key: cord-1020943-xnol6e6r authors: Aravindakshan, A.; Boehnke, J.; Gholami, E.; Nayak, A. title: Mask-Wearing During the COVID-19 Pandemic date: 2020-09-13 journal: nan DOI: 10.1101/2020.09.11.20192971 sha: c392e2f0606d28523cf841fdc6a5d3f9cce73796 doc_id: 1020943 cord_uid: xnol6e6r Background Masks have been widely recommended as a precaution against COVID-19 transmission. Several studies have shown the efficacy of masks at curbing droplet dispersion in lab settings. Using individual response data from the Imperial College London-YouGov personal measures survey, this study investigates reported mask use and its association with the spread of COVID-19. Methods We examine the association of reported mask-wearing in a population within a country with the growth rate of active cases COVID-19 cases. The analysis uses country-wide data from weekly surveys of individuals mask-wearing behavior in public places, as well as other concurrently implemented nonpharmaceutical interventions, mobility changes, new cases of COVID-19 data and other control variables. The mask-wearing data were obtained from the Imperial College-YouGov multi-country survey of peoples personal habits before and during COVID-19. Because mask-wearing showed substantial variation across countries and over time, we use a reduced form econometric model to relate population-wide changes in mask-wearing to the growth rate in COVID-19. Results The results indicate that reported mask-wearing could play an important role in mitigating the growth of COVID-19. Widespread mask-wearing within a country associates with an expected 7% (95% CI: 3.94%-9.99%) decline in the growth rate of daily active cases of COVID-19 in the country. This daily decline equates to an expected 88.5% drop in the growth of daily active cases over a 30-day period when compared to zero percent mask-wearing, all else held equal. The decline in daily growth rate due to the combined effect of reported mask-wearing, reduced social mobility, and non-pharmaceutical interventions averages 28.1% (95% CI: 24.2%-32%). These estimates remain robust across multiple sensitivity analyses. Conclusions Face mask usage as a preventative measure against the transmission of COVID-19 varies widely across countries. This observational study uses data from 24 countries and finds that reported face mask usage associates with a decline in new COVID-19 cases. Even though the model controls for a large variety of variables that affect spread, and we run a number of robustness checks, it remains possible that some of the decline associated with face masks is related to confounding variables not included in the model. In summary, due to the confounding variables and the variations in the type of mask and its usage, randomized control trials of mask usage in populations are needed to determine the true effect of mask wearing on mitigating the transmission of infectious respiratory diseases. In response to the COVID-19 pandemic, multiple countries curbed the spread of the disease by enforcing strict policy measures such as lockdowns and shelter-in-place orders [1] . The nonpharmaceutical interventions (NPIs) included closures of school, restaurants, bars, retail outlets and other non-essential businesses, as well as shelter-in-place policies and the prohibition of large gatherings (e.g., limited to 10 people) [2] . These institutional measures aimed to reduce the exposure of susceptible individuals to symptomatic and asymptomatic infected individuals by decreasing social mobility (e.g., going out to movies, concerts and restaurants, assembling in large groups) and encouraging social distancing. (e.g., 1m-2m physical distancing) [3, 4] . Unlike the widespread and proactive implementation of lockdowns and physical distancing measures, the usage of masks varied widely across countries. Some countries quickly adopted guidelines for mask usage (e.g., Malaysia, Singapore, Taiwan, and Thailand) while others did not recommend using face mask unless sick [5, 6, 7] . Indeed, the World Health Organization updated its mask-wearing guidelines only on June 5, 2020 [8] , to recommend that "The general public should wear non-medical masks where there is widespread transmission and when physical distancing is difficult, such as on public transport, in shops or in other confined or crowded environments.". Due to these changing guidelines and uneven implementations, mask-wearing varied dramatically across countries and over time [9] . Multiple studies have investigated the impact of various governmental NPIs [3, 10, 11, 12] , that encourage physical distancing and the wearing of face masks [13, 14, 15] . In each case they find that NPIs, physical distancing, and face mask usage reduce the transmission of COVID-19. However, the interventions, government policies, and individual measures seldom act in isolation. Treating these measures in isolation could lead to under-or over-estimation of their effectiveness at reducing the spread of the disease, biasing the assessments of the measure's impact. In this study, we investigate the association of population-wide mask-wearing with the number of COVID-19 cases, concurrent with other individual and institutional measures. Specifically, we expand on the current stream of research by simultaneously considering the effects of NPIs and social mobility along with a population's reported usage of face masks in public places in a reduced-form econometric model (see examples in [3] and [10] ). Using data from 24 countries, we identify the effect of each measure by exploiting the country-wise differences in the (1) percentage of the population who report wearing a face masks in public places (YouGov Survey Data [9]), (2) social mobility across multiple categories such as Parks and Transit Locations (using Google Mobility Reports [16]) and NPI implementations (using CoronaNet-Project [1]) This study is a cross-sectional analysis of the effects of personal and governmental measures across 24 countries on mitigating COVID-19 disease spread. The data used in this study were collected from February 21, 2020 to July 8, 2020, representing 139 days of data for each country. All analysis presented in this paper uses publicly available data. Subsequently, we first present the data and then the model-based analysis. . CC-BY-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) preprint The copyright holder for this this version posted September 13, 2020. . https://doi.org/10.1101/2020.09.11.20192971 doi: medRxiv preprint Mask-Wearing: Survey data released by the Institute of Global Health Innovation (IGHI) at Imperial College London and YouGov [9] provide reported mask-wearing across multiple countries. The survey covers 26 countries (as of July 8, 2020), with around 21,000 people interviewed each week. We restrict our analysis to 24 countries, because two countries -China and Hong Kongdo not have publicly available data on the social mobility dimensions we control for in this study. The data present global insights on people's reported behavior in response to COVID-19. The dataset provides the percentage of population in each country who report to wear a mask in public places. Because these surveys were conducted at an interval of several days, we interpolate (linearly) to estimate the percentage of the population that would wear masks in public spaces for days when the data were unavailable (Figure 1 ). We use the significant variation of mask-wearing across countries to measure the association of people reporting mask wearing and the spread of COVID-19. Social Mobility: Google Community Mobility Reports provide data on relative mobility changes with respect to an internal baseline across multiple categories namely, retail and recreation, groceries and pharmacies, parks, transit stations, workplaces, and residential ( Figure 2) . A summary for the community mobility is shown in Table S2 in the Supplement. Apart from the Google Mobility reports, we also utilize mobility data from Apple to test the robustness of the model to different measures of mobility. We note that neither Google nor Apple provide absolute measures of mobility, but rather present relative changes with respect to benchmarks they use internally. Finally, drops in mobility could be driven by both individual actions (e.g., cautious behavior) as well as institutional actions due to NPIs enacted by governments. To control for mobility declines due to institutional actions, we also include country-specific interventions enacted both nationally and provincially. . CC-BY-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) preprint The copyright holder for this this version posted September 13, 2020. . https://doi.org/10.1101/2020.09.11.20192971 doi: medRxiv preprint Figure S8 in Supplementary. Governments across the 24 countries enforced different policies to control the spread of COVID-19. Prior research has shown that these policies played a significant role in reducing human to human physical contacts and led to a slowdown in the spread of the disease. However, these policies were implemented in different ways, some nationally, some provincially. We use data from COVID-19 Government Response Event Dataset [1] to control for government policies in estimating the effect of masks. Figure S10 lists the types and counts of national and provincial government policies implemented across the 24 countries we consider in this study. The dataset contains 5,816 entries on policies at National and Provincial level. Finally, inclusion of these interventions helps control for some of the observed drops in mobility that are not necessarily associated with individual actions but by the presence of institutional policies. Detailed information about the interventions are included in the Supplement (Section S2.5). Because the data span multiple countries and weeks, we include time and country fixed effects in the model. The model controls for country-level heterogeneity using fixed-effects, where the variable for a country assumes a value of one if the data considered are specific to that country, and zero otherwise. This allows for control of country-level characteristics that are not in the model and helps reduce the errors due to omitted variables in our analysis. In addition to country-level differences, we also control for time-based differences (e.g., people are more aware and cautious over time) by incorporating time-fixed effects, where the variable Weekt takes a value of 1 if the data are from week 't' (where t =1 in the first week for a given country in the data). In addition, we control for each country's testing capability ( Figure 3a ) by accounting for the total number of daily tests in the country. Finally, we also control for actions people take to educate themselves by including the Google Trends ( Figure 3b ) data for the search term 'coronavirus'. . CC-BY-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) preprint The copyright holder for this this version posted September 13, 2020. . https://doi.org/10.1101/2020.09.11.20192971 doi: medRxiv preprint Figure S13 and Figure S14 respectively in Supplementary. Data for the number of active daily cases in each country were obtained from the Johns Hopkins University School of Public Health [17] . We use a seven-day moving average of cumulative confirmed cases and cumulative recovered cases to compute daily active cases and daily growth rate. The dataset aggregates this information across multiple national, state, and local health departments within each country. The daily growth rate is then related, by means of a reducedform econometric model, to the independent variables described earlier. We describe the derivation