03 · What You Need to Know
How to Decide Whether Individual Randomization Remains Defensible
Contamination risk is not an automatic reason to randomize groups
Contamination occurs when participants assigned to one condition receive elements of an intervention intended for another condition. It may involve shared materials, access to treatment resources, changes in provider behavior, or transmission of intervention-related knowledge.
Some degree of communication or interaction is common in educational, organizational, and community research. However, interaction alone does not establish that contamination is substantial.
For example, students may discuss an AI-supported writing application without obtaining access to the software or receiving the individualized feedback that constitutes the intervention.
The methodological concern is whether such interactions change the actual treatment contrast or influence outcomes through interference.
Researchers should distinguish plausible contamination pathways from speculative possibilities. The decision to adopt cluster allocation should be supported by an assessment of likely consequences rather than an assumption that any shared environment makes individual randomization invalid.
The circumstances in which contamination makes cluster randomization preferable depend on the mechanism, magnitude, and feasibility of preventing spillover.
Individual randomization usually offers greater statistical efficiency
When individuals are randomized independently within a setting, treatment allocation can vary among participants who share the same institutional environment.
This arrangement may allow comparisons within schools, clinics, or workplaces, reducing the influence of between-institution differences on the estimated treatment effect.
Cluster randomization instead assigns the same condition to everyone within an allocation unit. Participants within that unit may have correlated outcomes because they share organizational practices, resources, or contextual influences.
For a conventional equal-size parallel cluster-randomized design, a commonly used approximation to variance inflation is:
This formula illustrates why cluster allocation can require more participants or clusters to achieve comparable precision.
It does not mean that an individually randomized study within classrooms can ignore classroom-level dependence. Students may still share outcome-related influences, and appropriate stratification, covariate adjustment, or variance estimation may be necessary.
Moreover, individual randomization can create treatment-group clustering when interventions are delivered by shared providers or in treatment groups. The analysis must reflect the actual dependence structure rather than assuming every observation is independent merely because allocation occurred individually.
The broader costs of cluster randomization should be weighed against the contamination it is expected to prevent.
When the intervention can be delivered independently
Individual randomization is particularly attractive when the intervention can be assigned, delivered, and monitored separately for each participant.
Examples include individual access to a digital learning platform, personalized feedback, individually administered counseling, or a treatment that does not require changing shared organizational practices.
In an educational technology trial, separate user accounts may allow researchers to control access to a learning application without changing how an entire classroom operates.
However, technical independence does not necessarily imply causal independence. Students may still share feedback, teachers may change their instruction in response to intervention users, or peer interactions may influence outcomes.
The intervention mechanism and the surrounding social environment should therefore be assessed together.
When contamination is unlikely to affect the primary outcome materially
Not every instance of spillover meaningfully changes the intervention contrast.
Suppose students assigned to an AI-supported feedback tool occasionally describe its interface to comparison students. If the intervention's active component is individualized feedback generated through sustained use, merely hearing about the interface may have little effect on writing performance.
By contrast, if the intervention consists primarily of teaching transferable writing strategies, students may communicate those strategies to comparison participants. The same interaction could then represent meaningful contamination.
Researchers should identify the intervention's active components and determine whether they can be transmitted through ordinary participant interactions.
Judgments about materiality should be based on plausible causal mechanisms, pilot evidence, or relevant prior studies rather than unsupported numerical thresholds.
Practical safeguards may reduce contamination without changing allocation
Several safeguards may be feasible when an intervention is delivered individually. Their suitability depends on the intervention and setting.
| Potential safeguard |
Appropriate situation |
Important limitation |
| Individual access controls |
Digital interventions with separate user accounts |
Cannot prevent participants from sharing knowledge or outputs. |
| Separate intervention sessions |
Individually delivered training or counseling |
Participants may communicate outside the sessions. |
| Separate intervention providers |
Provider behavior could transfer treatment components |
May increase costs and introduce provider-related differences. |
| Standardized implementation protocols |
Providers deliver both intervention and comparison conditions |
Protocols do not guarantee that providers can maintain separation. |
| Exposure monitoring |
Intervention use or resource sharing can be documented |
Measurement does not itself prevent contamination. |
| Shorter exposure windows |
Intervention effects can plausibly be evaluated over a shorter period |
May change the substantive effect being estimated. |
Safeguards should not substantially change the intervention being evaluated. For example, prohibiting ordinary student collaboration might prevent some contamination but create an artificial learning environment that differs from routine educational practice.
Researchers should consider whether contamination prevention itself introduces burdens, changes behavior, or reduces external validity.
