A statistical framework to adapt traditional spatial and temporal methods to non-standard specialized large databases

Abstract

In this thesis, our aim is to develop and propose new statistical approaches to analyse spatial and spatio-temporal data collected by large administrative databases. In Chapter 2, the impact of a long latency period of mesothelioma cancer in Belgium is investigated, in a case-control setup accounting for the residential histories of the patients to provide accurate risk maps. The proportion of time spent by every patient within an area from a period of 20 to 40 years prior to disease diagnosis is included in a multiple membership model through weights. Pancreatic cancer is used as a control disease instead of a population at risk, traditionally used in disease mapping. Chapter 3 presents a case-control study for areal data within the disease-mapping context investigating the mesothelioma risk in Belgium, using pancreatic cancer as the control disease. This choice is done due to unavailability of a population registry dated 20 or 40 years, and to a change in population size and age/gender structure over a long period of time in a country. Instead of calculating the expected values as commonly done using the population at risk in that area, we propose to estimate this expected values using a control-disease. A conditional autoregressive convolution model is used in this context, incorporating both a spatially structured and unstructured heterogeneity amongst the relative risks per area. In Chapter 4 the spatial and spatio-temporal risk of COVID-19 disease is investigated in Flanders, Belgium. A validation of the results from the Intego database, a morbidity general practitioners registration network, is made, by comparing epidemiological insights gained from these data to the results from analyses of the Sciensano database, data provided by the Belgian Public Health Institute. While Sciensano is responsible for monitoring the epidemiological evolution, the daily COVID-19 data on the new cases hospitalizations, ICU patients and deaths lacks individual level information. The Intego morbidity registration network collects individual-level patient data from a sample of the total general practitioners in Flanders, Belgium. The analysis is done at two different levels of areal aggregation, at the health-sector level, since the COVID-19 policies are undertaken at this level, and at the municipality level, a smaller and thus a finer spatial scale. Chapter 5 presents a new method that quantifies the public health threat of a pandemic, namely the impact on hospitalization load or on mortality when the number of COVID-19 cases doubles. These indicators are calculated for North-Western European countries (the United Kingdom, Belgium, the Netherlands, Denmark, Norway), Central-Southern European countries (Spain, France, Germany), Eastern European countries (Estonia, the Czech Republic, Latvia, Croatia) and South Africa. Firstly, a country-by-country time-varying doubling model is proposed allowing time-varying relationships. Secondly, this method is extended to a multi-country time-varying doubling model by modelling all considered countries simultaneously.