{
  "id": "cluster-randomized-trial",
  "title": "Cluster-Randomized Trial",
  "lang": "en",
  "source_lang": "en",
  "summary": "A cluster-randomized trial assigns groups rather than individuals to study interventions, requiring design and analysis methods that account for similarities among members of each group.",
  "category": "health",
  "subcategory": "medicine",
  "type": "concept",
  "aliases": [
    "Cluster randomized trial",
    "Cluster randomised trial",
    "Group-randomized trial",
    "Group-randomised trial",
    "Community-randomized trial"
  ],
  "generated_at": "2026-11-04",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/cluster-randomized-trial/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/cluster-randomized-trial/",
      "json": "https://aiwiki.page/zh/cluster-randomized-trial.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A cluster-randomized trial is a type of randomized controlled trial in which groups of people, rather than individual participants, are randomly assigned to intervention or comparison conditions. These groups, called clusters, may be households, schools, medical practices, hospitals, villages, or other organizational or geographic units. Outcomes may be measured in individuals, in the clusters themselves, or at both levels. Because people within a cluster often share characteristics, environments, and exposures, their outcomes are generally not statistically independent; this dependence must be addressed in the trial’s design and analysis."
    },
    {
      "heading": "Purpose and defining features",
      "text": "Cluster randomization is used when an intervention operates at group level, when individual allocation is impractical, or when assigning individuals separately would lead to substantial contamination—exposure of participants in one study arm to the intervention assigned to another.\n\nFor example, a program that changes prescribing procedures throughout a medical practice cannot readily be restricted to randomly selected patients within that practice. Similarly, a school-based educational program may affect classmates who were not intended to receive it.\n\nThe design distinguishes three potentially different units:\n\n- Unit of randomization: the group assigned to an intervention.\n- Unit of intervention delivery: the organization, professional, household, or individual receiving the intervention.\n- Unit of outcome measurement: the person or group for which outcomes are recorded.\n\nA hospital may be randomized, its staff may receive training, and patient outcomes may be measured. The number of measured patients does not change the fact that randomization occurred at hospital level.\n\nCluster trials are common in public health, primary health care, education, and health-system research. They are often, but not necessarily, pragmatic clinical trials: cluster allocation describes how treatment is assigned, whereas pragmatism concerns how closely a trial reflects routine practice."
    },
    {
      "heading": "Reasons for using cluster allocation",
      "text": "Group-level interventions\n\nSome interventions inherently affect whole populations or institutions. Examples include community sanitation programs, hospital infection-control policies, and changes to service delivery. Randomizing the relevant organizational or geographic unit allows evaluation at the level where the intervention operates.\n\nContamination and spillover\n\nIndividual randomization can be unsuitable when clinicians apply new practices to all their patients, participants share intervention materials, or an intervention changes a shared environment.\n\nIn infectious-disease research, one person’s intervention can also affect another person’s outcome. A vaccine program may reduce exposure among unvaccinated residents through herd immunity. In such settings, a cluster trial can estimate the overall effect of assigning a community to a program rather than only the biological effect among direct recipients.\n\nCluster boundaries do not necessarily eliminate spillover. Movement between communities, shared staff, and contact networks extending across trial arms can still affect the treatment contrast.\n\nOperational considerations\n\nDelivering one program throughout an institution may be more feasible than operating several competing programs simultaneously. However, operational convenience comes with statistical and ethical costs, including a potentially larger required sample and limited individual choice about exposure to institution-wide changes."
    },
    {
      "heading": "Major designs",
      "text": "Parallel-group trials\n\nIn a parallel cluster trial, clusters are allocated to intervention or comparison conditions and remain in those conditions during follow-up. Allocation may be unrestricted or may use stratification, matched pairs, or constrained randomization to improve balance on important baseline characteristics.\n\nMatching and restrictions affect the randomization scheme and need to be reflected in the analysis.\n\nCluster crossover trials\n\nIn a cluster crossover trial, each cluster receives different conditions during different study periods, with the sequence randomized. Comparing a cluster with itself can reduce the influence of stable differences between clusters.\n\nInterpretation becomes more difficult when an intervention has lasting effects, when washout is inadequate, or when outcomes change over calendar time. Interventions that permanently alter staff behavior or infrastructure may not be suitable for crossover evaluation.\n\nStepped-wedge trials\n\nIn a stepped-wedge cluster-randomized trial, clusters switch from a comparison condition to an intervention according to randomized schedules, usually until all clusters have received the intervention.\n\nThe intervention effect must be distinguished from calendar-time effects because more clusters receive the intervention later in the study. Secular trends, changing intervention effects, and correlations between observations within and across periods complicate analysis. Eventual access to an intervention does not by itself establish the design’s ethical acceptability or show that the intervention is beneficial.\n\nAcross these designs, studies may follow a fixed cohort, recruit different individuals at successive periods, or use an open cohort with people entering and leaving."
