key: cord-0832830-l4y7v729
authors: Gakidou, E.; Team, IHME COVID-19 Forecasting
title: Global projections of potential lives saved from COVID-19 through universal mask use
date: 2020-10-11
journal: nan
DOI: 10.1101/2020.10.08.20209510
sha: f15836588bd767e4823f98ce432274e23e6279ed
doc_id: 832830
cord_uid: l4y7v729

Background: Social distancing mandates have been effective at reducing the health impacts of COVID19. The ensuing economic downturns and unemployment increases have led many nations to progressively relax mandates. As COVID19 transmission and deaths rise in many low and middle-income countries (LMICs), with continuing widespread transmission elsewhere, policymakers are searching for options to reduce COVID19 mortality without reimposing strict social distancing mandates. Methods: Using a Bayesian meta-regression of 40 studies measuring the impact of mask use on respiratory viral infections, we estimated the reduction in transmission associated with the use of cloth or paper masks used in a general population setting. We used data from daily surveys conducted by Facebook, YouGov, and Premise, on the proportion of people reporting always wearing a mask outside their home for nearly all countries. We predicted deaths and infections until January 1st 2021 under a reference and universal mask use scenario using a deterministic transmission dynamics model with categories for susceptible, exposed, infected and recovered (SEIR). In the reference scenario, we assume continued easing of mandates but with action to re-impose mandates for a period of six weeks, at a level of eight daily deaths per million population. The universal mask scenario assumed scaling up of mask use to 95% over a one-week period. Findings: Use of simple masks can reduce transmission of COVID19 by 40% (95% uncertainty interval [UI] 20% to 54%). Universal mask use would lead to a reduction of 815,600 deaths (95% UI 430,600 to 1,491,000 deaths) between August 26th 2020 and January 1st 2021, the difference between the predicted 3.00 million deaths (95% UI 2.20 to 4.52 million) in the reference and 2.18 million deaths (95% UI 1.71 to 3.14 million) in the universal mask scenario over this time period. Mask use was estimated at 59.0% of people globally on August 18th, ranging from 41.9% in North Africa and the Middle East to 79.2% in Latin America and the Caribbean. The effect of universal mask use is greatest in countries such as India (158,832 fewer deaths in universal mask scenario, 95% UI 75,152 to 282,838 deaths), the United States of America (93,495 fewer deaths; 95% UI 59,329 to 150,967 deaths), and Russia (68,531 fewer deaths; 95% UI 34,249 to 145,960 deaths). Interpretation: The rising toll of the COVID19 pandemic can be substantially reduced by the universal adoption of masks. This low-cost policy, whether customary or mandated, has enormous health benefits and likely large economic benefits as well, by delaying the need for re-imposition of social distancing mandates.

UI 1.71 to 3.14 million) in the universal mask scenario over this time period. Mask use was estimated at 24 59 .0% of people globally on August 18 th , ranging from 41.9% in North Africa and the Middle East to 25 79 .2% in Latin America and the Caribbean. The effect of universal mask use is greatest in countries such 26 as India (158,832 fewer deaths in universal mask scenario, 95% UI 75,152 to 282,838 deaths), the United 27 States of America (93,495 fewer deaths; 95% UI 59,329 to 150,967 deaths), and Russia (68,531 fewer 28 deaths; 95% UI 34,249 to 145,960 deaths). 29

Interpretation: The rising toll of the COVID-19 pandemic can be substantially reduced by the universal 30 adoption of masks. This low-cost policy, whether customary or mandated, has enormous health benefits 31 and likely large economic benefits as well, by delaying the need for re-imposition of social distancing 32 mandates. 33 34 Research in context 35 Evidence before this study 36 One meta-analysis of 21 studies reported a pooled reduction in the risk of respiratory virus infection of 37 47% (95% CI 36-79%) from a subset of eight studies reporting on mask use in non-health workers but it 38 did not distinguish type of mask. Another meta-analysis reported on 26 studies of mask use in health 39 workers and three studies in non-healthcare settings, reported a pooled effect of a 66% (55-74%) 40 reduction in infections and a reduction by 44% (21-60%) in the three non-healthcare setting studies. 41 Several survey series regularly measure self-reported mask use but results from these different sources 42 have not previously been pooled to derive daily estimates of mask use over the course of the epidemic. 43

