**3. Results**

Case fatality rate varied widely by country, as low as 0.6% and as high as 17.7% (M = 5.4%, SD = 4.3%). Means and standard deviations of candidate predictors are shown in Table 1, along with rate ratios from the univariable negative binomial model for each individual predictor. We found that percent population >70 years old, general mortality per 1000 individuals, percentage of population illiterate, percentage of population with HIV, and air pollution were significantly associated with case fatality rate in these univariable models. In addition, tests per 1000 individuals, percentage of population obese, smokers, tobacco users, and with HIV significantly interacted with population density.


**Table 1.** Descriptive statistics (mean & SD) for candidate predictor variables, rate ratios (SE) for univariable relationships between predictors and case fatality rate, and *p*-value for the interaction of each predictor variable with population density (square root transformed). For comparability, all rate ratios reflect the effect of the standardized predictor on case fatality rate.

† Log-transformed variable was used for Rate Ratio, \* *p* < 0.05, \*\* *p* < 0.01.

We also found several instances of collinearity. For example, we found high correlations between: smoking prevalence and tobacco use prevalence (r = 0.92, *p* < 0.001); air travel and GDP (r = 0.96, *p* < 0.001); and percentage of population >70 years old was correlated with several variables such as general mortality (r = 0.80, *p* < 0.001), prevalence of COPD (r = 0.71, *p* < 0.001), life expectancy (r = 0.74, p < 0.001), physicians per capita (r = 0.73, *p* < 0.001), tests per capita (r = 0.88, *p* < 0.001), and GDP (r = 0.71, *p* < 0.001). When collinear variables were included in the model, we sequentially added them in one-by-one and evaluated the model fit; the variable producing the best fit was retained in the model.

Our final model included time from 100th case to quarantine (dichotomized >14 days), hospital beds per 1000 individuals, percentage population over 70 years, CT scanners per 1 million individuals (log-transformed), and interaction between smoking prevalence and population density. This model had good agreement between observed and predicted CFR values (Figure 1). We found that countries waiting over 14 days from the 100th case to quarantine had 1.5 times the case fatality rate of those that did not wait as long (*p* = 0.045), and each percentage increase in the population over 70 years was associated with 1.15-time increase in the case fatality rate (*p* < 0.001). Though proportion of population over 70 years was correlated with a slew of health-related variables, there were some that

#### *Int. J. Environ. Res. Public Health* **2020**, *17*, 8189

were predictive of case fatality rate above and beyond the proportion of the population that is elderly. We found that each additional hospital bed per 1000 individuals reduced the case fatality rate by 15% (RR = 0.85, *p* < 0.001), and that a 1-unit increase in the log number of CT scanners per million was associated with half the case fatality rate (RR = 0.49, *p* < 0.001). The deleterious effect of smoking on case fatality rate was significant, but only in countries with higher density (*p*-interaction < 0.001). To aid in interpretation, we calculated the rate ratio at the mean, and 0.5 SD below and above the mean, of square root transformed population density. These results are presented in Table 2 (Model 1).

**Figure 1.** Final model (Model 1) of predicted values plotted against observed values of case fatality rate. These two variables were correlated at r = 0.84.

**Table 2.** Final multivariable negative binomial model predicting case fatality rate. Rate ratios and 95% confidence intervals are presented. Smoking prevalence is evaluated at the mean of (square root transformed) population density, 0.5 SD below (low, approximately 65 per km2), and 0.5 SD above (high, approximately 200 per km2). Model I contains our final estimates without imputation (*n* = 26), Model II additionally adjusts for date of 100th case, and Model III shows the results from our final model on imputed CT scanner data (*n* = 39).


We performed additional sensitivity analyses to explore the effect of (1) date of country being impacted by COVID-19 and (2) missing covariate data. First, we created a new model by including date of 100th case to examine any change in coefficients (Table 2, Model 2). It was suspected that CFR may be lower in countries that reached their 100th case later, as they may not have had sufficient time for the virus to act on individuals. We examined the correlation between date of 100th case and days-to-quarantine and found a negative relationship (r = −0.47, *p* = 0.003) after excluding China, which was impacted early but also had quick quarantine implementation (Figure 2). We also note that

#### *Int. J. Environ. Res. Public Health* **2020**, *17*, 8189

13 countries were missing data on CT scanners in our model. To determine possible impacts of this missing data, we performed multiple imputation with 25 data sets using all complete variables in the data that were not included in our final models. Our final model changed slightly but not appreciably (Table 2, Model 3). Date of 100th case was not significant when added to this model (*p* = 0.66; model not shown).

