Up-to-date prognostic modelling of more detailed population-based cancer data

Clinical prognostic models use information about a patient's characteristics and medical history to produce a prediction. The focus of this thesis is on models for time-to-event outcomes where the interest is in producing long-term survival predictions following a diagnosis of cancer. These types of models can be very useful in clinical practice to understand a patient's likely prognosis and assess the potential benefits of different treatment options.

Fitting these types of models requires patients who were diagnosed many years ago to be included in order to have sufficient long-term follow-up. However, survival outcomes in cancer have been improving. Therefore, including these patients and not taking into account these changes can lead to miscalibrated predictions that under-estimate the survival of newly diagnosed patients.

Although there are established methods available for recalibrating prognostic models using more recent data in external validation studies, these methods are often not applicable at the stage of developing prognostic models. The main aim of this thesis was therefore to develop new statistical methodology to adjust for improvements in survival that could be applied when developing or updating prognostic models in order to produce and maintain well-calibrated survival predictions. The proposed method of temporal recalibration is primarily illustrated using an example of colon cancer survival but can be applied to a range of applications and model types. Since missing covariate data is often an issue, the methodology is also assessed for use with multiple imputation. The application in a competing risks setting is also considered, where temporal recalibration can easily be applied to each of the cause-specific hazard models to produce more up-to-date risk predictions.