Most analyses in stroke clinical trials assess the primary outcome at a single time point and do not consider conducting a longitudinal analysis, even when the outcome has been recorded at several time points. Efforts to improve the quality of statistical analyses have previously been made, but have been directed at techniques that compare groups at a predetermined fixed time point. The aim of this thesis was to specifically investigate the use of longitudinal techniques, such as generalised linear mixed models and latent variable methods, to model serially correlated stroke outcome data, and to consider the challenge of having a score for death within the scale. The chosen outcome was the modified Rankin scale, which is a functional assessment outcome after stroke and is one of the most popular endpoints in clinical trials; often recorded repeatedly during follow-up.
Initial chapters describe the epidemiology of stroke and various stroke outcomes, with a review of what appears in the literature regarding current longitudinal modelling of mRS data. Subsequent chapters report how proportional-odds models are fitted to longitudinal repeated measures trial data, with the effect estimates compared to an analysis at a single time point. Simulation techniques assess how different patterns of drop-out affect treatment estimates derived using all the data. These are then compared to a Markov analysis that treats death as an absorbing state. Finally, clustering methods are considered in order to try and describe trajectories of the mRS, but it is highlighted that this is challenging given limited movement between states. Each of the methods have their own merits, and trials should be encouraged to consider longitudinal analyses when there is repeated measures data.