key: cord-253827-5vodag6c authors: Karaivanov, A.; Lu, S. E.; Shigeoka, H.; Chen, C.; Pamplona, S. title: Face Masks, Public Policies and Slowing the Spread of COVID-19: Evidence from Canada date: 2020-09-25 journal: nan DOI: 10.1101/2020.09.24.20201178 sha: doc_id: 253827 cord_uid: 5vodag6c We estimate the impact of mask mandates and other non-pharmaceutical interventions (NPI) on COVID-19 case growth in Canada, including regulations on businesses and gatherings, school closures, travel and self-isolation, and long-term care homes. We partially account for behavioral responses using Google mobility data. Our identification approach exploits variation in the timing of indoor face mask mandates staggered over two months in the 34 public health regions in Ontario, Canada's most populous province. We find that, in the first few weeks after implementation, mask mandates are associated with a reduction of 25 percent in the weekly number of new COVID-19 cases. Additional analysis with province-level data provides corroborating evidence. Counterfactual policy simulations suggest that mandating indoor masks nationwide in early July could have reduced the weekly number of new cases in Canada by 25 to 40 percent in mid-August, which translates into 700 to 1,100 fewer cases per week. When government policies to stem the spread of were introduced in early 2020, the best available evidence supporting them was provided by studies of previous epidemics, epidemiological modeling, and case studies (OECD, 2020). Even when the efficacy of a given precaution in reducing COVID -19 transmission has been established, significant doubts regarding the usefulness of specific policy measures may persist due to uncertainty regarding adherence to the rules and other behavioral responses. For example, even though several observational studies, mostly in medical setting, have shown that face masks reduce the transmission of and similar respiratory illnesses (see Chu et al. (2020) for a comprehensive review), a face mask mandate may not be effective in practice if it fails to increase the prevalence of mask wearing (compliance), or if it leads to increased contacts due to a false sense of security. It is therefore important to directly evaluate and quantify the relationship between various policy measures and the rate of propagation of . The low cost and high feasibility of mask mandates relative to other containment measures for has generated keen interest worldwide for studying their effectiveness. This attention has been compounded by substantial variation, across jurisdictions and over time, in official advice regarding the use of masks. Figure B1 in the Appendix plots self-reported mask usage in select countries (Canada, United States, Germany and Australia) in the left panel, and across Canadian provinces in the right panel. The figure shows large differences in mask usage, both across countries and within Canada. 1 We estimate and quantify the impact of mask mandates and other non-pharmaceutical interventions (NPI) on the growth of the number of COVID-19 cases in Canada. Canadian data has the important advantage of allowing two complementary approaches to address our objective. First, we estimate the effect of mask mandates by exploiting within-province geographic variation in the timing of indoor face mask mandates across 34 public health regions (PHUs) in Ontario, Canada's most populous province with a population of nearly 15 million or roughly 39% of Canada's population (Statistics Canada, 2020). The advantage of this approach is that it exploits variation over a relatively small geographic scale (PHU), holding all other province-level policies or events constant. In addition, the adoption of indoor face mask mandates in these 34 sub-regions was staggered over approximately two months, creating sufficient intertemporal policy variation across the PHUs. Second, we evaluate the impact of NPIs in Canada as a whole, by exploiting variation in the timing of policies across the country's ten provinces. By studying inter-provincial variation, we are able to analyze the impact of not only mask mandates, but also other NPIs, for which there is little or no variation across Ontario's PHUs (regulations on businesses and gatherings, schooling, travel and long-term care). In addition, our province-level data include both the closing period (March-April) and the gradual re-opening period (May-August), providing variation from both the imposition and the relaxation of policies. Our panel-data estimation strategy broadly follows the approach of Chernozhukov, Kasahara and Schrimpf (2020), hereafter CKS (2020), adapted to the Canadian context. We allow for behavioural responses (using Google Community Mobility Reports geo-location data as proxy for behaviour changes and trends), as well as lagged outcome responses to policy and behavioral changes. Our empirical approach also allows current epidemiological outcomes to depend on past outcomes, as an information variable affecting past policies or behaviour, or directly, as in the SIR model framework. We find that, in the first few weeks after their introduction, mask mandates are associated with an average reduction of 25 to 31% in the weekly number of newly diagnosed cases in Ontario, holding all else equal. We find corroborating evidence in the province-level analysis, with a 36 to 46% reduction in weekly cases, depending on the empirical specification. Furthermore, using survey data, we show that mask mandates increase self-reported mask usage in Canada by 30 percentage points, suggesting that the policy has a significant impact on behaviour. Jointly, these results suggest that mandating indoor mask wear in public places is a powerful policy measure to slow the spread of , with little associated economic disruption at least in the short run. 2 Counterfactual policy simulations using our empirical estimates suggest that mandating indoor masks nationwide in early July could have reduced weekly new cases in Canada by 25 to 40% on average by mid-August relative to the actually observed numbers, which translates into 700 to 1,100 fewer cases per week. We also find that the most stringent restrictions on businesses and gatherings observed in our data are associated with a decrease of 48 to 57% in weekly cases, relative to a lack of restrictions. The business/gathering estimates are, however, noisier than our estimates for mask mandates and do not retain statistical significance in all specifications; they appear driven by the smaller provinces and the re-opening period (May to August). School closures and travel restrictions are associated with a large decrease in weekly case growth in the closing period. Our results on business/gathering regulations and school closure suggest that reduced restrictions and the associated increase in business or workplace activity and gatherings or school re-opening can offset, in whole or in part, the estimated effect of mask mandates on case growth, both in our sample and subsequently. An additional contribution of this research project is to assemble, from original official sources only, and make publicly available a complete dataset of cases, deaths, tests and policy measures in all 10 Canadian provinces. 3 To this end, we constructed, based on official public health orders and announcements, time series for 17 policy indicators regarding face masks, regulations on businesses and gatherings, school closures, travel and self-isolation, and long-term care homes. Our paper relates most closely to two recent empirical papers on the effects of mask mandates using observational data. 4 CKS (2020) and Mitze et al. (2020) study the effect of mask mandates in the United States and Germany, respectively. CKS (2020), whose estimation strategy we follow, exploit U.S. state-level variation in the timing of mask mandates for employees in public-facing businesses, and find that these mandates are associated with 9 to 10 percentage points reduction in the weekly growth rate of cases. This is substantially smaller that our estimates, possibly because the mask mandates that we study are much broader: they apply to all persons rather than just employees, and most apply to all indoor public spaces rather than just businesses. Mitze et al. (2020) use a synthetic control approach and compare the city of Jena and six regions in Germany that adopted a face mask policy in early to mid April 2020, before their respective state mandate. They find that mandatory masks reduce the daily growth rate of cases by about 40%. Our paper has several advantages compared to the above two papers. First, we exploit both regional variation within the same province (like Mitze et al., 2020) and provincial variation in the whole country (like CKS, 2020), and find similar results, which strengthens the validity of our findings. Second, we show that self-reported mask usage has increased after introducing mask mandates. We view this "first-stage" result on mask usage as informative, as the effectiveness of any NPI or public policy critically depends on the compliance rate. Moreover, this result mitigates possible concerns that the estimated mask mandate effect on case growth may be caused by factors other than mask policy. Third, a key difference between our paper and CKS (2020) is that we evaluate the effect of universal (or community) mandatory indoor mask wearing for the public rather than the effect of mandatory mask wearing for employees only. 5 While other factors such as differences in mask wear compliance between Canada and the U.S. may contribute to the different estimated magnitude of the policy impact, our results suggest that more comprehensive mask policies can be more effective in reducing the case growth rate. Other Related Literature Abaluck et al. (2020) discuss the effectiveness of universal adoption of homemade cloth face masks and conclude that this policy could yield large benefits, in the $3,000-$6,000 per capita range, by slowing the spread of the virus. The analysis compares countries with pre-existing norms that sick people should wear masks (South Korea, Japan, Hong Kong and Taiwan) and countries without such norms. 