key: cord-0704851-w9rw9vc8
authors: Karaivanov, A.; Kim, D.; Lu, S. E.; Shigeoka, H.
title: COVID-19 Vaccination Mandates and Vaccine Uptake
date: 2021-10-25
journal: nan
DOI: 10.1101/2021.10.21.21265355
sha: 8dbb16fd2656084e199cc03e8ae4da06e4f6de87
doc_id: 704851
cord_uid: w9rw9vc8

We estimate the impact of government-mandated proof of vaccination requirements for access to public venues and non-essential businesses on COVID-19 vaccine uptake. We use event-study and difference-in-differences approaches exploiting the variation in the timing of these measures across Canadian provinces. We find that the announcement of a vaccination mandate is associated with large increase in new first-dose vaccinations in the first week (more than 50% on average) and the second week (more than 100%) immediately following the announcement. The estimated effect starts waning about six weeks past the announcement. Counterfactual simulations using our estimates suggest that these mandates have led to about 289,000 additional first-dose vaccinations in Canada as of September 30, 2021, which is 1 to 8 weeks after the policy announcements across the different provinces. Time-series analysis corroborates our results for Canada, and we further estimate that national vaccine mandates in three European countries also led to large gains in first-dose vaccinations (7+ mln in France, 4+ mln in Italy and 1+ mln in Germany, 7 to 12 weeks after the policy announcements). NOTE: The reported numbers may change with more data. Please see updated version when available.

Immunization with mRNA or adenoviral vaccines has proven a very effective preventive measure for reducing the spread and severity of COVID-19, with fully vaccinated people facing more than 10 times lower risk from severe outcomes. 1 Yet, following a quick-paced uptake in early 2021, COVID-19 immunization rates in many countries slowed down significantly at about 60% of the overall population during the summer months despite the vaccines' proven benefits (see Fig. B1 ). In addition, most jurisdictions -even those with relatively high vaccination rates -either experienced increased viral transmission or had to maintain or re-introduce non-pharmaceutical interventions such as mask-wearing mandates or indoor capacity limits because of the elevated reproduction rate of the COVID-19 Delta variant. 2 Therefore, in many places, a further increase in the COVID-19 vaccination rate remains essential to minimize the health and economic impacts of the virus and enable lifting of the remaining restrictions. Public health authorities around the world have actively sought the most effective strategy to increase vaccine uptake and provide incentives for hesitant people to become immunized. In response to this challenge, several governments have recently introduced proof of vaccination mandates, which allow only vaccinated persons to attend non-essential sports or social activities and settings such as stadiums, concerts, museums, restaurants, bars, etc. 3 The goal of these policies is to boost the vaccination rate and to reduce viral transmission in risky settings. 4 We evaluate and quantify the effect of proof of vaccination mandates on first-dose vaccine uptake in nine Canadian provinces and three European countries (France, Italy and Germany) that announced such mandates during July-September, 2021. We use first doses, as they most directly capture the impact of a vaccine mandate on the decision to be immunized. While requiring proof of vaccination is naturally expected to raise vaccine uptake, the magnitude of the increase is hard to predict, as it depends on how much of people's hesitancy or procrastination to become vaccinated stems from a lack of social or economic incentives, or from entrenched political or religious beliefs. where t is the date on the horizontal axis. The dashed red lines denote the vaccine mandate announcement dates (for the countries this is the date of a national mandate). Spain has not announced a national proof of vaccination mandate. We show the four most populated Canadian provinces totalling about 87% of Canada's population (see Fig. B3 for all nine provinces with mandates). Alberta also has a $100 debit card incentive for doses received between Sep. 3 and Oct. 14, 2021. Fig. 1 plots the weekly vaccine first doses for the four most populous provinces in Canada and four European countries. All except Spain have announced a proof of vaccination mandate. We observe a sizable boost in vaccine uptake within two weeks after the mandate announcement date (the dashed vertical line) in all four provinces and in France, Italy and Germany, often following a sharp decline in the pre-announcement weeks. In contrast, Spain exhibits a steady decline in weekly first doses over the displayed period.

Motivated by this evidence, we aim to answer two important policy questions. First, what is the magnitude and time profile of new first dose vaccinations after a vaccine mandate is announced? Second, if an increase in vaccinations is observed, as seen in the data, does the increase occur because of intertemporal substitution (people receive the vaccine sooner than otherwise intended), or is it a net gain in vaccinations (people that otherwise would not have become immunized do so in response to the mandate)? In both cases, quantifying the increase in vaccinations is important, as it can speed up progress toward herd immunity. To the best of our knowledge, this is the first paper to rigorously assess the impact of COVID-19 proof of vaccination mandates on vaccine uptake.

Using Canadian province-level data in event-study and difference-in-differences (DID) approaches, we document a rapid and statistically and economically significant increase in first-dose vaccine uptake following the announcement of a proof of vaccination mandate. 5 The average estimated increase in new weekly first doses relative to the absence of mandate is large -55% (44 log points) for the first post-announcement week and more than 100% for the second week (81% on average over all weeks, as of Sep. 30, 2021). The estimated increase in first dose vaccinations gradually decreases after the second week past the mandate announcement and dissipates after seven weeks.

The Canadian province-level data is key for our identification strategy as it allows us to use the time variation in mandate announcement dates (ranging from Aug. 5, 2021 for Quebec to Sep. 21, 2021 for Prince Edward Island) across different geographic units in the same country via a difference-in-differences approach. 6 In contrast, the French, Italian or German mandates or the forthcoming U.S. vaccine mandate for employees apply at the national level, which makes it difficult to separate the effect of the vaccine mandate from that of time trends or other concurrent events and policies.

We complement the results for Canadian provinces with structural break and time-series analysis including data from France, Italy and Germany. We find, using a Chow test, that we can reject the null hypothesis of no structural break at the announcement date for all locations. The time-series estimates confirm our DID results showing a significant rise in first-dose uptake in the first two weeks after a mandate announcement relative to the preannouncement trend (the estimated effect varies, from as low as 6% increase in Ontario to 5 We also collected the dates of mandate implementation and find no additional effect on top of the announcement effect. The time between a mandate announcement and its coming into force varies across jurisdictions from less than a week to more than a month, see Table C1 . 6 In Canada, the provinces are separate public health jurisdictions with extensive powers over health policy while the vaccines are procured by the federal government and allocated to the provinces in proportion to their population. All COVID-19 vaccines used in Canada are two-dose vaccines. Namely, BNT162b2 (Pfizer /BioNTech Comirnaty), mRNA-1273 (Moderna SpikeVax) and AZD1222 (Oxford-AstraZeneca Vaxzevria). as high as 125% in Alberta), followed by a waning of the effect.

Counterfactual simulations using our estimates suggest that, as of September 30, 2021, all provinces and countries have achieved net gains in first-dose vaccinations relative to the hypothetical scenario of absence of vaccine mandate. We estimate 289,000 additional first doses for Canada as a whole (0.9 percentage points of the eligible population) as of September 30, 2021, which is one to eight weeks after the policy announcements across the different provinces. The largest absolute gains are estimated for Alberta (91,000 additional vaccinations or 2.4 p.p. of all eligible), British Columbia (69,000 doses, 1.5 p.p.) and Ontario (64,000, 0.5 p.p.). We also estimate 7.8 mln additional first doses in France, 4.5 mln in Italy and 1.3 mln in Germany, relative to the no-mandate counterfactual, as of September 30, 2021 (seven to twelve weeks after the respective policy announcements).

To avoid potential bias from constrained vaccine supply affecting the pre-or post-mandate pace of vaccination, we limit our sample to the period after June 15, 2021, when it is safe to assume that any age-eligible person (12 or older in Canada) was able to receive a first dose of COVID-19 vaccine with zero or minimal delay. 7 June 15 is almost 3 weeks after the eligibility for a first dose of the Comirnaty vaccine was expanded to ages 12+ by Nova Scotia, the last Canadian province to do so (see Table C2 ).

Vaccination mandates have been controversial, as some people perceive them as restrictions on personal freedom. This could affect compliance and increase both the direct costs of implementing and enforcing the mandates, as well as the political costs of introducing them. In this paper, we do not address ethical considerations; our aim is instead to assess the mandates' effectiveness purely in terms of boosting vaccination rates, which can then be weighed against the various costs or compared to other policy approaches such as financial incentives (e.g., cash, gift cards, lotteries) or behavioral nudges (e.g., messages from medical experts, reminder calls for appointments). In related work, Campos-Mercade et al. (2021) report results from a randomized controlled trial in Sweden and find that a modest monetary payment of SEK 200 (USD 24) is associated with an increase in the vaccination rate by 4.2 percentage points, relative to a baseline rate of 71.6% in the control group, while none of three behavioral nudges had statistical significant impact on the vaccination rate. 8 2 Data and empirical methods

We use data on COVID-19 vaccination numbers, cases and deaths for all ten Canadian provinces, as well as for France, Italy, Germany and Spain. 9 Announcement dates and implementation dates of the proof of vaccination mandates were collected from the respective government websites and major newspapers (see Table C1 ).

The main variables in our empirical analysis are defined below. Everywhere, i denotes province for the Canada analysis or country otherwise, and t denotes time measured in days (date). We aggregate the data on vaccinations, cases and deaths on a weekly basis (totals for the week ending on date t, i.e., dates t − 6 to t) to reduce the influence of day-of-the-week effects and reporting artifacts. 10 Outcome, V it . Our main outcome variable is the log of the number of vaccine first doses administered, V it , for the week ending at t. We use first doses as they most directly reflect the impact of the vaccine mandate on the intent to be immunized. 11 Using the logarithm of weekly first doses allows us to interpret the panel regression estimates as percentages and they are invariant to scale effects and normalization, e.g., by population. Policy, P it . Denote byt i the announcement date of the proof of vaccination mandate in jurisdiction i. We construct a binary policy variable P it that takes the value 1 for all post-announcement dates t ≥t i and the value 0 for t <t i . Proof of vaccination mandates have been officially announced in nine Canadian provinces as of Sep. 30, 2021, first in Quebec on Aug. 5, 2021 and latest in Prince Edward Island on Sep. 21, 2021 (see Fig. B2 ). Several European countries, including France, Italy and Germany, also introduced proof of vaccination mandates in summer 2021 (see Table C1 ). Information, I it . We construct variables related to the underlying COVID-19 epidemiological situation, specifically log of weekly cases C it and deaths D it for the week ending at date t, dates t − 6 to t. 12 We refer to these variables jointly as "information", I it since 9 See Table C3 and the paper's Github repository for all sources and details. We collected the Canadian data from the official provincial dashboards or equivalent sources. We use the Our World In Data dataset for the country data. 10 Some locations (e.g., British Columbia, Nova Scotia) do not report vaccination data on weekends and then report the total for several days at once (e.g., Monday's number contains 3 days of data). In these cases we spread the reported total equally over the affected days. 11 In Table A5 we also report results using second doses as the outcome. We find no statistically significant effect on second dose uptake following announcement of vaccine mandate. 12 We also checked including hospitalization and ICU levels in the analysis (weekly moving averages for dates t − 6 to t). However, these variables are very strongly correlated with the COVID-19 cases and deaths and so we do not include them in our baseline specification to avoid multicollinearity issues. they are likely to play an information role and influence a person's COVID-19 exposure risk assessment and/or decision to be vaccinated. 13 Controls, W it . We include province fixed effects, date fixed effects and the day-of-week by province fixed effects. Province fixed effects account for any time-invariant province-level characteristics such as the sentiment towards vaccination and the age distribution. Date fixed effects control for nation-wide trends and events such as political campaigning for the September 2021 federal election. Day-of-week by province fixed effects account for remaining data reporting artifacts by province. Time period. We use the period June 15, 2021 to September 30, 2021 for our baseline analysis. The end date is based on data availability at the time of statistical analysis and writing. The start date was chosen so that possible constraints to obtain a first dose related to eligibility or vaccine supply are minimal or non-existent. 14 We explore different start dates in robustness checks. Fig. 2 shows that weekly first doses in the Canadian provinces and in France, Italy and Germany grow quickly after the proof of vaccination mandate announcement, reaching a peak at about 1 or 2 weeks after the announcement date, and then decline. 15 Some provinces, notably New Brunswick, Alberta, Nova Scotia and Saskatchewan, registered increases in firstdose vaccine uptake of over 100% relative to the pre-announcement week within 2 weeks of the announcement date. For other provinces, e.g., Quebec, Ontario and Manitoba the increase is more moderate (below 50%). In Quebec, as well as in France, Italy and Germany, for which we have data for a longer period after the mandate announcement, the number of weekly first doses falls below the pre-announcement week level about 6 weeks past the announcement. 13 To deal with zero weekly values, which sometimes occur in the smaller provinces for deaths or cases (4.4% of all observations for cases, and 10.7% for deaths), we replace log(0) with -1 as in Chernozhukov et al. (2021) and Karaivanov et al. (2021) .

14 In Canada, the provinces opened registration for first-dose vaccination for any person of age 12+ between May 10, 2021 in Alberta and May 27, 2021 in Nova Scotia, see Table C2 . First-dose availability in France, Italy and Germany was similar by mid-June, at least for the 18+ age group. 15 British Columbia and Ontario exhibit a second peak around 4 weeks after the announcement.

8 All rights reserved. No reuse allowed without permission.

(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 figure plots the observed weekly first doses of COVID-19 vaccine, V it for all dates t after the mandate announcement against the number of days after the respective announcement date (denoted by 0 on the horizontal axis). The weekly first doses for the week just prior to the mandate announcement are normalized to 100 for each respective province (on the left) or country (on the right).

We estimate a behavioral model in which the decision to receive a first dose of a COVID-19 vaccine (measured by new weekly doses, V it ) is affected by the policy setting, P it (whether a proof of vaccination mandate has been announced), and the current COVID-19 epidemiological and public health conditions, I it , proxied by weekly cases and deaths. Additionally, in the panel data analysis, we control for other potential confounding variables or unobserved heterogeneity by including time and/or location fixed effects and interaction terms. Based on the raw data patterns on Fig. 1 and Fig. B3 and the absence of vaccine supply constraints in the time period we analyze, we assume no lag between the mandate announcement and a person's ability to receive a first vaccine dose. 16 Finally, we assume that individuals take into account the current public health conditions information, I it (aggregate numbers for the week ending at date t) with no further lag.

We use the following difference-in-differences (DID) model:

where the controls W it include province fixed effects, date fixed effects and day-of-week-byprovince fixed effects and where ε it is an error term. The coefficient α reflects the average effect of the vaccine mandate announcement on weekly first doses. We also estimate a version of equation (1) where P it is broken into multiple binary variables, one for each week after the mandate announcement date, to capture dynamic effects. To account for the small number of clusters in the estimation due to there being only ten provinces, we use "wild bootstrap" standard errors and p-values as in Cameron et al. (2008) . 17 3 Results

Fig. 3 displays the results from an event study analysis of weekly first doses administered in Canadian provinces from eight weeks before to eight weeks after the announcement of a proof of vaccination mandate. In Fig. 3 , T = 0 denotes the announcement date, and the reference point is one week before the mandate announcement (T = -1). We make several observations. First, there is no mandate-associated pre-trend -the estimates before the mandate announcement are statistically indistinguishable from zero. This addresses the potential endogeneity concern that provinces that announced a vaccine mandate may have had a different trend in first dose vaccinations than provinces that did not announce a mandate. Second, the impact of a mandate announcement on first-dose vaccine uptake is relatively quick, with the largest estimated increase around one week after the announcement. Third, the effect on the pace of first-dose vaccinations is large in magnitude -an increase of 59 log points (81 percent) in weekly first doses on average. The observed large initial increase mitigates possible concerns that vaccine supply or administration constraints may be affecting our results (see also Fig. 2 , which plots the increase in first-dose vaccinations by province after mandate announcement in the raw data). Fourth, the average policy effect gradually decreases over time and dissipates by week 7 after the mandate announcement. Table 1 shows the estimates of equation (1) along with p-values from wild bootstrap 17 Alternative ways of computing the standard errors are explored in Section 3.1.2 and Table A4. 10 All rights reserved. No reuse allowed without permission.

(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 preprint this version posted October 25, 2021. ; Figure 3 : Canadian provinces -Event study analysis

weeks from vaccine mandate announcement The outcome variable is log weekly first doses administered for dates t − 6 to t inclusive. The figure plots the estimates from a version of equation (1) where the mandate announcement variable is replaced by the interaction of being in the 'treatment' group (announced vaccine mandate) with a series of dummies for each week ranging from exactly 8 weeks before (T = −8) to exactly 7 weeks after the announcement (T = +7), where T = 0 is the announcement date. The reference point is one week before the announcement (T = -1). The dotted lines correspond to 5-95% confidence intervals.

standard errors clustered by province in the square brackets. 18 Column (1) shows that the vaccine mandate announcement is associated with a 59 log points or 81 percent (p< 0.01) increase in weekly first doses on average. Adding information (log weekly cases and deaths) in column (2) has almost no effect on the estimate. Column (3) of Table 1 reports estimates from a variant of equation (1) where the policy variable P it is split into multiple binary variables, one for each week after the mandate announcement date, to account for dynamic effects. The results indicate a fast increase in weekly first doses of 44 log points (55%, p< 0.01) in the week starting at the announcement date (week 0), reaching a peak of 77 log points (116%, p< 0.001) increase in the following week (week 1). The estimate for week 2 is also statistically significantly different from zero (p< 0.01). The estimates for the later weeks remain economically large but are not statistically significantly different from zero, and the estimated mandate effect on vaccine uptake dissipates 7 weeks past the announcement. We note that the estimate on "weeks 7+," which captures the average effect for all dates from week 7 onward, is close to zero. This (3) is statistically significantly positive (p< 0.05), which is consistent with higher vaccine uptake when the observed number of deaths is higher (e.g., because of elevated fear of COVID-19), all else equal.