Global models of the impact of scaled up mask use have to our knowledge not been published. 44

Added value of this study 45 We combined the studies on mask use identified in the two meta-analyses and added one further study. 46

In a Bayesian meta-regression approach, we derived the effect of simple cloth or paper masks used 47 outside of a healthcare setting. In the meta-regression we make use of all the information provided by 48 all of these studies, rather than subsetting to just those studies that provided the direct comparison of 49 interest. Pooling estimates on the prevalence of self-reported mask use from three survey series 50 provides up-to-date information on trends in mask use in almost all countries. We use extensive survey 51 data covering nearly every country in the world to assess recent trends and current mask use. We then 52 use an SEIR transmission dynamics model with good predictive validity to assess the potential of scaled 53 up mask use to reduce global mortality from COVID- 19. 54 Introduction supermarket, going to a main road)" with responses "Always", "Frequently", "Sometimes", "Rarely", and 158 "Not at all". Respondents for "Always" were the numerator in our proportion. 159

Mask use for each location was estimated using a spline-based smoothing process. This smoothing 160 process averages each data point with five neighboring data points. To arrive at smooth, flat values at 161 the ends of the observed data, we computed the average of the change in mask use over the three 162 following days (left tail) and three preceding days (right tail). For locations without data on mask use, we 163 used, in order of preference, national level estimates (for subnational locations), regional estimates, and 164 super-regional estimates based on the regional groupings used by the Global Burden of Disease Study 165 (GBD). The only exception was for countries in Oceania, a region where no data are available through 166 any of the three survey platforms. In the GBD hierarchy, these countries are part of the East Asia and 167

Southeast Asia super-region; however, mask use in Oceania is likely to be more similar to mask use in 168

Australia and New Zealand and so the mask use from that region was used for countries in Oceania. 169

To construct our scenario of universal mask use, we assume that current mask use in all locations would 170 increase to 95% over the course of 7 days. We use 95% as that is the highest level of mask use reported 171 at the national level to date during the COVID-19 pandemic; this level was reported in Singapore. Our 172 "universal mask use" scenario assumes that this level of mask use can be achieved through the adoption 173 and enforcement of mask use mandates around the world. 174

COVID-19 SEIR model construction for each location 175 The IHME COVID-19 prediction model has been described in detail elsewhere. 64 For the results 176 presented in this analysis, the SEIR part of the model are most relevant. We construct an SEIR model for 177 each location we model; the Appendix shows the basic states included in the model and the transition 178 parameters. The critical driver of the epidemic is the rate at which susceptible individuals become 179 infected in each location which is assumed to be represented as: 180 . 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 October 11, 2020. . https://doi.org/10.1101/2020. 10 .08.20209510 doi: medRxiv preprint

Where βt is the transmission parameter for time period t, represents a mixing coefficient to account 181 for imperfect mixing within each location, S is the fraction of each location's population that is 182 susceptible, and I1 and I2 is the fraction that are infectious. Effective Rt, the number of new infections 183 caused by each case is a simple monotonic transformation of βt and the fraction of the population that is 184 susceptible. We use an efficient algorithm to directly estimate βt in the past -see appendix for details. 185

