Tabulations of COVID-19 deaths were obtained from the 21 countries with the highest recorded number of cases of COVID-19 as of April 12, 2020. Our data source for COVID-19 cases and deaths was the Johns Hopkins University, Center for Systems Science and Engineering Coronavirus Resource Center (CSSE). CSSE provides numbers of deaths and confirmed cases for each country across the globe [4]. The countries in our sample were (in alphabetical order): Austria, Belgium, Brazil, Canada, China, France, Germany, India, Iran, Israel, Italy, Netherlands, Portugal, Russia, South Korea, Spain, Sweden, Switzerland, Turkey, the United Kingdom, and the United States. We collected the total number of deaths each country over 6 weeks starting the day of a country’s fiftieth recorded COVID-19 death. On May 8, 2020, complete 6-weeks of data were available for 19 of the 21 countries. Age and sex distributions for COVID-19 deaths were procured for 16 of the 19 countries, predominantly from government ministries of public health. We estimated the number of COVID-19 deaths for each age and sex group for the 6-week totals of COVID-19 deaths for each country. COVID-19 mortality rates were determined using age and sex specific population sizes for each country using 2020 population estimates from the Central Intelligence Agency (CIA) World Factbook [5]. Table 1 provides variable definitions, all publicly available data sources, their provenance, and weblinks where these data can be extracted.
Statistical methods
We performed an ecological study to model the association between COVID-19 mortality as the outcome versus age and sex using a multilevel mixed-effects Poisson regression model. The predictor of particular interest, age, was categorized as: 0 to 54, 55 to 64 and 65 or older years. We modeled COVID-19 deaths (Deaths) as.
$$ \mathsf{\log}\left(\mathsf{E}\left[{\mathsf{Death}}_{\mathsf{i}\mathsf{jk}}\ |\ \mathsf{Age},\mathsf{Sex}\right]/{\mathsf{N}}_{\mathsf{i}\mathsf{jk}}\right)={\mathsf{b}}_{\mathsf{i}}+{\upbeta}_{\mathsf{0}}+{\upbeta}_{\mathsf{1}}\ {\mathsf{Age}}_{\left[\mathsf{55}-\mathsf{64}\right]}+{\upbeta}_{\mathsf{2}}\ {\mathsf{Age}}_{\left[\ge \mathsf{65}\right]}+{\upbeta}_{\mathsf{3}}\ {\mathsf{Sex}}_{\left[\mathsf{female}\right]} $$
where E[Deathijk | Age, Sex] denotes the mean number of COVID-19 deaths by age (Age) and sex (Sex) groups. We incorporated a normal, mean zero country-level random effect, bi; to account for correlated data at the country level. The model indices denote the country (i = 1,2,…,16), age groups (j = 1 ages [≤ 54 years], 2 ages [55–64 years], 3 ages [≥ 65 years]) and sex (k = 1 for female, k = 0 for male); the country population sizes by age and gender groups, Nijk, were modeled as an offset. The exponentiated regression coefficients are interpreted as incidence rate ratios (IRR’s) of COVID-19 mortality. Six-week total COVID-19 deaths were disaggregated for each country using their respective estimated age and sex distributions of COVID-19 deaths. Our analyses accounted for this source of variation using a non-parametric bootstrap procedure. We selected 10,000 bootstrap samples from the estimated age and sex distributions of COVID-19 deaths to generate 10,000 disaggregated estimates of the COVID-19 death totals by age and sex for each country. Bootstrap standard errors were used in testing whether there were differences in the COVID-19 mortality rates between age groups and for sex. Specifically, Wald statistics were used to test for associations between the age groups and for sex. We provide 95% bootstrapped confidence intervals using the tolerance interval method on the 10,000 bootstrap estimates. Analyses were conducted using the Stata statistical package (StataCorp. 2019. Stata Statistical Software: Release 16. College Station, TX: StataCorp LP).