Alternative initial dates. The top left panel of Fig. B4 shows that our estimates of the effect of a vaccine mandate announcement on first dose vaccine uptake are not sensitive to the choice of initial sample date between May 1 and Aug. 1, 2021. This gives us reassurance that vaccine supply constraints do not appear to be a major concern for our results.

Drop province. The top right panel of Fig. B4 shows that our estimates are robust to omitting any of the ten provinces from the regression, one at a time. This provides reassurance that our main results are not affected by a particular province.

Lags. In equation (1), we assume no lag between the mandate announcement P it and a person's ability to receive first vaccine dose (the outcome V it ). In practice, a few days of delay may occur (e.g., from making an appointment to receiving the vaccine) even though vaccine supply constraints were not present in the sample period. The bottom left panel of Fig. B4 displays the coefficient estimates on the announcement variable when assuming no lag (our baseline model) vs. assuming up to 7-day lag for the mandate variable. The policy effect estimate remains roughly stable when varying the lag length, with a slight decrease for longer lags. The bottom right panel of Fig. B4 shows that the R-squared value is the highest when the lag is zero, supporting our choice of baseline model (see Hsiang et al., 2020) . 19 Heterogeneous treatment effects. Several recent papers argue that the two-way fixed effects estimator (province and date in our case) is a weighted average of treatment effects, where some of the weights can be negative if the treatment effect is not constant across groups and over time. 20 In particular, Sun and Abraham (2021) develop an estimator that is valid under these conditions. We apply their estimator to our baseline specification. 21 We obtain an estimate of 0.602 (s.e. 0.039) for the vaccine mandate announcement coefficient, which is very close in magnitude to the two-way fixed effect estimate (0.593) from our baseline 19 We also checked a version in which we also lag the information variables I (cases and deaths) with the same lag as the policy P and the results are nearly identical. 21 We use the Stata command "eventstudyinteract" provided by the authors.

All rights reserved. No reuse allowed without permission.

(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 preprint this version posted October 25, 2021. ; specification, column (2) of Table 1 .

Other robustness checks. We perform additional robustness checks in Table A2 . For ease of comparison, column (1) replicates our baseline estimates from column (2) of Table 1 . Column (2) of Table A2 removes log weekly cases, the estimated effect of which is near zero in the baseline specification, as control variable. Column (3) allows the day-of-the-week fixed effects to have differential impact not just by province, but also by year-month for further flexibility (e.g., to account for possible changes in reporting patterns by day of the week). Column (4) replaces the date fixed effects by year-month fixed effects interacted with a dummy variable for each province to control more flexibly for province-specific changes potentially correlated with the vaccine uptake such as weather or news coverage. The estimates in Table A2 , columns (2) through (4) are slightly smaller in magnitude than our baseline estimate (column 1) but remain economically large. Column (5) of Table A2 reports the estimates of a specification with the provincial populations as weights and the baseline set of variables. The resulting estimate is a bit smaller and has weaker statistical significance than that from the unweighted baseline in column (1), suggesting that smaller provinces contribute slightly more to the the estimated effect of vaccine mandate announcements on vaccine uptake. Finally, column (6) of Table A2 uses the level of weekly first doses per 100,000 people as the outcome. The estimate indicates that the announcement of vaccine mandate increases weekly first dose uptake by roughly 555 per 100,000 people on average (p< 0.005), suggesting that our main results are not sensitive to choosing log of weekly doses as the outcome.

Randomization inference. Inference from the DID model in equation (1) relies on asymptotic approximations associated with the assumption that the number of provinces grows large. While we account for the fact that there are only 10 provinces in our sample in how we report standard errors, in addition, we implement a variant of Fisher's randomization (permutation) test (Fisher, 1935) . Specifically, we estimate equation (1) 5,000 times by randomly assigning the timing of the announcement for each of the 9 provinces that announced vaccine mandate. Fig. B5 provides a histogram of these placebo randomized inference results, along with 95% confidence bands. The figure shows that none of 5,000 placebo estimates is larger than our baseline estimate from column (2) in Table 1 (the solid vertical line on Fig. B5) , implying that the p-value of the randomization inference two-tailed test is zero and reassuring that our main results do not hinge on asymptotic inference approximations.

The provinces had different levels of first-dose vaccination coverage at the time of the respective announcement of vaccine mandate. To explore a "ceiling effect" hypothesis, i.e., that it can be more difficult to induce new vaccinations in places that already have high coverage, we add to equation (1) a variable interacting the announcement dummy P it with the percentage of eligible people that have not yet received a first dose at the time of the announcement. At the respective mandate announcement dates, this percentage ranged from 6.9% in Prince Edward Island to 21.2% in Alberta, with an average of 16.7%.

The estimate of the interaction term, shown in Table A3 , is 0.072 (p< 0.001). This suggests that the effect of a mandate announcement on the pace of weekly vaccinations is about 7.2 log points larger for each additional percentage point in the fraction of eligible people that have not yet been vaccinated. This is consistent with the "ceiling effect" idea, that provinces with more remaining eligible people at the time of announcement have room for a larger boost in vaccine uptake.

In Table A5 we use the same specifications as the baseline Table 1 , but with log of weekly second vaccine doses as the outcome. There are two potential mechanisms for the vaccine mandate announcement to have an impact on second dose uptake: (i) people that already had a first dose before the announcement decide to receive their second dose soon after the announcement (e.g., because most mandates require full vaccination) and (ii) people who took a first dose because of the announcement take their second dose roughly 4-5 weeks after the first.

Column (3) of Table A5 shows the dynamics of second-dose vaccine uptake by week from the announcement. We estimate an increase in uptake of 23.6 log points (27%) (p< 0.05) in the second week from the announcement, which is consistent with mechanism (i) but short-lived as it does not persist in the following weeks. We do not find evidence consistent with mechanism (ii): the estimates for weeks 4 to 7+ are relatively small and either not statistically significantly different from zero or negative. However, the latter finding is only preliminary because we currently have few provinces with more than 4 weeks of post-announcement data. The latter may also explain why the estimated average effect of a mandate announcement on second-dose uptake in columns (1) and (2) of Table A5 is not statistically significantly different from zero. 15 All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. 

