Space-time modelling of health data

With the increasing availability of spatially referenced health data over long period of times, it urges different methods to detect missing opportunities contained in them can be utlised to learn more about epidemiology of diseases. This thesis presents different methods of spatiotemporal analysis over health data and considers three different datasets. The three main parts of this thesis covers different applications of spatiotemporal analyses on three major health datasets: official tuberculosis data in Portugal, Surveillance, Epidemiology and End Results Program data and National Health Service open prescribing data. All statistical inferences are mainly based on the Bayesian framework including both exact and approximating methods. The first main contribution uncovered the sub-regional epidemiology of tuberculosis from 2000-2013 in Lisbon and Oporto Metropolitan Areas in Portugal. The analysis provided a Poisson mixed effect model approach on tuberculosis in Portugal at a finer geographical scale compared to past researches. Such model includes both spatial and temporal correlated effects, interaction terms and unstructured random effects. The inference is based on the Integrated Nested Laplace Approximation method. The analysis revealed that having adjusted to known risk factors based on previous research, there are still spatial clusters in both Metropolitan areas with similar but not identical temporal trends. High incidences are more focused around poverty zones which are not necessarily high population density areas. The second interest of this thesis falls in the area of survival studies. A two-way Bayesian spatiotemporal hazard model is introduced in the second main part of this thesis. This is a novel extension which resolves the ignorance of real calendar time when subjects enter the study by including it as a secondary timescale, on top of random effects. In many previous studies, the main interest focuses on the survival duration of study subjects and the time of entry is often eliminated, especially in hazard models. This is not always a suitable assumption when the study is carried over an extended period of time, e.g. decades. Such two-way spatial survival model knits the problem into a 3-dimensional one where both timescales and spatial random effects are considered at the same time. Even though the application of the model over breast cancer in New Mexico from the Surveillance, Epidemiology and End Results program did not present as satisfying results. Supported by simulation studies, the two-way spatial survival model does respect the fact that data are accumulated over time and in that sense, treats the data in a more natural way. It should still provide a different and more natural way to approach long-term data and allow comparisons of different behaviours between; for example population-based cancer registries. The last contribution of this thesis concentrates on the modelling of general practice (GP) open prescribing data. This source of data constitutes a rich time series covering drugs prescribed by all general practitioners across England. The full potential of such a source of data has not been exploited. The ultimate goal for this project is to detect missing opportunities in such open-source data to learn about the epidemiology of a disease. In this part of the thesis, we focused on the pregabalin prescribing data at both GP and Clinical Commissioning Group(CCG)-level across England. The work demonstrated a clear North-South divide in behaviour of pregabalin prescriptions adjusting for spatial effects and deprivation at units of milligrams. This shows that the cause of higher prescribing rate of pregabalin in the North is not just caused by higher deprivation rate. Temporal trend developed differently within each CCG and among GPs, some of which showed an increasing trend.