key: cord-0711115-ltt9lsaz
authors: Canelo-Aybar, C.; Beltran, J.; Santero, M.; Alonso-Coello, P.
title: Adjusted fatality rates of COVID19 pandemic: a comparison across countries
date: 2020-05-16
journal: nan
DOI: 10.1101/2020.05.13.20099796
sha: c5f60defee1e1bccfa7d875e19663b853f01c0ea
doc_id: 711115
cord_uid: ltt9lsaz

Background: A key impact measure of COVID-19 pandemic is the case fatality rate (CFR), but estimating it during an epidemic is challenging as the true number of cases may remain elusive. Objective: To estimate the CFR applying a delay-adjusted method across countries, exploring differences to simple methods and potential correlation to country level variables. Methods: Secondary analysis of publicly available data from countries with [≥]500 cases by April 30th. We calculated CFR adjusting for delay time from diagnosis to death and using simple methods for comparison. We performed a random effects meta-analysis to pooling CFRs for all countries and for those with high testing coverage and low positivity rate. We explored correlation of adjusted CFR with age structure and health care resources at country level. Results: We included 107 countries and the Diamond Princess cruise-ship. The overall delay adjusted CFR was 2.8% (95%CI: 2.1 to 3.1) while naive CFR was 5.1% (95%CI: 4.1 to 6.2). In countries with high testing coverage/low positivity rate the pooled adjusted CFR was 2.1% (95%CI: 1.5 to 3.0), there was a correlation with age over 65 years ({beta} = 0.12; 95%CI: 0.06 to 0.18), but not with number of physician or critical care beds. Naive method underestimated the CFR of the CFR with a median of 1.3% across countries. Conclusion: Our best estimation of CFR across countries is 2% and varies according to the aged population size. Modelers and policy makers may consider these results to assess the impact of lockdowns or other mitigation policies.

The coronavirus disease 2019 (COVID- 19) is an infectious illness caused by the severe 47

We considered data from countries with at least 500 of reported cases with the SARS-92

Cov-2 by April 30th. This threshold was based in our sample size calculation to estimate a single 93 population proportion of 3% with a 95% confidence interval and a margin of error of +/-1.5% 94 [16] . 95

We collected data per country on the total number of confirmed cases, total number of 97 cases recovered, and total number of deaths from January 22 to April 30th. Based on these three 98 variables we estimated the CFR using the following methods: 99

We estimated the delay adjusted CFR (dCFR) applying a correction method previously 100 described by Nishiura et al, which accounts for the delay from case confirmation to 101 death or recovery [17] . This method has recently been used to correct the CFR from 102 the Diamond Princess cruise ship [18] , and the programing code was made publicly 103 available by the authors (https://github.com/thimotei/cCFRDiamondPrincess) based in 104 the following formula [18] : 105

Where u t is the correction factor applied to the denominator (cumulative number of 108 cases), to estimate the proportion with known outcome, c is the daily case incidence at 109

time i and f is the proportion of cases with delay between onset or hospitalization and 110 death [18] . 111

To apply this method we assumed a similar time distribution from hospitalization (which 112 we assumed to be equivalent to the date of diagnosis in our analysis) to death, with a 113 mean of 13 days, based on an analysis of 39 cases of COVID-19 from the city of Wuhan, 114

China [19] . 115 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted May 16, 2020. 

To pool the CFR across countries, we performed a meta-analysis of proportions using a 122 generalized linear mixed random model (which assumes a distribution of different true 123 estimates across entities) with a logit transformation, and the Clopper Pearson method to 124 estimate the confidence intervals for each observation [20] . We assessed the heterogeneity in 125 the estimated CFR across countries by visual inspections of the forest plots, as the Q statistic 126 and the I-square parameter are not recommended for to evaluate inconsistency for proportion 127 meta-analysis. We performed the analysis using the package metaprop in the R statistical 128 platform, version 3.6.2. 129

We provided a pooled delay adjusted CFR estimate for countries with a good coverage 130 of COVID-19 testing, which we defined for purpose of our analysis as having performed more 131 than 10 tests per thousand habitants and maintaining a rate of positive tests lower than 10%, 132 and as a sensitivity analysis we performed the same analysis for countries with less than 5% of 133 positivity rate. Countries with these characteristics are more likely able to maintain an 134 acceptable surveillance of positive cases as the outbreak progresses and, thus typically provide 135 more reliable CFR estimates. 136

