Principle 1: Design type and justification of use of CRT

The most basic form of CRT design is the parallel CRT; however, this design has several variations. Detailed descriptions, advantages, and disadvantages of each design type are outlined in Table 1. Authors should report the rationale for why the design chosen is most appropriate for their study. Some acceptable examples of design justification include the desire to minimize contamination bias between ICUs or floors in different study groups within the same facility, the recognition that a unit-level intervention would be more generalizable than randomly assigning the intervention at the patient-level, and the need to conduct a study of sufficient size with the available resources.

Table 1.

Variations of the Cluster-Randomized Trial (CRT) Design

Systematic review findings

Of the 44 studies included in the review, 15 (34.1%) used a CRT with crossover, 11 (25.0%) used a parallel CRT design, 7 (15.9%) used a stratified CRT design, 4 (9.1%) used a CRT with stepped-wedge design, 3 (6.8%) used a matched CRT design, 2 (4.5%) used a CRT with crossover and multiple periods, and 2 (4.5%) used a stratified CRT design with crossover. Also, 22 of the included studies (50.0%) offered justification for their use of a CRT (Table 2). In a good example of justification for the CRT design, Huang et al (2016) stated that they “… chose this design to obtain results that could be generalized to the broadest set of hospitals, to use processes potentially adoptable by many hospitals, and to conduct a study of sufficient size with the available resources. Randomization of entire hospitals allowed us to recruit a broad array of hospitals” (see Appendix 2 online).

Table 2.

Compliance with 7 Key Epidemiological Principles Among Published Cluster-Randomized Trials (CRT) in Infection Control and Hospital Epidemiology (N=44)

Table 3.

Summary of Key Design and Analysis Considerations When Developing a Cluster-Randomized Trial (CRT) in Infection Control and Hospital Epidemiology

Principle 2. Accounting for clustering when estimating sample size and reporting of intracluster correlation coefficient or coefficient of variation

The correlation and thus nonindependence that exists among individual patients in a cluster must be accounted for when estimating sample size for such trials, yet many studies neglect to consider the within-cluster and between-cluster variation as measured by the intracluster correlation coefficient (ICC) or coefficient of variation (CV). A review by Simpson et alReference Simpson, Klar and Donnor ⁵ of primary prevention trials showed that only 4 studies (19%) accounted for between-cluster variation in their sample size or power calculation. ICC measures the degree of similarity among outcomes within a cluster.Reference Donner and Klar ⁶ Generally, the higher the ICC, the more similarity that exists within clusters resulting in a loss of precision estimating effect of intervention. Therefore, standard approaches for estimating sample size that do not consider clustering may increase the probability of a type II error, meaning that the study will be underpowered.Reference Donner and Klar ¹

In some studies, clustering may arise at >1 level; therefore, 2 ICCs should be defined, for example, when an ICU within a hospital and the hospital itself are randomized. Variation exists among hospitals in addition to variation among ICUs within a hospital. An additional source of variance arises when the crossover design is used and each cluster receives the intervention in a separate period of time. In this case, it is important to account for period variance.Reference Arnup, McKenzie, Hemming, Pilcher and Forbes ⁷

Cluster randomization is less statistically efficient than randomizing individuals. Increasing the number of clusters enrolled in a CRT has a greater impact than increasing the number of individuals enrolled within each cluster on statistical power.Reference Donner and Klar ^{6 ,} Reference Rothman, Greenland and Lash ⁸ Therefore, many investigators choose to enroll a subsample of individuals within each cluster. The numbers of individuals needed to enroll per cluster depends largely on the underlying value of the ICCReference Donner, Birkett and Buck ⁹ and the anticipated effect size. A paper by Rutterford et alReference Rutterford, Copas and Eldridge ¹⁰ provides detailed guidance on how to estimate sample sizes for CRT.

Principle 3. Consent

Randomization of groups rather than individuals presents unique ethical considerations. It may be appropriate for key decisions makers to act as surrogates for a community or cluster and consent to randomization.Reference Sim and Dawson ¹¹ For example, nurse managers may consent on behalf of their unit to participate in an intervention trial with the outcome of hand hygiene adherence. Although ethical approval may be given at the cluster level, the refusal of an individual patient or healthcare worker (HCW) to participate in a study must be Epidemiology respected. It can be logistically difficult and perhaps unfeasible to obtain individual consent from large clusters.Reference Edwards, Braunholz and Lilford ¹²

Principle 4. Level of inference

In epidemiology, inference refers to the statistical process of generalizing from sample data to a wider population. A key property of CRT is that inferences are frequently intended to apply at the individual level, whereas randomization occurs at the cluster or group level. For example, to evaluate the effectiveness of a hand hygiene improvement intervention, researchers may choose randomization to occur at the unit level but adherence with hand hygiene recommendations to be assessed for each individual HCW within each cluster.