6 In the medical literature, Prather et al. (2020) argue that masks can play an important role in reducing the spread of COVID-19. Howard et al. (2020) survey the medical evidence on mask efficiency and recommend public use of masks in conjunction with existing hygiene, distancing, and contact tracing strategies. Greenhalgh et al. (2020) provide evidence on the use of masks during non-COVID epidemics (influenza and SARS) and conclude that even limited protection could prevent some transmission of . Leung et al. (2020) study exhaled breath and coughs of children and adults with acute respiratory illness and conclude that the use of surgical face masks could prevent the transmission of the human coronavirus and influenza virus from symptomatic individuals. Meyerowitz et al. (2020) present a recent comprehensive review of the evidence on transmission of the virus and conclude that there is strong evidence from case and cluster reports indicating that respiratory transmission is dominant, with proximity and ventilation being key determinants of transmission risk, as opposed to direct contact or fomite transmission. Our paper also complements recent work on COVID-19 policies in Canada. Mohammed et al. (2020) use public opinion survey data to study the effect of changes in mask-wear policy recommendations, from discouraged to mandatory, on the rates of mask adoption and public trust in government institutions. They show that Canadians exhibit high compliance with mask mandates and trust in public health officials remained consistent across time. Yuksel et al. (2020) use an outcome variable constructed from Apple mobility data along period January 3 to February 6, 2020. 9 In Ontario, these location data are available for each of the 51 first-level administrative divisions (counties, regional municipalities, single-tier municipalities and districts). 10 We follow the approach of CKS (2020), but modify and adapt it to the Canadian context. The empirical strategy uses the panel structure of the outcome, policy and behavioral proxy variables, and includes lags of outcomes as information, following the causal paths suggested by the epidemiological SIR model (Kermack and McKendrick, 1927) . Specifically, we estimate the effect of policy interventions on COVID-19 outcomes while controlling for information and behaviour. In contrast to CKS (2020) and Hsiang et al. (2020) , who study variation in NPIs across U.S. states or across countries, our identification strategy exploits policy variation at the sub-provincial level (Ontario's PHUs) in addition to cross-province variation, and our data captures both the closing down and gradual re-opening stages of the epidemic. 5. Controls, W it -province or PHU fixed effects, growth rate of weekly new tests, and a time trend. To assess and disentangle the impact of NPIs and behavioral responses on COVID-19 outcomes, we estimate the following equation: where l denotes a time lag measured in days. Equation (1) models the relationship between COVID-19 outcomes, Y it , and lagged behaviour, B it−l , lagged policy measures, P it−l and information (past outcomes), I it = Y it−l . For case growth as the outcome, we use l = 14. For deaths growth as the outcome, we use l = 28. 11 The choice of these lags is discussed in Appendix D. By including lagged outcomes, our approach allows for possible endogeneity of the policy interventions P it , that is, the introduction or relaxation of NPIs based on information on the level or growth rate of cases or deaths. Also, past cases may be correlated with (lagged) government policies or behaviors that may not be fully captured by the policy and behaviour variables. In Appendix Table A18 , we also report estimates of the following equation: which models the relationship between policies P it , information, I it (weekly levels or growth of cases or deaths) and behaviour, B it . It is assumed that behaviour reacts to the information without a significant lag. We find strong correlation between policy measures and the Google mobility behavioral proxy measure. Equation (1) captures both the direct effect of policies on outcomes, with the appropriate lag, as well as the potential indirect effect on outcomes from changes in behaviour captured by the changes in geo-location proxy B it−l . In Appendix Tables A19 and A20, we also report estimates of equation (1) without including the behavioral proxy, that is, capturing the total effect of policies on outcomes. Since our estimates of the coefficient α in equation (1) are not significantly different from zero, the results without controlling for the behavioral proxy are very similar to those from estimating equation (1). Outcomes. Our main outcome of interest is the growth rate of weekly new positive cases as defined below. 12 We use weekly outcome data to correct for the strong day-of-the-week effect present in COVID-19 outcome data. 13 Weekly case growth is a metric that can be helpful in assessing trends in the spread of , and it is highlighted in the World Health Organization's weekly epidemiological updates (see, for example, World Health Organization (2020)). Specifically, let C it denote the cumulative case count up to day t and define ∆C it as the weekly COVID-19 cases reported for the 7-day period ending at day t: The weekly case (log) growth rate is then defined as: that is, the week-over-week growth in cases in region i ending on day t. 14 The weekly death growth rate is defined analogously, using cumulative deaths data. Policy. In the Ontario analysis, we exploit regional variation in the timing of indoor mask mandates staggered over two months in the province's 34 regions ("public health units" or PHUs). Figure 1 displays the gradual introduction of mask mandates across the 34 PHUs in Ontario. The exact implementation dates of the mask mandates are reported in Table C2 . Mandatory indoor masks were introduced first in the Wellington-Dufferin-Guelph PHU on June 12 and last in the Northwestern PHU on August 17. 15 12 We also report results using the growth rate of deaths as supplemental analysis in Section 4.6. 13 Figures B9 and B10 in the Appendix respectively display the weekly and daily cases, deaths and tests in each Canadian province over time. There are markedly lower numbers reported on weekends or holidays. 14 To deal with zero weekly values, which mostly occur in the smaller regions, as in CKS (2020), we replace log(0) with -1. We also check the robustness of our results by adding 1 to all ∆C it observations before taking logs, by replacing log(0) with 0, and by using population weighted least squares; see Tables A5 and A8 . 15 There is no PHU-wide mask mandate in Lambton as of August 31, but its main city, Sarnia, enacted a mask mandate on July 31. Figure 1 : Ontario -mask mandates over time 15 In the province-level analysis, we assign numerical values to each of the 17 policy indicators listed in Table C1 in Appendix C. The values are on the interval [0,1], with 0 meaning no or lowest level of restrictions and 1 meaning maximal restrictions. A policy value between 0 and 1 indicates partial restrictions, either in terms of intensity (see more detail and the definitions in Table C1 ) or by geographical coverage (in large provinces). The numerical values are assigned at the daily level for each region (PHU or province, respectively for the Ontario and national results), while maintaining comparability across regions. Many NPIs were implemented at the same time, both relative to each other and/or across regions (especially during the March closing-down period), which causes many of the policy indicators to be highly correlated with each other (see Appendix Table A4 ). To avoid multi-collinearity issues, we group the 17 policy indicators into 5 policy aggregates via simple averaging: (i) travel, which includes international and domestic travel restrictions and self-isolation rules; (ii) school, which is an indicator of provincial school closure; (iii) business/gathering, which comprises regulations and restrictions on non-essential businesses and retail, personal businesses, restaurants, bars and nightclubs, places of worship, events, 10 gyms and recreation, and limits on gathering; (iv) long-term care (LTC), which includes NPIs governing the operation of long-term care homes (visitor rules and whether staff are required to work on a single site) and (v) mask which takes value 1 if an indoor mask mandate has been introduced, 0 if not, or value between 0 and 1 if only part of a province has enacted such policy. 