Climate affects global patterns of COVID-19 early outbreak dynamics

Gentile Francesco Ficetola; Diego Rubolini

Therefore, for climatic variables, we included in models both linear and quadratic terms. 5 Humidity, population density, health expenditure and PM2.5 were log 10 -transformed to reduce skewness and improve normality of model residuals. We adopted a model selection approach to identify the variables most likely to affect the global variation of Covid-19 growth rate (29) . We built models representing the different combinations of independent variables, and ranked them on the basis of Akaike's Information 10 Criterion (AIC). AIC trades-off explanatory power vs. number of predictors; parsimonious models explaining more variation have the lowest AIC values and are considered to be the "best models" (29) . For each candidate model, we calculated the Akaike weight ω i , representing the probability of the model given the data (30). We then calculated the relative variable importance of each variable (RVI) as the sum of ω i of the models where each variable is included. RVI can 15 be interpreted the probability that a variable should be included in the best model (29, 31) . Model selection analyses and the calculation of RVI can be heavily affected by collinearity among variables. In our dataset, temperature and humidity showed a very strong positive correlation ( Fig. S1 and Table S7 ); furthermore, population density was strongly positively associated with PM2.5 ( Figure S1 and Table S7 ). Therefore, temperature and humidity, or population density 20 and PM2.5, could not be considered together in the same models (31, 32) . All other predictors showed weak correlations and should not cause collinearity issues (32) (Table S7) . We therefore repeated the model selection for different combinations of uncorrelated variables. First, we . CC-BY-NC-ND 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.