Hospitalists are often involved in quality and research efforts to improve care delivery in the inpatient setting.1 To assess the effectiveness of these efforts, new healthcare delivery initiatives may be evaluated by comparing clinical outcomes between patients who received the intervention and those who received usual care, which is commonly used as a control. Healthcare redesign and improvements often require complex interventions involving multiple components; therefore, simple study designs may not be appropriate.2 For example, in 2021, Schnipper et al.3 implemented a multifaceted intervention to improve the transition of care from hospital to ambulatory settings. This intervention had multiple components, including pharmacist-led medication reconciliation, patient education and coaching, post-discharge follow-up, and health information technology (HIT) enhancements. A traditional randomized controlled trial (RCT) at the individual level would have been impractical or impossible because some of the components required implementation at the hospital level (e.g., hospital-wide implementation of HIT) or primary care practice (e.g., education and coaching for all patients in a practice to avoid contamination). In situations when randomization at the individual level is impractical or not possible, randomization can be performed at the level of a group of individuals (cluster).2 Schnipper et al.3 used a stepped-wedge cluster randomized trial (SW-CRT) design, which is the primary focus of this methodological progress note. SW designs are becoming increasingly popular in hospital-based research studies, and hospitalists should be prepared to understand, participate in, and even lead large multisite CRTs. CRTs have become increasingly common in health services research as they are useful for evaluating interventions that represent new standards of care or new guidelines.4 The primary difference between traditional RCTs and CRTs is that, in CRTs, patients are randomized in groups or clusters rather than individually.5 Clusters are generally social or administrative units, which can range in size from small units, such as households or hospital floors, to larger units, such as clinics or entire healthcare systems.2 Although traditional RCTs are statistically more straightforward,6 there are many reasons why a CRT may be either more suitable than a traditional RCT or the only practical option for testing the clinical outcomes of an intervention (Table 1). One common reason is that some interventions can only be implemented at the system or cluster level, making individual randomization impractical.2, 6 For example, in the study discussed in the introduction, "health technology enhancements" may have to be implemented at the level of the hospital, eliminating the ability to randomize individuals in a single hospital to different technologies3 (see Table 1 for an additional example). CRT designs can also reduce the risk of contamination or the risk of the intervention "bleeding into" usual care.2, 6, 11 This is because the entire cohort is either implementing the intervention or continuing with the usual care. In the study by Schnipper et al.,3 contamination may occur when the same clinician is treating both an intervention and a control participant, allowing the "treatment" of one to influence the other. If a provider believes education and coaching are helpful, they may provide the same to a usual care participant. In a cluster trial, the entire hospital, clinic, or clinician is receiving the intervention, so a site can consistently treat all participants the same. One advantage of some CRTs compared to RCTs is the potential for waiving individual informed consent, particularly when the intervention is evidence-based with clear benefits and implemented as the new standard of care at the cluster level.2 As waiver of consent is based on the level of interaction with patients, the risk to patients, and the type of data collected, it is governed by local, national, and international ethical guidelines. Therefore, consultation with the local Institutional Review Board is essential. In a study by Oakeshott et al.,12 the research team aimed to assess the impact of providing laminated radiology guidelines to general practitioners on the rate of X-ray referrals. Because their primary outcome was the percentage of X-ray referrals received by the radiology department that conformed with the guidelines, the study did not require consent from individual patients or participating clusters. Simple cluster randomized design: Each cluster is randomized to either the intervention or control group, and the outcomes are compared between the two study arms. The advantage of this cluster design is its relative simplicity and potential ease of implementation. One disadvantage is that only half the clusters receive the intervention, which may not be well received if the intervention is likely to have a benefit.6, 13 Cluster with crossover randomized design: In this design, each cluster is randomized to either the intervention or control group, similar to the simple cluster design. However, after a certain period of time (usually halfway), the clusters switch or cross over to the other study arm. For this crossover design to work, the intervention must be able to be "turned on/off" so it does not carry over from the pre-crossover period into the post-crossover period (contamination across arms). For example, if the intervention leads to a permanent change (e.g., a vaccination), then a crossover study cannot be conducted as the vaccination cannot be undone. When a crossover trial is appropriate, a washout period after the crossover is often used to minimize contamination. When the intervention cannot be easily turned on/off, a cluster with a partial crossover design can also be used.6 A partial crossover occurs when only part of the clusters crossover to the other study arm (usually from usual care to implementation), while the remaining clusters stay in the usual care study arm for the entire duration of the study.6, 13 SW design: Can be considered a special type of cluster with