16 The five policy aggregates are constructed at the daily level and capture both the closingdown period (an increase in the numerical value from 0 toward 1) and the re-opening period (decrease in the numerical value toward zero). In comparison, the policy indicators compiled by Raifman (2020) for the USA used in CKS (2020) are binary "on (1)"/"off (0)" variables. 17 For consistency with the weekly outcome and information variables and the empirical model timing, we construct the policy aggregates P j it used in the regressions (where j denotes policy type) by taking a weekly moving average of the raw policy data, from date t − 6 to date t. Figure 2 plots the values of the 5 policy aggregates over time for each of the 10 provinces. Travel restrictions, school closures (including Spring and Summer breaks) and business closures were implemented in a relatively short period in the middle of March. There is some variation in the travel policy aggregate since some Canadian provinces (the Atlantic provinces and Manitoba) implemented inter-provincial domestic travel or self-isolation restrictions in addition to the federal regulations regarding international travel. Restrictions on long-term care facilities were introduced more gradually. In the re-opening period (May-August), there is also more policy intensity variation across the provinces, especially in the business and gatherings category, as the different provinces implemented their own re-opening plans and strategies. Mask mandates were first introduced in Ontario starting from June in some smaller PHUs and early July in the most populous PHUs such as Toronto, Ottawa and Peel (see Appendix Table C2 ). In Quebec, indoor masks were mandated province-wide on July 18. Nova Scotia and Alberta's two main cities implemented mask mandates on July 31 and August 1, respectively. There are two empirical challenges specific to our Canadian context and data. The first challenge is the presence of small provinces and sub-regions with very few COVID-19 cases or deaths. In Section 4.3, we perform a number of robustness checks using different ways of handling the observations with very few cases (in particular zero cases). The second data limitation is that there are only 10 provinces in Canada and 34 public health units in Ontario, unlike the 51 U.S. jurisdictions in CKS (2020). To account for the resulting small number of clusters in the estimation, we compute and report wild bootstrap standard errors and p-values, as proposed by Cameron et al. (2008) . 18 On the flip side, our data has the advantage of a longer time horizon (March to August) and non-binary, more detailed policy variables compared to Raifman et al. (2020). Behaviour proxy. We follow CKS (2020) and other authors in interpreting the location change indices from the Google Community Mobility reports as proxies for changes in people's behaviour during the pandemic, keeping in mind that location is only one aspect of behaviour relevant to . The general pattern in the data (see Figure B3 ) shows sharply reduced frequency of recorded geo-locations in shops, workplaces and transit early in the pandemic (March), with a subsequent gradual increase back toward the baseline (except for transit), and a flattening out in July and August. Several of the six location indicators (retail, grocery and pharmacy, workplaces, transit, parks and residential) are highly correlated with each other (see Tables A1 and A2 ) and/or contain many missing observations for the smaller provinces. To address these data limitations and the possible impact of collinearity on the estimation results, we use as proxy for behavioral changes the simple average of the following three mobility indicators: "retail", "grocery and pharmacy" and "workplaces". To be consistent with the weekly outcome variables and to mitigate day-of-week behavioural variation, we construct the Behaviour proxy B it by taking a weekly moving average of the 1 3 (retail + grocery and pharmacy + workplaces) data, from date t − 6 to date t. 19, 20 As a result, our empirical analysis uses weekly totals (for cases, tests and deaths) or weekly moving averages (for policies and the behaviour proxy) of all variables recorded on daily basis. 21 18 Alternative methods for computing the standard errors are explored in Section 4.3. 19 We drop the "transit", "parks", and "residential" location indicators because, respectively, 10.6%, 13.7%, and 2.8% of the observations are missing in the provincial data, and 20.7%, 52.1%, and 11.1% are missing in the Ontario data. The "transit" and "residential" variables are also highly correlated with the three indicators we include in our aggregate behaviour proxy B it . Furthermore, the "parks" indicator does not have clear implication for outcomes. 20 In the Ontario analysis, 1.4% of the B it values are imputed via linear interpolation. 21 In estimation equation (1), we take moving average from date t − 14 to date t − 20 for policies and behaviour when the outcome is weekly case growth, and from date t − 28 to date t − 34 if the outcome is Tables A3 and A4 display the correlation between our behaviour proxy B it and the five NPI policy aggregates P j it . Importantly, the behaviour proxy and mask mandate variables are not highly correlated, suggesting that the effect of mask mandates on COVID-19 outcomes should be independent of location behaviour changes. Information. We use the weekly cases and case growth variables defined above, ∆C it and Y it , to construct the information variables I it in equation (1) . Specifically, we use as information the lagged value of the weekly case growth rate Y it−l (= ∆ log(∆C it−l ) and the log of past weekly cases, log(∆C it−l ). We also use the lagged provincial (Ontario analysis) or national (Canada analysis) case growth rate and log of weekly cases as additional information variables in some specifications. A two-week information lag l = 14 is used in the baseline results. In the supplementary regressions using the death growth rate as the outcome, we use information on past deaths and a four-week lag (see Section 4.6). Control variables. In all regressions, we control for region fixed effects (PHU or province) and the weekly COVID-19 tests growth rate ∆ log(∆T it ), where T it denotes cumulative tests in region i until date t and ∆T it is defined analogously to ∆C it above. We include a time trend: our baseline uses a cubic polynomial in days, but we also report results with no time trend and with week fixed effects. Robustness checks also include news or weather variables as controls (see Section 4.3). Time period. We use the period May 15 to August 13 for the analysis with Ontario PHU level data and the period March 11 to August 13 for the national analysis with provincial data. The end date reflects data availability at the time of empirical analysis and writing. The start date for the Ontario sample (May 15) is approximately two weeks after the last restrictive measures were implemented and four weeks before the first mask mandate was introduced in Ontario. Robustness checks with different initial dates (May 1, June 1 and June 15) are reported in Section 4.3, with our results remaining stable. The initial date for the national sample (March 11) was chosen as the first date on which each province reported at least one COVID-19 test (so that cases could be potentially reported). Again, alternative initial dates are explored in Section 4.3. We start with a simple graphical illustration of the effect of mask mandates on COVID-19 cases growth. Figure 3 displays the average log case growth, Y it = ∆ log(∆C it ) in Ontario PHUs with or without mask mandates. It shows that, on average, the PHUs with a mask mandate two weeks prior have lower case growth than the PHUs without a mask mandate two weeks prior. 15 No mask mandate at t -14 Mask mandate at t -14 Notes: The figure plots the average log weekly case growth ∆log(∆C) in the PHUs with mask mandate (blue) vs. without (red) mask mandate 14 days prior. Table 1 shows the estimates of equation (1), in which we control for other policies, behaviour and information, as explained in Section 3.1. 22 We report wild bootstrap p-values clustered at the PHU level to account for the small number of clusters. 23 The odd-numbered 22 Mask mandates and regulations on business and gatherings vary at the PHU level. Long-term care policy changed only province-wide. The other policies (schooling and travel) do not vary during the sample period and hence are omitted from the regressions with Ontario PHU data. 23 Table A6 in the Appendix reports alternative standard error specifications: regular clustering at the PHU level (Stata command "cluster"), wild bootstrap standard errors clustered at the PHU level, and wild columns in Table 1 use lagged cases and lagged cases growth at the PHU level as information; the even-numbered columns also include lagged cases and lagged case growth at the province level as additional information variables. In the tables, Variable 14 indicates a 14-day lag of Variable. We present estimates of equation (1) from three specifications that handle possible time effects differently. Columns (1) and (2) in Table 1 are the most basic specifications, without including a time trend. The estimates in columns (1) and (2) suggest that, controlling for behavioural changes, mandatory indoor face masks reduce the growth rate of infections by 29-32 log points (p < 0.05), which is equivalent to a 25-28% reduction in weekly cases. 