Population-based cancer survival at small area level: methodological developments

Cancer survival is a key indicator of the overall effectiveness of a health system in man-aging treatment and care of cancer patients. At the national level, cancer survival statistics facilitate overall surveillance of strategic importance. At the local level, they provide valuable insights into the performance of local cancer services essential for public health planning. There are however recurring concerns regarding both the estimation and dis-semination of such survival outcomes, in particular at the smaller area level. The research presented in this thesis aimed to address some of these concerns. A summary indicator of cancer survival, named Index of Cancer Survival, is proposed for all cancers combined designed to act as an overall measure of the effectiveness of cancer services in England at both national and local level. To estimate the index a two-step analytical approach was implemented, in which cancer survival is first estimated for each small area separately followed by a joint smoothing and mapping technique to filter out excessive variation from the resulting cancer survival maps. Such smoothed maps were thought suitable for national health policy-makers to devise national surveillance strategies, as they display in a simple way, the overall patterns of survival for the whole country. Funnel plots were then extended to visualise the spread of individual small-area cancer survival outcomes, mostly thought suitable for local health managers as a tool for monitoring the performance of survival outcomes in their local areas. However, the estimation of cancer survival for small-health geographies remained challenging. The last part of this thesis explored how Bayesian approaches could be used to improve the estimation of cancer survival in the presence of sparse data, and when using more com-plex data structures, including spatially arranged and hierarchical data. The feasibility of an existing Bayesian model for the excess hazard using Poisson regression was explored to estimate small-area patterns in cancer survival accounting for the spatial structure of thedata. A Bayesian flexible excess hazard regression model was then proposed based on the full likelihood specification to improve the modelling of both the baseline excess hazard and the smooth effect of continuous covariates using a special type of splines. The new model also accommodates hierarchical data allowing more complex cancer data structures to be modelled, such as patient level data nested within area of residence or hospital of care level data. In summary, the cancer survival index and both data visualisation techniques for cancer survival greatly improved the interpretability and dissemination of such outcomes for non-technical audiences, in particular health policy-makers. Meanwhile, the Bayesian excess hazard model using Poisson regression improved the estimation when data were sparse by incorporating the spatial data structure. The Bayesian flexible excess hazard model in particular, enabled a better investigation of inequalities in cancer survival using a range of covariate effects and facilitated the study of more complex cancer data structures.