Within this group of latter group of countries, we also conducted an exploratory 137 ecological analysis to assess if there were a positive correlation between the dCFR and the 138 proportion of population over 65 years old, or a negative correlation with the number of critical 139 care beds and the number of physicians. We performed three independent bivariate linear 140 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted May 16, 2020. . https://doi.org/10.1101/2020.05. 13.20099796 doi: medRxiv preprint regression analysis after a logarithmic transformation of the delay adjusted CFRs, and a 141 multivariate regression analysis including variables with a p value <0.10. We tested the normality 142 of residuals by producing a kernel density plot and homoscedasticity by plotting residuals versus 143 fitted (predicted) values. 144

Finally, we graphically plotted the progression of the nCFR and the dCFR over time, 145 since the seventh day after the first reported death, for three countries that showed a flattened 146 epidemic curve over time (New Zealand, South Korea, and Germany). 147

We included data from a total of 107 countries with more than 500 confirmed cases 149 reported, most of them from Europe (39), and Asia (32) . We also included data from the 712 150 Table S1) . 157 We identified a subset of 26 countries, together with the Diamond Princess cohort as 158 having large number of tests performed and low positivity rate. Among these countries, the one 159 with the highest number of tests performed was Iceland (120.4 tests per thousand people), and 160 the lowest was South Korea (10.7 tests per thousand people). The highest positivity rate was 161 observed in Singapore (9.7%) and the lowest in Hong Kong (0.8%). 162

The dCFR ranged from 0.2% (Qatar) to 24% (United Kingdom), our overall pooled 163 estimation showed a dCFR across countries of 4.3% (95%CI: 3.5 to 5.1). There was important 164 heterogeneity, with a group of 12 countries showing a CFR above 15%, most of them from the 165 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 16, 2020. 3.3) across countries. The difference between the nCFR and the dCFR across countries had a 168 median of 1.3% and ranged between 0.1% (Hong Kong, Iceland, China) to than 12% (Mexico), 169 being correlated to the intensity of the outbreak progression. The recovery method showed a 170 much higher fatality rate (CFR 16.0%; 95%CI 11.9% to 21.3%), but many of the individual values 171 per country were implausible (over 50%), precluding the possibility of further analysis using this 172 method (Supplementary Figure S1) . 173

In the subgroup with large testing coverage, the dCFR ranged from 0.2% (Singapore) to 174 9.6% (Canada). Our initial pooled estimation was 2.3% (95% CI 1.6% to 3. The most remarkable result to emerge from the data is that our best approximation for 198 the COVID-19 CFR across countries with a good testing coverage is of 2%. This estimation is an 199 average of potentially different true CFR as they might vary depending on countries´ age 200 structure (higher in countries with aged populations). Accordingly, this estimation based on the 201 available data is lower than the reported by the WHO of around 3.4%.but still 20 times higher 202 than the reported CFR for other diseases like seasonal influenza (0.01%) [21] . Moreover, it is 203 lower than other recent coronavirus epidemics as the Severe Acute Respiratory Syndrome 204 (SARS) with a fatality rate of 10% [22] , or the Middle East Respiratory Syndrome (MERS) with an 205 even higher mortality rate of around 34% [23] . We selected countries with a high testing 206 coverage for our primary analysis to provide more robust results, however, further population 207 seroprevalence studies will be required to estimate the infectious fatality rate which in 208 simulation studies have been estimated to be lower than 1% [24] . 209 Importantly, we have verified a positive correlation with the population age structure. 210

Larger delay adjusted CFRs were observed in countries like Canada, Slovenia or Denmark which 211 have a high proportion of people over 65 years old [12] . In addition, it have been reported that 212 the outbreak specifically affected the older segment of the population in Canada, where more 213 than 90% of who have died were over 60 years old, and nearly half of the deaths occurred in 214 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 16, 2020. . https://doi.org/10.1101/2020.05.13.20099796 doi: medRxiv preprint long term care homes [25] , something that has been also reported in Slovenia [26] . On the other 215 side, Singapore had the smallest CFR, partly due to the smaller proportion of older people (11%), 216 and because the outbreak concentrated in younger migrant workers. 217

