Register for: WIMSIG2017

Details of talk

TitleInformation content of cluster-period cells in stepped wedge trials
PresenterJessica Kasza (Monash University)
Author(s)Jessica Kasza, Andrew Forbes
SessionBiostatistics and Bioinformatics
Time11:30:00 2017-09-26

Stepped wedge trials are a particular variant of multiple-period cluster
randomised trials, being used with increasing frequency in health research.
Instead of being randomised to one of (usually) two treatments as in parallel
cluster randomised trials, in SWs the clusters are randomised to particular
sequences of treatments. In a standard stepped wedge trial, all clusters
initially implement the control condition before eventually implementing the new
intervention, with an equal number of clusters switching to the new intervention
at regularly-spaced time points. These trials are often expensive to implement,
and require that data be collected from each of $K$ clusters at $T$ time
periods. To reduce the costs associated with the implementation of the trial it
may be desirable to restrict the total number of cluster-period cells in which
data is collected. Although some designs with missing cells have been proposed,
such as the so-called dog-leg design, there are no general guidelines regarding
which cells of a stepped wedge design contain the most information about the
treatment effect, and correspondingly which may be omitted with a minimum loss
of information. 

We consider the information content of the cluster-period cells and the
clusters of the standard stepped wedge design, where the information content of
a cell is quantified by the increase in the variance of the treatment effect
when that cell is omitted. The variance of the treatment effect will never
decrease whenever a cell is omitted, and the smaller the increase in variance,
the less information about the treatment effect a cell is said to contain. We
show that the most information-rich cells are those that occur immediately
before or after treatment switches, but there are also additional cells that
contribute almost as much to the estimation of the treatment effect. We also
show that the information content patterns depend on this correlation structure.