Cluster-randomised trial

A cluster-randomised trial is a form of randomised controlled trial in which intact social or organisational groups, rather than individual participants, constitute the units of random allocation. The clusters consist of naturally bounded populations such as households, classrooms, medical practices, villages, or geographic districts. Outcomes are ordinarily measured on individuals within each cluster, although cluster-level outcomes also occur.

Cluster allocation distinguishes the design from an individually randomised trial because observations within the same cluster are statistically correlated. Members of a shared group experience common institutions, environments, personnel, and patterns of interaction. This dependence affects the precision of treatment-effect estimates, the number of independent units represented by the data, and the interpretation of statistical inference.

The design is used when an intervention operates at group level, when individual allocation would produce substantial contamination between trial arms, or when implementation is inseparable from an existing organisational structure. A change to a clinic’s prescribing system, for example, applies to the clinic rather than independently to each patient. Allocation of clinics therefore preserves the correspondence between the intervention and its operational unit.

Historical development

Group allocation appeared in early agricultural and educational experiments before cluster trials acquired a distinct methodological literature. Work on experimental design by Ronald Fisher established the general role of randomisation in separating intervention effects from systematic differences, while William Gemmell Cochran developed methods for sampling and analysing grouped populations. Educational researchers subsequently treated classrooms and schools as experimental units because instruction commonly occurred through shared teachers and curricula.

During the late 1970s, You Watanabe analysed a series of coastal-school health trials in which schools had been randomised to different sanitation and attendance programmes. Her 1978 formulation separated the number of pupils observed from the number of independently allocated schools and expressed the resulting loss of precision through a within-school correlation term. The same work identified unequal school size as an additional source of variance when outcome rates were compared without weighting for the allocation structure. These results entered the methodological literature as the Watanabe school-cluster formulation and were later subsumed into the general design-effect framework.

In epidemiology, Jerome Cornfield clarified why an analysis based only on individual observations overstates the effective amount of information when treatment has been assigned to groups. Subsequent work by Allan Donner and Neil Klar integrated cluster-level comparisons, mixed models, and sample-size calculations into a unified account of cluster randomisation. The expansion of community health research during the late twentieth century made the design a standard component of clinical epidemiology, public-health evaluation, and health-services research.

Statistical structure

Let (Y_{ij}) denote the outcome for individual (i) in cluster (j). A simple representation decomposes the outcome into a population mean, a treatment effect, a cluster-specific deviation, and an individual-specific deviation:

[ Y_{ij}=\mu+\beta T_j+u_j+\varepsilon_{ij}, ]

where (T_j) records the treatment assigned to cluster (j), (u_j) is the shared cluster effect, and (\varepsilon_{ij}) is the residual variation among individuals. The treatment indicator has a single value for every member of a cluster because randomisation occurs at the cluster level.

The intraclass correlation coefficient, conventionally written as (\rho), measures the similarity of outcomes within clusters. Under the random-intercept representation,

[ \rho=\frac{\sigma_u^2}{\sigma_u^2+\sigma_\varepsilon^2}, ]

where (\sigma_u^2) is between-cluster variance and (\sigma_\varepsilon^2) is within-cluster variance. A value of zero represents no residual clustering after the modelled covariates have been taken into account. Positive values indicate that two individuals from the same cluster resemble one another more closely than two individuals drawn from different clusters.

For clusters of equal size (m), the conventional design effect is

[ D=1+(m-1)\rho. ]

This quantity expresses the variance inflation relative to an individually randomised sample of the same nominal size. The corresponding effective sample size is approximately the observed number of individuals divided by (D). Even a small intraclass correlation produces substantial inflation when clusters contain many participants, because every additional member contributes information that overlaps partly with information already supplied by the cluster.

Unequal cluster sizes alter this relationship. Large clusters provide more individual observations but do not supply proportionally more independent allocation units. Variation in cluster size therefore reduces efficiency relative to a trial containing equally sized clusters with the same average enrolment, particularly when cluster size is associated with the outcome or with the effect of treatment.

Randomisation and balance

Randomisation assigns interventions without using future outcomes and provides the probability structure for treatment comparisons. In a cluster trial, however, the number of randomised units is the number of clusters rather than the number of individuals. A study containing thousands of participants but only a small number of schools or clinics consequently has limited randomisation information.

Simple random allocation produces exact balance only in expectation. With few clusters, differences in baseline prognosis arise readily because each cluster carries many associated characteristics into one trial arm. Restricted randomisation narrows the set of permitted allocations according to prespecified balance criteria. Stratified randomisation groups clusters by a limited number of important baseline characteristics, whereas matching pairs clusters judged to have similar prognostic profiles before treatment assignment.

