Article
Robustness in open cohort stepped wedge cluster-randomized studies with binary response
Search Medline for
Authors
-
Jochem König
- Institut für Medizinische Biometrie, Epidemiologie und Informatik, Universitätsmedizin der Johannes Gutenberg Universität Mainz, Mainz, Germany
-
Philipp Mildenberger
- Universitätsmedizin der Johannes Gutenberg-Universität Mainz, Institut für Medizinische Biometrie, Epidemiologie und Informatik, Mainz, Germany
| Published: | September 6, 2019 |
|---|
Outline
Text
In studies on complex interventions in health services research, patients are frequently randomized in clusters, e.g. practices or hospitals, in order to preclude contamination of treatment groups. If it is felt desirable to introduce the intervention eventually in all participating units, if the simultaneous implementation is deemed difficult, or if the number of clusters is limited, the stepped wedge cluster randomized studies (SWCRT) may be preferred to the parallel group design. Unfortunately, this comes with the expense of more model complexity, in particular if patients are repeatedly observed. For study planning, a simplified model is frequently assumed (see e.g. [1], fixed effects for treatment and period and time-constant treatment specific random cluster-level effects, or Li et al. [2] who consider a nested exchangeable covariance structure with three correlation parameters depending on whether observations shared a cluster, a period or a patient.) The marginal modelling approach is robust to miss-specification of covariance structures. However it is not clear how precision of treatment effect estimates is affected.
We therefore investigated, for the setting of an ongoing study that evaluates specialized care for depressed patients in ten nursing homes, which type of marginal model analysis preserves type I error, and how relaxing of model assumptions affects precision and power. We consider sensitivity to change in correlation parameters, allowance for time-varying treatment effects, correlation structures induced if random cluster and treatment effects may change by period, and if patient level correlations are of autoregressive type. Furthermore, we have investigated the issues of time-varying treatment effects, unbalanced cluster sample sizes, and delayed response to intervention on patient level.
Fairly adjusted marginal model analyses seem to maintain validity over a range of deviations in covariance structures, and also with moderate delayed entry and attrition. Efficiency starts being compromised with unbalanced samples and more with delayed response to treatment at patient level.
The authors declare that they have no competing interests.
The authors declare that an ethics committee vote is not required.
