What is a stepped-wedge cluster randomised trial?

A stepped wedge cluster randomised trial (SW-CRT) is a type of CRT where clusters are randomised to sequences. All or most clusters begin the trial receiving the control condition. During the trial, each sequence switches to the intervention condition at a different time so that, by the end of the trial, all or most of the clusters are receiving the intervention condition. No clusters switch from intervention to control at any time.

A schematic of a typical SW-CRT (in this case with eight clusters, four sequences, and five steps) looks like the example below.

A related design is the cluster crossover trial. SW-CRTs are a modification of the crossover design. In a cluster crossover trial, clusters are randomised to two sequences. Most cluster crossover trials have two periods only. In one sequence, clusters receive the control condition followed by the intervention. In the other sequence, clusters receive the intervention condition followed by the control. Thus, SW-CRTs are essentially a one-way crossover trial. On this website, we focus on parallel CRTs and SW-CRT, but have included some references for cluster crossover trials in the SW-CRT resources pages.

Cluster and SW-CRT overviews

The following are well-known textbooks and journal articles on cluster and SW-CRTs:

Example SW-CRTs

Like with parallel CRTs, SW-CRTs are used in a range of settings, as demonstrated by the examples below:

SW-CRT examples

Health treatment and prevention programs

Vaccination programs

Behavioural interventions

School-based interventions

Why do a SW-CRT?

SW-CRTs have become popular in the last few years. The reasons for randomising clusters are the same as described in why do a CRT? The section below gives an overview of the reasons for performing a SW-CRT. There is disagreement about the validity of some reasons given, as discussed below:

  • Logistical reasons: Some researchers say that it is easier to implement the intervention in an SW-CRT because the intervention is only rolled out to a few clusters at one time. However, some trialists have found it difficult to manage the growing burden of maintaining the intervention in all clusters towards the end of the trial. A parallel CRT only requires maintenance of the intervention in half of all clusters, and can also incorporate a staggered roll out if necessary. In fact, in health services research staggered rollouts in cluster randomised trials are extremely common.

  • Recruitment of clusters: Researchers have found it easier to recruit clusters when they are giving the intervention to all clusters. Other designs can also fulfil this need, for example a waitlist design where control clusters are given the intervention after the trial has ended.

  • Ethics: Some researchers feel like it is a more ethical design when the intervention is expected to be beneficial. However, some have disputed the ethics of with-holding the intervention from clusters randomised to receive the intervention later in such situations.

  • Utilising the rollout of an intervention that is happening anyway: Some SW-CRTs are done as a means of evaluating an intervention as it is being implemented.

For more on the debate about why to do a SW-CRT, see the following papers:

Why do a SW-CRT?

Correlation in SW-CRTs

In SW-CRTs observations are not independent of one another. There are two types of correlation to account for in the design and analysis:

  1. Intracluster correlation: As with CRTs, observations in the same cluster will be more similar to one another than observations in different clusters. In a SW-CRT, this is the correlation of observations in the same cluster at the same time.

  2. Correlations over time within clusters: Unlike a CRT, in a SW-CRT we can get information about the effect of the intervention by comparing observations from the periods that a cluster is in the control condition to observations from periods that the cluster is in the intervention condition. Because of this, we need to account for how similar observations are over time within each cluster.

    In general, we would expect observations collected in the same period to be more similar (more correlated) than observations collected in different periods. The correlation between observations in the same cluster but from different periods is sometimes called the inter-period correlation.

    Designs with more than two periods may have different inter-period correlations for different pairs of periods. An example of this is autocorrelated data, where observations from periods closer in time are more correlated than observations from periods further apart in time.

Types of design

The schematic above shows a typical SW-CRT, but there are many variations of this design depending on the number of sequences, clusters, periods, and whether different individuals are observed in each period (a repeated cross-section design), or the same individuals are observed over time (a cohort design). Some designs do not collect observations from all clusters in all periods, for example some trials include transition periods, during which the intervention is embedded into the cluster and the cluster is considered neither exposed nor unexposed. See our resources for designing SW-CRTs for more on this.