Design and analysis of crossover clinical trials with non-normally distributed outcome variables

In a crossover trial, participants are randomised to receive a number of treatments across several time periods. For example, in a typical trial using the “AB/BA design”, half of the participants would receive an experimental treatment first, and subsequently receive a placebo, with the remainder first given the placebo, and then the experimental treatment. The main advantage of this approach is that it allows participants to act as their own control, which makes crossover trials often far more efficient than parallel group trials. Consequently, whilst crossover trials do have their limitations, they are the design of choice for trials in a host of disease areas, in particular many chronic conditions, such as asthma, arthritis, and hypertension. In fact, they play an important role in an array of research fields, including psychology, manufacturing, education and veterinary science.
Given their use is so frequent, it is important that we have a range of statistical methods available for crossover trial design and analysis. When the outcome data are normally distributed, this methodology is highly refined and generally well known, as summarised by Senn (1), and Jones and Kenward (2).
However, for non-normally distributed outcome variables many methodological challenges remain. This project will attempt to explore some of these issues. Options for exploration include investigating the optimal approach for analysis when outcome data is binary or categorical in nature, determining optimal sequences for treatment allocation, and developing methods for handling missing data.
The MRC-BSU also has expertise in adaptive trial design, which are increasingly of interest in the trials community. Examination of adaptive designs for crossover trials with non-normally distributed endpoints, including how to re-evaluate the trials required sample size, how to drop ineffective treatment arms, and how to response adaptively allocate treatments, could all be incorporated in to the project.
Such work is of great importance, as non-normally distributed outcomes arise regularly in crossover studies. For example, in arthritis trials, ordinal pain scores are often the endpoint of interest. Similarly, in hypertension research, a binary indicator of whether blood pressure was controlled by a treatment is frequently of most concern.
Overall, the focus of the project will be statistical. However, there is scope if a student is interested to perform a thorough literature review, or to focus more on computationally intensive procedures. The project supervisors are experienced in a range of programming languages, and will be happy to provide training as required. The project will suit someone who wants to determine their own research interests, who wants to develop novel methodology to help improve the efficiency of clinical research, and who is interested in efficient programming techniques.