To determine the strength of the relationship between βt and various covariates, we perform a log-186 linear regression using the open source mixed effects solver SLIME 5 . All covariates are assumed to have 187 fixed effects while the intercept is allowed to vary by location. For location , the regression is calculated 188 as: 189

where 0, is the random intercept for location , is a matrix with a column for each covariate in the 192 regression and a row for each day, and is the coefficient indicating the strength of the relationship 193 between log and each of the covariates. We tested many covariates and included the following in the 194 model: population density measured as the share of the population living in areas with more than 2,500 195 individuals per square kilometer, the fraction of the population living below 100 meters above sea level, 196 smoking prevalence, particulate matter air pollution (PM2.5 population-weighted annual average 197 concentration), mobility measured using cell phone apps, mask use, COVID-19 testing per capita, and 198 pneumonia seasonality. Pneumonia seasonality was constructed as an index using medical certification 199 of cause of death data on pneumonia deaths by week and normalized annually. In locations without or 5-star quality cause of death data 65 , we used latitude as a predictor of the pattern of pneumonia 201 seasonality. To avoid over-estimating the effects of pneumonia seasonality and mask use we used 202 constraints on each in the regression -see Appendix for details. Specifically, for mask use, we did not let 203 . 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 October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 doi: medRxiv preprint the regression estimate an effect size larger than what was consistent with the mask use meta-analysis 204 of the individual level effect. To capture uncertainty in the input data, model parameters and regression 205 coefficients linking βjt to covariates, we generated 1,000 models for each location -see Appendix for 206 details. 207

We evaluated out of sample predictive validity for this modeling approach by holding out the last five 208 weeks of data and compared predictions from the held-out data to what occurred; median absolute 209 percent error for cumulative deaths at five weeks was 7%. 66 We also compared this model to other 210 COVID-19 prediction models that make their estimates publicly available; overall, we find that our model 211 has the best performance at 5-and 6-weeks out-of-sample. 66 212

We used the set of 1,000 SEIR models for each location to generate two types of scenarios: a reference 213 scenario and a universal mask use scenario. In the reference scenario, or what we think is most likely to 214 occur, key drivers such as mobility and testing per capita evolve according to past trends -see Appendix 215

for details. In the universal mask use scenario, we assume that mask mandates and other campaigns 216 lead to scale-up of mask use to 95% within seven days of enactment. We also assume both in the 217 reference and in the universal mask use scenario that social distancing mandates would be re-imposed 218 when the daily death rate reaches eight per million people per day. This daily death rate represents the 219 90 th percentile across countries of the observed daily death rate in the past few months before each 220 country imposed the maximum number of social distancing mandates. This daily death rate also 221 represents the observed average daily death rate to date among the small number of locations that are 222 experiencing a resurgence and are re-imposing social distancing mandates. 223 . 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 October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 doi: medRxiv preprint The meta-regression suggested the benefits of non-medical masks in the general population to be a 40% 231 (UI 20% -54%) reduction in transmission. The benefits of wearing surgical or medical masks in the 232 general population were slightly larger, a 43% reduction (23% -59%) in transmission. Even larger 233 reductions in transmission were estimated for non-medical (54% [40%-64%] and medical (56% [48%-234 64%]) mask use amongst healthcare worker populations. More details are provided in the Appendix. 235

Based on survey data collected through smartphone hosted questionnaires, Figure 2 shows a map of 236 mask use by location as of August 18 th , the last date of fully observed data in the model. Mask use is 237 high in most parts of Latin America and South-East and East Asia. The highest mask use on August 18 th 238 was in Chile (93.6%), followed by Puerto Rico (93.5%), and Guatemala (92.2%). The lowest rates are seen 239 in Northern Europe (Sweden, Norway, and Denmark < 1%) and North Africa (Tunisia 6.5%). Lower rates 240 are seen in some parts of sub-Saharan Africa, Northern Europe, and states and provinces in the United 241

