Abstract
Background Social distancing mandates have been effective at reducing the health impacts of COVID-19. The ensuing economic downturns and unemployment increases have led many nations to progressively relax mandates. As COVID-19 transmission and deaths rise in many low and middle-income countries (LMICs), with continuing widespread transmission elsewhere, policymakers are searching for options to reduce COVID-19 mortality without re-imposing 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 COVID-19 by 40% (95% uncertainty interval [UI] 20% – 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 COVID-19 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.
Evidence before this study One meta-analysis of 21 studies reported a pooled reduction in the risk of respiratory virus infection of 47% (95% CI 36-79%) from a subset of eight studies reporting on mask use in non-health workers but it did not distinguish type of mask. Another meta-analysis reported on 26 studies of mask use in health workers and three studies in non-healthcare settings, reported a pooled effect of a 66% (55-74%) reduction in infections and a reduction by 44% (21-60%) in the three non-healthcare setting studies. Several survey series regularly measure self-reported mask use but results from these different sources have not previously been pooled to derive daily estimates of mask use over the course of the epidemic. Global models of the impact of scaled up mask use have to our knowledge not been published.
Added value of this study We combined the studies on mask use identified in the two meta-analyses and added one further study. In a Bayesian meta-regression approach, we derived the effect of simple cloth or paper masks used outside of a healthcare setting. In the meta-regression we make use of all the information provided by all of these studies, rather than subsetting to just those studies that provided the direct comparison of interest. Pooling estimates on the prevalence of self-reported mask use from three survey series provides up-to-date information on trends in mask use in almost all countries. We use extensive survey data covering nearly every country in the world to assess recent trends and current mask use. We then use an SEIR transmission dynamics model with good predictive validity to assess the potential of scaled up mask use to reduce global mortality from COVID-19.
Implications of all the available evidence Universal mask use can save many lives and avoid or, at least, delay the need for re-imposition of social mandates (such as stay-at-home orders, curfews, etc.), which would also contribute to ameliorating the negative effects of COVID-19 on the economy and unemployment. Until an affordable vaccine becomes universally available, mandating mask use is the most attractive policy option available to all countries, particularly if an expected increased transmission risk occurs in the Northern hemisphere’s fall and winter. Simple face coverings are cheap and effective; one of the few available interventions that is widely available to everyone. Countries with currently low mask use will need to determine the optimal balance between encouraging the use of masks through advocacy and information about their benefits, and governance of a compulsory use associated with penalties for non-compliance.
Introduction
COVID-19 has spread to all regions of the world and as of August 30th 2020, with over 25 million cases and 845,414 deaths have been reported globally1. This is undoubtedly an underestimate and only one direct measure, among many, that may be used to refer to the impact of the pandemic on health and health systems around the world. In March 2020, nearly all nations put in place a set of social distancing mandates that have contributed to blunting the effects of COVID-19. Economic downturns and the associated mass unemployment caused by these measures has led most nations to progressively relax social distancing mandates. As COVID-19 transmission and deaths rise in many low and middle-income countries (LMICs) and some high-income countries, policymakers are keen to identify policy options to reduce COVID-19 mortality without re-imposing strict social distancing mandates.
One attractive strategy is the imposition of mandates requiring the wearing of masks in public spaces when physical distancing is not feasible. Initially, the World Health Organization (WHO) discouraged mask use by public questioning of the evidence supporting general use and arguing it might lead to mask shortages for healthcare providers2,3. Several systematic reviews, however, have suggested that cloth and other non-medical masks worn by the general public can markedly reduce transmission.4,5 A growing number of national and local governments have recently adopted mask use mandates, with now over 150 countries mandating mask use either nationally, or in specific transmission hotspots, focused in enforcing their use within on public transit or indoor public spaces.6 In nations such as the United States (US) and Brazil, mask use has become a political issue, with widely varying viewpoints on the ethics, legality and enforcement of mask use mandates.7,8 Evidence of the individual benefits, population health impacts, and the potential economic benefits of increased mask use may be a useful input in these national debates.
