Geographically weighted regression of land cover determinants of Plasmodium falciparum transmission in the Ashanti Region of Ghana

BACKGROUND Malaria is a mosquito-borne parasitic disease that causes severe mortality and morbidity, particularly in Sub-Saharan Africa. As the vectors predominantly bite between dusk and dawn, risk of infection is determined by the abundance of P. falciparum infected mosquitoes in the surroundings of the households. Remote sensing is commonly employed to detect associations between land use/land cover (LULC) and mosquito-borne diseases. Due to challenges in LULC identification and the fact that LULC merely functions as a proxy for mosquito abundance, assuming spatially homogenous relationships may lead to overgeneralized conclusions. METHODS Data on incidence of P. falciparum parasitaemia were recorded by active and passive follow-up over two years. Nine LULC types were identified through remote sensing and ground-truthing. Spatial associations of LULC and P. falciparum parasitaemia rate were described in a semi-parametric geographically weighted Poisson regression model. RESULTS Complete data were available for 878 individuals, with an annual P. falciparum rate of 3.2 infections per person-year at risk. The influences of built-up areas (median incidence rate ratio (IRR): 0.94, IQR: 0.46), forest (median IRR: 0.9, IQR: 0.51), swampy areas (median IRR: 1.15, IQR: 0.88), as well as banana (median IRR: 1.02, IQR: 0.25), cacao (median IRR: 1.33, IQR: 0.97) and orange plantations (median IRR: 1.11, IQR: 0.68) on P. falciparum rate show strong spatial variations within the study area. Incorporating spatial variability of LULC variables increased model performance compared to the spatially homogenous model. CONCLUSIONS The observed spatial variability of LULC influence in parasitaemia would have been masked by traditional Poisson regression analysis assuming a spatially constant influence of all variables. We conclude that the spatially varying effects of LULC on P. falciparum parasitaemia may in fact be associated with co-factors not captured by remote sensing, and suggest that future studies assess small-scale spatial variation of vegetation to circumvent generalised assumptions on ecological associations that may in fact be artificial.