States, Canada, and Australia. Mask use is highest among people who live in cities and lowest in people 242 who live in rural communities (see Appendix). 243 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 October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 doi: medRxiv preprint Latin America and the Caribbean (Figure 3 ). Mask use has increased in some locations where mandates 248 have been put into place such as Australia, Belgium, and the United Kingdom (SI Figure 7) . Mask use has 249 declined in some settings where death rates are declining such as Poland, Czechia, and Italy (SI Figure 7) . . 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 October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 doi: medRxiv preprint Figure 5 shows a map of the percent reduction in expected deaths from the reference scenario 271 compared to the mask use scenario in the deaths from August 28 th to January 1 st . The largest percent 272 reduction in expected deaths occurred in Tanzania This represents a 26.5% reduction in the number of deaths expected from August 28 th to January 1 st 284 2021. The benefits of increased mask use are greatest in settings with ongoing substantial transmission 285 and low current levels of mask use regardless of sociodemographic status. Our models also show that in 286 many settings, increased mask use will delay the need re-imposition of social distancing mandates by 287 many weeks or even months; in addition to the lives saved, a delay in re-imposition of social mandates 288 might also be accompanied by substantial economic benefits. 289

The estimate of 40% effectiveness of non-medical mask use by the general public is based on 13 290 observations (from 9 studies) specific to the general population, from a total of 63 observations from 40 291 separate studies included in the meta-regression. The uncertainty interval is wide, from 20% to 54% 292 reduction in transmission. Even with this uncertainty interval, there was absolutely no indication of any 293 . 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 October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 doi: medRxiv preprint harmful impacts of mask use, such as tendencies for engaging in riskier behaviour or self-contamination 294 via more frequent face touching, as has been suggested. 67, 68 Further, the published studies of SARS-CoV-295 2 included in our analysis of mask effectiveness demonstrated reductions in relative risk of 30% -100%, 296

with the one study of non-medical mask use amongst the general population, indicating a reduction of 297 42% for any mask use and 70% for consistent mask use. 69 Mask effectiveness is also supported by 298 additional evidence from laboratory studies that report on the efficacy of masks in reducing exhalation 299 of both aerosols and droplets by those infected with SARS CoV-2. 70 Further, a recent case series 300 reported no secondary infections among 139 individuals exposed to two symptomatic hair stylists with 301 confirmed COVID-19 while both the stylists and their clients wore masks. 71 In addition, a study of Swiss 302 soldiers indicated that physical distancing and use of medical masks led to no COVID-19 symptoms 303 despite the presence of virus-specific antibodies 72 , while a study of healthcare workers indicated that 304 universal use of medical masks was associated with lower rate of SARS-CoV-2 positive tests 73 . Based on 305 all types of available evidence, it seems critical to encourage mask use throughout the world as the 306 benefits can be substantial with low to zero contraindications. Guidance for specific materials, handling 307 of face coverings and other considerations is rapidly evolving as additional evidence emerges 74 . 308 At the population level, the regression analysis of the determinants of the transmission parameter 309 suggests a much larger effect of mask use than the one seen in the published studies. To avoid the risk 310 of over-stating the benefits of mask use in this analysis, however, we have constrained the regression to 311 yield results that are consistent with the range of the effect sizes found in the individual level studies. In 312 other words, the benefits of a universal mask use mandate could be substantially larger and what we 313 report here can be seen as a conservative estimate of the impact of mask use on lives saved. 314

Given that the cost of masks is very low, mask mandates and/or the promotion of mask use seems 315 prudent, as the risk of adopting these policies is minimal and the potential benefits very large. In 316 recognition of this, over the last few months we have seen the number of countries and territories with 317 . 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 October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 doi: medRxiv preprint mask mandates in place increase substantially. Nevertheless, there remains reluctance to adopt mask 318 use and to impose mask mandates. In some settings, the current epidemiological context means that 319 mask wearing is not viewed as a necessary part of control, such as in Norway where the Norwegian 320

