Mask mandates save lives

Abstract

We quantify the effect of statewide mask mandates in the United States in 2020. Our regression discontinuity design exploits county-level variation in COVID-19 outcomes across the border between states with and without mandates. State mask mandates reduced new weekly COVID-19 cases, hospital admissions, and deaths by 55, 11, and 0.7 per 100,000 inhabitants on average. The effect depends on political leaning with larger effects in Democratic-leaning counties. Our results imply that statewide mandates saved 87,000 lives through December 19, 2020, while a nationwide mandate could have saved 57,000 additional lives. This suggests that mask mandates can help counter pandemics, particularly if widely accepted.

1. Introduction

Governments around the world implemented a range of policy measures to counter the rapid spread of COVID-19. Mandating the use of face masks was one such measure. Widespread mask wearing allows for continued human interaction and economic activity, unlike strict lockdown measures such as stay-at-home orders. However, mask mandates remain controversial in many countries. Evidence on the effect of mask mandates is important to inform decisions on their potential future deployment — both in the United States and abroad.

This paper asks the question: “Did statewide mask mandates save lives?” Mask mandates can reduce the transmission of COVID-19 both directly, by inducing greater mask usage, which is proven to lower the transmission of airborne diseases such as COVID-19,2 and potentially also indirectly, by inducing behavioral changes.

Importantly, this paper does not estimate the effect of wearing masks, but rather the effect of mandating mask wearing. For mask mandates to succeed in saving lives, they would need to reduce the transmission of COVID-19. This would put a dent on the number of COVID-19 cases, hospital admissions, and ultimately deaths. We estimate the effect of statewide mask mandates on these variables across counties in the United States in 2020.

To answer our research question, we use a regression discontinuity design. In the United States, mask mandates were implemented at the state or sub-state level. We rely on the variation between counties across state mask borders, that is a state border that separates two counties, in which one county is in a state with a mask mandate at a given time and the other county is in a state without a mask mandate at the same time. Variation in COVID-19 outcomes across these mask borders is much more likely to be driven by the existence of mask mandates and less by different local stages of the pandemic. This identification strategy allows estimating the effect of statewide mask mandates.

In this paper, we study the effect of mask mandates at the state- rather than county-level. This choice is motivated on both conceptual and methodological grounds. First, statewide mandates were a far more common policy tool during the pandemic – in 2020, close to 94 percent of county-week pairs that were subject to mask mandates had a statewide order.3 Second, mandates at the state level were more visible to the population and subject to less ambiguity. In particular, the enforceability of a mandate at the county level may be questioned if such a mandate does not exist at the state level. Relatedly, as state laws supersede county regulation it will only be possible to identify the effect of county-level mandates where no state-level mandates are in place. Finally, the issue of spillovers – where individuals travel to neighboring counties to avoid a mask mandate – is likely more pertinent for county-level mandates since the likelihood of neighboring a county with a different mandate is higher when the mandate is only countywide. However, we do show that controlling for the existence of county level mandates is not material for our results.

We find a significant and substantial effect of statewide mask mandates. Specifically, mask mandates on average reduced new weekly COVID-19 cases, hospital admissions, and deaths by 55, 11, and 0.7 per 100,000 inhabitants. Given existing evidence that attitudes towards mask wearing vary with political leaning (Milosh et al., 2020), we also estimate the effects of mask mandates across political leaning as measured by the vote shares in the 2020 Presidential Election—a variable that is readily available at the county level. Political leaning should be seen as a proxy for deeper-held attitudes. We find that estimated effects vary strongly with political leaning: the effect on cases and deaths is larger in Democratic-leaning counties and smaller in Republican-leaning counties. This is likely because mandates in more Democratic-leaning counties are more effective in promoting mask wearing. These conditional effects are important enough that they must be taken into account. Indeed, the estimated effect of mask mandates on weekly deaths per 100,000 inhabitants varies from around −2.5 to 0 depending on the political leaning of counties, compared to the average effect of −0.7 mentioned above.

Our results imply that statewide mask mandates saved 87,000 lives through December 19, 2020, while an additional 57,000 lives could have been saved in the same period if a nationwide mandate had been enacted starting in April 2020. Both numbers are chiefly driven by the estimated effects of mask mandates in urban Democratic-leaning counties. Lives saved by mask mandates are concentrated in urban Democratic-leaning counties in states that imposed mandates, while lives that could have been saved are concentrated in urban Democratic-leaning counties in states without mask mandates. The magnitude of these effects is large. For comparison, COVID-19 deaths in the same period amounted to around 309,000 in the United States.

