Improving disease prediction and modeling through the use of meteorological ensembles: Rainfall and human plague cases in Uganda

Background/Question/Methods

Climate and weather influence the occurrence, distribution, and incidence of infectious diseases, particularly those caused by vector-borne or zoonotic pathogens. Thus, models based on meteorological data have helped predict when and where human cases are most likely to occur. Such knowledge aids in targeting limited prevention and control resources and may ultimately reduce the burden of diseases. Paradoxically, localities where such models could yield the greatest benefits, such as tropical regions where morbidity and mortality caused by vector-borne diseases is greatest, often lack high-quality in situ local meteorological data. Satellite- and model-based gridded climate datasets can be used to approximate local meteorological conditions in data-sparse regions, however their accuracy varies. Most studies use only a single data source to examine the association between weather and disease occurrence, which can yield spurious model predictions. Here we investigate how the selection of a particular dataset can influence the outcomes of disease forecasting models. Our model system focuses on plague (Yersinia pestis infection) in the West Nile region of Uganda.

Results/Conclusions

Because good quality ground-based meteorological data for this region of Uganda is largely non-existent, we selected six gauge- and satellite-estimated or re-analysis meteorological datasets as well as meteorological data from the airport in Arua, Uganda, which is the closest meteorological station to our study region. The correlation between datasets in the standardized monthly days of rainfall varied widely. Using an ensemble rainfall dataset weighted based on the correlation strength of each rainfall dataset with eleven regional meteorological stations, the annual occurrence of suspected human plague was negatively associated with the number of days of >10mm of rainfall in the preceding dry season (December-February) and positively associated with the number of days with between 0.2-10mm of rainfall during June and July prior to the start of the main plague transmission season (August to December). Both rainfall variables from this model were statistically significant for only two of the seven rainfall datasets, and neither were significant for three of the seven datasets, which happen to have the coarsest spatial resolution. However, both variables remained statistically significant when their coefficient values were averaged across all potential meteorological datasets. The large range of results using individual datasets highlights the importance of employing ensemble approaches likes the one used here for disease model development and implementation, to increase the likelihood of robust results and to understand model uncertainty.