Institute of Public Health determined that, given their current low prevalence, "200,000 people would 321 need to wear facemasks to prevent one new infection per week." 75 In other settings however, despite 322 increasing cases, public sentiment towards mask wearing hinders universal utilization. Past messages 323 from some governments have not encouraged mask use and may have actually discouraged mask use. 76 -324 78 Early on, WHO stated "the wide use of masks by healthy people in the community setting is not 325 supported by current evidence and carries uncertainties and critical risks" 2 and only changed their official 326 position on June 5 th to encourage mask use. 79 This reluctance to embrace mask use given no real risks of 327 use and considerable potential health and economic benefits is hard to understand and justify. For those 328 decision-makers who are concerned with the economic effects of social distancing mandates, mask use 329 mandates provide a low-cost strategy to reduce the risk of a further round of social distancing mandates 330 and the associated unemployment and economic downturn. 331

While the effective R, the number of new infections created by a single infection under the auspices of 332 control, can potentially be reduced by one-third through universal mask use compared to no mask use, 333 mandates alone will likely be insufficient to control the epidemic in many locations. Even with universal 334 mask use, we expect the death toll due to COVID-19 to reach 2.18 million deaths by the end of the year, 335 and many more in 2021, assuming an efficacious vaccine is not discovered, licensed and widely deployed 336 in the interim. Countries will have to consider other policy strategies to reduce transmission, including 337 increasing testing, contact tracing, and isolation, along with "smart mandates", which refers to targeting 338 mandates or restrictions to particular subgroups of the population, such as specific age-groups or local 339 communities for short periods of time. A central policy challenge for many countries is understanding 340 . 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 October 11, 2020. . which of these mandates targeting strategies makes the most sense in a given context and at what level 341 of COVID-19 transmission. This is the focus of our ongoing work. 342

The findings of this study should be interpreted while taking into account its limitations. First, there is a 343 set of limitations related to the meta-regression, including the following. The number of published 344 studies on the protection provided by cloth masks worn by the general public is limited. With rapid 345 development of the COVID-19 literature, new data on the effectiveness of masks can quickly be 346 incorporated into our meta-regression model. Future studies could change our pooled estimate of the 347 effect size and/or the large uncertainty interval around it in the meta-regression. As we looked at 348 multiple observations per study, it was not really feasible to account for all possible clustering. We 349 performed sensitivity analyses (shown in the Appendix) and found minimal differences in investigating 350 the role of clustering in our results. Studies had different endpoints and while we controlled for that in 351 our meta-regression, it would be ideal to have more studies that focus on COVID-19 as an endpoint. 352

Second, related to the modeling framework, we use an SEIR model to predict the course of the epidemic 353 with and without universal mask use. In general, SEIR models have tended to overestimate the 354 infections and deaths associated with COVID-19. Over-estimation is likely due to the fact that individuals 355 change their behaviours as the epidemic gets worse around them and governments tend to react when 356 hospital systems are nearing capacity. We have built the government response into our reference 357 scenario and have tried to use empirically observed data on mobility and current mask use to reflect the 358 individual behavioural response. Further, our model makes a number of simplifying assumptions 359 associated with mixing and transmission heterogeneity and as such, our conclusions must be considered 360 with these assumptions in mind. Also, we use a log-linear mixed regression which does not take into 361 account the potential for non-linear relationships We acknowledge that there are likely non-linear 362 relationships between some of the drivers of transmission and transmission intensity. Moreover, we 363 expect there to also be complex lagged relationships between covariates and transmission (e.g., fatigue 364 . 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 October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 doi: medRxiv preprint related to duration of mandate altering its impact on transmission). Improving how our model uses 365 covariates to capture temporal variation in new infections is an open avenue of research. 366

Third, our out-of-sample predictive validity testing has shown that errors tend to progressively get larger 367 the longer the forecast, but has also shown that errors are much larger in settings where there are 368 fewer than 50 total deaths to date. For example, predictions from publicly available models for sub-369

