03 · What You Need to Know
Choosing a Research Design That Can Withstand Anticipated Contamination
Why Should Contamination Be Considered Before Randomization?
Randomization determines how participants or groups are assigned to intervention conditions. Once assignments are made, the opportunities for cross-group exposure are partly determined by the study's structure.
For example, individually randomizing students within the same classroom creates a different contamination environment from assigning entire classrooms to one condition.
If intervention materials are routinely shared, the observed treatment contrast may become smaller or differ from what the researchers intended to evaluate.
Trying to repair the design after implementation may require changes that affect recruitment, intervention delivery, or analysis.
Researchers should therefore identify plausible contamination pathways during protocol development and select a design consistent with the intended causal question.
Begin With the Intended Treatment Effect
Before deciding how to prevent contamination, researchers should specify what treatment effect they want to estimate.
Suppose a university wants to know whether offering a digital learning platform improves achievement under ordinary conditions where students share resources.
That question differs from asking what the platform's effect would be if control students had no exposure to any intervention components.
The first concerns the effect of an implementation strategy within a particular social environment. The second concerns a more restricted treatment contrast.
Both may be legitimate, but they require different design assumptions.
Researchers should not automatically eliminate all spillover if diffusion is part of the policy or implementation process they intend to study.
Map the Pathways Through Which Contamination Could Occur
A contamination risk assessment should identify who can transfer intervention components, how transfer could occur, and which participants may be exposed.
Relevant pathways may include participant communication, shared digital resources, common intervention providers, organizational practices, and social networks.
| Potential pathway |
Illustrative example |
Design consideration |
| Participant sharing |
Students exchange experimental worksheets. |
Access procedures, group separation, or cluster assignment |
| Shared providers |
Teachers use experimental strategies in control classes. |
Separate delivery teams or provider-level assignment |
| Organizational diffusion |
Schools adopt practices introduced through experimental training. |
School-level assignment or explicit spillover evaluation |
| Digital access |
Control participants use intervention accounts. |
Individual permissions and controlled distribution |
| Social influence |
Intervention participants teach strategies to peers. |
Network-aware design or measurement of indirect exposure |
The most consequential pathways depend on the intervention's causal mechanism. Sharing a peripheral document may matter less than transferring the activities believed to produce the intended outcome.
When Is Individual Randomization Still Appropriate?
Individual randomization may remain appropriate when cross-group exposure is expected to be limited or can be reduced through feasible procedures.
For example, an intervention delivered through individual appointments or private digital accounts may be implemented without substantial sharing.
Researchers might use separate scheduling, condition-specific materials, staff training, and controlled access to reduce unintended exposure.
However, these safeguards should not impose unrealistic restrictions or interfere with the intervention's intended operation.
Individual randomization can preserve statistical efficiency and may allow better balance across participants, but these advantages should be weighed against the expected loss of treatment separation.
The decision should depend on the likely consequences of contamination rather than its mere possibility.
When Should Cluster Randomization Be Considered?
Cluster randomization assigns intact groups, such as classrooms, schools, clinics, or workplaces, to intervention conditions.
It may reduce contamination when participants interact primarily within those groups or when providers cannot realistically deliver different conditions to individuals in the same setting.
For example, assigning entire classrooms to the same instructional condition can reduce cross-group sharing during classroom activities.
However, cluster randomization does not eliminate exposure between clusters.
It also requires accounting for correlated outcomes and an adequate number of independent randomized units.
Researchers should evaluate when cluster randomization is justified according to the actual interaction and delivery structure.
How Do You Choose the Appropriate Cluster?
The cluster should correspond to the boundaries within which intervention components are most likely to spread.
If students collaborate primarily within classrooms, classroom assignment may reduce contamination.
If teachers deliver instruction across several classrooms, teacher-level or school-level assignment may be more appropriate.
However, larger clusters may reduce the number of independent randomized units available within a fixed study population.
Researchers should therefore seek a balance between reducing cross-group exposure and maintaining adequate statistical information.