Masks can affect the transmission of respiratory pathogens in a number of ways, thereby providing a complement to stringent social distancing mandates without the associated severe economic impacts, and in settings such as indoor public spaces where distancing may not be possible. Simple cloth masks are also affordable and equitably accessible. At the individual level masks reduce transmission via a physical barrier whereby the force of exhalation, coughing, or sneezing leads to impaction and interception of viral droplets (> 5μm in diameter) and aerosol (≤ 5 μm in diameter, resulting from evaporation of expired droplets, also described as droplet nuclei) onto the fibers of the mask.9–11 Given greater impaction efficiency for droplets, masks are most effective in blocking outward exhalation12, although evidence from studies of air pollution aerosol indicates some protection of non-medical masks in blocking a small fraction of aerosol inhalation.13 Thus mask wearing is effective both for the infected and the uninfected14 and is also relevant to the recent discussions regarding airborne transmission of SARS-CoV-2.15 Airborne transmission is, by definition, caused by droplet nuclei (i.e. evaporated droplets) that may remain infectious when suspended in the air over extended distances or time periods.11 As masks can effectively block exhalation of droplets, they also reduce production of droplet nuclei by evaporation, and therefore offer a practical intervention against airborne transmission to complement ventilation and less widely accessible interventions such as air filtration and UV disinfection.
In this paper, we provide a meta-regression of the reduction in transmission associated with mask use, use survey data to assess current levels and trends in use of masks globally, and quantify the benefits of universal mask use (here defined as 95% of individuals always wearing a mask when outside their home) on COVID-19 mortality using a COVID-19 transmission dynamics model. We provide estimates of lives saved through universal mask use globally and at the national level between July 26th 2020 and January 1st 2021.
Methods
Our analysis is comprised of three main components: a literature scoping analysis and meta-regression of the benefits of mask use with an emphasis on non-medical mask use amongst the general (non-healthcare) population; an analysis of survey data on the levels and trends in mask use; and modeling two scenarios (reference and universal mask use) of COVID-19 incidence and death using a deterministic transmission dynamics model with categories for susceptible, exposed, infected and recovered (SEIR).
Each component is discussed in more detail below.
Estimating the effectiveness of masks in preventing transmission
We performed a meta-regression of 63 observations from 40 studies of the effectiveness of masks in preventing the transmission of respiratory virus infections. These studies were included in two published meta-analyses4,5 with one further study of mask use for COVID-19 in the general population added.16 The studies varied in setting (general population versus healthcare), type of mask (which we dichotomized into medical-grade masks, including surgical and N95 masks, and non-medical masks, including cloth masks), comparator group (no mask use or “occasional” mask use), type of diagnosis (clinical or laboratory), country of study (dichotomized into Asian and non-Asian countries) and type of respiratory virus (SARS-CoV-1 or 2 versus H1N1, influenza, or other respiratory viruses); details on the details of all included studies16–55 can be found in SI Table 2 in the Appendix. From the identified papers, we extracted all relevant observations that assessed mask effectiveness, allowing for multiple observations per study based on variations in mask type, virus studied, or comparison group. From the two meta-analyses, four identified studies were excluded where a relative risk was not available56,57, we were unable to extract mask use from general PPE use58, or because the comparison group was of less protective masks rather than no or infrequent mask use59. In order to derive the most relevant pooled estimate for the effect of mask use on preventing the spread of COVID-19 in populations, we performed a meta-regression of all 63 observations and their characteristics to predict the effect of non-medical mask use in a community setting to prevent laboratory-confirmed SARS-CoV-2 compared to no use of masks. For studies with zero counts in the numerator, we used a continuity correction of 0.001 to estimate the RR in order to ensure inclusion in our analyses; sensitivity analyses considering other continuity corrections produced highly consistent results (see Appendix). To generate summary estimates we used a custom Bayesian mixed-effects meta-regression tool (MR-BRT – “meta-regression – Bayesian, regularized, trimmed”)60 which accounted for between-study heterogeneity in the width of the uncertainty interval (see Appendix for more details).