Contamination and interference should be distinguished
Contamination involves exposure to intervention components across assigned conditions. Interference is broader: one participant's treatment assignment or exposure affects another participant's outcome.
For example, students receiving an intervention might improve the quality of group discussions, indirectly benefiting comparison students without directly sharing the intervention.
Such effects can complicate the interpretation of an individually randomized trial even when access controls prevent direct contamination.
The conventional no-interference assumption is often associated with the stable unit treatment value assumption. Its plausibility depends on the intervention and setting.
When indirect effects are central to the research question, researchers may need an alternative design or an estimand that explicitly accommodates interference.
Individual randomization can still estimate meaningful assignment effects under certain interference structures, but those effects may depend on the treatment allocation proportion and interaction environment.
Some contamination does not necessarily invalidate intention-to-treat analysis
Intention-to-treat analysis compares participants according to their randomized assignments, regardless of subsequent adherence or intervention receipt.
When contamination occurs, this analysis can still estimate the effect of assignment under the actual implementation conditions, subject to the study's assumptions and missing-data mechanisms.
However, the estimated effect may differ from the effect of receiving the intervention without spillover.
For example, an intervention assigned to selected students may produce a smaller contrast when comparison students obtain some of its benefits. The resulting intention-to-treat estimate may still be relevant to a policy question about offering the intervention under ordinary conditions.
Researchers should distinguish the effect of assignment from the effect of treatment receipt and avoid interpreting the former automatically as the latter.
Excluding participants who experienced contamination can introduce post-randomization selection bias. Supplementary exposure analyses require appropriate causal assumptions and should not replace the prespecified primary comparison without justification.
Why contamination does not always bias effects toward zero
It is tempting to assume that contamination simply makes intervention and comparison groups more similar, reducing the estimated treatment effect.
That may occur when comparison participants receive beneficial intervention components. However, spillover can have more complicated consequences.
Intervention participants might compete with comparison participants for shared resources, indirectly reducing comparison outcomes. Alternatively, changes in peer behavior might benefit both groups to different degrees.
The direction of the resulting difference depends on the causal mechanism and the intended estimand.
Researchers should not describe contamination as necessarily producing conservative estimates without establishing the assumptions required for that conclusion.
When group-level delivery does not require group-level randomization
An intervention may involve group sessions while still permitting individual randomization.
For example, participants may be randomized individually and then assigned to intervention-only training groups. Comparison participants may receive individual usual-care services or participate in separate comparison groups.
Such designs can create correlation among participants who receive treatment together. They may require analytical methods for individually randomized group-treatment trials or partially nested designs.
Thus, the presence of group-based delivery does not automatically mean that the groups must be the units of randomization.
The critical distinction is between how treatment is assigned and how treatment exposure or outcome dependence is structured.
When individual randomization becomes difficult to defend
Individual allocation becomes less attractive when intervention components are routinely transmitted across participants, when providers cannot realistically maintain separate conditions, or when the intervention changes shared organizational practices.
For example, a professional development intervention intended to transform whole-class teaching may be impossible to restrict to selected students within the same classroom.
Similarly, a school-wide policy or institutional learning management system configuration may affect all students regardless of individual assignment.
In such circumstances, randomizing intact groups may better preserve the intended intervention and comparison.
However, the chosen cluster level must still be justified, and the number of available clusters must be sufficient for the intended inference.
How to assess contamination before selecting the design
Researchers should identify the intervention's active components, likely communication pathways, shared providers, and organizational practices that could transmit exposure between trial arms.
Formative interviews, pilot studies, workflow observation, and prior research can help establish whether contamination is plausible and consequential.
For digital interventions, platform logs may help identify unauthorized access or shared use. However, access logs cannot capture every form of knowledge transfer or indirect influence.
Researchers should also consider whether the intervention's benefits are likely to spread through peer interactions even without direct access.
When uncertainty remains, design-stage sensitivity analyses can explore plausible contamination scenarios. Such analyses should be interpreted as conditional assessments rather than empirical proof of the amount of contamination that will occur.
Compare the expected information gained with the expected contamination avoided
The decision between individual and cluster randomization involves a trade-off.
Individual allocation may provide greater statistical efficiency and better control of shared institutional characteristics. Cluster allocation may reduce contamination or preserve group-level intervention delivery, but it generally requires more independent groups or participants.
The preferred design depends on whether the expected reduction in contamination justifies the loss of efficiency and additional operational requirements.
Guidance from the National Institutes of Health Research Methods Resources identifies low contamination risk as a circumstance favoring individual randomization, while substantial contamination risk supports consideration of group randomization.
Nevertheless, the threshold at which contamination becomes unacceptable is not universal. It depends on the intervention, estimand, and scientific consequences of spillover.
