Design and analysis of stepped wedge cluster randomised trials

This thesis explores aspects of the design and analysis of stepped wedge cluster randomised trials from a statistical viewpoint. It contains a review of current research practices when the design is used in the field and also explores the existing methodological research into the design. It was found that stepped wedge trials often have few clusters (45% < 10 clusters) and a binary outcome (62%). Following this there are three simulation studies presented that aim to explore the use of binary outcome measures in stepped wedge trials with few clusters. The first simulation study examines the different ways in which data from a stepped wedge cluster randomised trial with repeated cross-sections might be analysed. This study also explores the minimum number of clusters needed for consistent and reliable inference under ideal circumstances: such as equal cluster sizes, a time trend that is truly linear, and an intervention effect that is identical for every cluster. For a stepped wedge cluster randomised with 3 steps we found that randomising less than 6 clusters led to estimation problems for all methods of analysis. The second simulation study compares the existing power and sample size determination method for stepped wedge cluster randomised trials to the statistical power of simulated data using the same assumptions. The aim of this study was to see how well the use of Normal approximations in the case of a binary outcome worked for formula based approaches when few clusters were available for analysis. As these approximations became less appropriate, formula based approaches consistently overestimated stepped wedge trial power. The final simulation study is about stepped wedge cluster randomised trials in which cohorts of participants are repeatedly measured. Both closed cohorts, where the same participants are followed throughout the trial, and open cohorts, where participants may enter into the trial at any point, are considered. The effect on study power of different values for the level two and level three variation for both designs is explored. Given the same number of participants per cluster per time, there was generally little difference between the open cohort and closed cohort design for the correlation values we used.