The role of treatment crossover adjustment methods in the context of economic evaluation

This thesis investigates the problem of treatment crossover – where patients randomised to the control group of a clinical trial are permitted to cross over onto the experimental treatment at some point during follow-up. Methods commonly used to adjust for treatment crossover within health technology assessments are known to be prone to bias, and these may lead to inconsistent resource allocation decisions. The objective of the thesis is to identify which methods are most appropriate for adjusting for treatment crossover in an economic evaluation context.

If control group patients cross over and benefit from the experimental treatment, an intention to treat analysis will underestimate the “true” survival benefit associated with the new treatment – that is, the benefit that would have been observed had treatment crossover not been allowed. Simple methods for adjusting for crossover, such as excluding or censoring crossover patients, will lead to substantial bias when crossover is associated with prognosis. More complex crossover adjustment methods have been described in the literature and previous research has shown that some of these, such as the Rank Preserving Structural Failure Time Model, perform very well when their key methodological assumptions are satisfied. However, a full comparison of all relevant methods across a range of realistic scenarios – including scenarios where key assumptions are not satisfied – has not previously been undertaken. Approaches for incorporating these methods within an economic evaluation – specifically their use in combination with extrapolation modelling – have also not previously been investigated.

In this thesis I demonstrate the importance of the treatment crossover problem, review and assess relevant crossover adjustment methods, and provide an analysis framework to enable the most appropriate method to be identified on a case-by-case basis. Importantly, it is shown that no single method will be satisfactory in all circumstances. In order to identify the method that is likely to provide least bias, consideration must be given to the crossover mechanism, the available trial data, disease and patient characteristics, and the nature of the treatment effect.