**Figure 2.** Days from 100th case to quarantine plotted against date of 100th case (in days from 19 January 2020; China's 100th case). The two variables are correlated (r = −0.47, *p* = 0.003) after removing China (red square).

#### **4. Discussion**

Our analysis of 24 variables relative to COVID-19 mortality across 39 countries suggests that the case fatality rate is related to a variety of country-specific factors, including time to implement social distancing measures after the 100th case, hospital beds per 1000 individuals, percentage population over 70 years, CT scanners per 1 million individuals, and smoking prevalence with high population density.

Social distancing interventions, such as increased case isolation and community contact reduction, have been shown to be highly effective in slowing the spread of the virus [29]. According to our model, countries that waited over 14 days to implement social distancing interventions after their 100th reported case saw an increased CFR (RR = 1.54, *p* = 0.045), consistent across all population densities. As COVID-19 spread has been shown to occur during the asymptomatic incubation period [30,31], the promptness of local and national government response in implementing quarantine policies may have played a crucial role in limiting human-to-human transmission. As respiratory failure from Acute Respiratory Distress Syndrome appears to be the leading cause of mortality [32], surges in severe COVID-19 cases have the potential to overwhelm the capacity of a country's healthcare system to provide mechanical ventilation and other intensive resources [33,34]. Thus, timely implementation of social distancing measures may, in many cases, have delayed epidemic peak in regard to CFR by reducing exponential growth of cases [29]. However, we did see this effect attenuated when including date of 100th case in the model. While it may be possible that countries that were affected more recently may not have had time to fully experience the true extent of COVID-related deaths, we did find that countries affected by COVID-19 later were quicker to implement quarantine measures. Because of this, we cannot truly disentangle the effect of time-to-quarantine from date of 100th case.

With regards to comorbidities, our model predicts that a 10% increase in smoking prevalence more than doubles the CFR in countries with high population density (RR = 2.53, *p* < 0.001). There is a growing body of evidence associating increased risk of both increased severity of disease, ICU admission, and death among infected patients with smoking history, particularly active smokers [35,36]. Yet, it is unclear why our model correlated smoking prevalence with increased CFR strongly in high-density populations only. One possible explanation is that high density areas are more likely to suffer outbreaks sufficient to overwhelm the local hospital and ICU capacity, triggering mass casualty protocols with possible triaging of resource allocation in favor of patients with more favorable prognostic indicators [37,38]. Smoking history is often associated with other medical comorbidities, notably cardiovascular disease, which may contribute to a poorer overall prognostic presentation and thus less priority in a resource-poor scenario [36,39]. Finally, the effect that smoking may have as a potential risk factor is likely to be more exaggerated in densely populated regions where these vulnerable individuals interact with others in closer proximity and with higher frequency. Smoking prevalence may also be related to other factors that are impacted by population density. It will be necessary to further explore the effects of smoking and its sequalae on disease course to determine the additional considerations that should be given to patients with a history of smoking, particularly in areas of high population density.

Case fatality rate was reduced by 15% (RR = 0.85, *p* < 0.001) for each additional hospital bed per 1000 individuals, a reflection of resources available for delivering inpatient medical services. Studies on past influenza pandemics have shown that scarcity of healthcare resources and clinical infrastructure, particularly in rural or developing areas, is a major limitation to pandemic preparedness [40–42]. It is important to note that the capacity of a healthcare system is tied not only to infrastructure but also to the availability of providers; in order to increase capacity, it will be necessary to build a larger healthcare workforce to support more hospital beds and higher patient volume [41]. CT scanners represent another limited resource that appears to be a protective factor. Our model shows that a 1-unit increase in the log number of CT scanners per million was associated with half the case fatality rate (RR = 0.49, *p* < 0.001). A possible explanation for this protective effect is that CT scans have led to earlier detection of the disease, as early reports describe characteristic imaging features that are helpful in aiding diagnosis [43]. This may have been particularly advantageous in developing countries, where the capacity to develop and mass-distribute testing was limited during the first months of the pandemic [43–45], as well as for frontline providers in any country irrespective of testing ability as a means of providing earlier diagnosis [46,47]. Countries with more radiologists and CT scanners per capita are also likely to have increased availability of other health resources, and thus the variables may serve as proxy for other factors that reflect the robustness of a nation's healthcare system.