24 In order to control for potential province-wide factors affecting the spread of COVID-19 such as income support policies or adaptation to the pandemic over time (so-called COVID fatigue), we also estimate (1) with a cubic time trend in days from the beginning of the sample, in columns (3) and (4) of Table 1 , and with week fixed effects, in columns (5) and (6) . Columns (3)- (6) show that our estimates of the mask mandate policy remain robust to the inclusion of a cubic time trend or week fixed effects. The results indicate that, depending on the specification, mask mandates are associated with a reduction of up to 38 log points in weekly case growth or, equivalently, a 31% reduction in weekly cases. The magnitude of the mask policy estimate is not very sensitive to whether lagged province-level data are included as additional information. The results in Table 1 suggest that indoor mask mandates can be a powerful preventative measure in the COVID -19 context. Our estimates of the mask mandate impact across Ontario's PHUs are equivalent to a 25-31% reduction in weekly cases. This estimate is larger than the 9-10% reduction estimated by CKS (2020) for the U.S. One possible explanation is that Ontario's mask policy is more comprehensive: we evaluate the effect of universal indoor mask-wearing for the public rather than the effect of mask wearing for employees only in CKS (2020). Differences in the compliance rate may also contribute to this difference; we discuss this potential channel in Section 4.4. The results in Table 1 also show a statistically significant negative association between information (log of past cases, log(∆C) 14) and current weekly case growth (p < 0.01 in all specifications), indicating that a higher level of cases two weeks prior is correlated with lower current case growth. While B it allows for behavioural responses to information, the negative estimate on log(∆C) 14 in Table 1 suggests that our location-based proxy does not capture bootstrap standard errors clustered by both PHU and date. Our results are robust to alternative ways of calculating standard errors. 24 Using equation (3), a coefficient of x translates into a 1 − exp(x) reduction in weekly cases ∆C it /∆C it−7 . 1 and in Section 4.2's province-level results), unlike in CKS (2020). 25 In Appendix Table A18 , we find strong contemporaneous correlations between the policy measures, log cases, and the Google mobility behavioral proxy from estimating equation (2) . This suggests that the information (lagged cases) and the lagged policy variables included in equation (1) may absorb lagged behavioral responses proxied by B it−l or other latent behavioral changes not captured by B it−l . We next evaluate the impact of NPIs on COVID-19 cases growth in Canada as a whole by exploiting variation in the timing of policies across the 10 provinces. Here, we examine NPIs for which there is no variation across Ontario's PHUs (i.e., schooling, travel, and LTC) in addition to mask mandates. Also, provincial data contain variation in the timing of policy changes in both the closing and re-opening phases, allowing us to study both the imposition and relaxation of restrictions. Figure 4 : Canada -mask mandates and weekly case growth 15 No mask mandate at t -14 Mask mandate at t -14 Notes: The figure plots the average weekly case growth ∆ log(∆C) in the provinces with mask mandate (blue) vs. without mask mandate (red) 14 days prior. 25 We also tried including each location change measure separately and the results are similar (not shown). All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . As in the Ontario analysis, we begin with a graphical illustration of mask mandates and COVID-19 case growth across Canadian provinces, in the period March 11 to August 13, 2020. Figure 4 plots the average log weekly case growth in the provinces with vs. without mask mandates. While mask mandates are implemented relatively late in our sample period, average case growth in the provinces with a mask mandate (Ontario and Quebec) diverged from the average case growth in the provinces without a mandate begin roughly four weeks after the mandates are imposed. 26 Table 2 displays the estimates of equation (1) for weekly case growth, along with wild bootstrap p-values, clustered at the province level (see Table A9 for other methods of computing the standard errors). The odd-numbered columns use lagged cases and lagged case growth at the provincial level as information while the even-numbered columns include in addition lagged cases and case growth at the national level as additional information variables. As in the Ontario analysis, we present in Table 2 estimates from three specifications: no time trend (columns (1)-(2)), including cubic time trend in days (columns (3)-(4)) and including week fixed effects (columns (5)- (6)). The most robust result is the estimated effect of mask mandates: they are associated with a large reduction in weekly case growth of 45 to 62 log points, which is equivalent to a 36 to 46% reduction in weekly cases across the different specifications. The estimates are statistically significantly different from zero in all cases, with a p-value of less than 0.001 in columns (1)- (4) . It is reassuring that these results regarding mask mandates are consistent with the Ontario analysis in the previous section. Table 2 further shows that restrictions on businesses and gatherings are associated with a reduction in the weekly case growth of 65 to 85 log points or, vice versa, that relaxing business/gathering restrictions is associated with higher case growth. The estimate is equivalent to a 48 to 57% decrease in weekly cases in our sample period. The business/gathering estimates are, however, more noisy than our estimates for mask mandates and do not retain statistical significance in the specifications with week fixed effects (p = 0.15 and 0.14). Tables A8 and A15 further suggest that the results on business and gathering NPIs are driven by the smaller provinces and the re-opening period (May to August). Still, these results suggest that lowered restrictions and the associated increase in business/workplace activity or gatherings can be an important offsetting factor for the estimated effect of mask mandates on COVID-19 case growth, both in our sample and in the future. We also find that school closures (the School 14 variable in Table 2 ) can be negatively 26 Figure 4 assumes a July 7 mask mandate implementation date for Ontario (when its most populous PHU, Toronto, adopted a mask mandate, along with Ottawa), and July 18 for Quebec (province-wide mandate). All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . associated with case growth. However, the estimates are statistically significant from zero only in the specifications with cubic time trend (columns (3) and (4)). As seen in Figure 20 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . 2, provincial school closures occurred in a very short time interval during March, so we may lack statistical power to separately identify its effect from other NPIs (especially the travel-related). Hence, we interpret this result with caution. As in Table 1 , the level of past cases, log(∆C), is negatively and statistically significantly associated with current weekly case growth in columns (1)- (4) . Since the specification with cubic time trend in Tables 1 and 2 allows for possible nonmonotonic aggregate time trends in case growth in a parsimonious way, we choose it as our baseline specification with which to perform robustness checks in the next section. Robustness checks with the other specifications are available upon request. A possible concern about our data for the national analysis is that some NPIs (e.g. international travel restrictions or closing of schools) were implemented within a very short time interval. 27 Thus, we may lack enough regional variation to distinguish and identify the separate effect of each policy. 28 Collinearity could also affect the standard errors and the signs of the estimated coefficients. To check robustness with respect to potential collinearity in the NPI policies, Tables A7 and A10 report estimates from our baseline specification, omitting one policy at a time, for Ontario and Canada respectively. First, it is reassuring that the mask mandate estimates are hardly affected by omitting any of the other policies. This is expected since mask mandates were imposed during a period where other NPIs changed little (see Figure 2) . Similarly, the effects of business/gathering regulations and school closures in Table A10 are not sensitive to omitting other policies one at a time, which suggests that there is sufficient statistical power and variation to identify them in the national analysis. Another concern for our empirical strategy is that the usual formula for our dependent variable, ∆ log(∆C it ), cannot be applied when the weekly case total ∆C it is zero. We follow CKS (2020) and replace ln(0) with -1 in our baseline specifications in Tables 1 and 2 . We now check the robustness of our estimates to alternative treatments of zero weekly cases. For easier comparison, the first two columns in Table A5 repeat columns (3) and (4) 27 For example, Table A4 shows a correlation of 0.61 between the Travel and School policy aggregates. 