And yet, these are likely to be lower bound estimates, particularly after controlling for other policies enacted contemporaneously. Indeed, states imposing mask mandates are in principle reacting to worse outbursts of the pandemic. This means our estimates could be subject to simultaneity bias, leading us to find smaller differences between counties with and without mandates. Another potential source of bias towards zero is the presence of county-level mandates in states without a statewide mandate. However, we show that our findings are robust to controlling for county-level mandates. On the other hand, there are two additional potential sources of bias that go towards us finding effects. First, states enacted other policies at around the same time as statewide mask mandates, which could confound the estimated effects of mask mandates. We show that our findings are robust to controlling for other contemporaneous statewide public health measures. Second, spillovers could be another concern. The imposition of mask mandates could prompt individuals to travel more into neighboring counties without mask mandates, and thus increase the spread of COVID-19 in those counties—a negative spillover of mask mandates. But conceptually the opposite could also be true—individuals that prefer a safer environment may travel from counties without mandates to those with mandates, which would be a positive spillover of mask mandates. In the data, we find no significant relationship between mask mandates and mobility, consistent with Chernozhukov et al. (2021).

Our results hold important lessons for all countries aiming to counter the spread of COVID-19 and other pandemics in the years to come. First, not all countries are rolling out vaccines at the same pace, and many will struggle to reach large swaths of their population in the medium-term. Second, existing vaccines may not be as effective against either known or future variants, and there may well be future outbreaks of other airborne pandemics. At the same time, our results highlight the limitations of mask mandates and their dependence on individual’s attitudes. Mask mandates do not reduce the spread of a pandemic if they fail to engender greater mask usage and/or behavioral changes, such as incentivizing social distancing. Thus, mask mandates are likely to remain a highly relevant policy tool in the years to come, particularly if coupled with other complementary policies that educate, persuade, and incentivize the population to comply.

Our contribution to the existing literature is both conceptual and methodological. Some studies found an effect of statewide mask mandates in the United States, including Chernozhukov et al. (2021), Lyu and Wehby (2020), and Renne et al. (2020). Others analyzed the same question for Germany (Mitze et al., 2020) and Canada (Karaivanov et al., 2020). In contrast, Leech et al. (2021) do not find evidence that mask mandates reduce transmission within 92 regions of 56 countries. We add to these existing studies conceptually by showing that the effect of mask mandates depends crucially on the political leaning of voters. This complements existing evidence that political leaning shapes the use of masks, even when mask mandates are in place (Milosh et al., 2020). These conditional effects are critical to obtain unbiased estimates at the county-level. Our study is also related to Welsch (2020), which investigates the effect of mask usage on deaths at the county level using as an instrument the vote share in the 2016 election. Conceptually, our paper is also related to the strand of literature showing how deeper views shaped the responses to COVID-19. Simonov et al. (2020) shows how media consumption shapes compliance with social distancing rules. Gollwitzer et al. (2020) finds that social distancing in U.S. counties relates negatively to the Republican vote share in the 2020 election. Methodologically, we depart from previous studies by using a different empirical design, a new coding of state-level mask mandates, and by exploiting more data. First, our empirical strategy uses a regression discontinuity design based on granular county-level variation, while the aforementioned papers for the United States estimated state-level effects using an event study methodology (Lyu and Wehby, 2020) or a structural epidemiological model (Chernozhukov et al., 2021, Renne et al., 2020). Leech et al. (2021) estimate a Bayesian hierarchical model. We collected state mask orders to define statewide mask mandates consistently and more broadly than Chernozhukov et al. (2021) and Lyu and Wehby (2020), which analyze mask mandates for employees only. Finally, we make a data contribution by exploiting an additional COVID-19 outcome variable – COVID-19 hospital admissions – and by analyzing a longer period.

This paper is organized as follows: Section 2 details our approach, Section 3 shows key results and Section 4 concludes.

2. Methodology

To identify the effect of state-level mask mandates we rely on variation in the adoption of these mandates across time and on how COVID-19 outcomes vary across counties near mask borders. A mask border is defined as a state border dividing two counties, in which one of the counties is in a state with a statewide mask mandate and the other county is in another state without such a mandate.

