When Does Contamination Make Cluster Randomization Preferable?

When Does Contamination Become Serious Enough to Change the Randomization Design?

Suppose you are evaluating an AI-supported teaching strategy. Within each school, some teachers are randomly assigned to receive professional development on the strategy, while others continue their usual instructional practices. The intervention teachers begin sharing lesson plans, assessment prompts, and classroom techniques with their colleagues.

The comparison teachers may now be using elements of the intervention, even though they were not assigned to receive it. The study still has randomized groups, but the difference between what those groups actually experience has become less distinct.

Would randomizing entire schools instead of teachers solve the problem? It might reduce contamination within schools, but it would also increase the number of participants needed for comparable precision and introduce additional design requirements. The decision depends on how contamination occurs, how much it matters, and whether less costly alternatives are available.

Contamination Justifies Cluster Randomization When It Threatens a Credible Treatment Contrast

How Contamination Affects Randomized Trials and Design Decisions

What contamination means in an intervention trial

Contamination generally refers to exposure of participants or providers in one trial arm to intervention components intended for another arm. It can occur through direct access, communication, shared practices, organizational processes, or movement between intervention environments.

For example, students assigned to a conventional learning condition may receive access to an experimental learning application through classmates. Teachers assigned to a comparison condition may adopt instructional techniques learned from intervention teachers.

Contamination concerns what participants actually experience, not merely whether they interact. Ordinary social contact does not necessarily constitute contamination. The interaction must plausibly transmit intervention content, practices, resources, or effects relevant to the study's causal comparison.

It is also important to distinguish contamination from nonadherence. A participant who does not use an assigned intervention is nonadherent, but that does not necessarily mean the comparison group has been contaminated. The two problems may coexist and should be measured separately.

Contamination can occur through different mechanisms

Identifying the mechanism is more informative than simply stating that contamination is possible. Different mechanisms may require different preventive measures.

These mechanisms can overlap. An intervention introduced through professional development may spread through shared documents, informal conversations, and institutional expectations. A design that addresses only one route may leave other important routes open.

Why contamination can distort the estimated intervention effect

Randomization establishes an allocation contrast, but the actual difference in intervention exposure may become smaller when comparison participants receive intervention components.

If the intervention is beneficial and contamination exposes comparison participants to its benefits, the estimated intention-to-treat effect may be smaller than the effect that would have been observed without spillover. However, attenuation toward zero is not guaranteed. The direction and magnitude of the resulting difference depend on the intervention, exposure patterns, interference, and outcome mechanisms.

For example, a collaborative learning intervention could improve outcomes among comparison students who interact with intervention students. Alternatively, competition for shared resources could disadvantage one group. In either case, interference changes the relationship between assignment and outcomes.

Researchers should therefore avoid claiming that contamination always creates a conservative estimate. That conclusion requires assumptions that may not hold.

Contamination and interference are related but not identical

Interference occurs when one participant's treatment assignment or exposure affects another participant's outcome. Contamination is one possible mechanism of interference, but interference can occur without direct transfer of intervention materials.

Consider an intervention that improves the behavior of selected students. Their classmates may benefit from a more productive learning environment even without directly receiving the intervention. This is an indirect treatment effect, although the classmates may never use the intervention themselves.

Traditional causal interpretations often invoke a no-interference assumption, sometimes discussed as part of the stable unit treatment value assumption. Cluster randomization may make a partial-interference assumption more plausible when interactions occur predominantly within clusters, but this must be justified rather than assumed.

If indirect effects are themselves important, the study may need a design that explicitly estimates spillover or group-level effects rather than treating all cross-person influence as unwanted contamination.

When contamination is sufficiently consequential to favor cluster randomization

Cluster randomization becomes more defensible when three substantive conditions are present.

First, the intervention is likely to spread across individually randomized participants through routine interactions or shared delivery systems. Second, this exposure is expected to alter the outcome or undermine the intended comparison. Third, the spread cannot reasonably be prevented without changing the intervention, compromising implementation, or imposing disproportionate operational burdens.

For instance, a school-wide professional learning culture may make it unrealistic to expect trained teachers to withhold instructional practices from their colleagues. Randomizing schools could preserve a more meaningful contrast between institutions implementing the program and those continuing usual practice.

By contrast, if an individually delivered digital intervention can be restricted through separate accounts and does not substantially affect shared classroom practices, individual randomization may remain preferable.

The broader decision about when groups rather than individuals should be randomized should consider intervention delivery and the causal question alongside contamination risk.