We complement the panel data analysis for Canada by estimating a time series variant of our empirical model, for given country or province. While the difference-in-differences approach in Section 3.1 allows us to reliably estimate the average impact of mandate announcements on new first doses in Canada by exploiting the variation in the timing of the mandates across provinces, Fig. 1 and our "ceiling effect" results in Section 3.1.3 suggest that the impact may vary across provinces.

To explore this possible variation, we perform a time series analysis, which allows us to separately estimate the mandate announcement effect for each location, while being aware of the limitation that the required identification assumptions are stronger than those for the DID analysis. In particular, we need to assume that the time-series process of weekly first doses changed after the mandate announcement only due to the policy, i.e., that it would have continued to follow the same pre-trend if no mandate had been announced. Another strong assumption needed is exogeneity of the policy announcement date. We also cannot control for fixed effects in a flexible way beyond including a constant. Nonetheless, this is the only feasible approach for the small number of countries with national mandates. Under these assumptions, the coefficient on the policy variable P t captures the average level shift in first-dose uptake caused by the mandate after the announcement date.

We estimate the following time-series specification for each country or province:

where V t are log weekly first doses for the week ending at date t, P t is an indicator variable for the vaccine mandate being announced at or before t, c is a constant, η t is the error term, and I t are proxies for information (log weekly cases and deaths), analogous to the variables included in equation (1). We include 7-day and 14-day lagged values of V t since the outcome variable V t is a weekly total.

We perform a structural break test for known break point (Chow, 1960) using equation (2) and the first dose vaccination series, V t . We report the results in Table A1 . We use a bandwidth of 50 days before and 35 days after the mandate announcement date. 22 The before-after bandwidth was chosen to reduce the size distortion of the test since the outcome 22 If fewer than 35 days of data are available after the announcement, we use all post-announcement data for that location. The results are also robust to picking June 15, 2021 as the initial date for the test.

All rights reserved. No reuse allowed without permission.

(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 preprint this version posted October 25, 2021. ;

variable is a weekly sum and the error terms are serially correlated. 23 Columns (1) and (2) in Table A1 use the log of weekly first doses series, V t . For eight of the nine Canadian provinces and the three countries that have announced a proof of vaccination mandate at the time of the analysis, we strongly reject the null hypothesis that the mandate announcement date is not a structural break. There was insufficient post-announcement data to perform the test for Prince Edward Island.

We check the robustness of these results using the first difference of weekly first doses. The first-differenced series is stationary and the error terms are not serially correlated. This further alleviates possible concerns about size distortion in the Chow test. The results are displayed in column (3) of Table A1 . The null hypothesis of no structural break at the announcement date is rejected for all locations except Manitoba, Ontario and Nova Scotia. 24 Overall, the structural break test results suggest that a mandate announcement is associated with a trend break in first-dose vaccine uptake in most or all locations, depending on the specification used. Table 2 reports estimation results for the time series model in equation (2) . The p-values are calculated using the Newey-West HAC (heteroskedasticity and autocorrelation) robust estimator (Newey and West, 1987 ) with 3 lags. 25 In columns (1)-(3), we estimate a specification including the auto-regressive first-dose terms, the policy term P t , and log weekly deaths as information I t . We include only deaths since only log weekly deaths and not cases were statistically significant in the DID analysis (see Tables 1 and A2 ). In addition, in France and Manitoba, there is spurious positive correlation between the weekly cases and first doses time series over the sample period, which can bias our estimates.

Since we use level variables in equation (2) that are non-stationary, we first use the augmented Dickey-Fuller (ADF) unit root test to test whether the residuals η t are stationary to avoid a spurious regression problem. For all locations and all specifications in columns (1), (2) and (3) of Table 2 , we reject the null hypothesis that the residual contains a unit root at the 95% confidence level. The ADF test results for specification (1) with a single 23 As shown in Giles and Scott (1992) , the size of the Chow test is upward distorted (over-rejection of the null) in the presence of positive serial correlation in the error terms. The size distortion is mitigated if the sample split is imbalanced before vs. after the break point. 24 The power of the Chow test is necessarily weaker in this specification since the first differenced V t series is a growth rate being used to test for a level shift. The differenced series is also noisier, as it captures daily fluctuations. 25 The lag length for the HAC estimator is chosen following Greene (2012) as the closest integer to T 1/4 , where T = 108 is the sample size.

All rights reserved. No reuse allowed without permission.

(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 preprint this version posted October 25, 2021. ; policy dummy are shown in column (4) of Table 2 . 26 In column (1) of Table 2 , the coefficient estimates on the mandate announcement variable for the European countries are all positive and economically large but only statistically significant for France (0.186), which registers the largest increase in first doses in the raw data among the three countries. Among the Canadian provinces, we obtain statistically significantly positive estimates of the mandate announcement effect for most (the exceptions are Manitoba, Ontario and Nova Scotia). We also note that the estimates tend to be larger if the respective vaccine mandate was announced more recently (e.g., in Alberta, Saskatchewan and New Brunswick; see Table C1 ), that is, when the post-announcement data is from the early weeks after the announcement. This is consistent with our results in Section 3.1.

Since we expect from the raw data ( Fig. 1 and Fig. B3 ) and the DID analysis that the increase in first doses after a mandate announcement tends to dissipate over time, in columns (2) and (3) of Table 2 , we split the announcement policy variable P t into two binary variables corresponding respectively to the 'early' post-announcement period, from the announcement date to 2 weeks thereafter (column 2), and the 'later' period, from 2 weeks after the announcement to the end of the sample, Sep. 30, 2021 (column 3). The 'early response' results in column (2) show a statistically significantly positive and economically large effect for all countries and provinces (p< 0.06), except for Manitoba and Ontario, for which the estimates have a positive sign but are not statistically significantly different from zero. 27 For the sub-period of more than 2 weeks after the announcement (column 3), the estimated coefficients on the mandate announcement variable are smaller in magnitude and negative for some locations, consistent with our previous results that the boost in first doses observed after a vaccine mandate announcement dissipates over time. 26 These ADF test results were produced with 6 lags. We also checked using from 3 to 10 lags and we were still able to reject the null at the 95% confidence level everywhere except in Ontario and Manitoba, where we reject the null at least at the 90% level. 27 The estimates from the '≤ 2 weeks' specification without including log weekly deaths are statistically significantly positive for all provinces and countries (p< 0.07) (not displayed in the table).

