Article Text
Statistics from Altmetric.com
Definition and rationale
Cluster randomised trials (CRCTs), sometimes called group randomised, field, community-based or place-based trials, involve the random allocation of existing groups of individuals to study arms. Group (‘cluster’) size can vary, ranging from families or classrooms to clinics or communities. Because many injury prevention interventions (eg, community education, mass media and legislation) are delivered to groups, CRCTs are an important methodology for evaluation studies in this field.
Members of clusters tend to respond to interventions in ways that are more similar to others in the same cluster than to members of different clusters, because, first, like people choose to join the same cluster (eg, schools, churches); second, cluster members have common exposures (eg, living on a busy street); and, third, cluster members interact with each other in ways that influence their responses (eg, sharing information).1 Thus, in a CRCT, participant outcomes are usually correlated within clusters, which affects both sample size requirements and statistical analysis. Even minimal intracluster correlation can substantially increase required sample size (see accompanying paper in this issue2). Except in rare instances when there is no intracluster correlation, CRCTs will therefore have reduced statistical efficiency relative to trials that individually randomise the same number of people.3
Because CRCTs are more complex to design, implement and analyse than individually randomised controlled trials, as we discuss below and in our companion paper,2 alternative designs should be carefully considered and use of a clustered design justified. Justifications fall broadly into two categories: scientific and logistical.4
Scientific justification
Cluster-level intervention and action
The intervention is delivered to and affects groups of people rather than, or in addition to, individuals. For example, one experimental school-based violence prevention programme established a local task force at each school to implement the programme, propose policy changes, develop school–community liaisons and solicit community support (table 1).11 Training of clinicians, teachers or other professionals in order to influence their patients or students is another common type of cluster-level intervention, as in Gielen et al (2001).12 Media campaigns, organisational changes and legislation are examples of cluster-level interventions.
Interventions where individual- and cluster-level actions cannot be separated
‘Herd’ immunity, well understood for vaccines, applies to injury prevention interventions when the actions of one person affect the risk of many. A smoke alarm giveaway programme in dense urban housing could protect both alarm recipients and their neighbours, by detecting fires before they spread to neighbouring homes and by directly alerting adjacent neighbours, who can hear the alarm and escape spreading fires.9
Contamination between study groups
Contamination of control subjects by intervention subjects in the same social unit can occur because people often share information and resources. Allocating the entire unit to a study arm reduces contamination risk. One CRCT evaluated the provision of visibility aids for children to wear while walking, along with educational information.20 Random allocation of schools reduced the potential for communication between intervention and control families, and for observation by control children of visibility aids being worn by intervention children.
Logistical reasons
Efficiency and cost
Using clusters allows study personnel to concentrate their activities in fewer locations. One CRCT involved door-to-door distribution of 20 000 free smoke alarms within defined geographic areas.9 To enrol similar numbers of widely scattered households would have required substantially greater resources. If professionals (eg, teachers) are delivering the intervention, fewer need be trained when each can intervene with many individuals. Also, if access to subjects is available only through professionals, individual randomisation of subjects could require large numbers of individual professionals to be contacted and asked to inform potential subjects. When a whole-class curriculum is being taught, as in Farrell et al (2003),10 it would be logistically difficult to break up existing classes and individually allocate children to different classes to receive the experimental curriculum.
Access to routine data
To obtain outcome data from relevant administrative or archiving authorities, it is typically easier, and may better protect patient privacy, to request aggregate data for an entire social unit—for example, a classroom or clinical practice—than individual subject data. In DiGuiseppi et al (2002),9 randomisation by electoral ward facilitated data collection through a surveillance system that linked residential fires and fire-related injuries to postcodes in the intervention or control wards rather than to individual households.