Choosing the cluster level requires identifying where spillover occurs

Randomizing classrooms may reduce student-to-student contamination within classrooms, but it may not prevent teachers from sharing instructional practices across classrooms in the same school.

Randomizing schools could reduce that particular pathway, yet teachers may still participate in interschool professional networks. The appropriate cluster level depends on the interaction structure and intervention mechanism.

A cluster must be an identifiable unit of allocation, not merely a convenient grouping variable. Researchers should define cluster boundaries and consider whether those boundaries correspond to the pathways through which intervention effects spread.

Choosing increasingly large clusters is not automatically better. Larger allocation units may contain more potential spillover, but they also reduce the number of independently randomized units available from a fixed study population.

Cluster randomization imposes a statistical cost

Participants within the same cluster often have correlated outcomes because they share teachers, facilities, organizational practices, or other contextual influences.

For approximately equal cluster sizes, a common approximation to the design effect is:

This calculation illustrates the potential loss of statistical efficiency. It does not establish the required sample size or number of clusters for a particular trial. Unequal cluster sizes, outcome type, baseline adjustment, and other design features affect power and precision.

The costs and limitations of cluster randomization must therefore be weighed against the contamination it is expected to prevent.

Contamination should be assessed before randomization

Researchers can investigate contamination pathways through formative interviews, workflow observation, pilot implementation, stakeholder consultation, and examination of existing communication or resource-sharing practices.

For educational technology interventions, relevant questions include whether participants share devices or accounts, whether teachers collaborate on lesson planning, whether intervention materials are publicly accessible, and whether administrators require uniform instructional practices.

Where possible, the protocol should specify anticipated contamination mechanisms, preventive procedures, and indicators used to monitor exposure. A pilot may reveal that a supposedly isolated intervention is routinely shared across classrooms.

Such information supports a more credible design justification than an unsupported assertion that contamination is inevitable.

Measuring contamination during the trial is still necessary

Even when clusters are randomized, researchers should monitor whether intervention components reach comparison clusters.

Possible indicators include platform access logs, training attendance, teacher practice observations, participant reports, and records of resource distribution. Each has limitations. Self-reported exposure may be inaccurate, while access logs may show use without demonstrating meaningful engagement.

Contamination measurement should be planned before outcome analysis. Post hoc comparisons based on actual exposure may be confounded because exposure after randomization can be influenced by participant characteristics and other post-allocation factors.

Intention-to-treat analysis generally preserves the original allocation comparison. Additional analyses examining contamination or adherence can be informative, but their causal interpretation requires appropriate assumptions and methods.

When individual randomization may remain preferable

Some contamination risk may be acceptable when it is limited, measurable, and unlikely to materially affect the primary estimand. Individual randomization may also remain preferable when clustering would leave too few independent allocation units for reliable inference.

Researchers might reduce contamination through separate intervention providers, restricted digital access, standardized protocols, or distinct intervention sessions. These safeguards should be evaluated for feasibility and possible effects on external validity.

The relevant question is not whether contamination can be eliminated completely. It is whether an individually randomized design can still answer the intended causal question credibly and efficiently.

This trade-off is central to deciding when individual randomization remains preferable despite contamination risk.

Deciding Whether to Randomize Teachers or Schools

Interpretation: The observed collaboration patterns provide a substantive reason to consider school-level randomization. However, eight schools may be insufficient for the planned effect size and analysis. The presence of contamination does not automatically make an underpowered cluster-randomized trial an acceptable solution.

The research team must balance the credibility of the treatment contrast against the consequences of having fewer independently randomized units.

Misconceptions About Contamination and Cluster Randomization

Use Contamination Risk to Inform the Design, Not Dictate It Automatically

Before selecting the randomization unit, identify how intervention exposure could spread, whether that exposure would influence the outcome, and which allocation boundaries could realistically limit it.

Then compare the expected consequences of contamination under individual randomization with the statistical and operational costs of cluster allocation. Where evidence is uncertain, sensitivity analyses using plausible contamination and ICC assumptions may help clarify the trade-off.

For educational interventions involving teachers, students, and shared digital resources, a contamination assessment should be part of the design justification. It should explain why the selected allocation unit is likely to preserve a meaningful intervention contrast, rather than merely claiming that participants might communicate.

Before Using Contamination to Justify Cluster Randomization

Frequently Asked Questions About Contamination in Randomized Trials

Cluster Randomization Is Justified by Consequential, Difficult-to-Control Spillover

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.