Geographical study of the stroke incidence and mortality rates using Bayesian analysis

Abstract

There is large geographical variation in the stroke incidence rate worldwide. In addition, stroke is one of the diseases with highest mortality rate.

In this thesis, our main focus is to examine if there are any geographical variation in stroke incidence and mortality rates in Norway. We approach this study using models within the Bayesian framework. Before performing analysis, we first introduce the theory behind Bayesian statistics. Moreover, we take a look at the integrated nested Laplace approximation which is a computational framework within Bayesian analysis as an alternative to the traditional Markov chain Monte Carlo simulation techniques.

To study the stroke incidence rate, we first perform a preliminary analysis using classical hypothesis tests within the frequentist framework. It is well known that two of the main non-modifiable risk factors for stroke are age and gender. We use non-parametric hypothesis tests that allow for block design, namely age and gender as subgroups, within the geographical areas. Friedman test is used to see if the incidence rates between the areas are different from each other, and Wilcoxon test is used to perform pairwise test to find out which areas are different. The results of the hypothesis tests give us a better understanding of the data. However, hypothesis tests assume independence between the areas and we know that is not true. Therefore, we enhance the study by using a spatial model that takes into account the areas and their neighbours and allows the model to identify spatial trends. We find that there are no geographical variation in the stroke incidence rate.

We further look at stroke mortality by using Bayesian survival analysis. In addition to examining for geographic variation, we are also interested in effects of socioeconomic status. Since stroke is a condition where many of the patients have comorbidity, we look at the cause-specific mortality (death by stroke) and other death separately. We run the Cox proportional hazards model with the factors of interest to see their effects on stroke mortality. We find that the difference between the cause-specific and other death lies within the baseline hazards, but not so much in the risk factors. In both cases, there are no geographical variation in stroke mortality, however, those with higher income and education levels have lower mortality rate.