in the Supplement (Section S1.1). We illustrate the daily cases and growth rate for one country, Italy, in Figure 4 . Figure S6 and Figure S7 respectively in the Supplement. A reduced form econometrics model was used to relate the growth rate of daily active infections to the independent variables described earlier. Similar models have been used by [3] to determine the effect of anti-contagion policies on the spread of COVID-19. In brief, the model assumes that the daily growth rate (ratio of active infections today to active infections the day before) is affected by the institutional measures such as NPIs as well as individual measures such as social mobility and mask-wearing. The covariates listed above help control for other factors . CC-BY-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) preprint The copyright holder for this this version posted September 13, 2020. . https://doi.org/10.1101/2020.09.11.20192971 doi: medRxiv preprint that could affect growth over time. Because the epidemiological parameters for new diseases such as COVID-19 might not be well understood, reduced form techniques allow for the estimation of the impact of governmental and personal measures to help contain the spread of the virus. The model we use initializes when a country reaches 20% of peak new cases as observed by July 8, 2020 to filter out the high variation in growth rates when the number of cases is very low at the beginning of the pandemic. For robustness, we also test other starting times in the Supplement and find results in line with the ones presented here. The Supplement also provides further details about the methodological approach as well as information on the model formulation used. We provide some brief notes on the operationalization of the independent variables and the model initialization below: 1. Responses to the survey about mask-wearing are subject to biases. For example, individuals might overestimate the efficacy of their mask or their wearing pattern. To alleviate some of these concerns, we compute the natural log of the mask-wearing variable to discount its impact on the growth of cases. This transformation yields a curve that grows at a slower rate as the values of mask-wearing increase, thereby diminishing the impact of higher levels of mask-wearing. We also test for other functional forms (square-root and linear) and present those results in the Supplement (Table S7) to July 8, 2020, we check if a policy p was implemented in a country j on day t. If the policy was implemented, we assign a value of 1 to , , , where s represents the level of policy coverage. If the policy was introduced at a provincial level, we , ,  by the population of the state. Because several policies were introduced at the same time or close together, they too suffered from collinearity issues. To minimize multicollinearity issues, we choose only a specific set of policies to include in the analysis. The Supplement (Section S2.5) discusses this selection mechanism. 4. Due to uncertainty of the lag in COVID-19 incidences and the difficulties in detection during the early days of the disease (Woloshin et al 2020), similar to prior research we tested the focal model across multiple lag periods (shift) from zero to 14 days and for different initialization thresholds (th) for each country (zero percent to 20% of a country's peak daily cases). We chose the best shift and th values using a k-fold crossvalidation process (k=5). The chosen model had the highest maximum likelihood estimate of the data as well as the lowest prediction error. We discuss this procedure in the Supplement (Section S4.1). The results presented in the next section correspond to a model with a shift of nine days and a th of 20% of peak new cases by July 12, 2020. Finally, the model was estimated on 1,422 observations across 24 countries, averaging 60 observations per country. We restrict our analysis to the first 60 days. However, we test . CC-BY-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 this version posted September 13, 2020. . https://doi.org/10.1101/2020.09.11.20192971 doi: medRxiv preprint the robustness of the findings for other lengths of data. This allows for greater variation in mask usage within the data. In the next section, we describe our results and their policy implications. The results indicate that individual (mask-wearing, social mobility) and institutional measures (NPIs) could play a role in mitigating the growth of COVID-19. The estimates from the focal reduced form model for these measures and their corresponding confidence intervals are shown in Figure 5 . The full table of results, along with results for all robustness checks are provided in Supplement (Section 3). We first list the results of the key measures we consider and then discuss their implications. Masks. The model finds that a reported mask wearing of 100% is associated with an average 7% (95% CI: 3.94% -9.99%) drop in daily COVID-19 cases. While this daily effect appears small over a 30 day period, 100% reported mask-wearing leads to approximately 88.5% (95% CI: . CC-BY-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 this version posted September 13, 2020. . https://doi.org/10.1101/2020.09.11.20192971 doi: medRxiv preprint 68.7% -89.2% ) decline in active cases when compared to the situation where 0% of the people report wearing masks (all else remaining the same across the two scenarios). Modifying the functional form of the mask variable did not appreciably change the association. For example, in the linear model, masks are associated with an average 8.69% (95% CI: 5.63% -11.66%) drop in daily growth rate and for the square root model the expected daily drop in growth rate was 7.89% (95% CI: 4.81% -10.87%). The stability of the results seems to indicate that maskwearing can potentially play a significant