Our model also suggests a positive relationship between CFR and percentage of the population over age 70, with every added percent increasing the CFR by a factor of 1.15. This variable is highly associated with general mortality (r = 0.80, *p* < 0.001), life expectancy (r = 0.74, *p* < 0.001), and COPD (r = 0.71, *p* < 0.001) at the country level and could potentially be viewed as a proxy for indicating a country with older, more vulnerable population. Observations on age and increased case mortality are consistent with multiple retrospective studies identifying advanced age as a potential risk factor for more severe disease and worsened prognosis [48,49]. There are multiple possible explanations for this observation. In addition to frailty and increased risk of having multiple co-morbidities, some studies suggest age-related declines in T and B cell function alongside preserved innate immunity may be contributory, with the resultant cytokine 2-dominant response triggering a pro-inflammatory state which increases mortality [50,51].

#### *Limitations and Future Directions*

This modeling study used a cross-sectional, ecological dataset taken during the pandemic's first wave. As such, most of our model's limitations stem from the weaknesses of this approach. As with most data gathered in the first several months of the pandemic, accuracy was limited by the information available at that time and capacity to report cases in a timely manner may vary from country to country. This limitation affects estimates of CFR, which has been shown to require an assessment of the delay between infection and the reporting of case data as well as the extent to which death-related cases are underreported. Furthermore, CFR itself is best modeled as a continuous variable which changes over time as the disease spreads to areas which vary in risk due to population density and demographic, as well as how healthcare systems and government policies adapt to disease burden [52]. Finally, given the ecological nature of our study, it should be noted that our findings and discussion of risk factors do not necessarily reflect a relationship between these variables and probability of survival at the individual level.

In addition, our study was limited by the availability of datasets used for country-specific data. Some of the datasets used to generate data for the 24 variables did not include all countries. For example, although our univariate analysis did not find a significant correlation between CFR and total radiologists or radiologists per one million, our data on radiologists per country was limited to one study from 2008 that included only 26 of the 39 countries analyzed [53]. More data is needed on this subject, as early reports on the global response to COVID-19 have shown that computerized tomography (CT) of the chest may serve an integral role in the timely diagnosis of the disease, as well as in severity staging and monitoring of clinical course [43,44,47,54,55]. Another limitation arises from utilizing data pertaining to entire countries for risk factors that may vary within each country on a geographically smaller scale-among cities, for example. This is especially true for large countries such as the United States and China with wide variations in population density, resource availability, and environmental characteristics based on region. For the timing of isolation and quarantine measures, our study relied on multiple secondary news sources for specific dates (Table S2), which, despite a standardized and systematic approach, is inherently less reliable than documentation from a single primary source.

Future studies are also needed to evaluate the effect and timing of government interventions on disease spread and CFR. While our results suggest a possible relationship between early government implementation of social distancing measures and reduced CFR, it remains unclear whether differences in the type or stringency of these measures appreciably influences CFR. Timing of quarantine measures is also important, and a comparison of specific state-imposed measures as well as their timing relative to the date of first confirmed case and other milestones of cumulative case growth could represent a strong follow-up to our findings. We also suggest a closer look at the relationship between smoking and its association with severe disease, as well as the extent to which such risk factors affect comprehensive care under triage protocols in resource-restricted circumstances.

#### **5. Conclusions**

Using country-based multivariate modeling, our study found significant correlations between increased CFR and smoking prevalence, percentage of a population over the age of 70 years, increased time to implementation of social distancing or stay-at-home measures, as well as decreased CFR and hospital beds per 1000 and CT scanners per million. Notably, CFR appears to increase significantly for every day after the 100th documented case where governmental precautions are not put in place. More research is needed on the relationship between the timing and effectiveness of government precautions and case fatality rate, as well as the role that hospital bed capacity and CT scanner availability play in reducing case fatality. The relationship between population density and smoking prevalence, as well as the influence smoking has on disease course, is also worth further exploration.

**Supplementary Materials:** The following are available online at http://www.mdpi.com/1660-4601/17/21/8189/s1, Table S1: 24 country-level risk factors with potential influence on COVID-19 case fatality rates, Table S2: Date of first government intervention per aggregation of secondary news sources as indicated.

**Author Contributions:** J.P. and J.M.S.P. generated the dataset. B.K.K.F. and N.L.D. provided feedback and guidance during the data collection. T.A.P. performed the statistical analysis. B.D. coordinated interdepartmental outreach. J.P., J.M.S.P., and T.A.P. drafted the manuscript with support from A.G., who conceived the idea.

#### *Int. J. Environ. Res. Public Health* **2020**, *17*, 8189

All authors reviewed the manuscripts and provided edits prior to submission. J.P. and J.M.S.P. contributed equally to the work and should be considered as co-first authors. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work was supported by grants UL1TR001855 and UL1TR000130 from the National Center for Advancing Translational Science (NCATS) of the U.S. National Institutes of Health. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

**Conflicts of Interest:** The authors declare no conflict of interest.
