gms | German Medical Science

Background: Stepped wedge trials are cluster randomized trials (SW-CRTs) in which clusters are followed longitudinally, with different clusters randomized to cross over from the control to the active intervention condition at pre-specified staggered time points. This approach mirrors the natural process of implementing an intervention over many sites. According to current literature, SW-CRTs often do not reach their pre-planned target sample size due to several centre-related barriers to patient recruitment (e.g. lack of resources, lack of motivation to recruit on control condition, cluster-dropout) [1].

Methods: Given the substantial variation in the design of SW trials, we focus on a SW-CRT with a binary primary efficacy outcome, with participants being recruited continuously over time who are exposed rather briefly (continuous recruitment short exposure design [2]), and no “contamination” of treatments.

We illustrate the statistical challenges using as example an ongoing German multicentre SW-CRT with a quality improvement intervention in the ICU setting [3]. During three predefined steps, 12 hospitals were randomly selected to move in a one-way crossover from the control to the intervention condition. After a multifactorial training programme on quality indicators (QIs) on the level of the staff personnel, ICUs received an adapted, interprofessional protocol for a multi-component intervention comprising of daily tele-medical rounds at ICU. The primary objective was to evaluate the effectiveness of a telehealth program on the adherence to several QIs daily measured on the patient-level during the patient's ICU stay, compared to usual care. The target sample size was 1431 patients based on a rather naive approach.

As principal analysis, we consider a generalized linear mixed-effects model for binary outcomes that allows the use of continuous time parameterizations and also adjusts for a secular time trend and exposure time. This approach could also accommodate a non-linear (smooth or polynomial) time effect or a within-cluster correlation structure where the correlation between a pair of subjects' outcomes in a cluster depends on subjects' measurement times, e.g. decaying as the distance between measurement times increases.

Results: To reach the recruitment goal in our trial, the recruitment duration was extended from 12 to 19 months after the second step, while postponing the pre-specified last crossover date by three months to further enhance the number of patients treated on control condition (extension of the rollout period) and lengthening of the post-rollout period. Therefore, several problems with the adherence to the randomization schedule occurred during the course of the trial because some units implemented the intervention later than their allocated step (“late adopters”). Additionally, several imbalances between clusters emerged (large variability in cluster sizes).

We perform a simulation study on allocation characteristics to explore the statistical impact of cluster-level departures from the randomisation schedule on the attained power and possible bias regarding the effect of the intervention.

Conclusion: Simulation studies are essential tools in understanding the impact of protocol deviations in a complex SW-CRT. Results provided by these simulation studies enable to strengthen unbiased analyses in SW-CRTs. Structured reporting on how they were performed is necessary to make their influence on the principal analysis transparent.

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.