We showed that naïve CFR underestimates corrected estimates by a median of 1.3%, 218

with the magnitude of this difference being correlated to the epidemic curve growth. This 219 observation should be considered by policy makers when communicating the consequences of 220 the outbreak. This bias, due to the calculation method, has been previously reported during the 221 SARS epidemic, in which naïve fatality rate increased over time, leading some to conclude that 222 was more lethal than it resulted to be. The public health impact of inaccurate estimates, 223 resulting in misinformation and conflicting messages, can therefore exacerbate public alarm [7] . 224

Some previous analysis has also attempted to estimate the CFR with different limitations 225 that we have attempted to overcome. One study adjusted the delay time between diagnosis to 226 death an reported CFR for 82 countries outside China of 4.24%, but this was calculated using a 227 fairly simplistic method, based on the number of cases of 13 days previous to the assessment 228 date [27] . Another study collected data from several surveillance sources of cases from mainland 229

China only, adjusting for under-ascertainment, and time delay; the authors obtained a CFR of 230 1.38%, increasing to 6.4% in those older than 64 years [24] . The Centre of Evidence Based 231

Medicine used a meta-analytic approach to calculate a prediction interval across countries of 232 between 0.84% to 8.67%. The authors did not provide a pooled estimate due to the large 233 heterogeneity observed. However, by using a fixed effect model the authors assumed the 234 existence of an only true value, an assumption that is not appropriate to account for different 235 true CFR across countries, they relied in naive estimates, and did not considering testing 236 coverage in their assessment. 237

Recently, two non-peer reviewed seroprevalence studies reported estimates for the 238 infectious fatality rate. In the first one, researchers from the university of Bonn took serological 239 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 16, 2020. . https://doi.org/10.1101/2020.05. 13.20099796 doi: medRxiv preprint samples from approximately 1,000 inhabitants of the German town of Gangelt (population of 240 12,529 people), estimating an infection rate of 14%, and a fatality rate of 0.37%. Another study 241 conducted in the Santa Clara County, California, United States, found a population-weighted 242 prevalence that ranged from 2.5% to 4.2% and an infectious fatality rate of 0.12% to 0.2% [28] . 243

However, these results have received some criticism related to the potential large false positive 244 results from serology tests, and the recruitment method employed that might have 245 overestimated the number of infected people. 246

We are aware that our estimates are based on the reported number of cases with a 247 positive test result and not in the real number of infected people, overestimating the CFR. 248

Asymptomatic cases and mild symptomatic patients are not routinely tested as it is underscored 249 by recent publications. In a large-scale COVID-19 diagnostic testing in Iceland with 9,199 250 persons, found that 43% of positive cases were asymptomatic at the time of testing [22] . In 251 another study, from a total of 215 pregnant women, 29 of 33 (88%) who were SARS-CoV-2 252 positive at admission were asymptomatic [29] . Although, the number of tests performed might 253 differ from the number of individuals tested, and the distinction is not always clear in public data 254 (i.e. Hong Kong reports in number of tests, while Singapore in number of people) [11] , including 255 countries with a large number of tests performed and low positive rate in our main analysis, has 256 likely reduced the impact of the bias due to the ascertainment of cases. 257

We should also consider the bias due to underreporting of the cause of death, which 258 goes in the opposite direction (underestimating CFR). For example, China has recently revised 259 its death tolls, on April 17, adding 1,300 fatalities to its initial official count for the city of Wuhan; 260 due to the inclusion of deaths that occurred at home or at institutions [30] . Similarly, in Madrid, 261

Spain, around 4,100 death cases occurred among elder people in long term care homes who 262 reported symptoms compatible with COVID-19 and that were not accounted in official reports 263 [31] . One strategy to address this bias is comparing the all-cause mortality estimates with 264 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 16, 2020. . https://doi.org/10.1101/2020.05. 13.20099796 doi: medRxiv preprint previous years, as in the Guayas region of Ecuador in which an excess of more than 6,000 deaths 265 was observed, probably be related to the COVID-19 outbreak [32] . Based in this approach, 266 approximately a 25% of the comparative excess mortality could be also attributed to in countries like Germany or Portugal [33] . 268

We consider our calculations also have strengths, compared with previous estimation 269 across countries, because: 1) we included countries with more than 500 confirmed cases which 270 making calculations more stable, 2) we used an appropriate method for pooling proportions, 271 including a random model that considers the different true estimates across different settings, 272

3) instead of using a naïve estimation, we applied a correction method to consider the delay 273 time between diagnosis and death, and 4) we only included in our main analysis countries with 274 good testing coverage, which might reduce the impact of the different biases due to 275 ascertainment of cases. 276 277