Saharan Africa have been particularly bad. 66 Fifth, our models are sensitive to the trends in the last 7-14 days in deaths and somewhat sensitive to 377 the trend in cases. In settings where deaths and cases are steadily rising, the model will tend to have 378 large estimates of βt. If the rise in transmission is not captured by trends in mobility or other covariates, 379 the unexplained residual in the model increases and this is then reflected in the forecasts by day 380 through to January 1 st . The reverse relationship also holds true for when there is a consistent downward 381 trend. The sensitivity of our model to data trends is a strength in that it makes our models reflect the 382 on-the-ground realities; it is also a challenge in the sense that our model results will change when there 383 are changes in recent transmission that are not captured by the covariates. 384 Sixth, we rely on self-reported data which is collected via mobile phone app-based surveys on use of 385 masks. In addition to the usual biases that accompany self-reported data, in this case we do not know 386 whether in settings with mask mandates in place, respondents may be reporting their behavior 387 . 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 October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 doi: medRxiv preprint differently compared to settings without mask mandates. The respondents to app-based surveys are 388 also likely to not be a truly representative sample of the populations in each location. The degree to 389 which the respondents represent the general population varies across locations, and depends on the 390 prevalence of Facebook and other app use in each country. While this is a limitation of the data that are 391 currently available, we believe that given the samples tend to be biased towards more educated, urban, 392 and younger populations, the reported mask use is likely to be an over-estimate in these locations. If 393 this is true, then the estimates of the impact of expanding mask use to reach 95% coverage in these 394 populations would be an underestimate of the true effect of the intervention. 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 October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 doi: medRxiv preprint future. We expect more deaths in the second half of 2020 than were seen in the first six months. Not 411 only are there large epidemics unfolding in Latin America, the Middle-East and South Asia, but 412 seasonality suggests a second wave can be anticipated in the Northern Hemisphere. The rising toll of the 413 COVID-19 pandemic can be reduced by 0.82 million deaths in the next few months by the adoption of 414 universal mask mandates. This low-cost intervention that is available and accessible to all populations, 415 regardless of socio-economic status or other dimensions of inequity, has enormous health benefits and 416 might also lead to large economic benefits by delaying the need for re-imposition of social distancing 417 mandates. In global health we rarely encounter effective, low-cost, and universally available 418 interventions that can save lives: immediately, equitably and safely. Ensuring that individuals, as well as, 419 local, national and global decision makers are all doing everything in their power to achieve the highest 420 rates of mask use in all exposed populations is one of the best strategies available to us to mitigate the 421 toll of the pandemic in the months to come. 422 423 . 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 October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 doi: medRxiv preprint 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 October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 doi: medRxiv preprint 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 October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 doi: medRxiv preprint 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 October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 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.

The copyright holder for this preprint this version posted October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 doi: medRxiv preprint Beijing 0 (0 to 0) 0% (-0.1 to 0%) 10 (10 to 10)

Chongqing 0 (0 to 0) 0% (-0.1 to 0%) 7 (6 to 7)

Guangdong 0 (0 to 0) -0.1% (-0.1 to 0%) 10 (10 to 10)

Hainan 0 (0 to 0) 0% (-0.1 to 0%) 7 (7 to 7)

Hebei 0 (0 to 0) 0% (-0.1 to 0%) 7 (7 to 7)

Heilongjiang 0 (0 to 0) 0% (0 to 0%) 15 

Our study follows the Guidelines for Accurate and Transparent Health Estimate Reporting. All code used for these analyses is publicly available online (http://github.com/ihmeuw/).

Results for each scenario are accessible through a visualization tool at http://covid19.healthdata.org.

The estimates presented in this tool will be iteratively updated as new data are incorporated and will ultimately supersede the results in this paper.

We thank the various Departments of Health and frontline health professionals who are not only responding to this epidemic daily, but also provide the necessary data to inform this work -IHME wishes to warmly acknowledge the support of these and others (http://www.healthdata.org/covid/acknowledgements) who have made our COVID-19 estimation efforts possible. This work was supported by the Bill & Melinda Gates Foundation, Bloomberg

Philanthropies, as well as funding from the state of Washington and the National Science Foundation (2031096). We also extend a note of particular thanks to John Stanton and Julie Nordstrom for their generous support. We are grateful to Professor Wei Huang from Peking University for helping us extract information from scientific papers in Chinese.