All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (2) on "mandate announced", a binary variable that takes value 1 for all dates on or after the vaccine mandate announcement date and 0 otherwise. In columns (2) and (3), we instead use two binary variables, "≤ 2 weeks," which takes value 1 for all dates within 2 weeks of the announcement date (estimates reported in column 2) and "> 2 weeks," which takes value 1 for all dates 2 weeks or more after the announcement (column 3). P-values computed using Newey-West heteroskedasticity and autocorrelation robust standard errors with 3 lags are reported in the square brackets. *, ** and *** denote 10%, 5% and 1% significance respectively. †For AB, SK, NB and PE, we have insufficient data in our sample to estimate specification (3) . Column (4) reports the augmented Dickey-Fuller (ADF) test statistics and p-values for specification (1).

All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. We can use our difference-in-differences estimate of the average effect of mandate announcements on weekly first doses from Section 3.1 to evaluate the counterfactual of no vaccine mandate in each Canadian province. Using the definition of the mandate announcement variable P it , define the counterfactual weekly first dosesṼ did it in province i for the week ending at date t asṼ

whereV did it is the fitted value of log weekly first doses for the week ending at t in regression equation (1), andα is the coefficient estimate (0.593) on the vaccine mandate announcement P it in the baseline specification, column (2) of Table 1 .

Alternatively, we can use the time-series location-specific estimates of the mandate effect from Section 3.2 and Table 2 to construct counterfactual no-mandate simulations for each country or province separately. Accounting for the lagged terms in equation (2), define the counterfactual weekly doses,Ṽ ts t , as V ts t =V ts t , for all t <t and

wheret is the mandate announcement date,V ts t is the fitted value of log weekly first doses for the week ending at t in regression equation (2),π is the coefficient estimate on the vaccine mandate announcement dummy P t displayed in column (2) of Table 2 , andλ 1 andλ 2 are the coefficient estimates on the lagged terms V t−7 and V t−14 in equation (2) (not displayed in the table). The estimatesπ,λ 1 andλ 2 are specific for each respective country or province.

Note that using equation (3), the counterfactual for each province is computed using the same 'average' estimate of the mandate announcement effect. Nonetheless, the advantage of using the DID estimate to evaluate counterfactuals is that, given the event-study results in Section 3.1, this estimate is less likely to be subject to endogeneity concerns than the time-series estimates. Conversely, computing counterfactuals using the time-series estimates, equation (4), allows us to use location-specific mandate effects and potentially gain precision. However, the time-series estimation approach admits a much more limited set of control variables and requires stronger identification assumptions, as discussed earlier.

20 All rights reserved. No reuse allowed without permission.

(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 preprint this version posted October 25, 2021. ;

In Fig. 4 , we display results obtained using our DID estimates and the counterfactual series for log weekly doses,Ṽ did it , defined in equation (3). We compute the total first dose vaccinations for province i and date t under the counterfactual of no proof of vaccination mandate, starting from the actual total doses as of the day before the announcement and adding up forward in time. We then compare these counterfactual total doses (the red dotted line) with the actually observed total doses (the black diamonds). 28 The number next to the province name in each panel of Fig. 4 reports the estimated difference between the observed (with mandate) total first doses and the mean counterfactual value (no mandate), as of September 30, 2021. Adding up these numbers, we estimate that the provincial proof of vaccination mandates have led to 289,000 additional first doses in Canada (about 0.9 percentage points of the eligible 12+ population). The largest absolute gains in new vaccinations are in Alberta (91,000 additional doses, or 2.4 p.p. of all eligible people), British Columbia (69,000, 1.5 p.p. of all eligible) and Ontario (64,000, 0.5 p.p. of all eligible). The estimated gains in vaccine uptake control for all factors in equation (1), that is, log weekly cases and deaths and the various time and province fixed effects (see Table 1 ). Fig. 4 shows a relatively quick divergence between the observed and counterfactual cumulative first-dose numbers in all provinces. However, several weeks past the announcement where such data exist, the actual-counterfactual gap is closed in some provinces, e.g., Quebec and Manitoba, consistent with intertemporal substitution. For other provinces, e.g. Alberta or Saskatchewan, the estimated total gains in first doses are still growing as of Sep. 30, 2021.

We can estimate a lower bound for the amount of intertemporal substitution by comparing the largest difference between actual and counterfactual cumulative doses over all dates after the announcement with the difference at the sample end date reported next to each province name on Fig. 4. 29 If, over time, the gap between actual and counterfactual total doses first grows but then starts to decline, this means that vaccinations have been pulled forward relative to the no-mandate counterfactual (intertemporal substitution). Applying this procedure, we estimate, as of Sep. 30, 2021, lower bounds of about 30,600 first doses in Quebec, 2,500 in Ontario, 1,000 in Manitoba and 340 in British Columbia administered earlier than in the no-mandate counterfactual. These numbers can be considered part of the 

All rights reserved. No reuse allowed without permission.

(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 figure plots the observed (black diamonds) and the expected "no mandate" counterfactual (dotted red line) cumulative first doses (in millions) for each province by date. The counterfactual is computed using the estimate 0.593 from column (2) of Table 1 . The numbers next to the province names are the estimated total additional first doses relative to the counterfactual rounded to the nearest thousand, as of September 30, 2021.

All rights reserved. No reuse allowed without permission.

(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 preprint this version posted October 25, 2021. ; total effect of vaccine mandate announcements on vaccine uptake, together with the respective estimated net gains of 7,000, 64,000, 9,000 and 69,000 first doses for these four provinces as of Sep. 30. In contrast, for the remaining five provinces, we do not find intertemporal substitution as of Sep. 30, possibly because of their later mandate announcement dates.