Past, current and future prevalence of mask use
We used three main sources of data on self-reported use of masks: the Facebook Global Symptom Survey61, the PREMISE surveys62, and the YouGov COVID-19 Behaviour Tracker surveys.63 Between April 23rd and August 18th, Facebook has surveyed 23.3 million Facebook users from 102 countries using an instrument with multiple items on behaviours related to COVID-19, including mask use. For the US, we used data collected through PREMISE. There were 190,216 total PREMISE responses representing all 50 states and the District of Columbia with responses collected between April 21st and August 21st. The YouGov surveys cover 29 countries and have interviewed around 493,400 individuals since March 1st and up until August 8th 2020. From the Facebook surveys, we used the item: “In the last 7 days, how often did you wear a mask when in public?” to which there are the following responses “All of the time; Most of the time; About half of the time; Sometimes; Never; I have not been in public during the last 7 days”. Respondents for “All of the time” were the numerator in our proportion. From the PREMISE surveys, we use the following question: “When you leave your home do you typically wear a face mask (SELECT_ONE)” with responses “Yes, always; Yes, sometimes; No never”. Respondents for “Yes, always” were the numerator in our proportion. From YouGov, we use the following question: “Thinking about the last 7 days, have you worn a face mask outside your home (e.g. when on public transport, going to a supermarket, going to a main road)” with responses “Always”, “Frequently”, “Sometimes”, “Rarely”, and “Not at all”. Respondents for “Always” were the numerator in our proportion.
Mask use for each location was estimated using a spline-based smoothing process. This smoothing process averages each data point with five neighboring data points. To arrive at smooth, flat values at the ends of the observed data, we computed the average of the change in mask use over the three following days (left tail) and three preceding days (right tail). For locations without data on mask use, we used, in order of preference, national level estimates (for subnational locations), regional estimates, and super-regional estimates based on the regional groupings used by the Global Burden of Disease Study (GBD). The only exception was for countries in Oceania, a region where no data are available through any of the three survey platforms. In the GBD hierarchy, these countries are part of the East Asia and Southeast Asia super-region; however, mask use in Oceania is likely to be more similar to mask use in Australia and New Zealand and so the mask use from that region was used for countries in Oceania.
To construct our scenario of universal mask use, we assume that current mask use in all locations would increase to 95% over the course of 7 days. We use 95% as that is the highest level of mask use reported at the national level to date during the COVID-19 pandemic; this level was reported in Singapore. Our “universal mask use” scenario assumes that this level of mask use can be achieved through the adoption and enforcement of mask use mandates around the world.
COVID-19 SEIR model construction for each location
The IHME COVID-19 prediction model has been described in detail elsewhere.64 For the results presented in this analysis, the SEIR part of the model are most relevant. We construct an SEIR model for each location we model; the Appendix shows the basic states included in the model and the transition parameters. The critical driver of the epidemic is the rate at which susceptible individuals become infected in each location which is assumed to be represented as:
Where βt is the transmission parameter for time period t, α represents a mixing coefficient to account for imperfect mixing within each location, S is the fraction of each location’s population that is susceptible, and I1 and I2 is the fraction that are infectious. Effective Rt, the number of new infections caused by each case is a simple monotonic transformation of βt and the fraction of the population that is susceptible. We use an efficient algorithm to directly estimate βt in the past – see appendix for details. To determine the strength of the relationship between βt and various covariates, we perform a log-linear regression using the open source mixed effects solver SLIME5. All covariates are assumed to have fixed effects while the intercept is allowed to vary by location. For location l, the regression is calculated as:
where α0,l is the random intercept for location l, Xl is a matrix with a column for each covariate in the regression and a row for each day, and α is the coefficient indicating the strength of the relationship between log β and each of the covariates. We tested many covariates and included the following in the model: population density measured as the share of the population living in areas with more than 2,500 individuals per square kilometer, the fraction of the population living below 100 meters above sea level, smoking prevalence, particulate matter air pollution (PM2.5 population-weighted annual average concentration), mobility measured using cell phone apps, mask use, COVID-19 testing per capita, and pneumonia seasonality. Pneumonia seasonality was constructed as an index using medical certification of cause of death data on pneumonia deaths by week and normalized annually. In locations without 4- or 5-star quality cause of death data65, we used latitude as a predictor of the pattern of pneumonia seasonality. To avoid over-estimating the effects of pneumonia seasonality and mask use we used constraints on each in the regression – see Appendix for details. Specifically, for mask use, we did not let the regression estimate an effect size larger than what was consistent with the mask use meta-analysis of the individual level effect. To capture uncertainty in the input data, model parameters and regression coefficients linking βjt to covariates, we generated 1,000 models for each location – see Appendix for details.