States implemented mask mandates at different points in time (Fig. 1). In fact, face covering requirements loosely defined varied even more significantly across time and states. Some states merely recommended wearing masks, others mandated face coverings in select indoor spaces (e.g. state government buildings), yet others mandated much stricter face coverings. We code statewide mask mandates as a binary variable. In particular, we code a state to have a mask mandate in place if face coverings are at least required in most public indoor spaces if social distance cannot be maintained. We collected individual state orders and documented whether mandates adhere to this definition as well as the day in which they become effective.4 For the detailed coding please see the Online Appendix E.

Fig. 1.

Statewide Mask Mandates in the 48 Contiguous States and the District of Colombia . Note: State-level mask mandates are defined as the requirement that face coverings are worn at least in most public indoor spaces if social distance cannot be maintained, see Online Appendix for details.

Crucial to our research design is the variation at the county level across state mask borders which we will exploit using a regression discontinuity design. Counties close to each other are more likely to be in similar stages of the pandemic. Thus, mask mandates are likely less endogenous to the local state of the pandemic for nearby counties across state borders. To focus squarely on the distance to a mask border, we compute the minimum distance between all pairs of counties at every point in time, in which one county has a mask mandate and the other county does not. 5Online Appendix A.1. Thus, each county is assigned a minimum distance to a mask border. Counties in states with mask mandates are assigned a positive distance while those without a mask mandate are assigned a negative. Finally, we exclude any absolute minimum distances greater than 150 miles from our regressions. This choice is based on the actual distribution of absolute minimum distances, and ensures that all counties by state borders are included while avoiding comparing counties too far from a mask border. See Online Appendix A.1 for more details.

Fig. 2 illustrates our approach. New Mexico was the first state west of the Mississippi to adopt a mask mandate. The figure takes the situation in June 1st, 2020 as an example. Counties with a statewide mask mandate that are within 150 miles of a county without a mask mandate are shown in green (dark gray if viewing in gray-scale). As mentioned, on that day these are all in New Mexico for this region. Two counties in New Mexico’s interior are excluded because they are too far from the mask border. Counties without a mask mandate within 150 miles of a county in New Mexico are shown in red (dark gray if viewing in gray-scale). These are in the neighboring states of Arizona, Utah, Colorado, Kansas, Oklahoma and Texas. Most of the counties in those neighboring states, though, are colored white because they are too far from counties in New Mexico. Our regression discontinuity design exploits variation in COVID-19 outcomes across green and red (dark and light gray) counties.

Fig. 2.

Mask Mandates in Counties in and around New Mexico on June 1, 2020. Note: The map shows counties in New Mexico and the neighboring states of Arizona, Utah, Colorado, Kansas, Oklahoma and Texas. Counties are colored: (i) white if they are beyond 150 miles from the mask border; (ii) red (light gray) if they are within 150 miles from the mask border and do not have a mask mandate; and (iii) green (dark gray) if they are within 150 miles from the mask border and have a mask mandate.(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

This approach is, if anything, biased towards not finding an effect of statewide mask mandates, particularly after controlling for other policies enacted contemporaneously. State mandates are very likely to respond to recent changes in COVID-19 outcomes. Since outcomes are auto-correlated, our estimates could be subject to simultaneity bias against finding an effect of mask mandates. While mitigated by the use of county-level variation around the border, this bias could still be present in our design. The fact that Republican states are less likely to impose mandates all else equal only diminishes the strength of this bias, but do not change its sign. A second potential bias arises from omitting controls for county-level mask mandates. Such mandates are also likely to bias our results away from finding an effect, as county mandates are prone to be most relevant in counties without statewide mandates.

An argument that runs counter to our estimates being a lower bound for the effect of mandates is that the imposition of mask mandates could prompt individuals to travel more into neighboring counties without mask mandates, and thus increase the spread of COVID-19 in those counties—a negative spillover of mask mandates. But conceptually the opposite could also be true—individuals that prefer a safer environment may travel from counties without mandates to those with mandates, which would be a positive spillover of mask mandates. In reality, none of these possibilities seems to be borne by the data, as we find no significant relationship between mask mandates and mobility. This is consistent with the evidence from Chernozhukov et al. (2021). Finally, estimates could be biased by the simultaneous imposition of other state-level public health measures. To ensure that this is not driving our results we include controls for a large set of other such measures, like school closures or stay-at-home orders among many others. Our results are robust to including these controls. We also run placebo regressions to show that discontinuities in these other measures, mobility, and political leaning do not drive our results.

