Dynamic Prediction, Mediation and Communication for Survival Outcomes, with applications to Cystic Fibrosis

Patients with chronic diseases and their clinicians want accurate and up-to-date information about risk, prognosis, and survival. The overall aim of this thesis is to advance statistical methods available for providing such information. The methods are motivated by analysis of cystic fibrosis (CF), a genetic life-shortening disease, and illustrated using longitudinal data from the UK CF Registry. First, dynamic models, that update predicted survival probabilities as new measurements become available, are studied. Although machine learning methods are established for prediction problems, they have not been widely used in dynamic survival prediction. Here, the combination of a machine learning ensemble with the landmarking approach is developed. Predictive performance of this method is compared to that of the most commonly-used statistical techniques: joint modelling and landmarking. A simulation study investigates cases where a machine learning ensemble may improve predictive accuracy. This thesis then provides a review of literature on communicating survival predictions, focusing on preferred graphical formats, comprehension by a broad audience, and best practices in survival communication. Based on this literature and semi-structured interviews conducted by qualitative research partners, an online tool was created. This provides life expectancy information sensitively and according to an individual's characteristics. In the final part of the thesis, CF-related diabetes (CFRD), a common comorbidity of CF, and its role in survival are investigated. Using multi-state models, the relationship between CFRD and survival is described. Mechanisms through which CFRD affects survival are explored using two methods that can accommodate longitudinal mediators for survival outcomes. Each method is applied to a stacked dataset, constructed in similar fashion to a landmark dataset, designed to maximally use the longitudinal registry data. A simulation study investigates the sensitivity of these two methods to model misspecification and data availability issues.