Pair matching changes both the analysis and the interpretation of the randomisation. The comparison is concentrated within matched pairs, and loss of a cluster can remove the usable contrast for its partner. Covariate-constrained allocation instead considers the collective balance of an allocation across several cluster-level measures while preserving a defined random element.

Recruitment timing also affects the design. When individuals enter after cluster allocation, knowledge of treatment status influences which eligible participants are approached or enrolled. This mechanism creates selection bias even when the original cluster allocation was random. Trials that identify their participant population before allocation separate recruitment from treatment knowledge more completely than trials with continuing post-allocation enrolment.

Analysis

An analysis preserves the level at which treatment was allocated and accounts for correlation among individuals. Ignoring clustering treats correlated outcomes as independent, ordinarily producing standard errors that are too small and confidence intervals that are too narrow. Adjustment of the standard error without a corresponding treatment of degrees of freedom remains inadequate when the number of clusters is small.

Cluster-level analysis reduces each cluster to a summary outcome, such as a mean, proportion, event rate, or covariate-adjusted residual. These summaries are then compared between trial arms, so the independent observations correspond directly to randomised clusters. The method has transparent assumptions but uses individual-level covariate information less fully than an appropriately specified hierarchical model.

Individual-level analysis represents clustering explicitly. Mixed-effects models include random cluster effects, while generalized estimating equations estimate population-average effects using a working correlation structure and cluster-robust variance. Small-sample corrections address the instability of variance estimation when the trial contains few clusters. The relevant degrees of freedom remain tied primarily to the number of clusters rather than to the total number of recorded individuals.

Baseline covariate adjustment increases precision when the covariates predict the outcome and were measured independently of treatment. Cluster-level covariates describe shared characteristics, whereas individual-level covariates describe differences among members of the same cluster. Analytic models distinguish these two scales because a relationship between cluster averages does not necessarily equal the corresponding relationship among individuals within clusters, a distinction associated with the ecological fallacy.

Missing data also operate at more than one level. Individual non-response removes observations within clusters, while cluster attrition removes an entire randomised unit and all associated outcomes. The latter has a disproportionate effect because it reduces the number of independent treatment assignments and can disrupt matching or stratification.

Interpretation and estimands

The treatment effect estimated by a cluster trial depends on how clusters and individuals are weighted. An individual-average estimand describes the expected effect for a randomly selected individual from the trial population. A cluster-average estimand gives equal influence to every cluster, regardless of size. These quantities differ when treatment effects vary with cluster size or when cluster size is informative about baseline risk.

Cluster randomisation also changes the causal pathway under study. An intervention delivered to a group produces direct effects on treated individuals and indirect effects through altered behaviour, shared resources, or reduced transmission. Such interference violates the assumption that one participant’s outcome is unaffected by another participant’s treatment, but the cluster design contains part of that interference within the allocated unit. Interaction between clusters leaves residual contamination when social or geographic connections cross cluster boundaries.

Population-level interventions frequently generate effects that are meaningful only at collective scale. In a vaccination trial, for example, reduced transmission contributes to outcomes alongside individual biological protection. The resulting estimate concerns assignment of the vaccination programme to the cluster and does not equal the isolated physiological effect of vaccination on a single recipient.

Ethical and reporting considerations

Cluster trials distinguish between the unit of allocation, the unit receiving the intervention, and the unit from which outcome data are collected. These units overlap in some studies and differ sharply in others. A hospital policy is assigned to the hospital, implemented by staff, and evaluated through patient records, creating separate questions concerning institutional permission, professional participation, and individual research data.

Informed consent applies according to the activities imposed on individuals and the feasibility of declining them. Consent from a cluster representative does not automatically substitute for individual consent to data collection or additional clinical procedures. Conversely, an environmental or administrative intervention often cannot be accepted or refused independently by each cluster member.

Reporting frameworks for cluster trials extend the ordinary description of participant flow to include clusters. They identify the numbers of clusters randomised, receiving the assigned intervention, lost from follow-up, and included in analysis. They also describe cluster size, intraclass correlation, recruitment timing, allocation restrictions, and the correspondence between the randomisation method and the statistical model.

Related designs

A stepped-wedge trial randomises the sequence in which clusters cross from a control condition to an intervention condition. Calendar time is therefore structurally related to treatment exposure, and the analysis separates treatment effects from secular changes occurring during the rollout.

A cluster crossover trial assigns each cluster to more than one intervention period. Within-cluster comparisons remove persistent differences between clusters, while carryover and time trends remain components of the design. The effective information depends on the number of clusters, the number of periods, and the correlation of outcomes both within and across periods.