Reporting the rationale for adopting a cluster designi
The trial report should state the scientific or logistical justification for choosing a cluster randomised design—for example, ‘We randomised by ward in order to evaluate an area wide distribution … and to reduce contamination between intervention and control households.’9
Ethical issues in CRCTs
Recruitment and informed consent
There are several levels at which consent might be, and ideally is, sought in a CRCT,4 including randomisation, intervention and data collection. Consent for randomisation of the clusters tends not to be obtained from every individual within clusters, due to the cost and time required and the fact that ‘there is no accepted way … of allowing some cluster members' decisions to influence the chances of others participating.26 Ethical review boards may waive the right to individual informed consent under certain circumstances, which often apply to CRCTs. In the USA, for example, the requirement for informed consent may be waived if the research involves minimal risk, the waiver will not adversely affect subjects' rights and welfare, the research could not practicably be carried out without the waiver and, whenever appropriate, the subjects are provided with additional pertinent information after participation.27
Edwards et al (1999)28 divide CRCTs into two categories relevant to informed consent: ‘cluster–cluster trials', in which the intervention is aimed at clusters (eg, mass media campaign), and ‘individual–cluster trials’, in which individuals within clusters receive the intervention directly (eg, safety counselling for clinic patients). Eldridge et al (2005),26 after reviewing 199 CRCTs, proposed a more complex typology involving four different types (or levels) of interventions. Further, two-thirds of the trials had multifaceted interventions, precluding their dichotomisation as cluster–cluster or cluster–individual interventions. Nevertheless, to simplify our discussion, we have dichotomised interventions as in Edwards et al (1999),28 referring the interested reader to Eldridge et al (2005)26 for further information.
Cluster–cluster interventions
With cluster–cluster interventions, cluster members cannot act independently if the cluster participates, potentially contravening the ethical principle of autonomy.29 In Rowland et al (2003),23 for example, intervention schools received expert consultation and established working groups to develop school travel plans. Every student received the intervention whether or not the individual student chose to. Individuals not wishing to participate could theoretically leave the cluster (eg, change schools), but in practical terms, avoiding participation may be impossible. With cluster–cluster interventions, therefore, assessing the interests of the cluster and obtaining appropriate ‘cluster consent’ is paramount.
Consent to cluster randomisation is usually provided by a decision-maker (or ‘guardian’28) with administrative responsibility for the cluster (eg, headteacher or clinic director).30 The guardian's “scope of authority should encompass interventions of the type in question if provided outside of a research project”.30 The guardian must assess and weigh cluster interests to ensure that risks do not outweigh potential benefits. How best to assess these interests is not established.28 The guardian may consult with community representatives before deciding to allow cluster participation.30 In CRCTs based on geographic areas, there may be no one individual or group with overall administrative responsibility, in which case broad community consultation is appropriate. In DiGuiseppi et al (2002),9 the smoke alarm giveaway trial was developed by a working group representing local government, health agencies and community groups. The Medical Research Council (UK) recommends establishing a ‘cluster representation mechanism’ (CRM), comprising an individual, body or other mechanism, independent from the research team, to safeguard the interests of cluster members.31 The CRM volunteers its cluster when trial entry is perceived to be in the cluster's best interests and similarly may withdraw the cluster. When the intervention may be controversial, the CRM should consider directly examining cluster members' interests through focus groups or surveys prior to consenting. Members of participating clusters should generally be provided with information about the trial, including intended data use, to allow them to give their opinion to the CRM and, if possible, opt out of participation or data use.30 31 The study protocol should detail CRM selection and the information to be provided to the CRM and to cluster members.30
Individual–cluster interventions
With an individual–cluster intervention, after cluster enrolment, individual cluster members can still decide to participate or not. Hence, individual consent may, and generally should, be obtained. Some injury prevention trials have obtained consent for intervention from individuals in the intervention arm but not the control arm, usually because of cost and logistical considerations. In DiGuiseppi et al (2002),9 households in the intervention wards were contacted, informed about the smoke-alarm giveaway programme and given the opportunity to accept or decline a free alarm. Households in control wards were not contacted, although aggregated outcome data were collected from all wards. This is akin to the Zelen design for individual RCTs, in which consent for intervention is sought only from those in the experimental arm, after randomisation has occurred.32 The Zelen design has been criticised because control group members are participating in a study without consent.33 In addition, obtaining consent only from the intervention group members risks selection bias unless data are collected on all subjects. Therefore, if at all possible, informed consent should be sought from all potential subjects within study clusters, ideally prior to cluster randomisation. Kendrick et al (2005) randomised general practices, but delivered education discouraging baby walker use to individual patients.16 Within both intervention and control general practices, pregnant women were enrolled and gave informed consent.
Occasionally, individual consent may not be sought in an individual–cluster trial (eg, Cumming et al (2008)8 and Kendrick et al (1999)15) for reasons of science—for example, the trial combines cluster and individual interventions and consent cannot be obtained for the former—or economics. The Medical Research Council (UK) states that in such situations it is essential that no cluster member be individually disadvantaged by the cluster's participation in the trial, a requirement likely to be met only by trials that offer an additional (minimal risk) option to members of intervention clusters.31 For example, Kendrick et al (1999), in a study providing safety advice and low-cost safety equipment to intervention families through home visitors, collected medical record data from all subjects without consent.15 In such trials, informing cluster members in general terms about the nature of the research and potential uses of their data is recommended.
