Improving Statistical Methods To Understand Differences In Cancer Survival

Cancer survival varies substantially across population groups. For instance, there are differences across socioeconomic groups that persist irrespective of how deprivation is defined. The underlying determinants are not well understood as they are driven by complex mechanisms. Identifying drivers of variation is important and can lead to targeted interventions to improve survival. This thesis involves the development and application of statistical methods to understand and report population variation; largely focussing on cancer-related differences through a relative survival setting. The developed methodology is applied using English registry data for several cancer types. Differences in all-cause survival arise from both cancer-related and other-cause factors: the advantage of using the relative survival framework is the possibility, under assumptions, to isolate differences due to cancer-related factors. There have been past examples of exploring differences across population groups, and this thesis sets those approaches into an appropriate causal framework. Causal inference and mediation analysis are extended to the relative survival framework and marginal measures of interest are defined. Contrasts between subgroups in terms of net and all-cause measures are introduced and shown to be identifiable under assumptions. Mediation analysis allows the possibility to delve deeper into observed differences and explore the role of intermediary explanatory factors. The potential impact of removing differences is explored and quantified as the number of avoidable deaths under hypothetical interventions. Marginal estimates are obtained using regression standardisation, inverse probability weighting, or doubly robust standardisation. Methodology that allows excess mortality to be partitioned into components due to specific non-cancer causes is also provided. Finally, additional reporting measures such as loss in life expectancy are utilised to help understand the lifetime impact of a cancer diagnosis. The extensions proposed in this thesis, and the focus on a broad range of intuitive metrics, could have wide-ranging impact in cancer (and other disease) epidemiology.