We evaluated out of sample predictive validity for this modeling approach by holding out the last five weeks of data and compared predictions from the held-out data to what occurred; median absolute percent error for cumulative deaths at five weeks was 7%.66 We also compared this model to other COVID-19 prediction models that make their estimates publicly available; overall, we find that our model has the best performance at 5- and 6-weeks out-of-sample.66
We used the set of 1,000 SEIR models for each location to generate two types of scenarios: a reference scenario and a universal mask use scenario. In the reference scenario, or what we think is most likely to occur, key drivers such as mobility and testing per capita evolve according to past trends – see Appendix for details. In the universal mask use scenario, we assume that mask mandates and other campaigns lead to scale-up of mask use to 95% within seven days of enactment. We also assume both in the reference and in the universal mask use scenario that social distancing mandates would be re-imposed when the daily death rate reaches eight per million people per day. This daily death rate represents the 90th percentile across countries of the observed daily death rate in the past few months before each country imposed the maximum number of social distancing mandates. This daily death rate also represents the observed average daily death rate to date among the small number of locations that are experiencing a resurgence and are re-imposing social distancing mandates.
Results
Figure 1 shows the relative risk of viral respiratory infection among users of different types of facemasks. Studies of general population use (13 observations of 63 total) included cluster randomized trials of household member mask use living with an infected individual, cohort analyses of close contacts of infected individuals, and case control analyses of mask use prior to infection amongst the general population or of secondary infections amongst household members of an infected individual. A total of 19 observations were of non-medical masks, with eight of these focused on use by the general public. The meta-regression suggested the benefits of non-medical masks in the general population to be a 40% (UI 20% – 54%) reduction in transmission. The benefits of wearing surgical or medical masks in the general population were slightly larger, a 43% reduction (23% - 59%) in transmission. Even larger reductions in transmission were estimated for non-medical (54% [40%-64%] and medical (56% [48%-64%]) mask use amongst healthcare worker populations. More details are provided in the Appendix.
This analysis includes only variables addressing facemask type (non-medical versus medical) and population type of mask user (general population versus healthcare setting). The size of the box is proportional to the precision of the estimate, based on number of observations, with more precise studies having larger boxes.
Based on survey data collected through smartphone hosted questionnaires, Figure 2 shows a map of mask use by location as of August 18th, the last date of fully observed data in the model. Mask use is high in most parts of Latin America and South-East and East Asia. The highest mask use on August 18th was in Chile (93.6%), followed by Puerto Rico (93.5%), and Guatemala (92.2%). The lowest rates are seen in Northern Europe (Sweden, Norway, and Denmark < 1%) and North Africa (Tunisia 6.5%). Lower rates are seen in some parts of sub-Saharan Africa, Northern Europe, and states and provinces in the United States, Canada, and Australia. Mask use is highest among people who live in cities and lowest in people who live in rural communities (see Appendix).
Proportion of the population that self-report always wearing a facemask when outside the home on July 21, 2020.
Figure 3 shows the global and super-region trends in mask use since the beginning of the epidemic. Mask use data start only in the beginning of April so the rapid expansion from likely a very low baseline pre-COVID-19 outside of East Asia most probably occurred in March. 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 (Figure 3). Mask use has increased in some locations where mandates have been put into place such as Australia, Belgium, and the United Kingdom (SI Figure 7). Mask use has declined in some settings where death rates are declining such as Poland, Czechia, and Italy (SI Figure 7). Mask use varied by gender in some regions, with higher use by females in all regions except for South Asia, (globally 54.2% in females vs to 52.3% in males) and by age with generally lower use amongst 18-24 year olds (50.0%) and those over 65 years (46.4%) compared to 55-54 year old adults (56.5%) who were the most likely to wear masks (percentages on July 21, see Appendix).
Proportion of the population that self-report always wearing a facemask when outside the home by Global Burden of Disease study super-region between March 1, 2020 and September 30, 2020. Values after the last date of observed survey data are dashed lines and horizontal projections from the last observed values.