role in mitigating the spread of the disease. Figure 6 plots the ratio of active cases under different proportions of respondents who claim to wear masks as against no mask wearing. Mobility and NPIs. As expected, the model finds that a rise in mobility links with a rise in the number of cases. Specifically, the selected mobility variables associate with a combined 8.1% (95% CI: 5.6% -10.6%) drop in daily case numbers. Similarly, we find that the implementation of NPIs also associate with a drop in daily growth rates across countries. After accounting for mobility declines, the NPI measures 'Quarantine, 'External Border Restrictions,' and 'Closure and Regulation of Schools' link with the highest declines in the growth of cases. Overall, all NPIs included in the model led to a decrease in the growth rate of COVID-19. This finding confirms multiple studies that investigated the effects of NPIs at limiting the spread of COVID-19 [10, 18] . Overall, we find that if the NPIs were enacted uniformly across the whole country, then the combined association of the NPIs with the decline of growth in daily cases of COVID-19 would average 13% (95% CI: 9.2% -16.2%). We determine the combined effect using the Krinsky-Robb method, a Monte Carlo simulation used to draw samples from a multivariate normal distribution. Supplement S3.1 provides more details on this method. Due to nearly concurrent enactments and blanket coverage of policies and precautionary behaviors within countries, the individual (e.g., masks, limiting mobility) and institutional (NPIs) . CC-BY-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) preprint The copyright holder for this this version posted September 13, 2020. . https://doi.org/10.1101/2020.09.11.20192971 doi: medRxiv preprint measures correlate in time. This precludes causal identification of each measure's individual effect on disease mitigation. In other words, because mask-wearing, mobility reductions, and NPIs occur at similar times, their effects are intertwined and difficult to determine separately. For some variables such as mobility and NPIs, we lack the necessary data to fully control for some of these issues. In the case of mask-wearing, even though we cannot eliminate the simultaneity, we attempt to reduce the effects of confounding variables by employing control functions [19] . In addition, [20] note that control functions make the intervention exogenous in a regression equation. To create a control function, we use country mortality data for SARS, H1N1, and MERS to predict the propensity of mask-wearing in each country (see Supplement S4.5.2 for more details). Next, we compute the control function by determining the predicted mask-wearing residuals (Predicted Mask-Wearing minus Actual Stated Mask-Wearing), allowing for better identification of the effect of reported mask-wearing on COVID-19 case numbers. Using this procedure, we find that if 100% of the population claimed to wear masks, then maskwearing relates with an average 4.95% (95% CI: 2.26% -7.53%) drop in daily growth rate (nine days later), when compared to zero percent reported mask-wearing. While we are careful to note that this estimate could still be affected by confounding variables, this result lends further support to the estimate presented earlier. In summary, widespread mask-wearing could lead to a significant decline in the growth of COVID-19. To help determine the accuracy of the results we run several robustness checks (see Supplement S4): (1) We vary the lag period (shift) from 0 to 14 days. The results show that the estimates of the individual and institutional measures are relatively stable. (2) We also vary the length of time we consider in the analysis. The model considered 60 days of data for each country. We vary this to estimate the model on 35, 45, 55, 65, 75 and 85 days of data. We find that the results remain consistent to these changes. (3) We replace Google mobility data with Apple mobility data. The model estimates remain robust to this change. (4) We vary the functional form of how mask-wearing relates to the growth of COVID-19. The results do not vary appreciably with respect to this to this change. (5) We also test the robustness of the analysis by modifying the data using exponential smoothing. Specifically, for any day t, the focal model in Equation (1) ignores the value of the independent variables from days t-shift+1 to t (discussed in Figure S1 ). In the model we use for the robustness check, we do not ignore values between t-shift and t and use exponential smoothing to average the intervening data. Finally, we also modify the interpolation method from linear (current) to quadratic. We find that the results are stable to all these modifications. The Supplement details all the robustness checks and simulations as well as their results. . CC-BY-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) preprint The copyright holder for this this version posted September 13, 2020. . Over the past few months, several studies have investigated the efficacy of masks at minimizing droplet dispersion [21, 22] and the potential consequences of their use [14, 23] in the general population. Although no randomized control trial on the efficacy of face mask usage in the general population exists, [15, 14] provide compelling evidence for the use of face masks through a systematic review of the multiple observational studies and the evidence thus far. Both studies project that the use of face masks reduces the transmission of disease. While the type of face mask as well as the timing and length of use can affect its efficacy, its use as a precautionary principle has been strongly advised [24] . Despite this preponderance of evidence, face mask use, in the US remains controversial [25, 26] . Additionally, as observed in the data, even in countries where masks