Containing Interference with Cluster Randomization

33 concepts28 papers1 proposal

When designing an experiment, we rely on a fundamental assumption: what happens to participant A should not affect participant B. In statistics, this is known as the Stable Unit Treatment Value Assumption (SUTVA). If you are testing a new medication, giving it to one patient does not inherently change the biology of another patient down the hall.

But when interventions involve information, behavior, or social environments, this assumption quickly breaks down. If you randomize individual students within the same classroom to receive a new health education program, those students will naturally talk to their friends in the control group. The control group becomes "contaminated" with the treatment. As a result, the difference between the two groups shrinks, and the experiment understates the true impact of the intervention.

The Cluster as a Containment Boundary

To solve this problem, researchers use cluster randomization. Instead of randomizing individuals, they randomize intact social, organizational, or geographic units—known as clusters.

By assigning entire schools (Lubman et al. (2016); Werner-Seidler et al. (2025); Wassenaar et al. (2021)), villages (Zeng et al. (2008)), or community groups (Scull et al. (2020); Trost et al. (2021)) to either the treatment or control arm, the social interactions that would otherwise cause spillover are safely contained within the cluster.

If parents at a specific service center learn a new behavioral strategy, they share it with other parents at that same center—all of whom are already in the treatment group (Havighurst et al. (2019)). Meanwhile, participants in a separate, control-assigned community center remain unexposed (Tao et al. (2015)). This geographic or administrative partitioning allows researchers to cleanly evaluate community-wide interventions against standard care guidelines (Iwuji (2016); Liu et al. (2015)).

Preventing Institutional Contamination

Spillover isn't just a byproduct of people talking; it also occurs through institutional workflows, shared resources, and practitioner habits.

Consider a clinical trial evaluating a new digital decision-support tool for emergency department clinicians (Melnick et al. (2019)) or a specialized training program for general practitioners (Sharifi et al. (2019)). If a doctor learns a more effective way to manage a condition or reduce low-value care (Moore et al. (2023)), they cannot easily "unlearn" that behavior when walking into the room of a control-group patient. They will naturally apply their new knowledge to all of their patients.

Individual Randomization (High Contamination):
[Doctor] ---> Learning Intervention ---> Applies to Patient A (Treatment)
        ---> Cannot "Unlearn"       ---> Applies to Patient B (Control) ❌

Cluster Randomization (Contained):
[Hospital Clinic A] ---> All Doctors Trained ---> All Patients Treated  ✅
[Hospital Clinic B] ---> Standard Care       ---> All Patients Control  ✅

By randomizing at the organizational level—such as entire clinics (Bailey et al. (2023)), hospitals (Gallagher et al. (2023)), or workplaces (Hu et al. (2023))—the intervention isolates the practitioners themselves. This ensures that system-level changes or behavioral shifts do not bleed into the care of the control group (Jafar et al. (2022)).

Advanced Topologies and Digital Networks

While physical boundaries like schools or clinics are intuitive clusters, modern experimentation often deals with fluid, non-physical networks. In online marketing or social platforms, users interact constantly, making individual A/B testing highly susceptible to network interference.

To counter this, the authors of Cai et al. (2021) developed LinkLouvain, a method that uses graph neural networks to predict user interactions and cluster them into tightly knit groups based on interaction likelihood. By randomizing at the level of these algorithmic clusters, digital platforms can contain network interference just as effectively as a physical school boundary.

Furthermore, cluster designs can be adapted to study the spillover itself rather than just preventing it. In a two-stage factorial cluster-randomized design, the authors of Oles et al. (2024) assigned varying treatment dosages to primary clusters and then randomly selected units within those clusters for intervention. This structure allows researchers to observe how an intervention influences untreated individuals within the same social environment, turning a methodological nuisance into a measurable feature of community-wide impact.

The Statistical Cost of Clustering

If cluster randomization is so effective at preventing contamination, why don't we use it for every experiment? The answer lies in statistical power.

When individuals belong to the same cluster, their behaviors and outcomes are no longer independent. Students in the same school share the same teachers, neighborhood, and socioeconomic background. This correlation means that adding more individuals from the same cluster provides less new information than adding individuals from entirely new clusters.

To account for this nested data structure, researchers cannot use standard statistical tests. Instead, they must employ specialized frameworks like linear mixed models (Trost et al. (2021)) or multi-level randomization policies (Raji et al. (2025)) to correctly estimate treatment effects without inflating their confidence. Cluster randomization requires a deliberate trade-off: you accept a loss in statistical efficiency per participant in exchange for a clean, uncontaminated measurement of the treatment's true effect.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 28 papers
  2. Version 1 · 2026-10-04 · 29 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_cluster_randomization_496780cd,
  title = {cluster randomization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cluster-randomization/rdr_496780cdc4b84b8eba0432ebb7d0a631},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 2, frozen text):

@misc{lacuna_cluster_randomization_496780cd_v2,
  title = {cluster randomization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cluster-randomization/rdr_496780cdc4b84b8eba0432ebb7d0a631/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake