Geostatistical methods are increasingly used in low-resource settings where disease registries
are either non-existent or geographically incomplete. In this thesis, which is comprised
of four papers, we address some of the common issues that arise from analysing
disease prevalence data. In the first paper we consider the problem of combining data
from multiple spatially referenced surveys so as to account for two main sources of variation:
temporal variation, when surveys are repeated over time; data-quality variation,
e.g. between randomised and non-randomised surveys. We then propose a multivariate
binomial geostatistical model for the combined analysis of data from multiple surveys.
We also show an application to malaria prevalence data from three surveys conducted
in two consecutive years in Chikwawa District, Malawi, one of which used a more economical
convenience sampling strategy. In the second paper, we analyse river-blindness
prevalence data from a survey conducted in 20 African countries enrolled in the African
Programme of Onchocerciasis Control (APOC). The main challenge of this analysis is
computational, as a binomial geostatistical model has to be fitted to more than 14,000
village locations and predictions carried out on about 10 millions locations across Africa.
To make the computation feasible and efficient, we then develop a low rank approximation
based on a convolution-kernel representation which avoids matrix inversion. The
third paper is a tutorial on the use of a new R package, namely “PrevMap”, which provides
functions for both likelihood-based and Bayesian analysis of spatially referenced
prevalence data. In the fourth paper, we present some extensions of the standard geostatistical
model for spatio-temporal analysis of prevalence data and modelling of spatially
structured zero-inflation. We then describe three applications that have arisen through
our collaborations with researchers and public health programmers in African countries.