    },
    {
      "heading": "Intracluster correlation and sample size",
      "text": "People within a cluster may have similar outcomes because they share clinicians, social conditions, infrastructure, or exposure to the intervention. This resemblance is commonly summarized by the intracluster correlation coefficient, or ICC.\n\nFor a simple random-intercept model,\n\n\\[Y_{ij}=\\mu+\\tau T_j+b_j+\\varepsilon_{ij},\\]\n\n\\(Y_{ij}\\) is the outcome for person \\(i\\) in cluster \\(j\\), \\(T_j\\) indicates treatment allocation, \\(b_j\\) represents a cluster-level random effect, and \\(\\varepsilon_{ij}\\) represents individual-level variation. If the two variance components are \\(\\sigma_b^2\\) and \\(\\sigma_w^2\\), respectively, then\n\n\\[\\rho=\\frac{\\sigma_b^2}{\\sigma_b^2+\\sigma_w^2}.\\]\n\nHere, \\(\\rho\\) is the ICC. Its value depends on the outcome, setting, population, and measurement schedule.\n\nFor a simple parallel design with equally sized clusters containing \\(m\\) participants, a common sample-size inflation factor, or design effect, is\n\n\\[DE=1+(m-1)\\rho.\\]\n\nFor example, with 50 participants per cluster and an ICC of 0.02, the design effect is 1.98. Under the assumptions behind this approximation, nearly twice as many individual observations would be needed as in a comparable individually randomized trial.\n\nThis formula is not a universal sample-size calculation. Unequal cluster sizes, few clusters, repeated measurements, missing data, outcome distributions, and the allocation scheme require additional consideration. Adding clusters generally provides more information than adding the same number of people to existing clusters. A large number of participants cannot fully compensate for having very few independent randomized units."
    },
    {
      "heading": "Statistical analysis and interpretation",
      "text": "Treating all participants as independent usually underestimates uncertainty when within-cluster correlation is positive. The resulting confidence intervals can be too narrow and statistical tests too permissive.\n\nCommon analytical approaches include:\n\n- Cluster-level analysis: calculating an outcome summary for each cluster and comparing those summaries between arms.\n- Mixed-effects models: modeling individual outcomes while including cluster-level random effects.\n- Generalized estimating equations: estimating population-average effects while accounting for correlated observations.\n- Randomization-based methods: evaluating results against the allocation procedure actually used.\n\nThese methods are not interchangeable. Their suitability depends on the outcome, number of clusters, design, and target effect. With few clusters, standard large-sample methods can provide unreliable inference unless appropriate small-sample procedures are used.\n\nDefining the treatment effect\n\nA trial may target the effect for the average individual or the effect for the average cluster. These differ when clusters vary in size: an individual-average effect gives larger clusters more influence, whereas a cluster-average effect gives clusters equal weight.\n\nThe distinction becomes especially important when cluster size is related to outcomes or treatment effects. For binary outcomes, population-average and cluster-specific model estimates can also differ because they represent different statistical quantities.\n\nAn intention-to-treat analysis ordinarily compares clusters and their eligible participants according to randomized allocation, irrespective of adherence. This principle does not remove the need to address clustering, missing outcomes, or changes in cluster membership."
    },
    {
      "heading": "Bias and practical limitations",
      "text": "Recruitment after randomization\n\nIf individuals are recruited after their cluster’s allocation becomes known, knowledge of treatment assignment may influence whom recruiters approach or who agrees to participate. This can create systematic differences between trial arms despite valid randomization of clusters.\n\nPre-randomization identification of participants, objective eligibility rules, and allocation-masked recruitment are design features used to limit this problem. Allocation concealment before cluster assignment is distinct from preventing recruitment bias afterward.\n\nBaseline imbalance and masking\n\nWith relatively few randomized clusters, chance imbalances in important characteristics are more likely. Prespecified covariate adjustment can improve precision, but cannot automatically remedy every imbalance.\n\nBlinding of participants and intervention providers is often difficult for organizational interventions. Outcome assessors and analysts may nevertheless be masked, and routinely collected or objectively defined outcomes may reduce some forms of measurement bias.\n\nMissingness and generalizability\n\nLoss of an entire cluster can have a substantial effect on precision and balance. Missing individual outcomes may also be related to allocation or prognosis.\n\nInstitutions willing and able to participate may differ from those that do not participate. Consequently, cluster randomization does not by itself guarantee that results generalize to other organizations, communities, or health systems."
    },
    {
      "heading": "Ethics and reporting",
      "text": "Cluster trials are a form of human-subjects research and require ethical review appropriate to their activities. Ethical questions include who counts as a research participant, who receives the intervention, whose data are collected, and when individual consent is required.\n\nPermission from a hospital administrator, school leader, or community representative may authorize institutional participation, but does not automatically replace individual informed consent. An institutional review board may approve a waiver or modification of consent under applicable criteria; cluster allocation alone does not justify one.\n\nReporting guidance includes the cluster-trial extension to the CONSORT statement. Important reporting elements include the rationale for cluster randomization, eligibility criteria at both levels, allocation methods, numbers of clusters and individuals recruited and analyzed, timing of individual recruitment relative to allocation, sample-size assumptions, and analytical handling of clustering. Clear reporting also distinguishes cluster withdrawal from individual loss to follow-up and identifies the population to which the estimated effect applies."
    }
  ],
  "links": [
    "randomized-controlled-trial",
    "public-health",
    "primary-health-care",
    "health-system",
    "pragmatic-clinical-trial",
    "sanitation",
    "infectious-disease",
    "vaccine",
    "herd-immunity",
    "crossover-trial",
    "mixed-effects-model",
    "intention-to-treat-analysis",
    "allocation-concealment",
    "blinding-clinical-trials",
    "human-subjects-research",
    "institutional-review-board",
    "consort-statement"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}