When Should You Use Cluster Randomization to Reduce Contamination?

Should You Randomize Entire Groups When Participants Are Likely to Share an Intervention?

You are evaluating a new digital learning program across several university classrooms. If students within each classroom are individually assigned to different conditions, those receiving the program may share materials, demonstrate its features, or explain the strategies they learned to classmates assigned to conventional instruction.

Teachers may also struggle to deliver two distinct instructional approaches within the same classroom without allowing elements of one condition to influence the other.

One proposed solution is to randomize entire classrooms rather than individual students. Every student in a classroom would then receive the same assigned instructional condition.

Would that solve the contamination problem? More importantly, when does reducing cross-group exposure justify the additional methodological demands of a cluster-randomized trial?

Use Cluster Randomization When Group-Level Assignment Meaningfully Reduces Contamination

Deciding Whether Cluster Assignment Is the Right Response to Contamination

What Is Cluster Randomization?

Cluster randomization assigns intact groups of participants, rather than individual participants, to intervention conditions.

Clusters may be classrooms, schools, clinics, hospitals, workplaces, communities, or other organizational or social units.

For example, a study involving twenty classrooms might randomly assign ten classrooms to a digital instructional intervention and ten to conventional instruction. Students within each classroom would generally receive the condition assigned to their classroom.

The unit of randomization is the classroom, even if outcomes are measured separately for individual students.

This distinction has important implications for sample size, analysis, recruitment, and the interpretation of treatment effects.

Cluster randomization is particularly relevant when an intervention cannot be delivered independently to individuals or when individual assignment would create substantial opportunities for cross-group exposure.

Why Can Individual Randomization Create Contamination?

Individual randomization assigns different participants within the same setting to different conditions.

When participants interact frequently, intervention components may cross group boundaries.

For example, students assigned to an experimental learning platform may share access credentials, downloaded resources, or strategies learned through the program with control students.

Teachers may also transfer instructional practices from one condition to another, particularly when they deliver both interventions.

These pathways can reduce the distinction between assigned conditions and complicate interpretation of the treatment comparison.

However, not every interaction between participants constitutes contamination. The relevant concern is whether the interaction transfers intervention components or influences outcomes in a way that matters for the intended estimand.

How Does Cluster Randomization Reduce Contamination?

Cluster randomization can reduce contamination by placing participants who interact closely within the same assigned condition.

If every student in a classroom receives the experimental intervention, sharing materials among classmates no longer transfers those materials between randomized conditions.

Similarly, assigning an entire school to one instructional approach may reduce the need for teachers to maintain different treatment protocols within the same institution.

The design effectively moves the boundary between treatment conditions from individual participants to larger social or organizational units.

Its success depends on whether the chosen cluster boundaries correspond to the pathways through which intervention components would otherwise spread.

If students regularly interact across classrooms or teachers share resources across schools, contamination may still occur.

When Is Cluster Randomization Particularly Appropriate?

Cluster randomization may be justified when the intervention operates at the group level or when individual allocation would substantially interfere with its delivery.

These examples illustrate plausible reasons for clustering, not automatic prescriptions. The appropriate randomization unit depends on intervention delivery, interaction patterns, and the scientific question.

Should the Cluster Match the Intervention Delivery Unit?

Often, but not necessarily.

If an intervention is delivered to entire classrooms, classroom randomization may align naturally with implementation. If teachers deliver the intervention across several classrooms, however, assigning classrooms independently may still create contamination through teacher behavior.

In that case, teacher-level or school-level randomization might better reflect the relevant exposure pathways.

Researchers should map who delivers the intervention, who interacts with whom, and how materials or practices may spread.

The cluster should be large enough to reduce important cross-group exposure but not unnecessarily large.

Increasing cluster size can reduce the number of independent randomized units available within a fixed study population.

Does Cluster Randomization Eliminate Contamination?

No. It may reduce contamination within the chosen cluster boundaries while leaving opportunities for exposure between clusters.

Students from different classrooms may collaborate outside class. Teachers from different schools may attend common professional development sessions. Clinicians may work across several facilities.

Cluster randomization can therefore shift the contamination problem rather than eliminate it.

Researchers should examine whether clusters are sufficiently separated in the ways relevant to the intervention.

Where cross-cluster interaction remains substantial, additional implementation procedures or alternative designs may be necessary.

Why Does Cluster Randomization Usually Require More Participants?

Participants within the same cluster often resemble one another because they share teachers, institutional conditions, environments, or other characteristics.

Their outcomes may therefore be correlated.

When outcomes are positively correlated within clusters, observations provide less independent information than the same number of observations from independently randomized individuals.

This dependence is commonly summarized using the intracluster correlation coefficient, or ICC.

For equal-sized clusters under a simple exchangeable correlation model, the design effect is often approximated as follows:

This formula is an approximation, not a complete sample size calculation. Unequal cluster sizes, covariate adjustment, outcome type, and the number of clusters can materially affect precision and power.

Researchers should not simply multiply a conventional sample size by the design effect without considering the full design and analysis.

Why Does the Number of Clusters Matter?

A study with many participants but very few randomized clusters may provide limited information about the treatment effect.

