SPATIAL STATISTICAL ANALYSIS OF MALARIA PREVALENCE DATA IN BOTSWANA

Abstract

ISEE-300 Introduction: Recent years have seen increased application of spatial statistical methods in modelling and mapping the distribution of vector-borne diseases. In sub-Saharan Africa the MARA (Mapping Malaria Risk in Africa) project has been working towards a malaria risk atlas to guide rational and targeted control. Historical malaria prevalence data have been collated from malarious countries and a number of malaria risk maps have been produced at different scales, using different data sets and analytical approaches. For malaria risk maps to be useful to control managers, they have to depict the malaria situation with certain real-life accuracy, accepting that some residual variation will always remain unexplained. Retrospective analysis of observational data requires many decisions concerning the explanatory and outcome variables and the statistical methods, which impact on the model outcome and the resulting risk map. Aim: In this paper we grapple with the issue of biological plausibility versus statistical accuracy by investigating the sensitivity of malaria risk maps to model formulation. We focus on Botswana, which lies at the southern edge of the historical malaria distribution. Methods: The 1063 age-specific malaria prevalence rates collected for this country differed with regard to age groups, month of survey, area covered and subject selection method. To reduce bias from various sources we focussed on the 1961/62 national survey, involving 125 prevalence results for the 1-14 year age group for 116 unique locations fairly well distributed across the whole country. Potential explanatory variables were generated from various environmental data sets available for the African continent. Simple, non-spatial uni-variate and multi-variate logistic regression analysis was carried out against the available environmental co-variates. Spatial statistical analysis was also carried out using the most plausible non-spatial multi-variate models as a starting point. We used both Bayesian estimation implemented via Markov chain Monte Carlo methods, and a combination of Bayesian and geo-statistical methods. Risk maps were generated using the model results. Results: Of over 60 potential explanatory variables, the large majority were strongly associated with malaria prevalence in uni-variate analysis. Different co-variate selection methods resulted in several very different multi-variate regression models. Though good overall fit could be achieved, the best-fitting models did not produce the most plausible risk maps. Conclusion: The choice of co-variates had a greater impact on the risk map outcome than using spatial versus non-spatial statistical methods. Different model outcomes will be presented and the issue of biological plausibility versus statistical accuracy will be discussed.