crossover design5 and it is discussed in detail in the section below. In a SW-CRT design, the clusters are randomized into a few groups or waves that determine when each group will begin the intervention.4 All clusters begin the trial in the control arm and each cluster or wave of clusters cross over to the intervention on a staggered randomized schedule (steps). Once a cluster crosses over to the intervention, the intervention remains in place until the end of the study (Figure 1).5, 6 One advantage of the SW-CRT design is that all clusters eventually receive the intervention.4, 5, 15 This is particularly important if the intervention is believed or proven to be beneficial based on prior effectiveness research. This may allow the SW-CRT to focus on implementation outcomes in addition to effectiveness outcomes.16 Each cluster receiving the intervention may also help with stakeholder buy-in. Spacing the interventions over time allows for control of external temporal trends (secular trends).6 Since each cluster has pre- and postimplementation outcomes, one can make between and within-cluster comparisons. Therefore, compared to a single CRT, SW-CRTs tend to have increased statistical power, necessitating fewer clusters.15 SW-CRT designs are particularly suitable for situations in which it is not possible to roll out the intervention to all clusters at the same time. Reasons for rolling the intervention in steps may be related to financial, logistical, or other practical constraints.15 For example, the Non-pharmacological Options in Postoperative Hospital-Based and Rehabilitation Pain Management (NOHARM) trial assessed the effectiveness of an intervention bundle that included health technology clinical decision support and a portal-based conversation guide. The study had 22 clusters, randomized into five staggered steps, with the SW design advantageous to reduce the resourcing burdens associated with this large-scale implementation (e.g., staffing and training).14 Therefore, a small group of staff could concentrate their implementation efforts on one cluster at a time. Even though there are many advantages to using SW-CRTs, they are not always the best choice, so thoughtful consideration of all aspects of the design should be a priority. For example, informed consent can pose challenges for the SW-CRT designs because, in the United States, consent regulations were created with individual-level randomization in mind.4 In a SW-CRT, the randomization occurs at the cluster level rather than at the individual level, complicating the process of identifying research participants. Sometimes, a SW-CRT has both cluster-level participants (e.g., providers) and individual-level participants (e.g., patients). Informed consent may be required from clusters, individuals, or both.2 When individual consent is required for the completion of a SW-CRT, it can significantly hamper study recruitment (and, therefore, generalizability) and it can also introduce selection bias.15 To prevent selection bias one would have to enroll all members of a cluster or a randomly selected sample.11 Alternatives to traditional informed consent are waiver of consent or waiver of documentation of consent.4, 15 Finally, SW-CRTs are more likely to experience contamination compared to simple CRTs due to the complex coordination of care delivery involved. It can be difficult to fully control contamination both within and between clusters.15 The statistical design and analysis of CRTs, including SW-CRTs, can be complex with many special considerations that need to be considered (e.g., cluster size, variability within and between clusters, risk of co-interventions, or major temporal trends). Having an experienced statistician involved in every phase of the study is extremely valuable and often necessary.6 In general, CRTs require larger individual participant sample sizes compared to traditional RCTs to account for the correlation between cluster members. Members of the same cluster are more likely to have similar outcomes as compared to a randomly selected sample from the population of interest (e.g., all members of the cluster treated by the same provider or living in the same region with similar socioeconomic characteristics).2, 11 This tendency is known as "within-cluster homogeneity" and must be considered in the design and analysis of a CRT.2 Another general consideration when designing CRTs is trying to maximize the number of clusters that are being randomized. It is extremely important to select the lowest level cluster possible while keeping the risk of contamination at a minimum. For example, investigators may be more likely to randomize at the clinic level; however, to increase the number of clusters in the study and, therefore, improve statistical power, they may be able to randomize at the provider level.6 Finally, in healthcare CRTs the cluster size is often variable between clusters (e.g., different size hospitals). Having unequal cluster sizes will add additional statistical constraints as it decreases the power of the study as compared to a more balanced cluster design.6 CRT designs, including the SW design, can be great alternatives to traditional RCTs in health services research; however, they are not always the best choice. When considering the most suitable design for an intervention, investigators should consider the advantages and challenges of each design and provide a clear justification that supports their final choice. The SW design may be preferred over other CRT designs when all participants could benefit from the intervention; there is a need for buy-in from stakeholders when it is not possible to roll out the intervention to all clusters at the same time, and when trying to control for secular trends. Amanda Ullman's employer (UQ) has received investigator-initiated research grants from vascular access product manufacturers (3M, Becton Dickinson, and Medline) on behalf of her research, unrelated to the submitted project. David C. Brousseau receives consultancy funding from CSL Behring unrelated to this work. The remaining authors declare no conflicts of interest.
