Bayesian approach to disease model calibration

This thesis is concerned with the calibration of disease models in order to inform decisions about public health problems. These models are used to make predictions about the likely outcomes of disease spread (e.g. morbidity and mortality) in a population, given different interventions (e.g. vaccination, screening and treatment). These models necessitate simpli cations of the underlying mechanisms of disease progression, but have become essential tools in areas such as cost-effectiveness research. Although there are many types of models that can be used, all of them involve parameters, which affect the model's outputs, for example, the age of onset of illness, duration of particular health states or disease specifc mortality. Identifying areas of the model's parameter space that are consistent with data is referred to as model calibration. One of the areas of concern when searching a model's parameter space to get a prediction from the model (that best agrees with observed data) is the propagation of uncertainty. If one has a prediction about an outcome, the question of how uncertain we are about this prediction should be raised. In model calibration, the uncertainty we are interested in is uncertainty in parameters[5]. There are many approaches to model calibration with little consensus on the best practice. Additionally, within the current approaches, some signifi cant problems, that one could face when attempting calibration, merit further examination. We highlight this with a literature search. We believe, that many popular approaches have not yet addressed fundamental issues and more research may be needed to tackle these issues. The model calibrations in this thesis are carried out in a Bayesian framework, using Markov chain Monte Carlo (MCMC) techniques. Although the ultimate goal of calibration should be predictions about some quantities of interest, the emphasis of this work is on the techniques and methodology of model calibration in general. Many possible difficulties of using MCMC and some other sampling approaches are presented and discussed. Avenues for additional research are identified.