Community-level determinants of the geographic distribution of Wuchereria bancrofti infection in Leogane Commune, Haiti

Lymphatic filariasis (LF) is caused by the mosquito-transmitted parasitic worms Wuchereria bancrofti, Brugia malayi and Brugia timori. The geographic distribution of W. bancrofti infection, which is responsible for approximately 90% of LF worldwide, is non-uniform and its community-level determinants have not been identified. We conducted a school-based assessment of the geographic distribution of W. bancrofti infection in Leogane Commune, Haiti. Using generalized linear mixed models (GLMMs) with school-specific non-spatial random effects, we evaluated potential associations between infection status and geographic, socioeconomic and demographic risk factors, as well as landscape elements derived from Landsat satellite imagery. In addition, we analyzed the spatial patterns in the infection prevalence data using semi-variograms and correlograms, and compared the GLMM results to results from standard logistic regression models (GLMs) and Bayesian hierarchical models (BHMs) with spatial random effects. W. bancrofti infection status was strongly associated with topographic zone, administrative section, altitude and several landscape elements (agricultural vegetation, rock/rocky soil, native grasses and trees), and moderately associated with age; there was no association between infection status and gender, tuition (a marker of socioeconomic status) or nutrition program availability. The tuition and geographic variable results have important implications for LF elimination program community motivation/education and mapping activities, respectively. Although the utility of multivariate landscape element models (models containing >2 variables) was limited because of collinearity, our results suggest that combinations of landscape elements may be important in explaining the geographic distribution of W. bancrofti infection (more so than the individual elements alone). We discuss the challenges associated with processing and interpreting Landsat image data and incorporating image-derived landscape element variables into standard epidemiologic analytic frameworks. We also provide evidence of a clear spatial pattern in the W. bancrofti infection data, indicating that accounting for spatial correlation between outcomes is probably important and using GLMs to analyze these data is likely inappropriate. GLMM assumptions may be violated by such spatial correlation, suggesting that standard GLMMs may be inappropriate for the analysis of spatially correlated data as well. Bayesian hierarchical modeling is introduced as a readily-implementable framework wherein spatially structured random effects can be explicitly specified.