28 Aggregating the 17 basic policy indicators into five groups mitigates this issue. Here, we test whether any remaining collinearity poses a problem. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . from Table 1 for Ontario. 29 Our main results on mask mandates across Ontario PHUs are robust to replacing log(0) with 0 and to adding 1 to all ∆C it observations before taking logs, as shown in columns (3)-(6) of Table A5 . Another way to mitigate the issue of PHUs with very few cases is to estimate a weighted least squares regression where PHUs are weighted by population. Columns (7) and (8) in Table A5 show that the resulting mask estimate has a slightly smaller magnitude and, due to the reduced effective sample size, weaker statistical significance. Similarly, Table A8 shows that our province-level estimates, in particular for mask mandates, are also robust to the same manipulations as above. 30 In columns (9) and (10) of Table A8 , we restrict the sample to only the largest 4 provinces (British Columbia, Ontario, Quebec and Alberta), which have only 0.3% (2 out of 624) zero observation cases. Again, the estimated mask effects are little changed. Figure B4 shows that our estimates and confidence intervals for the effect of mask mandates in the Ontario baseline regressions do not vary much by the initial date of the sample. Similarly, Figure B5 shows that, in the national analysis, our results about mask mandates and business/gathering restrictions are also robust to alternative sample start dates. We explore alternative time lags, either shorter or longer in duration, centered around the baseline value of 14 days. Figure B6 (with Ontario data) and Figure B7 (with province-level data) plot the estimates and confidence intervals from the baseline regressions and show that our mask effect estimates remain fairly consistent for different lags. Our behaviour proxy variable (Google geo-location trends) likely misses some aspects of behaviour relevant for COVID -19 transmission. One factor that may meaningfully impact behaviour is weather. For example, good weather could entice more people to spend time outside, lowering the chance of viral transmission. Columns (3) and (4) in Table A11 report national estimates with lagged weather variables (daily maximum and minimum temperatures and precipitation for the largest city in each province 31 ) as additional regressors. Our NPI estimates, in particular mask mandates, are little changed from the baseline results in columns (1) and (2) . All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . Another possible concern is that our information variables, lagged cases and lagged case growth, may not fully capture the information based on which people react or adjust their behaviour, possibly affecting the observed weekly case growth. Columns (5) and (6) in Table A11 add a national-level "news" variable to the baseline specification. The news variable is defined as the number of daily search results from a news aggregator website (Proquest Canadian Newsstream) for the terms "coronavirus" or "COVID-19" (see Appendix C for more details). In column (6), the lagged news variable approaches the 10% significance level (p = 0.103). Our estimates on masks and business/gathering remain very close to those in the baseline. The effectiveness of any NPI or public policy crucially depends on whether it affects behaviour. In this section, we use self-reported data on mask usage to examine whether mask mandates indeed increase mask use in Canada ("first-stage" analysis). We use data from the YouGov COVID- 19 Public Monitor, which includes multiple waves of public opinion surveys fielded regularly since early April in many countries. 32 Here, we focus on inter-provincial comparison within Canada. Our variable of interest is based on responses to the question "Thinking about the last 7 days, how often have you worn a face mask outside your home (e.g. when on public transport, going to a supermarket, going to a main road)?" The answer choices are "Always", "Frequently", "Sometimes", "Rarely", and "Not at all". We create a binary variable taking value 1 if the response is "Always" and 0 otherwise, as well as another variable taking value of 1 if the respondent answered either "Always" or "Frequently" and 0 otherwise. We begin with a simple illustration of self-reported mask usage in Canada from April to August 2020. Figure B2 plots the average self-reported mask usage (the response "Always") in the provinces with and without mask mandates. 33 The figure clearly shows that selfreported mask usage is higher, by up to 50 percentage points, in the provinces with a mask mandate than in the provinces without mask mandates. Since Figure B2 does not account for compositional changes in the data, we formally estimate equation (2), using self-reported mask usage as the behavioral outcome. 34 Notes: The data source is YouGov. The outcome is a binary variable taking value 1 if the respondent respectively answered "Always" (in the left panel) or "Always" or "Frequently" (in the right panel) to "Thinking about the last 7 days, how often have you worn a face mask outside your home?" The figure plots the estimates from a version of equation (2) where the mask policy variable is replaced by the interaction of the variables corresponding to being in the treatment group (imposed mask mandate) and a series of dummies for each week, ranging from 6 weeks before the mask mandate to 6 weeks after (T = -6 to +5, where T = 0 is the mandate implementation date). The reference point is 1 week before the implementation (T = -1). Wild bootstrap (cgmwildboot) standard errors clustered by province with 5000 repetitions are used to construct the confidence intervals. Sample weights are used. Figure 5 shows a graphical event study analysis on mask mandates and changes in mask usage. The event study approach is appropriate for the mask usage outcome variable, since the policy impact is expected to be immediate, unlike the other outcomes we study, for which any impact is expected to occur with a lag and we use weekly totals or moving averages. We replace the mask policy variable in equation (2) by the interaction of variables corresponding to being in the treatment group (i.e. under a mask mandate), and a series of dummies for each week, ranging from 6 weeks before the mask mandate to 5 weeks after the mask mandate (T = -6 to +5, where T = 0 is the implementation date of the mask mandate). The reference point is one week before the implementation of the mask mandate (T = -1), and we use the same y-axis scale on both panels. The left and right panels of Figure 5 present the results from the event study analysis for the "Always" and "Always" or "Frequently" mask usage answers, respectively. We make several observations. First, neither panel shows a pre-trend -the estimates are close to zero before the mask mandates. This addresses the potential concern that provinces that implemented mask mandates may have had a different trend in mask usage than provinces that did not. Second, the effect of mask mandates on mask usage is immediate: an increase 24 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . https://doi.org/10.1101/2020.09.24.20201178 doi: medRxiv preprint of roughly 20 percentage points as soon as the mask policy is implemented at (T = 0). Third, the effect appears persistent rather than transitory, since mask usage after T = 0 does not revert to its level before T = 0. Notes: The time period is April 2 to August 13, 2020. P-values from wild bootstrap (cgmwildboot) standard errors clustered by province with 5000 repetitions are reported in the square brackets. NC denotes national total cases. The data source is YouGov. The outcome is a dummy which takes value 1 if the respondent answered "Always" to the survey question "Thinking about the last 7 days, how often have you worn a face mask outside your home?" Sample weights are used. Individual characteristics include a gender dummy, age dummy (in years), dummies for each household size, dummies for each number of children, and dummies for each employment status. ***, ** and * denote 10%, 5% and 1% significance level respectively. Table 3 displays the estimates on self-reported mask usage (answer "Always") in equation (2) along with wild bootstrap p-values clustered at the province level. The odd-numbered columns use lagged cases and lagged case growth at the provincial level as information while the even-numbered columns include in addition lagged cases and case growth at the national level as additional information variables. As in Table 1 and Table 2 , we present estimates 25 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . without time trend, including cubic time trend (in days), and including week fixed effects. Our preferred specification with cubic time trend, column (4) of Table 3 , shows that mask mandates are associated with 31.5 percentage point increase in self-reported mask usage (p < 0.001), from a base of self-reported mask usage without mask mandate of 29.8%. 