researchers may need to recruit additional institutions, modify the intervention, revise the target effect or precision objective, use a more efficient design, or reconsider whether a definitive trial is feasible.

Any change to the target effect should be scientifically justified. Increasing the assumed effect merely to reduce the required sample size can produce an inadequately powered trial.

The financial and methodological costs of cluster randomization should therefore be considered before committing to a particular number of institutions.

Estimating How Many Schools a Trial Might Need

Interpretation: The preliminary calculation provides a starting point, not a definitive requirement of ten schools. A complete analysis may require more clusters because of small-sample inference, uncertainty in the ICC, attrition, and other features of the intended design.

The example also illustrates why the number of clusters cannot be determined from participant count alone. The calculation begins with the outcome, target effect, correlation structure, and inferential objective.

Common Mistakes When Deciding How Many Clusters to Recruit

Determine Cluster Requirements From the Research Objective, Not an Arbitrary Threshold

Begin by specifying whether the trial aims to test an intervention effect, estimate an effect with a desired degree of precision, or assess feasibility.

For a definitive effectiveness trial, use the primary outcome, scientifically justified target effect, ICC assumptions, expected cluster sizes, allocation ratio, and planned statistical method to calculate the required number of clusters.

Then evaluate whether the resulting design is operationally feasible. If not, revise the study transparently rather than treating the available number of clusters as automatically sufficient.

In educational research, schools or classrooms may be the appropriate randomized units because interventions operate through shared teaching practices. However, the number of institutions available should not determine the statistical justification after the fact.

A well-designed cluster trial establishes what it needs to estimate, calculates what information is required, and then evaluates whether the necessary clusters can realistically be recruited.

Before Deciding How Many Clusters Are Enough

Frequently Asked Questions About the Minimum Number of Clusters

Enough Clusters Means Enough Independent Information for the Intended Question

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.