Global projections of COVID-19 deaths with and without a universal mask use mandate are shown in Figure 4. In our reference scenario, we expect daily COVID-19 deaths to increase in October through December leading to 3.00 million deaths (95% UI 2.20 to 4.52 million) by January 1st 2021. Uncertainty in daily deaths and cumulative deaths widens steadily over time such that there is considerable uncertainty in global deaths by January 1st. In contrast, the universal mask use scenario leads to a mean estimate of 2.18 million deaths (95% UI 1.71 to 3.14 million) by January 1st. The difference in mortality between the reference and universal mask use scenario suggests that 0.82 million lives (95% UI 0.43 to 1.49 million) could be saved over this time period if 95% of people were to always wear masks when outside their home. Table 1 provides estimates by location of the expected deaths in the reference and universal mask use scenario along with the number of deaths saved through the mask use mandate. The countries where mask mandates would have the largest effect are populous 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). The greatest magnitude difference in mortality rate occurred in the Netherlands (23,282 fewer deaths, 136 deaths per 100,000), Switzerland (6,062 fewer deaths, 86 deaths per 100,000).
The difference in cumulative deaths between the universal mask use scenario and the reference (current mask use) scenario globally and for each GBD super-region, region, country, and first administrative subnational on January 1, 2021.
Projected global deaths due to COVID-19 up to January 1, 2021 in the scenario for current projections of mask use and the scenario with 95% coverage of mask use. Projections are shown for the world and for the seven super-regions as defined in the Global Burden of Disease Study.
Figure 5 shows a map of the percent reduction in expected deaths from the reference scenario compared to the mask use scenario in the deaths from August 28th to January 1st. The largest percent reduction in expected deaths occurred in Tanzania (80.4% difference, 95% UI 63.4 to 94.4%), Poland (78.8% difference, 95% UI 60.4 to 93.5%), and Algeria (78.4% difference, 95% UI 61.9 to 92.2%). Expected deaths were greater in the universal mask use scenario than in the reference scenario in some locations due to the delay in re-imposition of mandates such as in Kazahkstan (more deaths in mask use scenario, 95% UI 7.2 to 45.3% increase). There are large variations in the number of lives saved globally, regionally, nationally, and subnationally and these are detailed in full in Table 1.
Areas have “no estimates” either because of no available data or because the population size, cases or deaths are so small that the SEIR models do not run.
Discussion
Wearing a mask reduces the risk of contracting COVID-19 for individuals by 40% (20%-54%). Given global mask use is currently at 59.0%, increasing mask use to 95% through mandates could decrease cumulative COVID-19 deaths by January 1st 2021 by 0.82 million deaths (95% UI 0.43 to 1.49 million). This represents a 26.5% reduction in the number of deaths expected from August 28th to January 1st 2021. The benefits of increased mask use are greatest in settings with ongoing substantial transmission and low current levels of mask use regardless of sociodemographic status. Our models also show that in many settings, increased mask use will delay the need re-imposition of social distancing mandates by many weeks or even months; in addition to the lives saved, a delay in re-imposition of social mandates might also be accompanied by substantial economic benefits.
The estimate of 40% effectiveness of non-medical mask use by the general public is based on 13 observations (from 9 studies) specific to the general population, from a total of 63 observations from 40 separate studies included in the meta-regression. The uncertainty interval is wide, from 20% to 54% reduction in transmission. Even with this uncertainty interval, there was absolutely no indication of any harmful impacts of mask use, such as tendencies for engaging in riskier behaviour or self-contamination via more frequent face touching, as has been suggested.67,68 Further, the published studies of SARS-CoV-2 included in our analysis of mask effectiveness demonstrated reductions in relative risk of 30% - 100%, with the one study of non-medical mask use amongst the general population, indicating a reduction of 42% for any mask use and 70% for consistent mask use.69 Mask effectiveness is also supported by additional evidence from laboratory studies that report on the efficacy of masks in reducing exhalation of both aerosols and droplets by those infected with SARS CoV-2.70 Further, a recent case series reported no secondary infections among 139 individuals exposed to two symptomatic hair stylists with confirmed COVID-19 while both the stylists and their clients wore masks.71 In addition, a study of Swiss soldiers indicated that physical distancing and use of medical masks led to no COVID-19 symptoms despite the presence of virus-specific antibodies72, while a study of healthcare workers indicated that universal use of medical masks was associated with lower rate of SARS-CoV-2 positive tests73. Based on all types of available evidence, it seems critical to encourage mask use throughout the world as the benefits can be substantial with low to zero contraindications. Guidance for specific materials, handling of face coverings and other considerations is rapidly evolving as additional evidence emerges74.