Refusal, early stopping and withdrawal
If the cluster guardian declines participation (or withdraws the cluster from the trial), cluster members cannot choose (or continue) to participate even if they want to. Cluster members could potentially transfer to a participating cluster if they have been informed about the trial. If their individual consent was not obtained, however, they may be unaware that their participation has ended. Informing cluster members of cluster withdrawal may therefore be appropriate.
A cluster member might wish to leave a study even when their cluster guardian wishes cluster participation to continue. With cluster–cluster interventions, the cluster member may need to transfer to a non-participating cluster in order to withdraw. When the intervention is at the organisational or environmental level (eg, community-wide media campaigns), it may be impossible for cluster members to withdraw. In situations where the desire to withdraw comes because the subject has recognised adverse effects, the benefits of an independent CRM become apparent, since the CRM can withdraw the entire cluster if it is perceived to be in its members' best interests.31
Reporting ethical issues
CRCTs should report ethical review board approval and how, from whom and at what point in the trial informed consent was obtained.3 Information to be provided includes how cluster-level decision-makers were identified, what steps were taken (if any) to understand cluster members' interests, what information was provided and to whom, and what opportunities (if any) were offered to cluster members to refuse participation.
Study design issues
Selection bias
To reduce the risk of systematic differences between experimental and control groups, clusters should be randomised and allocation concealed using recommended methods.34 35 Because CRCTs often involve relatively few clusters, simple randomisation may result in chance imbalances between groups. Restricted randomisation procedures—for example, stratification or minimisation—or a matched-pair design can help assure balance.31 Using restricted randomisation in unblinded trials can, however, increase the ability to predict cluster assignments, potentially introducing selection bias. The use of matching can be limited by difficulty choosing appropriate matching variables, loss of information on between-cluster variability and loss of power if one pair member withdraws.31
The risk for selection bias may persist in CRCTs even when randomisation processes are adequate.36 37 One reason is that CRCTs often do not (or cannot) mask treatment allocation.38 If cluster members are recruited after cluster randomisation, without masking allocation at the time of recruitment, a strong potential for postrandomisation recruitment bias is created. In Kannus et al (2000),13 eligible subjects living in intervention care units were informed that trial participants would have to wear hip protectors. These residents were substantially less likely to participate than were control unit residents (69% vs 91%), presumably because of reluctance to wear hip protectors. This bias can be avoided if clusters, and individuals within clusters, are identified and recruited and consent is obtained before the clusters are randomly allocated.18 24 When cluster members cannot be recruited until after cluster allocation—for example, when high turnover of cluster membership or prospective recruitment over time is anticipated—recruitment and consent should be blind to group allocation, as in Thoreson et al (2009).25 Selection bias can also be avoided if data are included from every cluster member, as in studies such as those discussed above, where individual informed consent is waived or considered unnecessary.5 8 9 15 22 For example, in Ray et al (2005),22 aggregate (anonymous) outcome data on falls among all residents of consenting nursing homes were obtained from a hospital discharge data set.
Systematic differences between groups can also occur if intervention and control clusters or cluster members drop out at different rates (‘attrition bias’). Control clusters often drop out at greater rates because they receive fewer resources and less attention. Using an active (‘attention’) control condition can reduce attrition bias,11 14 as well as attention or performance bias.1 Flay et al (2004)11 provided a control curriculum of equal intensity to his violence prevention curriculum; all schools remained in the study all 4 years. Wait-listing of control clusters to receive the intervention after completion of the study is sometimes used to discourage control attrition6 7 17 but may be less effective than active controls. In Cox et al (2008),7 cluster attrition was substantially greater in the control group than the intervention group (55% vs 17%) despite a promise to provide the fracture prevention programme one year later. Shorter waiting periods may be more effective.6 17 Providing active or wait-list control conditions may require substantial resources to implement, given the large sample sizes typically involved. Intervention clusters sometimes drop out at a greater rate than controls, usually when the demands made on them exceed perceived benefits. Ensuring that the cluster ‘guardian’ understands what the intervention involves before the cluster is enrolled is essential. Attrition bias at the cluster level can be avoided if individual clusters cannot withdraw, such as when the clusters are geographic units within a consenting community.9 Within CRCTs, attrition bias can also act at the individual level for similar reasons and with similar solutions. Intention-to-treat analysis can reduce the influence of attrition bias at the analytic stage if sufficient data are available or can be imputed.
