Stepped-Wedge Design

A cluster-randomised design where all clusters begin in the control condition and then cross over to the intervention one step (time period) at a time, in a randomly allocated sequence, until every cluster is intervened. Each cluster acts as its own control before and after it crosses over.

A one-directional staircase: clusters move from control to intervention step by step in a randomised sequence, never switching back. Early steps contribute mostly control data, late steps mostly intervention data — so intervention exposure is tied to calendar time.

Each cluster contributes both control and intervention person-time, so comparisons are made both between clusters at a given step and within each cluster across its crossover. Because intervention periods cluster at the later (right-hand) steps, any underlying trend over time must be modelled out before the intervention effect can be trusted.

It is ethical and pragmatic when a rollout is going to happen anyway, or when staff/resources can only be deployed in stages — no cluster is denied an intervention believed to help. The price is that the design is confounded by time: secular trends masquerade as treatment effect unless explicitly adjusted for.

• Randomise the order/timing of crossover, not whether a cluster is treated
• Every cluster is its own control → intervention tied to calendar time
• Analysis must adjust for time period (secular trend)

The Gambia Hepatitis Intervention Study — the design that popularised the stepped wedge. Hepatitis B vaccine was rolled out to infants across the country in a randomised stepwise order through the existing immunisation programme, rather than withholding it from any region, because the vaccine was already believed to work. Every team eventually delivered it — randomisation governed only when each step started. Classic ED-relevant uses are implementation trials — rolling out a sepsis bundle, an early-warning score or a handover checklist hospital-by-hospital. Because uptake and outcomes drift over the rollout months, the secular-trend adjustment is what protects the result.

• Forgetting that time is confounded with the intervention — no time-period term invalidates the effect estimate.
• Underestimating the design and analysis complexity (correlated, clustered, time-varying data — not a simple before/after).
• Ignoring lag / transition effects — an intervention may take weeks to bed in, blurring the step.