This study was funded by the Bill & Melinda Gates Foundation and Bloomberg Philanthropies. The funders of the study had no role in study design, data collection, data analysis, data interpretation, writing of the final report, or decision to publish. The corresponding author had full access to all of the data in the study and had final responsibility for the decision to submit for publication.

. 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 October 11, 2020. . https://doi.org/10.1101/2020.10.08.20209510 doi: medRxiv preprint

An interactive web-based dashboard to track COVID-19 in real time

Advice on the use of masks in the context of COVID-19: Interim guidance

Face masks for the public during the covid-19 crisis

Efficacy of face mask in preventing respiratory virus transmission: A systematic review and meta-analysis

Physical distancing, face masks, and eye protection to prevent personto-person transmission of SARS-CoV-2 and COVID-19: a systematic review and meta-analysis

Community Initiatives. What Countries Require Masks in Public or Recommend Masks? #Masks4All

Judge Scolds Brazil's President for Violating Mask Law. The New York Times

How did face masks become a political issue in America? the Guardian

Removal of Nanoparticles from Gas Streams by Fibrous Filters: A Review

Filtration Mechanisms of Particulate Respirators

World Health Organization. Transmission of SARS-CoV-2: implications for infection prevention precautions

Particle sizes of infectious aerosols: implications for infection control

Airborne transmission of SARS-CoV-2: The world should face the reality

Reduction of secondary transmission of SARS-CoV-2 in households by face mask use, disinfection and social distancing: a cohort study in Beijing, China

Seroprevalence of 2009 H1N1 Virus Infection and Self-Reported Infection Control Practices Among Healthcare Professionals Following the First Outbreak in

Preliminary Findings of a Randomized Trial of Non-Pharmaceutical Interventions to Prevent Influenza Transmission in Households

Facemasks and Hand Hygiene to Prevent Influenza Transmission in Households

A cluster randomized clinical trial comparing fit-tested and non-fit-tested N95 respirators to medical masks to prevent respiratory virus infection in health care workers

Rapid awareness and transmission of severe acute respiratory syndrome in Hanoi French Hospital, Vietnam

Asymptomatic SARS Coronavirus Infection among Healthcare Workers

Factors associated with transmission of severe acute respiratory syndrome among health-care workers in Singapore

Effectiveness of personal protective measures in prevention of nosocomial transmission of severe acute respiratory syndrome

Factors associated with the transmission of pandemic (H1N1) 2009 among hospital healthcare workers in Beijing

Probable secondary infections in households of SARS patients in Hong Kong

Risk Factors for SARS among Persons without Known Contact with SARS Patients

Risk factors for SARS infection among hospital healthcare workers in Beijing: a case control study

Association between 2019-nCoV transmission and N95 respirator use

Epidemiological characteristics of COVID-19 in medical staff members of neurosurgery departments in Hubei province: A multicentre descriptive study

Risk factors for SARS infection within hospitals in Hanoi, Vietnam

Factors associated with nosocomial SARS-CoV transmission among healthcare workers in Hanoi

Effectiveness of precautions against droplets and contact in prevention of nosocomial transmission of severe acute respiratory syndrome (SARS)

Mild illness associated with severe acute respiratory syndrome coronavirus infection: lessons from a prospective seroepidemiologic study of health-care workers in a teaching hospital in Singapore

Investigation of the influencing factors on severe acute respiratory syndrome among health care workers

Surveillance of the Middle East respiratory syndrome (MERS) coronavirus (CoV) infection in healthcare workers after contact with confirmed MERS patients: incidence and risk factors of MERS-CoV seropositivity