do not face similar headwinds and as support for mask usage gathers further evidence, face mask use is not as common place (e.g., Denmark, Norway, Sweden, Finland, Australia), even as a precautionary principle. This study links the growth in active cases of COVID-19 in a country, to a population's reported wearing of face masks in public places over time. The model also includes other measures that could simultaneously impact the growth of the disease as face mask usage changes over time. After accounting for these measures and controlling for other covariates, the results indicate that reported face mask use is associated with a decline in the growth of COVID-19. More precisely, if 100% of the population claimed to wear masks, then mask-wearing is associated with an average 7% decline in the growth of COVID-19. This association persists across multiple robustness checks and model formulations. A decline of 7% corresponds to an 88.5% drop in the number of active cases 30 days later. Taken together with the other measures (mobility changes, NPIs), the combined association obtained via the Krinsky-Robb method of individual and institutional measures on the decline in the daily growth of COVID-19 is 28.1% (95% CI: 24.2%-32%). This study has several limitations. First, countries enacted multiple NPIs simultaneously. This precludes us from identifying the effectiveness of NPIs separately. Second, the mobility data provided by Google and Apple are only indicative of the relative changes from a benchmark, so their association with disease spread should be interpreted with precaution. Third, even though we consider multiple NPIs and mobility measures and control for populationwide awareness, testing, and country and time differences, there could be several variables that we do not or cannot include due to the lack of data. For example, one of the NPIs we consider relates to health announcements (e.g., handwashing, coughing into one's elbow, etc.). However, the model only notes the NPI and not the adoption of these practices in the general population. While some data (e.g. sale of soap, cleansers, etc.) could approximate this, multiple confounding variables still exist and thus limit the extent to which we can apply any causal interpretation of the results. We attempt to provide better insights using a control function approach, however, stop short of making causal claims. . CC-BY-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) preprint The copyright holder for this this version posted September 13, 2020. . https://doi.org/10.1101/2020.09.11.20192971 doi: medRxiv preprint Fourth, our analysis was conducted at the country level due to data restrictions. However, mask usage, mobility changes as well as the implementation of NPIs most likely vary within the country. Nevertheless, the analyses in this study are helpful at understanding the practical implementation of mask usage policy within a country. Finally, we rely on the accuracy of data collected by third parties. Inconsistencies in testing, reporting, and recording of the data could lead to errors in the results obtained. Additionally, mask-types and mask-wearing patterns could vary across countries, individuals, and over time. This limitation affects all observational COVID-19 population-based studies. Randomized control trials of mask usage in populations are needed to determine the true effect of maskwearing on mitigating the transmission of infectious respiratory diseases. The population-wide usage of face masks as a preventative measure against the transmission of COVID-19 varies widely across countries. Using data from 24 countries, this study finds that reported face mask usage associates with a decline in the growth rate of COVID-19. Even though we control for multiple variables that could affect spread and include multiple robustness checks, it remains possible that some of the decline associated with face masks may be related with other confounding variables not included in our model. . CC-BY-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) preprint The copyright holder for this this version posted September 13, 2020. . https://doi.org/10.1101/2020.09.11.20192971 doi: medRxiv preprint COVID-19 Government Response Event Dataset (CoronaNet v.1.0) Association of Public Health Interventions with the Epidemiology of the COVID-19 Outbreak in Wuhan The Effect of Large-Scale Anti-Contagion Policies on the COVID-19 Pandemic Rationing social contact during the COVID-19 pandemic: Transmission risk and social benefits of US locations Rational Use of Face Masks in the COVID-19 Pandemic Imperial College COVID-19 Response Team Estimating the effects of non-pharmaceutical interventions on COVID-19 in Europe Association between statewide school closure and COVID-19 incidence and mortality in the US Strong Social Distancing Measures In The United States Reduced The COVID-19 Growth Rate Community Use Of Face Masks And COVID-19: Evidence From A Natural Experiment Of State Mandates In The US Physical distancing, face masks, and eye protection to prevent person-to-person transmission of SARS-CoV-2 and COVID-19: a systematic review and meta-analysis An interactive web-based dashboard to track COVID-19 in real time Effects of nonpharmaceutical interventions on COVID-19 cases, deaths, and demand for hospital services in the UK: A modelling study Control function methods in applied econometrics Alternative Methods for Evaluating the Impact of Interventions: An Overview Face cCoverings and Mask to Mini9mise Droplet Dispersion and Aerosolisation: A Video Case Study Estimation of Viral Aerosol Emissions from Simulated Individuals With Asymptomatic to Moderate Coronavirus Disease Face mask use in the general population and optimal resource allocation during the COVID-19 pandemic Face masks for the public during the covid-19 crisis