In conclusion, assessing the fatality rate of COVID-19 is critical to determine the 279 appropriateness of mitigation strategies, as well as to enable forecasting of healthcare 280 requirements as the epidemic unfolds. Thus, one urgent need is to conduct large high quality 281 seroprevalence studies [34] . Meanwhile, our findings contributes to the literature providing a 282 good approximation of the true fatality rate across countries based in more appropriate 283 methods and taking into account the testing coverage and positivity rate. Until more robust 284 estimates are available, a CFR of 2% might be used for policy makers for comparability. Our 285 study shows that naive estimation of CFR underestimates the potential threat of COVID-19, and 286 although use of such methods are clearly easier to communicate to policy makers and the public, 287 their use could be misleading for the deployment of health systems responses. 288

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. This research did not receive any specific grant from funding agencies in the public, commercial, 295 or not-for-profit sectors. 296

The authors declare that there are no conflicts of interest 298 299 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 16, 2020. . https://doi.org/10.1101/2020.05.13.20099796 doi: medRxiv preprint Figure 1 . Meta-analysis of the delay adjusted CFR across countries with high testing coverage and low positivity rate* *Analysis excluding Canada due to having an outlier value.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted May 16, 2020. . https://doi.org/10.1101/2020.05. 13.20099796 doi: medRxiv preprint Figure 2 . Overall meta-analysis across countries of the naive and delay adjusted CFR.

A: Naive method.

B: Delay adjusted method . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted May 16, 2020. . https://doi.org/10.1101/2020.05. 13.20099796 doi: medRxiv preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted May 16, 2020. . https://doi.org/10.1101/2020.05. 13.20099796 doi: medRxiv preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted May 16, 2020. . https://doi.org/10.1101/2020.05. 13.20099796 doi: medRxiv preprint 

A novel coronavirus outbreak of global health concern

World Health Organization. WHO announces COVID-19 outbreak a pandemic

WHO Director-General's opening remarks at the media briefing on COVID-19 -17

World Health Organization

Coronavirus Travel Restrictions, Across the Globe

A third of the global population is on coronavirus lockdown -here's our constantly updated list of countries and restrictions

Potential Biases in Estimating Absolute and Relative Case-Fatality Risks during Outbreaks

Methods for estimating the case fatality ratio for a novel, emerging infectious disease

SARS becoming deadlier: officials

World Health Organization. World Health Organization

An interactive web-based dashboard to track COVID-19 in real time. The Lancet Infectious diseases

COVID-19) Testing -Statistics and Research

Critical Care Bed Capacity in Asian Countries and Regions. Critical care medicine

The variability of critical care bed numbers in Europe

Statulator: An online statistical calculator. Sample Size Calculator for Estimating a Single Proportion

Early epidemiological assessment of the virulence of emerging infectious diseases: a case study of an influenza pandemic

Estimating the infection and case fatality ratio for coronavirus disease (COVID-19) using age-adjusted data from the outbreak on the Diamond Princess cruise ship

Incubation Period and Other Epidemiological Characteristics of 2019 Novel Coronavirus Infections with Right Truncation: A Statistical Analysis of Publicly Available Case Data

Metaprop: a Stata command to perform meta-analysis of binomial data. Archives of public health = Archives belges de sante publique

Centers for Disease Control and Prevention

Spread of SARS-CoV-2 in the Icelandic Population. The New England journal of medicine

Middle East respiratory syndrome coronavirus (MERS-CoV)

Estimates of the severity of coronavirus disease 2019: a model-based analysis. The Lancet Infectious diseases

Nearly half of Canada's COVID-19 deaths linked to long-term care facilities: Tam

Global News

Afternoon 15 April: Numbers; Gloves No Longer Mandatory

Case-Fatality Risk Estimates for COVID-19

Calculated by Using a Lag Time for Fatality. Emerging infectious diseases

Universal Screening for SARS-CoV-2 in Women Admitted for Delivery. The New England journal of medicine

Coronavirus: China outbreak city Wuhan raises death toll by 50%

Spain's coronavirus deaths pass 14,500, but real toll may be bigger

Ecuador's death rate soars as fears grow over scale of coronavirus crisis

Guardian News and Media

Understanding excess mortality [Internet]. The Health Foundation

Antibody tests suggest that coronavirus infections vastly exceed official counts

Nature News