For example, assigning one school to intervention and one school to control leaves treatment condition completely confounded with school identity in the observed design.

Differences between the two schools cannot be reliably separated from differences associated with treatment assignment.

Adding more students within those schools does not create additional independently randomized schools.

Researchers should therefore prioritize an adequate number of clusters and use sample size methods appropriate to cluster randomization.

The number required depends on the ICC, cluster sizes, expected effect, allocation ratio, and analytical approach.

How Should Researchers Choose Between Classroom and School Randomization?

The choice should reflect the likely contamination pathways and the number of independent units available.

Classroom randomization may be efficient when teachers and students interact mainly within classrooms and intervention delivery can remain separate.

School randomization may be preferable when the intervention changes school-wide practices or teachers routinely share resources across classrooms.

However, school-level assignment may produce fewer randomized units and require more institutions to achieve adequate precision.

The decision therefore involves a trade-off between reducing contamination and retaining statistical information.

Researchers should consider feasible design strategies for settings with substantial contamination risk before selecting the cluster level.

Can Cluster Randomization Introduce Recruitment Bias?

Yes. Recruitment bias can arise when individuals are identified, approached, or enrolled after their cluster's assignment is known.

For example, a school assigned to an attractive intervention may encourage more motivated students to participate, while a control school may recruit a different group of students.

These differences can undermine comparability between participants in the randomized conditions.

Where feasible, researchers should identify and recruit participants before cluster allocation or use recruitment procedures that minimize allocation-related selection.

When recruitment after randomization is unavoidable, eligibility criteria, recruitment procedures, and participant flow should be carefully standardized and documented.

Cluster randomization does not automatically prevent selection bias merely because assignment is randomized.

How Should Outcomes Be Analyzed?

The analysis must account for the dependence among participants within clusters.

Common approaches include mixed-effects models, generalized estimating equations with appropriate small-sample considerations, and cluster-level analyses.

The appropriate method depends on the outcome, number of clusters, cluster sizes, and target estimand.

Analyzing all participants as though they were independently randomized can underestimate standard errors and produce misleading confidence intervals or p-values.

Researchers should specify the analytical approach in advance and ensure that the sample size calculation corresponds to the intended model.

Does Cluster Randomization Change the Treatment Effect Being Estimated?

Potentially. Assigning an intervention to an entire classroom may produce different effects from offering it to selected students within a classroom.

Group-level assignment can influence peer interaction, teacher behavior, institutional support, and intervention uptake.

Consequently, the effect of a classroom-wide program may not be identical to the effect of providing the same program to individual students within mixed-treatment classrooms.

Researchers should therefore consider whether the cluster-level intervention strategy matches the policy or practice decision they intend to inform.

Cluster randomization is not merely a technical adjustment to prevent sharing. It may change the intervention context and the causal contrast.

When Might Individual Randomization Still Be Preferable?

Individual randomization may remain preferable when contamination is unlikely, limited, or manageable through less costly procedures.

For example, an individually delivered intervention using private accounts and independent facilitators may be implemented without substantial cross-group exposure.

Researchers might also reduce contamination through separate scheduling, access controls, staff training, or clearer intervention procedures.

These approaches may preserve individual randomization and its statistical advantages.

The question is not whether any contamination could occur, but whether the anticipated contamination is sufficiently consequential to justify changing the unit of randomization.

There is no universal acceptable contamination percentage that determines this choice.

What Should Be Included in the Protocol?

The protocol should explain why clusters were selected, how they will be identified and randomized, and how contamination is expected to be reduced.

It should describe recruitment timing, allocation procedures, intervention delivery, outcome measurement, and safeguards against differential recruitment or attrition.

Sample size assumptions should include the expected ICC and relevant cluster-size characteristics.

The statistical analysis plan should specify how clustering will be addressed and how the treatment effect will be interpreted.

Researchers should also plan to measure important cross-cluster contamination where feasible, because cluster assignment does not guarantee complete separation.

Reporting should follow the relevant CONSORT extension for cluster-randomized trials and other applicable guidance.

Choosing Classroom Randomization for a Digital Learning Intervention

The design may reduce contamination because students who routinely collaborate within classrooms receive the same assigned condition.

However, the study still needs to address possible sharing between classrooms and the statistical dependence among students taught in the same class.

Whether classroom randomization is preferable to school-level or individual randomization depends on the actual interaction patterns and the number of independent clusters available.

Common Misconceptions About Cluster Randomization and Contamination

Choose Clusters According to How the Intervention Spreads

Before selecting cluster randomization, map the pathways through which participants, providers, and organizations could transfer intervention components between conditions.

Then identify the smallest practical cluster that meaningfully reduces those pathways while allowing sufficient independent randomized units.

Researchers should also consider whether intervention materials can be distributed in ways that reduce contamination caused by sharing with control participants.

The objective is not simply to create larger randomized groups. It is to align the randomization unit with the intervention's delivery structure and the pathways through which contamination could affect the intended comparison.

Before Choosing Cluster Randomization to Reduce Contamination

Frequently Asked Questions About Cluster Randomization and Contamination

Randomize Clusters When Group Boundaries Meaningfully Protect the Comparison

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.