Dilution bias
In CRCTS, the study intervention may not be received by some members of intervention clusters due to postrandomisation refusal or to outmigration from the cluster.38 Low rates of uptake by cluster members can dilute intervention effects, reducing effect size.39 In a CRCT in which 22% of the intervention group did not actually receive the home safety intervention for young children, the reported absence of intervention effect may have resulted in part from dilution bias.15 Recruiting individuals and obtaining their consent prior to cluster randomisation reduces dilution (and selection) bias and is therefore recommended. In Swart et al (2008),24 where 20% of eligible households were excluded prior to randomisation due to refusal or failure to make contact, more than 95% of intervention cluster households received the intervention. Inflating pre-trial sample size estimates may be necessary to account for potential dilution effects.39 Use of statistical methods to estimate intervention effects that take into account variable levels of exposure to the intervention should be considered.40 41 Compliance with the intervention, at both cluster and individual levels, should be carefully measured to aid in interpreting results.1
Information/detection bias
Differential outcome assessment can be an important source of bias in CRCTs, particularly when outcome assessors and participants are not blinded to group allocation. In Cumming et al (2008), those collecting outcomes regarding falls were not blinded to intervention status.8 Reporting of falls may have been more diligent in intervention than in control wards, hence underestimating any potential intervention benefit. Blinded assessment of outcomes6 9 25 and use of objectively measured outcomes can reduce this bias. Even when cluster-level blinding is impossible, it may be possible to blind individual cluster members, particularly when both study arms receive relatively similar interventions (eg, extensive versus brief advice). Individuals can be informed about the general nature of the study intervention and the participation of their cluster in a trial, but not to which arm their cluster has been assigned.
Reporting study design
Unlike individually randomised trials, CRCTs vary widely in their units of randomisation, experimentation and observation, all of which may differ from each other.4 The unit of randomisation might include an existing social unit either directly (eg, a clinic) or indirectly (ie, unit members with a specific characteristic or condition, such as clinic patients aged under 2 years). The unit of experimentation, which receives the intervention, may include the whole cluster, individuals within the cluster, or both. Although Prinz et al (1994)21 randomly allocated school classrooms, the intervention was delivered only to eight children per class. The unit of observation, on which outcomes are measured, might be the same as the unit of randomisation or experimentation, or may be individuals within units who could benefit indirectly. In Gielen (2001),12 the intervention was delivered to randomly allocated clinicians, but the outcomes (safety behaviours) were measured on their patients. Outcomes are often measured at multiple levels.12 16 19 Reports of CRCTs should specify the units of randomisation, experimentation and observation. Eligibility criteria for both clusters and participants within clusters should be described. Additional design features to be reported include who identified and recruited study participants, whether recruitment occurred before or after cluster randomisation and whether recruiters or cluster members were aware of the cluster allocation status at the time of recruitment. The flow both of clusters and of individual participants from initial recruitment to final analysis, including the total cluster population in each study arm and attrition at each stage, should be reported. Numerous examples of flow diagrams of varying complexity have been published.9 18 19 23
Conclusions
Cluster randomised trials are often used in injury prevention research, usually because of the nature of the intervention and its expected action or to improve efficiency and reduce costs. Due to the complexities of CRCT design, implementation and analysis, the investigator must consider numerous potential issues. The investigator must justify adequately the choice of this design and address multiple levels of consent and risks for selection, attention, dilution and information bias. Nevertheless, for many types of injury prevention interventions, the potential benefits of this design outweigh its risks. In such cases, careful consideration must be given to ensure appropriate trial design, implementation, analysis and reporting. In a companion paper,2 we report on other important aspects of the design and analysis of cluster randomised controlled trials in evaluating the effectiveness of injury prevention interventions.
References
Footnotes
Funding Dr DiGuiseppi is funded in part through Grant No R49-CCR811509 from the Centers for Disease Control and Prevention.
Competing interests None.
Provenance and peer review Commissioned; externally peer reviewed.
↵i At the end of each section, we provide guidelines and examples for reporting the issue discussed, based on a published extension of CONSORT guidelines for the reporting of individually randomised controlled trials that addresses unique aspects of reporting CRCTs.42