How Do You Handle Unequal Cluster Sizes in a Cluster-Randomized Trial?

What Happens When Randomized Schools or Clinics Have Different Numbers of Participants?

Suppose you are evaluating an instructional intervention across 20 schools. Some schools contribute 40 students, while others contribute more than 300. All schools are randomly assigned to intervention or comparison conditions, but the numbers of participating students differ considerably.

Should every school contribute equally to the treatment-effect estimate? Should larger schools receive more weight because they contain more students? And can you calculate the sample size using the average number of students per school?

These questions matter because unequal cluster sizes affect more than the appearance of a dataset. They can influence statistical precision, the effectiveness of randomization, the relative contribution of different clusters, and even the meaning of the estimated intervention effect.

Equal cluster sizes are convenient for statistical formulas, but they are rarely guaranteed in real-world research. The methodological task is to anticipate size variation and handle it consistently with the research question.

Account for Cluster-Size Variation in Both Design and Analysis

Why Unequal Cluster Sizes Matter and How to Address Them

Unequal cluster sizes are common and do not invalidate randomization

A cluster-randomized trial assigns intact groups to intervention conditions. These groups may differ substantially in their numbers of eligible or participating individuals.

Schools have different enrollments. Clinics serve different patient populations. Communities vary in population size, and workplaces employ different numbers of staff.

These differences do not invalidate randomization. A school with 50 participating students and a school with 500 participating students can each be a valid randomized cluster.

However, the clusters contribute different amounts of individual-level outcome information. Researchers must distinguish the unit of randomization from the number of observations collected within that unit.

The implications depend on whether cluster sizes are fixed before allocation, change during follow-up, or are themselves influenced by the intervention.

Why equal-size assumptions are often unrealistic

Many introductory sample-size formulas assume that every cluster contributes the same number of participants. This simplifies calculations because the variance contribution of each cluster can be represented using one common cluster size.

In practice, researchers may recruit institutions of different sizes or obtain different participation rates within institutions.

For example, two schools may each enroll 200 eligible students, but one may contribute 180 outcome observations while the other contributes only 110 because of consent, attendance, or missing-data differences.

Researchers should therefore distinguish between eligible cluster size, recruited cluster size, and analyzed cluster size.

Their distributions may differ, and each can have different implications for study planning and interpretation.

How to measure cluster-size variability

A useful descriptive measure is the coefficient of variation (CV) of cluster sizes. It expresses the standard deviation of cluster size relative to the mean cluster size.

The CV is useful because two studies may have identical average cluster sizes but very different distributions.

A study with 20 clusters of approximately 100 participants each differs from one with the same average but clusters ranging from 10 to several hundred participants.

Researchers should also inspect the actual distribution, including unusually large or small clusters. A single CV may conceal substantial skewness or extreme values.

How unequal sizes affect the design effect

For equal cluster sizes, the familiar approximate design effect is:

When cluster sizes differ, a commonly used approximation for a participant-weighted mean under a simple exchangeable-correlation model is:

The example shows how cluster-size variation can increase variance inflation under the stated assumptions.

However, different estimators, weighting schemes, covariance structures, and relationships between cluster size and outcomes can produce different efficiency patterns. A single design-effect formula should not replace a calculation matched to the intended analysis.

The broader sample-size requirements of cluster randomization must also account for the number of independently randomized clusters.

Unequal cluster sizes can change which treatment effect is estimated

One of the most important distinctions concerns whether the researcher wants an effect averaged across individuals or an effect averaged across clusters.

A participant-average treatment effect gives equal conceptual weight to individuals in the target population. Larger clusters consequently contribute more individuals to that population.

A cluster-average treatment effect gives equal conceptual weight to clusters. A small school and a large school each contribute one cluster-level effect to the average.

These estimands can differ when intervention effects vary across clusters and cluster size is associated with those effects.

For example, an intervention may be more effective in small schools with flexible teaching arrangements than in large schools with complex administrative structures.

Giving every school equal weight answers a different population-level question from giving every student equal weight.

Researchers should specify the intended estimand before selecting an analysis method rather than allowing software defaults to determine the weighting implicitly.

Informative cluster size requires particular attention

Cluster size is informative when it is associated with outcomes or treatment effects in a way relevant to the analysis and estimand.

For example, larger schools may have different baseline achievement, resources, or intervention responses from smaller schools. A clinic serving many patients may differ systematically from a smaller clinic in patient characteristics or care delivery.

Under informative cluster size, statistical procedures that weight observations or clusters differently may estimate different effects.

The issue is not necessarily bias in the ordinary sense. An estimator may be appropriate for one target estimand but inappropriate for another.

Researchers should investigate whether cluster size is associated with important baseline characteristics and prespecify how cluster-size-related differences will be addressed.

Should larger clusters receive more statistical weight?

There is no universal answer because weighting depends on the target estimand and statistical model.

If the objective is to estimate the average effect across all participating individuals, larger clusters naturally represent more members of that population. However, simply weighting cluster means by their sizes may not be the most efficient or robust estimator under every correlation structure.

