Spread of SARS-CoV-2 Coronavirus likely to be constrained by climate

Miguel B. Araujo; Babak Naimi

Metrics of model performance on the test data were generally high (mean/SD AUC=0,76/0,03, see supplementary Figure S4 ). However, models with high performance in the test data can still generate projections that are uncertain when used for forecasting (17) . While improvements in the data and the models can reduce uncertainties, an issue we are now exploring at light of the recently published standards for applied ENMs (18) , characterizing the uncertainty of existing models is a first step towards understanding the limitations of projections and highlighting areas of concern. We generated 200 models by varying the initial conditions and model classes. As such, we were able to quantify and map some of the methodological uncertainties associated with projections ( Figure 5 ). Using this approach, we show that variation associated with changing initial conditions is negligible, which is unsurprising given that splits are randomly performed. In contrast, consistently with previous studies addressing climate change forecasts (19) , uncertainty associated with use of different modeling techniques is high (20, 21) . Models are considerably variable in areas projected to have low seasonal variability in climate suitability across much of Latin America, sub-Saharan Africa and South East Asia. Specific areas in India, China, Western and Central Europe, Coastal Australia, and Central USA also 6 have high levels of model variability. Such variability is probably a consequence of the sparse nature of the positive cases of COVID-19 in those areas, causing different models to adjust differently to the data. In contrast, inter-model variability across northern higher latitudes is lower and this is likely because most records between January and March are found there. Such areas of low inter-model variability coincide with areas where seasonal variation is also greater.