Cancer survival is a key metric for monitoring improvement in awareness, early diagnosis and access to effective treatments for cancer patients. For the majority of cancers, survival has been increasing for a number of decades, as a result of successful health policies and the availability of more effective treatment. Nevertheless, there is an unavoidable delay between policy implementation and impact. In parallel, the measure of survival requires follow-up information, adding to the delay in quantifying health benefits. Predictions of cancer survival for cohorts of patients most recently diagnosed could help fill the gap in our knowledge of the likely effects of cancer policies. In this thesis, I modelled the excess hazard of death as a function of predictors available in linked cancer registry data in the UK. These include age, stage and year of diagnosis, levels of deprivation, type of diagnosis, and access to curative treatment. In such contexts, selecting the form of the model, the predictors, the shape of their effects, and potential interactive effects is challenging. Several model selection strategies are compared and their performance assessed in simulations. I provide practical guidelines for the modelling of the excess hazard of death, in particular in relation to cancer lethality, model complexity and impact of model mis-specification. Besides, these multi-variable regression models offer opportunities for predicting cancer-related death rate, for cohorts of patients most recently diagnosed, and for whom follow-up is not yet available. Along with model selection algorithms, I explore strategies based on information criteria and model averaging. Inference is therefore conditional on a pool of models of equivalent support, rather than a uniquely selected model. Advantages include absence of multiple testing, and allowance for model selection uncertainty in inference. Finally, a measure of explained variation, RE, is extended to the relative survival data setting. It is part of the model validation toolkit, and can provide estimates of how much variation in excess mortality due to cancer is explained by the models, and the variables that compose them. There are several methodological assets from the work presented here. First, excess hazard model selection is well formalised. Furthermore, the way RE is adapted to the relative survival data setting will most certainly nurture ideas for the adaptation of other validation tools, commonly used in prognosis research. Lastly, multi-model inference using model averaging is paving the way for the utilisation of ensemble learning in the prediction of excess hazard of death due to cancer. Scenario modelling is a public health application that naturally follows the work done in this PhD thesis. With well-crafted set of predictive models, simulated scenarios can be designed to identify areas for improvement in policy, prevention or treatment. Those generating largest increase in survival can lead to actual recommendations. Methodological advances and public health go hand in hand here. This work emphasises the importance of developing, assessing, and validating excess hazard models. It offers a toolkit so that accurate survival predictions help design effective policies.