If the objective is to estimate the average effect across institutions, equal cluster weighting may be appropriate.

Inverse-variance weighting may improve precision under suitable assumptions, but it can change the implicit weighting of clusters and may not correspond to the desired estimand when cluster size is informative.

The weighting strategy should therefore be chosen deliberately and justified in relation to the research question.

How unequal sizes affect allocation balance

Randomizing equal numbers of clusters to intervention and comparison conditions does not guarantee equal numbers of participants in each arm.

For example, ten schools may be assigned to each condition, but the intervention arm could contain 2,000 students while the comparison arm contains 1,200.

Such differences can affect precision and may be associated with other institutional characteristics.

When cluster size is known before randomization, researchers may consider balancing on size through stratification, matching, or covariate-constrained randomization.

However, these procedures should be prespecified, and the analysis should account for relevant allocation restrictions.

The choice between matching and stratifying clusters depends on the number of clusters, prognostic characteristics, and allocation objectives.

Cluster size should not be manipulated after allocation to improve results

Researchers should distinguish genuine variation in cluster size from differential recruitment after randomization.

If intervention schools recruit more students because teachers know their allocation, the resulting cluster-size differences may reflect post-randomization selection rather than pre-existing institutional size.

Similarly, excluding participants after observing outcomes can alter cluster sizes in ways that compromise comparability.

Prespecified eligibility criteria, recruitment procedures, and missing-data rules are therefore essential.

When participants are recruited after cluster allocation, researchers should assess whether the assigned condition influences enrollment and analyzed cluster sizes.

Analysis options must reflect the estimand and dependence structure

Cluster-randomized trials with unequal sizes may be analyzed using cluster-level summaries, mixed-effects models, generalized estimating equations, or randomization-based approaches.

Each approach requires attention to assumptions, weighting, small-sample inference, and the target treatment effect.

Cluster-level analysis can be relatively transparent because each randomized cluster contributes a summary outcome. However, equal weighting of cluster means generally targets a cluster-average contrast, while other weighting strategies may target different quantities.

Individual-level mixed-effects models can account for within-cluster correlation, but the interpretation of model coefficients and the implicit weighting of clusters require care, particularly with informative cluster size.

Generalized estimating equations may target marginal effects, but working correlation assumptions, weighting, and finite-sample corrections remain relevant.

No analytical method should be selected solely because it accommodates unequal cluster sizes computationally.

Why very small clusters and very large clusters both deserve attention

Very small clusters may contribute limited information about within-cluster outcome variability. Very large clusters may dominate participant-weighted estimates or provide diminishing statistical gains when the ICC is positive.

Extreme size imbalance can also increase sensitivity to the outcomes of a few large institutions.

Researchers should examine whether the analysis is disproportionately influenced by particular clusters and consider prespecified sensitivity analyses where appropriate.

However, excluding unusually large or small clusters simply to create a more uniform dataset can change the target population and introduce selection concerns.

How to plan for unequal sizes before the trial begins

Planning should use the anticipated distribution of eligible and analyzed cluster sizes, not only the average.

Relevant information may come from administrative records, previous trials, pilot studies, or institutional enrollment data.

Researchers should consider plausible changes due to recruitment, attrition, transfers, and missing outcomes.

When the distribution is highly unequal or the analytical model is complex, simulation-based power analysis may be preferable to a simple design-effect approximation.

Simulation should reproduce the intended allocation procedure, cluster-size distribution, ICC assumptions, outcome model, and planned statistical analysis.

The number of randomized clusters remains an important design parameter even when cluster-size variation is modeled carefully.

How Unequal School Sizes Can Change the Estimated Effect

Interpretation: Neither result is inherently incorrect. The five-point effect answers a question about the average school, while the 3.2-point effect answers a question about the average student in this hypothetical population.

The difference arises because school size is associated with the school-specific intervention effect. This illustrates why cluster-size weighting should be determined by the intended estimand rather than selected after observing which result is more favorable.

In an actual randomized trial, researchers would need sufficient clusters and an appropriate analysis to estimate either quantity reliably.

Common Mistakes When Clusters Have Different Sizes

Choose the Estimand First, Then Plan for Cluster-Size Variation

Before randomization, specify whether the primary effect should represent the average participant or the average cluster. This decision should reflect the substantive research question and target population.

Next, estimate the distribution of expected cluster sizes and assess whether unusually large or small clusters could affect allocation balance, precision, or analysis.

Finally, select a sample-size calculation and statistical model consistent with the intended estimand and realistic cluster-size variation.

For school-based educational technology research, differences in enrollment may coincide with differences in infrastructure, staffing, and implementation capacity. Treating all schools as interchangeable may therefore obscure important features of the intervention effect.

The aim is not to force equal cluster sizes. It is to ensure that the study's design, analysis, and interpretation remain consistent with the population and effect the researcher intends to investigate.

Before Analyzing a Trial With Unequal Cluster Sizes

Frequently Asked Questions About Unequal Cluster Sizes

Unequal Cluster Sizes Are Manageable, but Their Consequences Must Be Planned

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.