At the population level, the regression analysis of the determinants of the transmission parameter suggests a much larger effect of mask use than the one seen in the published studies. To avoid the risk of over-stating the benefits of mask use in this analysis, however, we have constrained the regression to yield results that are consistent with the range of the effect sizes found in the individual level studies. In other words, the benefits of a universal mask use mandate could be substantially larger and what we report here can be seen as a conservative estimate of the impact of mask use on lives saved.
Given that the cost of masks is very low, mask mandates and/or the promotion of mask use seems prudent, as the risk of adopting these policies is minimal and the potential benefits very large. In recognition of this, over the last few months we have seen the number of countries and territories with mask mandates in place increase substantially. Nevertheless, there remains reluctance to adopt mask use and to impose mask mandates. In some settings, the current epidemiological context means that mask wearing is not viewed as a necessary part of control, such as in Norway where the Norwegian Institute of Public Health determined that, given their current low prevalence, “200,000 people would need to wear facemasks to prevent one new infection per week.”75 In other settings however, despite increasing cases, public sentiment towards mask wearing hinders universal utilization. Past messages from some governments have not encouraged mask use and may have actually discouraged mask use.76–78 Early on, WHO stated “the wide use of masks by healthy people in the community setting is not supported by current evidence and carries uncertainties and critical risks”2 and only changed their official position on June 5th to encourage mask use.79 This reluctance to embrace mask use given no real risks of use and considerable potential health and economic benefits is hard to understand and justify. For those decision-makers who are concerned with the economic effects of social distancing mandates, mask use mandates provide a low-cost strategy to reduce the risk of a further round of social distancing mandates and the associated unemployment and economic downturn.
While the effective R, the number of new infections created by a single infection under the auspices of control, can potentially be reduced by one-third through universal mask use compared to no mask use, mandates alone will likely be insufficient to control the epidemic in many locations. Even with universal mask use, we expect the death toll due to COVID-19 to reach 2.18 million deaths by the end of the year, and many more in 2021, assuming an efficacious vaccine is not discovered, licensed and widely deployed in the interim. Countries will have to consider other policy strategies to reduce transmission, including increasing testing, contact tracing, and isolation, along with “smart mandates”, which refers to targeting mandates or restrictions to particular subgroups of the population, such as specific age-groups or local communities for short periods of time. A central policy challenge for many countries is understanding which of these mandates targeting strategies makes the most sense in a given context and at what level of COVID-19 transmission. This is the focus of our ongoing work.
The findings of this study should be interpreted while taking into account its limitations. First, there is a set of limitations related to the meta-regression, including the following. The number of published studies on the protection provided by cloth masks worn by the general public is limited. With rapid development of the COVID-19 literature, new data on the effectiveness of masks can quickly be incorporated into our meta-regression model. Future studies could change our pooled estimate of the effect size and/or the large uncertainty interval around it in the meta-regression. As we looked at multiple observations per study, it was not really feasible to account for all possible clustering. We performed sensitivity analyses (shown in the Appendix) and found minimal differences in investigating the role of clustering in our results. Studies had different endpoints and while we controlled for that in our meta-regression, it would be ideal to have more studies that focus on COVID-19 as an endpoint.
Second, related to the modeling framework, we use an SEIR model to predict the course of the epidemic with and without universal mask use. In general, SEIR models have tended to overestimate the infections and deaths associated with COVID-19. Over-estimation is likely due to the fact that individuals change their behaviours as the epidemic gets worse around them and governments tend to react when hospital systems are nearing capacity. We have built the government response into our reference scenario and have tried to use empirically observed data on mobility and current mask use to reflect the individual behavioural response. Further, our model makes a number of simplifying assumptions associated with mixing and transmission heterogeneity and as such, our conclusions must be considered with these assumptions in mind. Also, we use a log-linear mixed regression which does not take into account the potential for non-linear relationships We acknowledge that there are likely non-linear relationships between some of the drivers of transmission and transmission intensity. Moreover, we expect there to also be complex lagged relationships between covariates and transmission (e.g., fatigue related to duration of mandate altering its impact on transmission). Improving how our model uses covariates to capture temporal variation in new infections is an open avenue of research.