We also use the province-specific time-series estimates as an alternative way to compute counterfactual log weekly doses,Ṽ ts t , defined in equation (4) and counterfactual total first doses up to any post-announcement date t for each province. The results, displayed in Fig. B7 , are very similar to the DID-based results on Fig. 4 for the Atlantic provinces, Alberta and British Columbia. The main differences relative to Fig. 4 are for Quebec, Ontario and Manitoba, which reflect the differences between the province-specific estimates in Table 2 and the average mandate effect estimate in Table 1 . Nevertheless, for Canada as a whole, Fig. B7 implies 288,000 additional first dose vaccinations, which is very close to the 289,000 additional vaccinations estimated using the DID approach (Fig. 4 ).

We use our country-specific time-series estimates to compute counterfactual log weekly doses, V ts t , and the counterfactual total first doses for any post-announcement date t for each country. The results are displayed in Fig. 5 . The estimated number of additional first doses after the mandate announcement is economically meaningful; it is especially large in France (7.79 mln) and Italy (4.53 mln), and smaller in Germany (1.31 mln). The differences could be due to the timing of mandate announcements: a few weeks earlier in France and Italy than in Germany (see Table C1 ), or the baseline vaccination rate at the time of announcement (lower in France).

All counterfactual simulations discussed above assume that all explanatory variables except the vaccine mandate announcement (e.g., deaths or time fixed effects) would remain fixed at their observed values, and that the estimated model coefficients remain stable. This is a strong assumption that is more plausible over relatively short periods. These results should therefore be interpreted with caution and only as an illustration of the estimated impact of the proof of vaccination mandates on vaccine uptake as of September 30, 2021. 23 All rights reserved. No reuse allowed without permission.

(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 numbers next to the country names are the estimated total additional first doses relative to the no-mandate counterfactual, as of September 30, 2021.

We estimate the impact of government-mandated proof of vaccination requirements on COVID-19 vaccine uptake in nine Canadian provinces, as well as in France, Italy and Germany. We document a statistically significant large increase of first-dose vaccine uptake in the first two weeks after the announcement of a proof of vaccination mandate, relative to in absence of mandate (more than 100% increase for the second post-announcement week). The rate of weekly first dose vaccinations then gradually decreases and dissipates after 7-8 weeks. We obtain similar results from time-series analysis for each province and for France, Italy and Germany -a significant initial rise in first-dose uptake within two weeks after a mandate announcement (the estimated effect varies, from as low as 6% increase in Ontario to as high as 125% in Alberta over the post-announcement period), followed by a gradual waning. Using simulations based on our econometric results about the mandates' effect on vaccine uptake, we estimate that, as of September 30, 2021, all provinces and countries achieved net gains in first-dose vaccinations relative to the hypothetical scenario of absence of vaccine mandate (289,000 estimated additional first doses for Canada, 7.8 mln for France, 4.5 mln for Italy and 1.3 mln for Germany).

To quantify the costs of overcoming vaccine hesitancy, it may be useful to convert the increase in vaccinations induced by the mandates into a monetary equivalent. For example, 24 All rights reserved. No reuse allowed without permission.

(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 preprint this version posted October 25, 2021. ;

Campos-Mercade et al. (2021) -a randomized controlled trial (RCT) study in Swedenreport a 4.2 p.p. higher vaccination rate among participants who were offered a monetary payment of SEK 200 (USD 24) if they are vaccinated within 30 days of vaccines becoming available to them, relative to a control group that was not offered the monetary incentive. It is, however, very hard to map these or other similar results into our setting. The vaccination rate at the time of the provincial mandate announcements in Canada was 83.3% on average, which is 11.7 percentage points higher than the observed baseline vaccination rate of 71.6% in Campos-Mercade et al.'s control group in May-July. This suggests that the 'marginal' person deciding whether to be vaccinated in Canada in August or September, after 83% of eligible people had already done so, is likely very different in terms of attitude toward vaccination (and requires a much larger monetary incentive) from the average member of the Swedish RCT control group. Nonetheless, if we take the Campos-Mercade et al. estimate at face value and assume the effect is linear in the payment, the estimated cost of using monetary incentives to induce as many additional first doses as the provincial vaccine mandates, 0.9 p.p. of the eligible population, would be around USD 150 mln (= $24 × (0.9/4.2) × 29.2 mln people), assuming that, as in the Campos-Mercade et al. experiment protocol, all people with a first dose (29.2 million in Canada as of September 30, 2021) would be paid. However, because of the selection problem discussed above, the actual monetary equivalent of the mandates is likely significantly larger.

In terms of external validity, in June 2020, Lazarus et al. (2021) conducted a survey on vaccine hesitancy across 19 countries comprising around 55% of the world population. 30 Canada, Italy and Germany placed around the middle in self-reported vaccine hesitancy rate (29% to 31%), while France had a higher hesitancy rate (41%). In this regard, our results may be useful to public health authorities in other countries looking for the most effective strategy to increase vaccine uptake.

Importantly, our objective is not to engage in normative or ethical arguments related to vaccination mandates, but to provide a benchmark number for their effectiveness in terms of new first dose uptake, which can be compared to other approaches such as monetary incentives or behavioural nudges. Monetary incentives for vaccination have been criticized by many for the 'optics' of putting a (low) dollar value on being vaccinated 31 or, if payments are unavailable to those vaccinated early, because of potential moral hazard problems (e.g., 30 The participants were asked: "If a COVID-19 vaccine is proven safe and effective and is available, will you take it?" 31 For example, U.S. President Joseph Biden advocated in July 2021 that state and local governments pay $100 to anyone willing to receive a first dose, however, the public health value of vaccination is likely much higher (see Castillo et al., 2021) .

All rights reserved. No reuse allowed without permission.

(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 preprint this version posted October 25, 2021. ; expecting future payments) or the perceived unfairness of rewarding those that delayed their first dose. Others have argued that, given the already high vaccination rates in the developed countries, behavioural nudges may not be very effective (Thaler, 2021), consistent with the findings of Campos-Mercade et al. (2021) .

