Why Does Cluster Randomization Usually Require a Larger Sample?

Why Are 1,000 Students in Randomized Schools Not Equivalent to 1,000 Independently Randomized Students?

Suppose you are planning an educational intervention involving 1,000 students. If students are randomized individually, each student's treatment assignment can vary independently of the assignments of other students, subject to the allocation procedure. If you instead randomize 20 schools containing 50 participating students each, every student within a school receives the same assigned condition.

The participant count remains 1,000, but the statistical information available for estimating the intervention effect changes.

Students within the same school may share teachers, instructional resources, administrative policies, and socioeconomic environments. Their outcomes may consequently be more similar than outcomes from students attending different schools. Treating all 1,000 observations as independently randomized would overlook this dependence.

This explains why cluster-randomized trials often require larger samples. The more difficult question is how much larger, and whether researchers should recruit more participants, more clusters, or both.

Correlated Outcomes Reduce the Information Contributed by Each Participant

How Clustering Changes Sample Size, Precision, and Statistical Power

Why observations within a cluster are not fully independent

In an individually randomized trial, treatment assignment occurs at the participant level. In a cluster-randomized trial, intact groups such as classrooms, schools, clinics, or communities are assigned to conditions.

Members of these groups often share contextual influences. Students taught by the same teacher may experience similar instructional quality. Patients attending the same clinic may receive care shaped by shared clinical procedures. Employees in one workplace may experience common organizational policies.

These shared influences can produce correlation between outcomes from participants within the same cluster.

The distinction concerns statistical dependence, not whether participants provide separate measurements. Fifty students in a classroom still contribute fifty observations, but those observations may contain overlapping information about the intervention effect.

The unit that was randomized must therefore be distinguished from the individuals whose outcomes were measured.

What the intracluster correlation coefficient measures

The intracluster correlation coefficient, commonly abbreviated ICC, quantifies the similarity of observations within clusters under a specified statistical model.

For a simple random-intercept model with a continuous outcome, the ICC can be expressed as the proportion of total variance attributable to differences between clusters.

An ICC of 0.05 does not mean that 5% of participants have identical outcomes. It describes a model-based relationship between observations within the same cluster.

The ICC is outcome-specific and context-dependent. Estimates may differ across populations, measurement instruments, organizational levels, and time periods. ICC definitions and estimation methods also require care for binary, count, or other noncontinuous outcomes.

Researchers should not automatically transfer an ICC from one study to another merely because both involve schools or hospitals.

How the design effect inflates the required sample

The design effect describes how clustering changes the variance of an estimator relative to a specified comparison design.

For a simple parallel cluster-randomized trial with approximately equal cluster sizes, a frequently used approximation is:

In a simplified planning exercise, an individually randomized sample-size requirement may be multiplied by the design effect to approximate the participant count needed under clustering.

For example, if an individually randomized comparison requires 400 participants, applying a design effect of 2.56 gives approximately 1,024 participants.

This is an illustration of sample-size inflation, not a complete cluster-trial calculation. The number of clusters, finite-sample inference, covariate adjustment, cluster-size variability, and outcome characteristics must also be considered.

In particular, multiplying the participant count by a design effect does not establish that the resulting number of clusters will be sufficient for valid or adequately powered inference.

Why even a small ICC can have a substantial effect

The design effect depends jointly on the ICC and cluster size. A small positive ICC can produce substantial variance inflation when each cluster contains many participants.

The following values illustrate the relationship under the equal-size approximation.

These are calculated examples, not empirical ICC estimates or recommended planning values.

At an ICC of 0.05, increasing cluster size from 10 to 100 raises the approximate design effect from 1.45 to 5.95. The relationship demonstrates why adding participants within a cluster may become increasingly inefficient.

The appropriate planning ICC should be justified using relevant prior research, pilot data, or a defensible sensitivity range.

Effective sample size is a useful approximation, not an actual participant count

Researchers sometimes express the consequences of clustering through an approximate effective sample size.

A trial with 20 randomized schools still has 20 randomized clusters, regardless of whether an effective sample-size approximation produces 400, 600, or another value.

The effective sample-size calculation should not be used as a substitute for formal power analysis or as justification for treating observations as independent.

Why adding clusters can be more valuable than adding participants

When the ICC is positive, the precision gained from increasing the number of participants within a fixed number of clusters eventually diminishes.