It is important to correctly identify whether the unit of inference will be at the individual or cluster level early in the planning stage of the trial. If randomization, variable collection, and analysis are all conducted at the cluster level, then sample size estimates and statistical analyses can be done as a standard randomized controlled trial.Reference Donner and Klar ⁶

Principle 5. Matching and/or stratification

Although matching can provide a simple method to consider potential confounders at the design stage, this approach may be overused and effective matching may be especially difficult in smaller studies.Reference Martin, Diehr, Perrin and Koepsell ¹³ Recruiting a large number of pairs provides statistical advantage only if the pairs represent different levels of baseline risk.Reference Donner and Klar ¹ Furthermore, if a single member of a matched pair drops out of the study, this requires that both members of the pair be dropped from the analyses, thereby possibly rendering the study underpowered. Matching in a CRT should therefore be adopted with caution. Stratification is another approach that is commonly used to ensure that there is balance in cluster size per intervention and control groups within strata.Reference Donner and Klar ^{1 ,} Reference Donner and Klar ⁶

Principle 6. Reducing the potential for bias and/or contamination

The goal of randomization is to minimize bias or to ensure that the baseline characteristics of the various clusters are balanced in different intervention groups. When conducting a study in the healthcare setting, the “transmission” of behaviors, attitudes, or knowledge among HCWs who are in regular contact can result in similar responses. This is sometimes referred to as a “herd effect.” Similarly, the Hawthorne effect can be an issue in CRT. Intervention groups may benefit from increased attention and not solely from the intervention itself. To mitigate this, instead of studying only the standard of care in the control group, researchers may consider using a “minimal intervention” or “active controls.” Puffer et alReference Puffer, Torgerson and Watson ¹⁴ found potential recruitment bias in 14 of 36 CRT reviewed (14%). There are several additional ways to reduce the potential for bias and contamination when conducting a CRT in the field of ICHE. For example, the study can be implemented in areas where clusters are distinct and well separated, and control-group clusters can be used that are external to the experimental trial; randomizing different locations within a hospital to control and intervention groups may be problematic. If a crossover design is used, it may be appropriate to employ multiple crossover periods and a wash-out period that is long enough to ensure that there are no residual effects. Furthermore, employing the CRT with crossover design is only appropriate if there is no carryover, which is rare in ICHE.

Principle 7. Accounting for clustering in the analysis

The lack of independence among individual patients or HCWs in the same cluster, creates special methodological challenges. If between-cluster variation is not taken into account, a false claim of statistical significance may result via an increase in the probability of a type I error. Therefore, a main concern in CRT is internal validity. Many CRT fail to account for between-cluster variation at both the design and analysis stage. The aforementioned review by Simpson et al of primary prevention trials showed that only 12 (57%) accounted for clustering in their analyses.Reference Simpson, Klar and Donnor ⁵

To obtain unbiased estimates of the effect of the intervention, analyses must be based on data from all cluster members or must be based on a random subsample of cluster members. It is necessary to decide whether to model the predictor variables as either fixed or random.Reference Fitzmaurice, Laird and Ware ¹⁵ In many CRT, the cluster effect is modeled as random and the intervention effect is modeled as fixed. Several different approaches can be used to ensure that all comparative analyses allow for the clustered nature of the data and that correct confidence intervals and type I error rates are calculated. For example, a generalized estimating equation (GEE) can accommodate cluster-level and individual-level covariates. Similarly, proportional-hazards models with shared frailties can account for clustering within hospitals.

Systematic review findings

Of 33 studies, 29 (87.9%) accounted for clustering in their analyses. Only those in which the level of inference was the individual were included in the denominator of this calculation. Most of these studies used mixed-effects regression models or fixed-effects regression models to account for clustering. Rupp et al (2008) explain how they accounted for clustering at the analysis stage as follows: “… GEE were used to analyze hand hygiene adherence rates over time and their relationship to job category and hand gel availability, appropriately accounting for the potential correlation among observations” (Appendix 2 online).