The appropriate cluster is not necessarily the largest available organizational unit.
Can Separate Intervention Providers Reduce Contamination?
Using separate providers for experimental and comparison conditions may reduce the transfer of intervention practices between groups.
For example, instructors delivering an experimental feedback strategy might not teach control classes during the intervention period.
However, provider separation can introduce differences in experience, motivation, or teaching quality that become associated with treatment condition.
Researchers should therefore consider provider allocation, training, supervision, and analytical implications.
Separate delivery teams may be useful, but they should not be treated as an automatic solution to every contamination problem.
How Can Intervention Delivery Be Structured to Reduce Sharing?
Researchers may reduce contamination through condition-specific schedules, restricted distribution of experimental materials, individual digital accounts, and clear staff responsibilities.
For example, experimental tutorials could be delivered through accounts available only to assigned participants until the primary outcome assessment is completed.
However, participants may still discuss intervention content or demonstrate learned strategies to others.
Access controls also introduce practical considerations involving usability, accessibility, privacy, and participant burden.
Researchers should select proportionate safeguards and avoid assuming that technical restrictions eliminate social spillover.
Should Intervention Materials Be Withheld From Control Participants?
Temporary restriction may be appropriate when the materials are experimental components intended to distinguish treatment conditions.
However, withholding resources must be consistent with informed consent, ethical approval, and participants' legitimate access to usual services.
Researchers should not deny established care or essential educational support merely to preserve treatment separation.
Where appropriate, delayed access to intervention materials may be considered after the primary outcome assessment.
Such arrangements should be evaluated according to the research question and the consequences of delayed treatment.
Researchers should also consider how intervention material sharing causes contamination before selecting distribution restrictions.
Can Researchers Design a Study to Estimate Spillover Rather Than Prevent It?
Yes. When diffusion is expected and scientifically important, researchers may design the study to estimate effects extending beyond directly treated participants.
For example, a professional development program may influence untrained teachers through collaboration with trained colleagues.
A conventional individually randomized comparison may not separately identify direct and indirect effects when participants influence one another.
Designs involving two-stage randomization, variation in treatment saturation across clusters, or other explicitly structured exposure arrangements may support investigation of spillover under appropriate assumptions.
These designs are more demanding than simply recording that some participants shared materials.
Researchers must specify the relevant exposure mapping, interference assumptions, randomization structure, and causal estimands.
The work of Hudgens and Halloran provides a foundational framework for causal inference under interference.
What Is Partial Interference?
Partial interference is an assumption that outcomes may depend on treatment assignments within defined groups but not on assignments outside those groups.
For example, researchers might assume that students influence classmates within the same classroom but do not influence students in other classrooms.
This assumption can support certain cluster-based spillover analyses.
However, it must be substantively plausible. Students may interact across classrooms, and teachers may share resources across schools.
Researchers should not assume that organizational boundaries automatically correspond to the boundaries of causal influence.
When cross-cluster interference is likely, the design and interpretation must address that possibility.
Can a Stepped-Wedge Design Solve Contamination?
A stepped-wedge cluster-randomized design introduces the intervention to clusters at different randomized times, with clusters typically transitioning from control to intervention.
It may be useful when the intervention is intended to be implemented sequentially across organizations.
However, it is not a universal contamination solution.
Intervention knowledge may spread from clusters that have already transitioned to those awaiting implementation.
Time trends, implementation learning, and other temporal factors also require careful design and analysis.
Researchers should select a stepped-wedge design because it fits the scientific and implementation objectives, not simply because contamination is anticipated.
Should Researchers Consider a Crossover Design?
Crossover designs expose participants or clusters to different treatment conditions in successive periods.
They may be useful for some interventions with reversible effects and manageable carryover.
However, educational and behavioral interventions often produce learning that cannot simply be removed when participants switch conditions.
A student who learns a research strategy during the first period may continue using it during the second.
Such carryover can undermine the intended within-participant comparison.
