Stepped wedge cluster randomized trial
Why this level?
Participants or clusters are randomly allocated to concurrent comparison groups. NHMRC 2009 sourceEvery cluster eventually crosses over, and it is the ORDER of crossing that is randomised. That randomisation is what puts it at level II. A rollout whose order was decided operationally is not randomised, whatever the protocol calls it.
What question does this help answer?
It suits an intervention that is going to be implemented everywhere regardless, where the practical question is what it achieves, and where a phased rollout is happening anyway for operational reasons.
When to use it
Use it when withholding the intervention permanently would be unacceptable or impractical, when resources force a staged rollout, and when the rollout ORDER can genuinely be randomized. Every cluster contributes both control and intervention periods, which can be more efficient than a parallel design when clusters are few.
When not to use it
Avoid it when a parallel cluster trial is feasible, because the stepped wedge is confounded with calendar time by construction and buys that complexity for nothing when it is not needed. It is a poor fit when the intervention's effect takes longer to appear than the step length, since periods just after crossover will then understate it.
How it works in plain language
Clusters start in the control condition and switch to the intervention at randomly assigned times until all have switched. Because later periods contain more intervention clusters, anything else changing over calendar time -- a seasonal effect, a concurrent policy, a gradual improvement -- is entangled with the intervention and must be modelled explicitly rather than assumed away.
What the rollout looks like
Why Mallard may recommend it
Mallard may recommend a stepped wedge when the investigator describes rolling an intervention out site by site over months, and when the order of that rollout could be assigned at random.
Key assumptions and design requirements
- Secular time trend is modelled, because the design confounds treatment with calendar time by construction.
- The rollout order is genuinely RANDOMIZED. An operational or administrative order is not randomization, and a plan should say which it had.
- Both within-period and between-period correlation are considered; a single intracluster correlation is usually not enough.
- The design is sized by SIMULATION rather than a closed-form two-group formula.
- Any lag between crossover and full effect is stated, and periods affected by it are handled explicitly.
The usual analysis model
The standard analysis, described by Hussey and Hughes, is a mixed model with a fixed effect for each period to absorb the secular trend, an indicator for whether a cluster was in the intervention condition in that period, and a random intercept for each cluster. That model is written for repeated cross-sectional samples, where different people are measured in each period; in a cohort stepped wedge, where the same participants are measured repeatedly, it also needs a participant-level random effect or an equivalent correlation term, or its standard errors are too small. The intervention effect is then estimated from comparisons between clusters within periods and within clusters over time. Two of its assumptions deserve stating. A single random intercept implies that two measurements from one cluster are equally correlated whether they come from the same period or years apart; when that correlation decays over time, the model's intervals can be too narrow, and a cluster-by-period effect or a decaying correlation structure should be considered. And it assumes the intervention's effect is immediate and constant after crossover; a delayed or growing effect needs a term for time since crossover. With few clusters, which is typical, inference needs a small-sample correction such as Kenward-Roger.
What to watch out for
The design is confounded with time by construction, because every cluster is in the control condition early and the intervention condition later, so any secular trend must be modelled rather than assumed away. An operational or administrative rollout order is NOT randomization, and a design whose sequence was set by an IT schedule should say so.
Clinical example
Question: Does a sepsis early-warning bundle reduce mortality? Design: Nine wards adopt the bundle one at a time over twelve months, with the order randomly assigned, and the analysis compares outcomes under the intervention and control conditions while adjusting for the underlying time trend.
How to report it
CONSORT extension for stepped wedge cluster randomised trials is the guideline for this design. It asks for the sequences clusters were randomised to and when each switched to the intervention, how the analysis accounted for time and for correlation within clusters, and the numbers of clusters and participants in each period. A reporting guideline is not a design manual: use it to make the study transparent, not as a substitute for careful design. Check the target journal's own requirements as well, which this page cannot know.
- The stepped wedge cluster randomised trial: rationale, design, analysis, and reporting (opens in a new tab) · BMJ · doi 10.1136/bmj.h391
- Reporting of stepped wedge cluster randomised trials: extension of the CONSORT 2010 statement with explanation and elaboration (opens in a new tab) · BMJ · doi 10.1136/bmj.k1614
- Design and analysis of stepped wedge cluster randomized trials (opens in a new tab) · Contemp Clin Trials · doi 10.1016/j.cct.2006.05.007