What Is a Stepped-Wedge Cluster Randomized Trial?

A cluster randomized trial (CRT) is an experimental design where naturally occurring clusters such as hospitals or clinics, rather than individuals themselves, are randomized to intervention or control conditions. CRTs are useful when the intervention directly applies at the group level or when individual randomization is unfeasible. Stepped-wedge CRT (SW-CRT) is a recent variation of CRT; in an SW-CRT, all clusters start from the control condition, and a randomly selected subset of clusters will cross over to receive intervention at prespecified time points. SW-CRTs compare outcomes from individuals exposed to intervention to outcomes from individuals under the control condition, both within clusters across different periods and between clusters in the same period. This feature effectively leverages both the within-cluster and between-cluster comparisons to disentangle the treatment effect from the underlying time effect or secular trend.

SW-CRTs are increasingly common and span across different settings, including primary care clinics, nursing homes, and communities, among others.¹ They are useful for evaluating interventions with perceived benefit, as they ensure complete rollout during the study period. Kullgren et al² implemented an SW-CRT in 8 primary care clinics to investigate whether the Committing to Choose Wisely (CCW) behavioral economic intervention could better engage clinicians and older patients to reduce low-value care.

Use of Method

Description of SW-CRTs

SW-CRTs include 3 sampling structures. A cross-sectional structure involves different individuals seeking health care at different time periods in the same cluster, whereas a closed-cohort structure identifies 1 set of individuals in each cluster at baseline for longitudinal follow-up. An open-cohort structure mixes the 2, allowing for attrition from the original cohort and addition of new individuals as the trial progresses (as in Kullgren et al²).

In a prototypical SW-CRT (Figure, A), there is 1 baseline period and at least 1 cluster randomized to receive intervention in each follow-up period. Each row in Figure, A indicates a treatment sequence, and each column indicates a period. Figure, B shows the design in Kullgren et al² by including an additional implementation period to account for the time required to fully transition each cluster to the intervention phase. This implementation period recognizes that immediate intervention uptake is often unrealistic due to practical constraints. Variations of SW-CRTs exist, such as designs extending the trial phase to include additional maintenance periods and incomplete designs without data collection in certain periods.³

Figure. Stepped-Wedge Cluster Randomized Trial Designs With and Without Implementation Period.

A schematic of stepped-wedge designs with 8 clusters (A), such as 8 clinics in Kullgren et al,² randomized to 1 of 8 sequences (or waves; B). Each white cell indicates a control cluster-period, and each gray cell indicates an intervention cluster-period. The implementation period (in dark blue) can either be excluded from analysis or considered a different state from either control or intervention condition.

Statistical considerations for SW-CRTs are complicated by the need to account for secular trend and the clustering of observations. The secular trend describes the outcome trajectory over time in the absence of intervention and, without adjustment, would confound the estimation of treatment effect since more clusters are in the intervention condition at later periods. Accounting for clustering is a critical aspect in SW-CRTs. For example, in the sample size calculation stage, it is recommended to account for both the within-period and between-period intracluster correlation coefficients (ICCs).⁴ The within-period ICC describes the similarity between 2 observations collected from 2 individuals in the same cluster-period, whereas the between-period ICC describes the similarity between 2 observations collected from 2 individuals in different periods but same cluster.⁵ An additional within-individual ICC—describing similarity between repeated observations collected for the same individual—is needed for cohort designs. Reviews of existing sample size methods and software for SW-CRTs can be found in Li et al⁶ and Ouyang et al.⁴

Why Is a SW-CRT Design Used?

Hemming and Taljaard⁷ provided 4 key considerations in choosing a SW-CRT. First, as all clusters are scheduled to receive the intervention, this design facilitates cluster recruitment when the intervention is perceived as beneficial for stakeholders. Second, staggered rollout allows for more feasible management of resources, as the intervention is introduced in a staggered fashion. Third, it provides a means to conduct a randomized evaluation whenever there is a desire to ultimately roll out the intervention. Finally, this design allows each cluster to serve as its own control and leverages both within-cluster and between-cluster comparisons to potentially gain statistical efficiency over a traditional parallel-arm design.

Limitations

There are potential challenges when pursuing SW-CRTs. First, SW-CRTs can increase the total study duration compared to traditional parallel-arm CRTs and can thus be vulnerable to unanticipated external interruptions. Second, SW-CRTs often require a high data collection burden, although incomplete SW-CRTs have been proposed and can be a promising solution.³ Third, although randomization eliminates systematic bias that could arise if the order of intervention rollout was determined by factors such as readiness to proceed, the benefit of randomization in SW-CRTs with a very small number of clusters may not always be clear. Regardless, SW-CRTs with very few clusters typically require additional statistical adjustments to ensure comparability and valid inference about the treatment effect.^3,8 Fourth, blinding may not always be possible such that individuals and health care professionals in the SW-CRT are aware of the intervention status. The lack of blinding can introduce behavior change or reporting bias and may require the use of more objective outcome assessment methods. Finally, it can be challenging to ensure that all clusters adhere to the implementation schedule. In addition, certain types of interventions require additional time to fully develop and may exhibit a delayed effect.⁹ In the presence of an effect that may grow over exposure time with intervention experience, or a duration-specific treatment effect, oversimplification such as assuming a constant effect can lead to severe bias.^9,10 To this end, it is recommended to carefully think through the plausible treatment effect structure in SW-CRTs—an important caveat that is less relevant in traditional parallel-arm CRTs with baseline randomization and a single data-collection period.

How Has the SW-CRT Design Been Used?

Kullgren et al² evaluated whether the CCW intervention could reduce the use of certain low-value services for diabetes, insomnia or anxiety, and prostate cancer screening among older adults by engaging clinicians and patients. The SW design was chosen because (1) staggered rollout is more logistically feasible and (2) all clinics can eventually receive the intervention. The primary outcome was patient-months with low-value services. The intervention was rolled out to a new clinic each month, and all physicians, nurse practitioners, and physician assistants were invited to participate.

How Should a SW-CRT Design Be Interpreted?

Kullgren et al² demonstrated a statistically significant reduction in low-value care due to CCW (odds ratio, 0.79; 95% CI, 0.65–0.97). In the exploratory analysis, they examined whether treatment effects differed over months since study start and found that the intervention was less effective over calendar time. In secondary analyses, they also evaluated CCW’s impact on specific subpopulations with diabetes, insomnia or anxiety, and prostate cancer. Subgroup analyses did not show statistically significant effects in reducing low-value care. Unreported were results from analyses that could assess whether CCW was more or less effective over exposure time, which is a different time scale and quantifies the duration of the experience with intervention since first adoption.^9,10