For an equal-size design with K clusters and m participants per cluster, the total sample size is K × m. Combining this with the simplified design-effect approximation gives:

In this hypothetical comparison, distributing the same number of participants across more clusters yields greater approximate information.

However, the calculation does not account for additional recruitment costs, cluster-level variation in treatment effects, or all finite-sample considerations. More clusters may be statistically desirable but operationally expensive.

The practical issue is to determine how many clusters are required and how many participants should be recruited within each.

Unequal cluster sizes complicate sample-size planning

Schools, hospitals, and communities rarely have identical numbers of eligible participants. Some clusters may contribute many more observations than others.

Unequal cluster sizes can reduce statistical efficiency relative to equal-size clusters with the same average size, although the precise consequences depend on the estimand and analysis method.

A commonly used approximate design effect for certain unequal-size settings is:

This formula is not universal. Different estimators, cluster-size mechanisms, and trial designs may require different approaches. Researchers should use a method consistent with their intended estimand and analysis.

Substantial variation in cluster sizes warrants explicit attention rather than assuming that the average cluster size adequately represents the study.

Further design considerations arise when clusters contribute unequal numbers of participants.

Baseline measurements and covariate adjustment can improve efficiency

The simple design-effect calculation assumes a relatively basic comparison. More efficient designs may use baseline outcome measurements, prognostic covariates, or other prespecified information.

For example, adjusting for school-level baseline achievement may explain some between-school variation and improve the precision of the estimated intervention effect.

The potential gain depends on how strongly baseline measurements predict follow-up outcomes, the number of clusters, and the analytical model.

Researchers should not assume that covariate adjustment eliminates clustering or guarantees sufficient power. The planned adjustment should be incorporated into sample-size calculations when credible assumptions are available.

Attrition must be considered at both participant and cluster levels

Cluster trials can lose individual participants through nonresponse, withdrawal, or incomplete follow-up. Entire clusters may also withdraw because of administrative changes, implementation difficulties, or loss of institutional cooperation.

These losses have different consequences. Losing one student reduces the available participant information, whereas losing an entire school reduces both the participant count and the number of clusters contributing to the comparison.

Sample-size planning should therefore account for plausible participant attrition and cluster attrition separately. Simply increasing the number of students per school may not protect the study against the loss of a school.

Power calculations must match the intended analysis

Formal power calculations should reflect the primary outcome, target effect, significance level, desired power, allocation ratio, number of clusters, cluster-size distribution, ICC assumptions, and planned analysis.

For binary or count outcomes, repeated-measures designs, matched clusters, or small numbers of clusters, specialized analytical or simulation-based approaches may be appropriate.

Simulation can be particularly useful when conventional formulas do not adequately represent the intended design. However, simulation results remain dependent on the assumptions used to generate the data.

The objective is not to obtain the largest possible sample. It is to identify a feasible allocation of clusters and participants that provides adequate precision for the research question.

Comparing Two School-Randomized Designs With the Same Student Count

Interpretation: Under the simplified assumptions, Option B provides substantially greater statistical efficiency despite having the same number of students. The difference arises because the observations are distributed across more independently randomized schools.

This does not establish that Option B is adequately powered or universally preferable. Recruiting 30 schools may be more expensive, and the appropriate design depends on the intended estimand, implementation requirements, and available resources.

Common Errors in Cluster-Randomized Sample-Size Planning

Plan the Number of Clusters and Participants Together

When planning a cluster-randomized trial, begin with the primary outcome and target effect rather than an available participant count. Establish the feasible number of clusters, expected cluster sizes, and plausible ICC values.

Then compare alternative designs. Recruiting more clusters may provide greater statistical efficiency, while recruiting more participants within existing clusters may be less expensive. Neither strategy should be chosen without considering power, implementation, and the causal question.

The justification should be documented in the protocol, including the source of ICC assumptions, the target effect, the number and expected sizes of clusters, attrition allowances, and the planned analysis.

For educational technology interventions, institutional characteristics may strongly influence outcome correlation. A sample-size calculation based solely on student counts risks overlooking the shared conditions under which learning occurs.

Before Finalizing a Cluster-Randomized Sample Size

Frequently Asked Questions About Cluster Trial Sample Sizes

Cluster Trials Need More Information Because Their Observations Are Correlated

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.