35, 36 These "first-stage" results show that mask mandates exhibit significant compliance in Canada and establish a basis for the significant impact of mask mandates on the spread of COVID-19 that we find. That said, given that mask mandates do not change everyone's behaviour, our estimates in Tables 1 and 2 represent intent-to-treat effects. The full effect of the entire population shifting from not wearing to wearing masks is likely significantly larger. 37 There is a heated debate on whether community use of masks may create a false sense of security that reduces adherence to other preventive measures. We also investigate this question using YouGov survey data. As Tables A13 and A14 indicate, we find no evidence that mask mandates in Canada have had an offsetting effect on other preventive measures such as hand washing, using sanitizer, avoiding gatherings, and avoiding touching objects in public during the period we study. On the contrary, mask mandates may slightly increase social distancing in one out of the eight precaution categories (avoiding crowded areas) (p < 0.10). 38 We evaluate several counterfactuals corresponding to replacing the actual mask policy in a province or Canada-wide with a counterfactual policy, including absence of mask mandate. Letting t 0 be the implementation date of a counterfactual policy, we set the counterfactual weekly case count, ∆C c it , equal to ∆C it for all t < t 0 . For each date t ≥ t 0 , using the definition of Y it from (3), we then compute the counterfactual weekly cases, ∆C c it and the counterfactual 35 Similarly, in Table A12 , column (4) shows that "Always" or "Frequent" mask usage increases by 21.5 percentage points. The finding that the increase in mask usage among the "Always" respondents is larger than among the "Always" or "Frequent" respondents is consistent with some people switching from wearing masks "frequently" to "always." 36 Hatzius et al. (2020) document that state mask mandates in the US increased mask usage roughly by 25 percentage points in 30 days. The compliance with mask mandates may differ across countries or regions based on social norms, peer effects, political reasons or the consequences of noncompliance (e.g., fines). 37 If we take the increase of about 30 percentage points in reported mask usage induced by mask mandates at face value, the full effect of mask wearing (treatment-on-the-treated effect) would be roughly triple our estimates. It could be larger still if there is desirability bias in answering the mask usage survey question, so that the actual increase in mask use may be smaller than our estimate. 38 Consistent with this result, Seres et al. (2020) find that wearing masks increased physical distancing based on a randomized field experiment in stores in Germany. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . case growth rate, Y c it , as follows: whereŶ it is the regression-fitted value of weekly case growth; β M ask 14 is the coefficient estimate on the mask mandate variable Mask 14 in baseline specification (4) in Table 1 or 2, depending on the counterfactual; Mask c 14 is the counterfactual mask policy (e.g. different implementation date, wider geographic coverage or absence of mask mandate); and β log∆C 14 is the coefficient estimate (-0 .227 or -0.209) on lagged cases log(∆C) 14 in Table 1 or 2, column 4. The coefficient β log∆C 14 adjusts the counterfactual case growth rate for the negative statistically significant association between the weekly case total two weeks prior and time-t case growth. This effect may be due to people being more careful when they perceive the risk of infection to be higher or less careful vice versa. Notes: The left panel assumes that mask mandates were adopted in all PHUs on June 12 (date of the first mask mandate in ON). The right panel assumes that mask mandates were not adopted in any PHU. We use the mask estimate (-0.376) from column (4) of Table 1. 27 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Notes: The left panel assumes that mask mandates were adopted in all provinces on July 7 (the adoption date in Toronto and Ottawa). The right panel assumes that mask mandates were not adopted in any province. We use the mask estimate (-0 .376) from column (4) of Table 1 . Notes: The left panel assumes that mask mandates were adopted in all provinces on July 7 (the adoption date in Toronto and Ottawa). The right panel assumes that mask mandates were not adopted in any province. We use the mask estimate (-0 .613) from column (4) of Table 2 . 28 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . Figures 6, 7 and 8 show results from two counterfactual policy evaluations. The first exercise, depicted in the left-hand side panel of the figures, assumes that masks are adopted everywhere at the earliest date observed in the data. Specifically, Figure 6 considers the counterfactual of all Ontario PHUs adopting mask mandates on June 12, while Figures 7 and 8 assume that all provinces adopt a mask mandate on July 7. 39 Using our mask policy estimate from Table 1, Figure 6 shows that an earlier face mask mandate across Ontario PHUs could have lead to an average reduction of about 300 cases per week as of August 13, holding all else equal. For Canada as a whole, a nation-wide adoption of mask mandates in early July is predicted to reduce total cases per week in the country by 700 to 1,100 cases on average as of August 13, depending on whether we use the more conservative mask estimate (-0 .376) from column (4) of Table 1 (see Figure 7 ) or the larger estimate (-0 .613) from column (4) of Table 2 (see Figure 8 ). In all cases, the indirect feedback effect via β log∆C 14 (lagged cases as information) starts moderating the decrease in cases two weeks after the start of the counterfactual mask policy. In the right-hand side panel of Figures 6, 7 and 8, we perform the opposite exercise, namely assuming instead that mask mandates were not adopted in any Ontario PHU or any Canadian province. Our estimates imply that the counterfactual absence of mask mandates would have led to a large increase in new cases, both in Ontario and Canada-wide, especially when using the larger mask coefficient estimate from Table 2 (see Figure 8 ). Finally, in Figure B11 in the Appendix, we also evaluate the counterfactual in which British Columbia and Alberta, the third and fourth largest Canadian provinces by population, adopt province-wide mask mandates on July 15. The results, using the Mask 14 estimate from Table 2 , suggest a reduction of about 300 cases per week in each province by mid-August. The counterfactual simulations assume that all other variables, behaviour and policies (except the mask policy and t − 14 cases) remain fixed, as observed in the data. This is a strong assumption, but it may be plausible over the relatively short time period that we analyze. Moreover, the counterfactuals assume that regions without a mask mandate would react in the same way, on average, as the regions that imposed a mandate. Therefore, these results should be interpreted with caution and only offer a rough illustration and projection of the estimated effect of mask mandates on COVID-19 cases. 39 June 12 is the date of the earliest mask mandate in Ontario. For the national analysis, July 7, the effective date for Toronto and Ottawa, is considered Ontario's first significant date of mask mandate enactment: PHUs with earlier mandates account for less than 10% of Ontario's population. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . https://doi.org/10.1101/2020.09.24.20201178 doi: medRxiv preprint Closing and re-opening sub-periods We investigate whether policy impact varied in different phases of the pandemic by splitting the full sample period into two sub-periods: "closing" (March 11 to May 14) and "reopening" (May 15 to August 13). The dividing date of May 15 (referring to the NPIs in place around May 1) was chosen because very few policies were relaxed before May 1, and very few non-mask policies were tightened after May 1 in our sample period (see Figure 2) . In Table A15 , we report estimates and wild bootstrap standard errors using our baseline specification with cubic time trend, separately for the closing and re-opening periods. We find that the imposition of school closures and travel restrictions early in the closing period is associated with a very large subsequent reduction in weekly case growth, as can be also seen on Figure B8 -the average observed log growth rate of cases ∆ log(∆C) falls from 2.4 (ten-fold growth in weekly cases) to −0.4 (33% decrease in weekly cases) between March 15 and April 5. Long-term care restrictions are also associated with reduced case growth two weeks later during the March to May closing period. We interpret these results with caution, however, since many of these policy measures and restrictions were enacted in a brief time interval during March and there is not much inter-provincial variation (see Figure 2 ). No mask mandates were present in the closing period. In the re-opening period, our results in Table A15 are in line with our full-sample results for mask mandates and business/gathering regulations (Table 2) , with slightly larger coefficient estimates and less statistically significant p-values, possibly due to the smaller sample. Travel and school closures are not statistically significant in the re-opening period. This is unsurprising: relaxation of travel policies was minor and endogenous (only re-open to safe areas within Canada), and the schools that re-opened (in parts of Quebec and, on a part-time basis, in British Columbia) did so on voluntary attendance basis, yielding smaller class sizes. We also examine the weekly death growth as an outcome. We only have access to disaggregated deaths data at the province level (not at PHU levels in Ontario). We thus estimate regression equation (1) using Y it = ∆ log(∆D it ) for each province i as the dependent variable. In addition, we use a 28-day lag for the policy, behaviour proxy, and information variables to reflect the fact that deaths occur on average about two weeks after case detection; see Appendix D for details and references. 40 40 In Table 4 , Variable 28 denotes the Variable lagged by 28 days. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . Table 4 reports the estimates from the same specifications as those for case growth in Table 2 . In all specifications, mask mandates are associated with a large reduction in the observed weekly deaths growth rate four weeks later (more than 90 log points, or equivalently more than 60% reduction in weekly deaths). These results are larger than our case growth results, but consistent with them given the substantial uncertainty. See also Figure B12 , which plots the average weekly death growth in the provinces without a mask mandate four weeks prior vs. that for Ontario, the only province with mask mandate four weeks prior in our sample period. The robustness checks in Table A16 , however, show that, unlike for case growth, the mask mandate estimates in Table 4 are not robust to weighing by population or to restricting the sample to the largest 4 provinces. This suggests that the estimated effect is largely driven by observations from the small provinces, which have a disproportionately larger number of zero or small weekly death totals. 41 Furthermore, given the 28-day lag, there are only 9 days with observations (from Ontario only) for which the mask mandate variable takes value of 1. Due to these serious data limitations, the relation between mask mandates and COVID -19 deaths in Table 4 is suggestive at best, and we urge caution in interpreting or extrapolating from these results. That said, our main findings about the growth in cases may have implications about future growth in deaths, particularly if the affected demographics become less skewed toward the young in later periods. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The wearing of face masks by the general public has been a very contentious policy issue during the COVID-19 pandemic, with health authorities in many countries and the World Health Organization giving inconsistent or contradictory recommendations over time. "Conspiracy theories" and misinformation surrounding mask wear abound in social media, fuelled by some individuals' perception that mask mandates constitute significant restrictions on individual freedoms. Given the absence of large-scale randomized controlled trials or other direct evidence on mask effectiveness in preventing the spread of , quantitative observational studies like ours are essential for informing both public policy and the public opinion. We estimate the impact of mask mandates and other public policy measures on the spread of in Canada. We use both within-province and cross-province variation in the timing of mask mandates and find a robust and significantly negative association between mask mandates and subsequent COVID-19 case growth -25 to 46% average reduction in weekly cases in the first several weeks after adoption, depending on the data sample and empirical specification used. These results are supported by our analysis of survey data on compliance with the mask mandates, which show that the mandates increase the proportion of reporting as always wearing a mask in public by around 30 percentage points. However, our sample period does not allow us to determine whether their effect lasts beyond the first few weeks after implementation. We conclude that mask mandates can be a powerful policy tool for at least temporarily reducing the spread of Mask mandates were introduced in Canada during a period where other policy measures were relaxed, as part of the economy's re-opening. Specifically, we find that relaxed restrictions on businesses or gatherings are positively associated with subsequent COVID-19 case growth -a factor that could offset and obscure the health benefits of mask mandates. Past case totals were also found to matter for subsequent outcomes, suggesting that riskier behaviour based on favourable lagged information may limit how low mask mandates and other restrictions -short of a lockdown -can push the number of new cases. We have deliberately abstained from studying the direct economic impacts of , focusing instead on the unique features of the Canadian data for identifying the effect of NPIs, in particular mask mandates, on COVID-19 case growth. Future research combining epidemiological finding with the economic benefits and costs of various public policies or restrictions would enrich the ongoing policy debate and provide further guidance. 33 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. 38 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (1) and (2) repeat columns (3) and (4) from Table 1 where we replace log(0) with -1. Columns (3) and (4) replace log(0) with 0, and columns (5) and (6) add 1 to all ∆C it observations. Columns (7) and (8) report estimates from a weighted least squares regression with weights equal to the PHU population sizes. ***, ** and * denote 10%, 5% and 1% significance level respectively. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . 40 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . (1) and (2) repeat columns (3) and (4) from Table 1 . We drop each policy at at time in columns (3)- (8) . ***, ** and * denote 10%, 5% and 1% significance level respectively. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . The time period is April 2 to August 13, 2020. P-values from wild bootstrap (cgmwildboot) standard errors clustered by province with 5000 repetitions are reported in the square brackets. NC denotes national total cases. The data source is YouGov. The outcome is a dummy which takes one for the respondent who answers "Always" or "Frequently" to the survey question "Thinking about the last 7 days, how often have you worn a face mask outside your home?" Sample weights are used. Individual characteristics include a gender dummy, dummies for each age (in years), dummies for each household size, dummies for each number of children, and dummies for each employment status. ***, ** and * denote 10%, 5% and 1% significance level respectively. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . Notes: The time period is April 2 to August 13, 2020. P-values from wild bootstrap (cgmwildboot) standard errors clustered by province with 5000 repetitions are reported in the square brackets. NC denotes national total cases. The data source is YouGov. The outcome is a dummy which takes value 1 if the respondent answered "Always" or "Frequently" to each survey question in Table C4 . Sample weights are used. Individual characteristics include a gender dummy, age dummy (in years), dummies for each household size, dummies for each number of children, and dummies for each employment status. ***, ** and * denote 10%, 5% and 1% significance level respectively. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The time period is Feb 26 to July 30 (two weeks before the March 11 -August 13 sample period). Daily province-level data. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . Notes: The data source is YouGov. The figure plots the average self-reported mask usage by week (the fraction of respondents who answered "Always" to the survey question "Worn a face mask outside your home") in the provinces with vs. without mask mandates. Sample weights used to compute the averages. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . Figure B3 : Canada -Behaviour Notes: The Behaviour proxy B it is the average of the "retail", "grocery and pharmacy", and "workplaces" Google mobility indicators. Province-level 7-day moving averages are plotted. