*2.2. Statistical Analysis*

We began by examining the distributions of and correlations among all variables. We found that population density, GDP in 2017, scientific production, total tests, air travel, percentage illiterate, and air pollution were highly positively skewed. To reduce the influence of extreme observations, we transformed these variables on the log scale, except for population density for which a square root transformation was sufficient. We next examined the univariable relationships between independent variables and case fatality rates to find candidate variables for our final multivariable model. We found the negative binomial model for case fatality rate to be appropriate, as the overdispersion parameter α was statistically significant in each of these models. We noted the significance of each variable in these univariable models and selected variables for a preliminary multivariable model at *p* < 0.15. We additionally wanted to examine whether the effect of candidate predictors was moderated by population density and time to quarantine. We therefore examined models with each independent variable and (1) an interaction with population density and (2) an interaction with days from 100th case to quarantine (dichotomized as >14 days). Days from 100th case to quarantine was dichotomized as it showed a nonlinear relationship with CFR, with higher CFR rates in the highest quartile (>14 days). Our preliminary multivariable model was created by including all univariable candidate predictors and then adding all significant interaction terms. A hands-on guided approach was used to check for any anomalies that arose when adding or removing variables from the final model. Our preliminary final model contained all multivariable predictors significant at *p* < 0.05. Before arriving at our final model, each variable excluded in the univariable step was added back one-by-one to ensure they were still nonsignificant predictors of case fatality rate. All analyses were performed in Stata (v15.0).