Third, our out-of-sample predictive validity testing has shown that errors tend to progressively get larger the longer the forecast, but has also shown that errors are much larger in settings where there are fewer than 50 total deaths to date. For example, predictions from publicly available models for sub-Saharan Africa have been particularly bad.66
Fourth, we assume cases and deaths are accurately reported by JHU except in Ecuador, Kazakhstan and where excess mortality analyses have indicated substantial underreporting of COVID-19 deaths. Other countries may not be detecting or reporting deaths and cases due to lack of testing or other considerations. Overall, though, we believe that our prediction model performs well and where antibody tests have been conducted at the population level, we have found our model based on deaths has matched these results.64
Fifth, our models are sensitive to the trends in the last 7-14 days in deaths and somewhat sensitive to the trend in cases. In settings where deaths and cases are steadily rising, the model will tend to have large estimates of βt. If the rise in transmission is not captured by trends in mobility or other covariates, the unexplained residual in the model increases and this is then reflected in the forecasts by day through to January 1st. The reverse relationship also holds true for when there is a consistent downward trend. The sensitivity of our model to data trends is a strength in that it makes our models reflect the on-the-ground realities; it is also a challenge in the sense that our model results will change when there are changes in recent transmission that are not captured by the covariates.
Sixth, we rely on self-reported data which is collected via mobile phone app-based surveys on use of masks. In addition to the usual biases that accompany self-reported data, in this case we do not know whether in settings with mask mandates in place, respondents may be reporting their behavior differently compared to settings without mask mandates. The respondents to app-based surveys are also likely to not be a truly representative sample of the populations in each location. The degree to which the respondents represent the general population varies across locations, and depends on the prevalence of Facebook and other app use in each country. While this is a limitation of the data that are currently available, we believe that given the samples tend to be biased towards more educated, urban, and younger populations, the reported mask use is likely to be an over-estimate in these locations. If this is true, then the estimates of the impact of expanding mask use to reach 95% coverage in these populations would be an underestimate of the true effect of the intervention.
These limitations highlight important areas for improvement across the entire spectrum of research related to COVID-19. There is a critical need for more and higher quality data across the spectrum of information required to understand the trajectory of this pandemic, starting with better and higher-quality information on the numbers of cases and deaths, to improved and more representative data on use of masks among populations across the world, to additional studies that quantify the effect of face coverings on transmission probability of SARS-COV2 among the general population, to better data on social distancing mandates implemented in each location and the extent to which they are enforced. On the modeling side, continued work on making modeling frameworks more flexible and including alternative model specifications, as well as improving on how covariates are used to capture temporal variation in new infections might lead to improvements in the performance of long-range forecasts. A greater understanding of how the drivers of this pandemic interact with each other and affect transmission probabilities is critical for influencing its trajectory over the next few months.
Conclusion
The COVID-19 epidemic is far from over. While a selection of therapeutics is showing promise, a vaccine that can be deployed at global scale does not exist and is unlikely to be widely available in the near future. We expect more deaths in the second half of 2020 than were seen in the first six months. Not only are there large epidemics unfolding in Latin America, the Middle-East and South Asia, but seasonality suggests a second wave can be anticipated in the Northern Hemisphere. The rising toll of the COVID-19 pandemic can be reduced by 0.82 million deaths in the next few months by the adoption of universal mask mandates. This low-cost intervention that is available and accessible to all populations, regardless of socio-economic status or other dimensions of inequity, has enormous health benefits and might also lead to large economic benefits by delaying the need for re-imposition of social distancing mandates. In global health we rarely encounter effective, low-cost, and universally available interventions that can save lives: immediately, equitably and safely. Ensuring that individuals, as well as, local, national and global decision makers are all doing everything in their power to achieve the highest rates of mask use in all exposed populations is one of the best strategies available to us to mitigate the toll of the pandemic in the months to come.
Code Availability Statement
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/).
Online content
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.
Competing interests
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.
Acknowledgments
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.