Transmission among healthcare worker contacts with a Middle East respiratory syndrome patient in a single Korean centre

Prevention of nosocomial transmission of swine-origin pandemic influenza virus A/H1N1 by infection control bundle

Factors associated with transmission of Middle East respiratory syndrome among Korean healthcare workers: infection control via extended healthcare contact management in a secondary outbreak hospital

Risk of transmission via medical employees and importance of routine infection-prevention policy in a nosocomial outbreak of Middle East respiratory syndrome (MERS): a descriptive analysis from a tertiary care hospital in South Korea

Lack of SARS Transmission among Public Hospital Workers

SARS transmission in Vietnam outside of the health-care setting

SARS among Critical Care Nurses

Illness in Intensive Care Staff after Brief Exposure to Severe Acute Respiratory Syndrome

Transmission of COVID-19 to Health Care Personnel During Exposures to a Hospitalized Patient

Lack of SARS Transmission among Healthcare Workers, United States

Lack of SARS transmission and U.S. SARS case-patient

Enhanced Contact Investigations for Nine Early Travel-Related Cases of SARS-CoV-2 in the United States

Pilot Randomised Controlled Trial to Test Effectiveness of Facemasks in Preventing Influenza-like Illness Transmission among Australian Hajj Pilgrims in 2011

The role of facemasks and hand hygiene in the prevention of influenza transmission in households: results from a cluster randomised trial

Universal Mask Usage for Reduction of Respiratory Viral Infections After Stem Cell Transplant: A Prospective Trial

Protection by Face Masks against Influenza A(H1N1)pdm09 Virus on Trans-Pacific Passenger Aircraft

Risk Factors for Middle East Respiratory Syndrome Coronavirus Infection among Healthcare Personnel

Transmission of 2009 pandemic influenza A (H1N1) virus among healthcare personnel-Southern California

Health care worker contact with MERS patient, Saudi Arabia

Study on correlation between nosocomial respiratory virus infection and use of disposable respirator in medical staff

Hand Hygiene, and Influenza among Young Adults: A Randomized Intervention Trial

Seroprevalence of Middle East respiratory syndrome coronavirus (MERS-CoV) in public health workers responding to a MERS outbreak in Seoul, Republic of Korea

A case-control study on the risk factors of severe acute respiratory syndromes among health care workers

Trimmed Constrained Mixed Effects Models: Formulations and Algorithms

COVID-19 World Survey Data API

Are Americans Wearing Face Masks? Premise

Imperial College London Big Data Analytical Unit, YouGov PIc

COVID-19 scenarios for the United States

Global, regional, and national age-sex-specific mortality for 282 causes of death in 195 countries and territories, 1980-2017: a systematic analysis for the Global Burden of Disease Study

IHME COVID-19 Model Comparison Team. Predictive performance of international COVID-19 mortality forecasting models

Report on face masks' effectiveness for Covid-19 divides scientists. The Guardian

Face Masks for the General Public

Reduction of secondary transmission of SARS-CoV-2 in households by face mask use, disinfection and social distancing: a cohort study in Beijing, China

Respiratory virus shedding in exhaled breath and efficacy of face masks

Absence of Apparent Transmission of SARS-CoV-2 from Two Stylists After Exposure at a Hair Salon with a Universal Face Covering Policy

Social distancing alters the clinical course of COVID-19 in young adults: A comparative cohort study

Association Between Universal Masking in a Health Care System and SARS-CoV-2 Positivity Among Health Care Workers

Use of masks to help slow the spread of COVID-19

Norwegian Institute of Public Health. Should individuals in the community without respiratory symptoms wear facemasks to reduce the spread of COVID-19?

Avoid busy places and stay 1.5 metres away from others

The Public Health Agency of Sweden. FAQ about COVID-19

Coronavirus: Why some countries wear face masks and others don't -BBC News

Advice on the use of masks in the context of COVID-19: Interim guidance