Naturally, a complete cost-benefit analysis of vaccine mandates is beyond the scope of the current paper. In particular, the costs of imposing and enforcing the mandates -economic, political, or personal -are very hard to estimate, as is the social value of vaccinating an additional person. 32 Our results are one step toward quantifying the benefits of proof of vaccination requirements. (2) but without controlling for cases and deaths; column (2) uses log weekly first doses as the outcome variable and includes cases and deaths; column (3) uses the first difference of log weekly first doses and includes cases and deaths. We use a bandwidth of 50 days before and 35 days after the announcement date. If fewer than 5 weeks of data exist after the announcement, we use all post-announcement data. † denotes that H 0 cannot be rejected at the 90% confidence level.

All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Notes: "mandate announced" is a binary variable which takes value 1 on or after the proof of vaccination mandate announcement date and 0 otherwise. The cases and deaths variables are log weekly totals for dates t − 6 to t. P-values from wild bootstrap (cgmwildboot) standard errors clustered by province with 5,000 repetitions are reported in the square brackets. Column (1) repeats the baseline estimates from column (2) of Table 1 . Column (2) uses only deaths as information. Column (3) replaces d.o.w by province fixed effect by d.o.w by province-year-month fixed effects to allow for differential day-of-the-week effects by province-year-month. Column (4) replaces date fixed effects by year-month fixed effect interacted by each province dummy. Column (5) reports the estimates from equation (1) where the population is used as weight. Column (6) uses the level of weekly first doses per 10,000 people as the outcome variable. ***, ** and * denote 10%, 5% and 1% significance respectively. FE = fixed effects, d.o.w = day-of-week.

All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Notes: "mandate announced" is a binary variable which takes value 1 on or after the proof of vaccination mandate announcement date and 0 otherwise; "P it × % not yet vaccinated" is the interaction term between the vaccine mandate announcement variable and the % of eligible persons that are not yet vaccinated on the mandate announcement date. The cases and deaths variables are log weekly totals for dates t − 6 to t. P-values from wild bootstrap (cgmwildboot) standard errors clustered by province with 5,000 repetitions are reported in the square brackets. *, ** and *** denote 10%, 5% and 1% significance respectively. FE = fixed effects.

All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Notes: P-values from standard clustering by province in the ( ) parentheses, wild bootstrap with one-way clustering by province and 5000 repetitions in the [ ] square brackets, and wild bootstrap with two-way clustering by province and day with 5000 repetitions in the { } curly braces. "week n" where n = 0, 1, 2, ... is a binary variable which takes value 1 for the days in the n-th week immediately after the announcement date (week 0 is the week of the announcement) and value 0 otherwise. The cases and deaths variables are log weekly totals for dates t − 6 to t. P-values from wild bootstrap (cgmwildboot) standard errors clustered by province with 5,000 repetitions are reported in the square brackets. *, ** and *** denote 10%, 5% and 1% significance respectively. FE = fixed effects. d.o.w. = day-of-week.

All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Table C1 for the exact dates of mandate announcement in each province.

All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (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 estimate equation (1) 5000 times by randomly assigning the timing of vaccine announcement for each province with vaccine mandate. The figure provides graphical illustration (histogram) of the placebo inference results, along with values at 5th and 95th percentiles. The solid vertical line corresponds to our baseline estimate from column (2) in Table 1 .

All rights reserved. No reuse allowed without permission.

(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 figure plots the observed and counterfactual (assuming that vaccine mandate was not announced) weekly first doses per 100,000 people. We use the estimate 0.539 from column (2) of Table 1 to compute the counterfactual, as in equation (3). The diamonds plot observed weekly first doses per 100,000, the dotted lines plot the counterfactual mean value on and after the announcement date, and the shaded areas are 5-95 percentile confidence bands.

All rights reserved. No reuse allowed without permission.

(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 figure plots the observed (black diamonds) and the expected "no mandate" counterfactual (dotted red line) cumulative first doses (in millions) for each province by date. The counterfactual is computed using the estimates from column (1) of Table 2 for each respective province. The numbers next to the province names are the estimated total additional first doses relative to the counterfactual rounded to the nearest thousand, as of September 30, 2021.

All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Notes: 1. The data source for BC and NS vaccinations data is covid19tracker.ca, an aggregator website that gathers daily data from official provincial dashboards that do not provide historical data.

All rights reserved. No reuse allowed without permission.

(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 preprint this version posted October 25, 2021. ; https://doi.org/10.1101/2021.10.21.21265355 doi: medRxiv preprint

Experimental Evidence on the Effectiveness of Non-Experts for Improving Vaccine Demand

Vaccine effectiveness and duration of protection of Comirnaty, Vaxzevria and Spikevax against mild and severe COVID-19 in the UK

Design-based Analysis in Difference-In-Differences Settings with Staggered Adoption

Monetary incentives increase COVID-19 vaccinations

Difference-in-Differences with Multiple Time Periods

Bootstrap-based improvements for inference with clustered errors

Market design to accelerate COVID-19 vaccine supply

Causal Impact of Masks, Policies, Behavior on Early COVID-19 Pandemic in the U.S

Tests of equality between sets of coefficients in two linear regressions

Two-Way Fixed Effects Estimators with Heterogeneous Treatment Effects

The Design of Experiments

Some consequences of using the Chow test in the context of autocorrelated disturbances

Econometric Analysis

The effect of large-scale anti-contagion policies on the COVID-19 pandemic

Face Masks, Public Policies and Slowing the Spread of COVID-19: Evidence from Canada

Incentives can spur COVID-19 vaccination uptake

A global survey of potential acceptance of a COVID-19 vaccine

A Simple, Positive Semi-definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix

Monitoring Incidence of COVID-19 Cases, Hospitalizations, and Deaths, by Vaccination Status -13

Estimating dynamic treatment effects in event studies with heterogeneous treatment effects

Notes: At the time of writing, the mandates of Alberta, Saskatchewan, France, Italy and Germany also allow for a recent negative COVID-19 test in lieu of vaccination. Some jurisdictions implemented their mandates in stages, both in terms of the extent of vaccination required (at least one dose or full vaccination) and in terms of the scope of the mandate (e.g., only for very large events or across a wide array of settings); in these cases, we report the date on which the first stage with wide application started to be enforced.