Modeling Schistosoma japonicum Infection under Pure Specification Bias: Impact of Environmental Drivers of Infection

Uncertainties in spatial modeling studies of schistosomiasis (SCH) are relevant for the reliable identification of at‐risk populations. Ecological fallacy occurs when ecological or grouplevel analyses, such as spatial aggregations at a specific administrative level, are carried out for an individual‐level inference. This could lead to the unreliable identification of at‐risk populations, and consequently to fallacies in the drugs’ allocation strategies and their cost‐effectiveness. A specific form of ecological fallacy is pure specification bias. The present research aims to quantify its effect on the parameter estimates of various environmental covariates used as drivers for SCH infection. This is done by (i) using a spatial convolution model that removes pure specification bias, (ii) estimating group and individual‐level covariate regression parameters, and (iii) quantifying the difference between the parameter estimates and the predicted disease outcomes from the convolution and ecological models. We modeled the prevalence of Schistosoma japonicum using group‐level health outcome data, and city‐level environmental data as a proxy for individual‐level exposure. We included environmental data such as water and vegetation indexes, distance to water bodies, day and night land surface temperature, and elevation. We estimated and compared the convolution and ecological model parameter estimates using Bayesian statistics. Covariate parameter estimates from the convolution and ecological models differed between 0.03 for the nearest distance to water bodies (NDWB), and 0.28 for the normalized difference water index (NDWI). The convolution model presented lower uncertainties in most of the parameter estimates, except for NDWB. High differences in uncertainty were found in night land surface temperature (0.23) and elevation (0.13). No significant differences were found between the predicted values and their uncertainties from both models. The proposed convolution model is able to correct for a pure specification bias by presenting less uncertain parameter estimates. It shows a good predictive performance for the mean prevalence values and for a positive number of infected people. Further research is needed to better understand the spatial extent and support of analysis to reliably explore the role of environmental variables.