Researchers should therefore avoid choosing crossover designs when intervention effects are likely to persist and cannot be adequately addressed.
How Should Contamination Risk Influence Sample Size Planning?
Expected contamination may reduce the treatment contrast under certain assumptions, making the target effect harder to estimate precisely.
Researchers should consider plausible exposure scenarios during sample size planning rather than assume perfect separation.
For cluster-randomized trials, planning must also account for intracluster correlation, cluster sizes, and the number of randomized clusters.
For designs explicitly investigating interference, sample size and power depend on the allocation structure and effects being estimated.
Simulation may be useful when the exposure process is complex or conventional formulas do not represent the design adequately.
There is no universal sample size inflation factor that corrects every contamination problem.
How Should Contamination Be Monitored During Implementation?
Researchers should establish a measurement plan corresponding to the anticipated contamination pathways.
Possible indicators include intervention material access, session attendance, staff delivery records, participant reports, and observed transfer of practices.
The assessment should distinguish the proportion of participants exposed from the intensity and timing of exposure.
For example, a control participant who briefly views a worksheet should not necessarily be treated as equivalent to someone who completes the full experimental program.
Researchers should specify how contamination will be measured before outcomes are examined.
Monitoring should be proportionate and consistent with participant privacy and ethical requirements.
Should the Protocol Specify a Contamination Threshold?
An operational threshold may be useful for triggering an implementation review.
For example, a research team might specify that substantial cross-group exposure will prompt an audit of material distribution and staff delivery procedures.
However, no universal percentage determines whether contamination is acceptable or whether a trial is valid.
A small amount of exposure to a highly influential intervention component may be more consequential than widespread access to peripheral materials.
Researchers should distinguish monitoring triggers from statistical or causal judgments about the treatment effect.
How Should the Primary Analysis Handle Contamination?
The primary analysis should follow the prespecified estimand and randomization design.
For a treatment-policy estimand, participants may remain analyzed according to randomized assignment despite subsequent cross-group exposure.
Excluding contaminated participants or reclassifying them according to treatment received can introduce selection bias.
Contamination may be associated with motivation, social networks, previous outcomes, or other post-randomization factors.
Estimating an effect under hypothetical absence of contamination requires additional causal assumptions and potentially specialized methods.
Researchers should not treat ordinary regression adjustment for contamination status as an automatic correction.
How Does the Design Affect the Meaning of the Treatment Effect?
Different contamination-management strategies can change the treatment being evaluated.
Assigning an entire classroom to a digital learning program may influence peer learning and teacher behavior differently from assigning selected students within a mixed-treatment classroom.
Likewise, restricting material sharing may create implementation conditions that differ from ordinary practice.
Researchers should therefore consider whether the selected design estimates the effect relevant to their intended policy or practice decision.
Preventing contamination is not automatically preferable when doing so changes the intervention into something that would not realistically be implemented.
What If Contamination Cannot Be Prevented?
When substantial contamination is unavoidable, researchers should reconsider whether the original treatment contrast remains meaningful.
Possible responses include changing the randomization unit before implementation, selecting more appropriate comparison conditions, or explicitly designing the study to estimate spillover effects.
In some circumstances, the research question may need to focus on the effect of offering the intervention within a naturally connected environment.
Researchers should not assume that contamination can be eliminated through increasingly restrictive participant instructions.
The design should reflect the actual social and organizational processes through which the intervention operates.
What Should Be Included in the Research Protocol?
The protocol should describe the expected contamination pathways, selected randomization unit, intervention delivery procedures, and rationale for the treatment contrast.
It should identify permitted cross-group interactions, restricted intervention components, monitoring procedures, and responses to consequential deviations.
The statistical analysis plan should specify the target estimand, appropriate analysis for the randomization design, and any planned sensitivity or spillover analyses.
Researchers should also explain the assumptions supporting the design, particularly when interference between participants is expected.
Transparent reporting should describe what contamination occurred and how it affects interpretation of the findings.
