The Gold Standard of Experimental Design | Stepped-Wedge Trials, Secular Trends and Time Confounding

1. Define the sequence before the wedge

Back to contents

Specify the clusters, periods, crossover sequences, transition interval and whether individuals are repeatedly followed or newly sampled each period. Randomisation concerns the timing or order of crossover, not usually individual treatment. The estimand must state whose outcome, over what post-transition window and under what implementation state is being compared. A diagram helps reveal where information actually comes from.

2. Recognise the built-in treatment–time association

Back to contents

At the start almost every observation is control; near the end almost every observation is intervention. Even with randomised sequences, exposure is correlated with calendar time by design. If outcomes improve because winter ends, staff gain experience or a national campaign begins, an analysis that ignores period can attribute that change to the intervention. Randomisation does not eliminate this structural problem.

3. Choose clusters and periods around the mechanism

Back to contents

Clusters should minimise contamination while remaining numerous enough for credible randomisation and variance estimation. Period length must allow recruitment and measurement yet distinguish crossover steps. If effects are delayed, transient or cumulative, a simple immediate on/off indicator is inadequate. Plan transition periods, implementation fidelity and exposure duration before power calculations.

4. Randomise sequences and conceal the schedule where possible

Back to contents

Generate allocation to crossover sequences using a reproducible method, stratifying or restricting only for prespecified reasons. Future crossover dates can change behaviour before implementation, so concealment may matter even when blinding is impossible. Document deviations, early adoption and delays. A cluster that crosses late or incompletely should not be silently coded as fully treated from the planned date.

5. Model calendar time explicitly

Back to contents

A basic analysis usually includes period effects alongside the intervention indicator and accounts for clustering. Treating time as linear is fragile when secular trends are seasonal, abrupt or nonlinear. Prespecified categorical period effects are often safer, though they cost degrees of freedom. More complex models may allow cluster-specific trends, treatment-by-time interactions or exposure-time effects, but they require enough clusters and periods.

6. Worked example: eight clinics and a winter trend

Back to contents

Eight clinics cross to an electronic triage tool over four quarterly steps. Waiting times fall by ten minutes across all clinics from winter to summer because respiratory demand declines. A crude before–after comparison credits the full decline to the tool because most summer observations are treated. A mixed model with categorical quarter effects estimates a smaller intervention contrast using treated and control clinics observed in the same periods. Sensitivity analyses then test a cluster-specific trend and exclude transition weeks. The wedge contributes evidence only through contemporaneous comparisons, not the diagonal picture alone.

7. Respect the correlation structure

Back to contents

Outcomes within a cluster are correlated, and repeated measures on the same people add another layer. Correlation may weaken as observations are farther apart in time. A single random intercept can understate uncertainty when the true structure includes cluster-period shocks or individual follow-up. Use design-informed correlation models, small-sample corrections where appropriate and simulation-based checks during planning.

8. Changed-context transfer check: a school curriculum rollout

Back to contents

A district rolls out a new curriculum to schools in random order. Examination scores also rise as teachers become familiar with a new assessment framework. The transferable rule is to compare schools within the same terms and adjust for period. The context adds cohort turnover, school holidays and anticipatory preparation. If teachers train before the official crossover, exposure must reflect actual implementation rather than the published timetable.

9. Know when the wedge is a poor choice

Back to contents

Stepped wedges are weak when the intervention effect is urgent, harmful exposure cannot be withdrawn, few clusters are available, secular trends are unpredictable or implementation dates cannot be maintained. Universal eventual rollout may be politically attractive, but it is not an ethical or statistical guarantee. A parallel cluster trial can be simpler, shorter and less model-dependent. Choose the wedge for a justified reason, not because phased rollout looks inevitable.

10. Audit design, implementation and analysis together

Back to contents

Publish the sequence schedule, randomisation code, cluster eligibility, recruitment timing, transition rules, deviations, fidelity measures and a time-stamped analysis plan. Plot outcomes by cluster and period before fitting models. Compare categorical and plausible nonlinear time effects, inspect cluster-specific trends, vary transition and exposure-lag assumptions and report how much identification comes from each period.

11. Frequently asked questions

Back to contents

Does every cluster receive the intervention? Usually yes by design, but that does not guarantee unbiased estimation. Is a before–after analysis acceptable? No; it ignores secular time and randomised contemporaneous contrasts. Should transition periods count as treated? Only according to a prespecified exposure model. Can individual randomisation be added? Sometimes, but it changes the estimand and design rather than repairing a weak wedge automatically.

12. The gold standard separates rollout from causation

Back to contents

A stepped wedge is a randomised rollout schedule, not a licence to compare early control with late intervention data. Credible inference comes from prespecified sequences, contemporaneous comparisons, explicit period effects, faithful exposure measurement and robust correlation modelling. The gold standard asks whether the treatment effect survives reasonable descriptions of time; if it does not, the rollout has been observed but causation has not been isolated.