RESUMO
Approximately 1.5 million people, mostly children, die annually due to disease attributed to diarrhea reflecting urgent needs for improved understanding of associations between the disease and potential risk factors. Numerous epidemiological studies found spatially varying (non-stationary) disease associations attributable to changing geographic or demographic context. Spatial non-stationarity implies that average relationships from statistical models fitted to the whole study area might be inappropriate since they do not reflect local conditions. Spatial modeling techniques such as geographically weighted regression (GWR) have limitations in providing statistically robust analysis of spatial non-stationarity. Thus, there is a need for development or expansion of modeling techniques to address this issue. Using data for pediatric diarrheal mortality in Brazil in 2000, and different risk factors, we develop an analytical framework to determine regions of similar (stationary) local associations by combining GWR and max-p regionalization. We fit statistical models to these regions, and compare goodness-of-fit and regionally varying coefficients to the national-scale model measures. The proposed framework allows us to examine (a) impact of non-stationarity for regions of different geographic extent with acceptable statistical power, (b) the explanatory power of each risk factor in each region, and (c) if these regions reflect changing data quality or truly existing variations in putative associations.