Statistical analysis of self-assessed quality of life in cancer clinical trials
The assessment of quality of life as a primary outcome in cancer clinical trials is now almost universal. Such data are necessarily longitudinal and multidimensional, and are often severely unbalanced by missing values or early patient death. However, to date, their reporting in the applied literature has generally used simple descriptive summaries that ignore many of these complexities. Not only can these be misleading, but they generally do not allow firm conclusions to be drawn about a major endpoint. The aim of this thesis is to assess the practical application of recent developments in statistical methodology for the analysis of quality of life data collected using self assessment questionnaires within cancer clinical trials. Its emphasis is on the use of relatively simple and flexible tools that will allow more reliable and powerful inferences to be drawn from the data than is done at present. The principal statistical tools considered are random coefficient and marginal models. It is shown that these can be successfully used for the analysis of continuous, binary and ordinal responses. In particular, they offer a simple approach to the analysis of repeated multivariate outcomes and can be very easily extended to model the complex patterns of response that are often seen in following cancer treatment. In relation to the problem of censored quality of life as a result of patient death, analyses that attempt to combine the survival and quality of life endpoints in a single variable are contrasted with those that consider the two endpoints as a multivariate problem. It is shown how this latter model can provide a summary of the quality of life response conditional on patient survival that with further work should have great application to such quality of life data. Finally, the problem of intermittent missing data is reviewed. The implications of missing data for some of the analyses presented in the thesis are assessed, and two models that attempt to determine the nature of intermittent missing data are developed. It is concluded that the problem of non-ignorable intermittent missing data presents a very challenging area of further research.