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Notes: We plot the coefficient estimates on mask policy, with 95% confidence intervals, from equation (1), for different initial dates of the sample. The initial sample date in the baseline specifications reported in Table 1 Notes: We plot the coefficient estimates on mask policy, with 95% confidence intervals, in the upper panel and the estimates on business/gathering policy in the lower panel, from equation (1) for different initial dates of the sample. The initial date in our baseline specification (Table 2) is March 11. The left panels correspond to column (3) in Table 2 ; the right panels correspond to column (4) in Table 2 . All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Table 2 ; the right panels correspond to column (4) in Table 2 . All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . https://doi.org/10.1101/2020.09.24.20201178 doi: medRxiv preprint Figure B8 : Canada -Weekly cases, deaths and tests (growth rate) 27 week ending preprint (which was not certified by peer review) 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 25, 2020. . https://doi.org/10.1101/2020.09.24.20201178 doi: medRxiv preprint Figure B9 : Canada -Weekly cases, deaths and tests (level) 27 60 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . https://doi.org/10.1101/2020.09.24.20201178 doi: medRxiv preprint Figure B10 : Canada -Daily cases, deaths and tests 27- All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . https://doi.org/10.1101/2020.09.24.20201178 doi: medRxiv preprint Figure B11 : Counterfactuals -Mask No mask mandate at t-28 Mask mandate at t-28 Notes: Average log weekly death growth in provinces with vs. without mask mandates 28 days prior. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . https://doi.org/10.1101/2020.09.24.20201178 doi: medRxiv preprint Used hand sanitiser i12 health 6 Avoided going out in general i12 health 12 Avoided small social gatherings (not more than 2 people) i12 health 13 Avoided medium-sized social gatherings (between 3 and 10 people) i12 health 14 Avoided large-sized social gatherings (more than 10 people) i12 health 15 Avoided crowded areas i12 health 20 Avoided touching objects in public (e.g. elevator buttons or doors) Notes: the data source is YouGov. Possible responses to each survey item are "Always", "Frequently", "Sometimes", "Rarely", and "Not at all". For Table A13 , we create a binary variable taking value 1 if the response is "Always" and 0 otherwise. For Table A14 , we create a binary variable taking value of 1 if the respondent answered either "Always" or "Frequently" and 0 otherwise. All data used in the paper are available at https://github.com/C19-SFU-Econ/data. As discussed in Section 3.1, we assume a lag of 14 days between a change in policy or behaviour and its hypothesized effect on weekly case growth, and a lag of 28 days between such a change and its effect on weekly death growth. First, we consider the lag between infection and a case being reported. As most identified cases of in Canada are symptomatic, we focus on symptomatic individuals. For most provinces cases are listed according to the date of report to public health. In provinces where the dates instead refer to the public announcement, we shifted them back by one day, as announcements typically contain the cases reported to public health on the previous day. The relevant lag therefore has two components: 2. Time between symptoms onset and reporting of the case to public health: the Ontario data contain an estimate of the symptom onset date ("episode date") for each case. For our sample period the average difference between the date of report and the episode date is 4.8 days (median: 4 days) including only values from 1 to 14 days, and 6.3 66 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . days (median: 5 days) including only values from 2 to 28 days. We assume that the lags in Ontario and in other provinces are similar, and use a value of [5] [6] days between symptom onset and report to public health authorities. Adding these together implies that the typical lag between infection and a positive case being reported to public health is around 11 days. Second, we consider the effect of weekly averaging on the appropriate lag for our analysis. Suppose a policy or behavioural change starts on date t, impacting the daily growth in infections between dates t − 1 and t and in each subsequent day. Then, assuming a lag of 11 days between infection and case reporting, case counts C are affected from date t+11 onward. Our outcome variable ∆ log(∆C) thus would react to the original policy or behavioral change on date t + 11. The change is complete on t + 23, when the week from t + 17 to t + 23 is compared to the week from t + 10 to t + 16. The midpoint of the change is t + 17. Choosing a lag of l days implies that the policy/behaviour variable phases in from t + l to t + l + 6. To match the midpoint of this phase-in to the midpoint of the change in the outcome variable, we set l = 14. The chosen lag matches the lag used by other authors who study policy interventions, e.g., CKS (2020). We explore sensitivity to alternative lags in Section 4.3. With respect to deaths, our data are, in most cases, backdated (revised by the authorities ex , that is, two weeks longer than our estimate of the time from symptom onset to reporting of a positive test result. We correspondingly set the lag used in our analysis of the death growth rate (Section 4.6) to 28 days. 67 All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) 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 25, 2020. . https://doi.org/10.1101/2020.09.24.20201178 doi: medRxiv preprint The Case for Universal Cloth Mask Adoption and Policies to Increase Supply of Medical Masks for Health Workers Incubation period of 2019 novel coronavirus (2019-nCoV) infections among travellers from Wuhan, China Bootstrap-based improvements for inference with clustered errors Causal Impact of Masks, Policies, Behavior on Early COVID-19 Pandemic in the 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 Face masks for the public during the covid-19 crisis The effect of large-scale anti-contagion policies on the COVID-19 pandemic A contribution to the mathematical theory of epidemics The Incubation Period of Coronavirus Disease 2019 (COVID-19) From Publicly Reported Confirmed Cases: Estimation and Application Respiratory virus shedding in exhaled breath and efficacy of face masks Early Transmission Dynamics in Wuhan, China, of Novel Coronavirus-Infected Pneumonia Incubation Period and Other Epidemiological Characteristics of 2019 Novel Coronavirus Infections with Right Truncation: A Statistical Analysis of Publicly Available Case Data Community Use Of Face Masks And COVID-19: Evidence From A Natural Experiment Of State Mandates In The US Transmission of SARS-CoV-2: A Review of Viral, Host, and Environmental Factors Face Masks Considerably Reduce COVID-19 Cases in Germany Public Responses to Policy Reversals: The Case of Mask Usage in Canada during COVID-19 Flattening the COVID-19 peak: Containment and mitigation policies Reducing transmission of SARS-CoV-2 COVID-19 US state policy database High Contagiousness and Rapid Spread of Severe Acute Respiratory Syndrome Coronavirus 2 Face Mask Use and Physical Distancing before and after Mandatory Masking: Evidence from Public Waiting Lines Table 17-10-0009-01 Population estimates, quarterly Estimates of the severity of coronavirus disease 2019: a model-based analysis Weekly Epidemiological Update, Coronavirus disease 2019 (COVID-19) Estimating clinical severity of COVID-19 from the transmission dynamics in Wuhan, China log(∆C) -2.545 *** -2.062 *** -1.935 *** -1.537 *** -1.387 ** -0.912 * -1.956 *** -1.504 *** 1 We show mask usage for the U.S. and Germany because related work by Chernozhukov et al. (1) and (2) repeat columns (3) and (4) from Table 2 where we replace log(0) with -1. Columns (3) and (4) replace log(0) with 0, and columns (5) and (6) add 1 to all ∆C it observations. Columns (7) and (8) report results from a weighted least squares regression with the province populations as weights. Finally, columns (9) and (10) (3) and (4) report estimates with lagged weather variables as additional controls. Columns (5) and (6) add a "news" variable to the baseline specification (see Appendix C for more details). ***, ** and * denote 10%, 5% and 1% significance level respectively. Weather -we downloaded historical weather data for the largest city in each province from the Weather Canada website. The data provide daily information on 11 variables: maximum temperature (C), minimum temperature (C), mean temperature (C), heating degreedays, cooling degree-days, total rain (mm), total snow (cm), total precipitation (mm), snow on the ground (cm), direction of maximum wind gust (tens of degrees), and speed of maximum wind gust (km/h). We only use the temperature and precipitation data in Table A11 as possible factors determining outside vs. inside activity.News -we collected data from Proquest Canadian Newsstream, a subscription service to all major and small-market daily or weekly Canadian news sources. We recorded the number of search results for each day from Feb 1, 2020 to Aug 20, 2020 by searching the database for the keywords "Coronavirus" or "COVID-19". We only counted the results with source listed as "newspaper" since other sources, such as blogs or podcasts, tend to duplicate the same original content.