Moving Students: How Transfer Students Affect Power and Type I Error in Stepped Wedge Designs
64th ISI World Statistics Congress - Ottawa, Canada
Format: CPS Abstract
Keywords: cluster-randomized, contamination, stepped-wedge
Session: CPS 40 - Aspects of official statistics III
Tuesday 18 July 8:30 a.m. - 9:40 a.m. (Canada/Eastern)
The stepped wedge cluster randomized trial (SW-CRT) design begins with all clusters in the control arm and randomizes the time at which each cluster switches to the intervention arm. Each cluster ends the trial in the treatment arm. The cross-sectional version of the SW-CRT design enrolls a new set of participants at each step of the trial. For example, a cross-sectional SW-CRT studying students in a particular grade level would have a new group of participating students in each year of the study, so the step length for the study would be one year. The SW-CRT design is a popular trial design choice for educational settings because it allows all participating schools to receive the intervention during the course of the study.
Research in schools can be challenging for a variety of reasons. One understudied area is the impact of student transfer between schools, called school mobility. School mobility is prevalent in the United States, with one-third of fourth graders in the US having changed schools in the past two years and rates much higher in some student sub-populations. How this may affect the viability of research in schools is a concern for public health and education researchers conducting randomized trials in schools, including those who employ the SW-CRT design.
This contribution examines how misspecification and contamination caused by student transfer between schools under intervention and control can affect stepped wedge cluster randomized trials. Simulation studies were used to assess the impacts of changes in school mobility rate on power and Type I error. Data were generated using a novel cross-sectional SW-CRT model that allowed for random transfer between clusters. The data were then modeled using a linear mixed model naïve to any transfers and contamination. Across a variety of ICC levels, power decreased linearly and Type I error remained stable as school mobility rate increased. This presentation will address design choices that can mitigate this loss in power.