Medical Statistics

Previous and current research
An important role of our group is to give statistical support to others in the organisation. Much of the work of the group is thus carried out together with other Cancer Research UK groups, with a strong emphasis on clinical rather than laboratory research. Our own statistical research programme is influenced by the types of problem that arise in those collaborative projects. Because many of these relate to studies of patient prognosis, we have consistently been interested in aspects of survival analysis, including both non-technical and methodological work. For example, we have carried out a review of the use and presentation of survival analysis in clinical cancer journals and produced guidelines for presentation, and have carried out an in depth examination of the practical problems in using Cox regression models for data with updated patient information ("time-dependent" models). We have taken considerable interest in the difficulties associated with the use of continuous variables (such as tumour marker levels) in regression models, and have studied, and heavily criticised, the wide use in oncology of "optimal cut-points" to create prognostic groups. Other research related to continuous variables has led to the development of a new family of models - fractional polynomials - to allow greater flexibility in modelling curved relationships. We have also carried out simulation studies to compare various categorisation strategies in logistic regression models. Another area of special interest is measurement, including studies of measurement error (and its consequences) and method comparison studies.

Future projects
Several of the themes mentioned require further development. Production of prognostic models generates a large literature in all types of cancer. Yet many of these studies are too small, with poor study design and analysis. Areas needing further research include standard methods for investigating the value of new tumour markers, and a strategy for developing reliable models for patient prognosis. Further, there is a clear need to carry out meta-analysis of prognostic factor studies, and we wish to investigate methods for doing this. We will continue investigations of appropriate methods for handling continuous variables and in particular the creation of prognostic groups for clinical use. Future research of the group will also include more general aspects of meta-analysis and further reviews of the use of statistical methods in general.