Let us describe our dataset in more detail. We focus on the period between January 20, 2020 and December 20, 2020, when the COVID-19 pandemic hit the U.S. and mask mandates were being implemented and vaccines were not yet widely available. Daily data is aggregated to weekly frequency. Our dataset covers all counties in the 48 contiguous U.S. and the District of Columbia (more than 3000 matched units). As outcome variables, we use new COVID-19 cases and deaths from Centers for Disease Control and Prevention (CDC) compiled by https://usafacts.org/. Hospital admissions related to COVID-19 are from US Department of Health and Human Services (HHS). This data is measured at the facility level, and we aggregate to the county level. We express all outcome variables in percent of the county population. As controls, we use other containment and health policy measures from Hallas et al. (2020). We also control for mobility as measured by the simple average of individual indices for retail and recreation, grocery and pharmacy, transit stations and workplaces, provided by Google. For data on political leaning at the county level we use data on the 2020 presidential election collected by https://github.com/tonmcg. All variables and their sources are detailed in Online Appendix A.

Note that COVID-19 hospital admissions have conceptual drawbacks relative to new COVID-19 cases and deaths on how they are assigned to counties. First, some (smaller) counties do not report data, which could reflect that there are no hospitals located within their boundaries, and thus residents with COVID-19 need to be admitted in nearby counties. Second, hospitals admit patients from other counties more generally, even if their home county has hospitals. Third, hospitals without spare capacity may send patients to be admitted in other hospitals across county-lines. These factors make it harder to interpret admissions as a purely county-specific outcome. Beyond these conceptual drawbacks, the time period covered by the data on admissions is also more limited, starting in July 31, 2020.

A first look at the raw data on COVID-19 cases and deaths around mask borders suggests that cases and deaths are much lower in counties with mask mandates. Fig. 3 plots the distance of each county to a mask border (x-axis) against new weekly COVID-19 cases (panel a) and deaths (panel b) per 100,000 inhabitants on the $y$-axis. The red diamonds to the left of the mask border are the counties without a mask mandate. The green circles to the right of the mask border are the counties with a mask mandate. The figures show a discontinuity at the mask border. When moving from a county without a mask mandate (but close to a county with a mandate) the number of cases and deaths drops markedly. Our identification below exploits and formally estimates the size of this discontinuity.6

Fig. 3.

New weekly COVID-19 cases and deaths per 100,000 inhabitants . Note: These charts show raw data of new weekly COVID-19 cases and deaths and thus do not account for county or time fixed effects as done in our econometric analysis. Data is binned in intervals of three miles. The distance to a county with a different mask policy is negative(positive) for counties without(with) statewide mask mandates.

To explore the effects of mask mandates formally, we use the following regression discontinuity (RD) design:

$$y_{c,s,t+k}=\delta mas{k}_{s,t}+\beta dis{t}_{c,t}+{\beta }^{+}\mathbb{1}\{mas{k}_{s,t}>0\}\times dis{t}_{c,t}+\theta X_{c,s,t}+{\tau }_t+{\tau }_c+{ϵ}_{c,s,t}$$ (1)

where $y_{c,s,t+k}$ stands for new COVID-19 cases, new COVID-19 hospital admissions, and new COVID-19 deaths per 100,000 inhabitants in county $c$, state $s$, week $t+k$. We use $k=1$ for cases, $k=2$ for hospital admissions and $k=4$ for deaths. Our “running variable” is the minimum distance to a county with a different statewide mask policy ($dis{t}_{c,t}$). This variable is negative (positive) for counties without (with) a statewide mask mandate. This way, there is a discontinuity in treatment at $dis{t}_{c,t}=0$. This discontinuity in treatment is captured by $mas{k}_{s,t}$ which takes the value of 1 if a statewide mask mandate existed in state $s$ at week $t$.7 $\mathbb{1}\{mas{k}_{s,t}>0\}\times dis{t}_{c,t}$ is the interaction of the running variable and the dummy variable indicating whether or not the county is subject to a statewide mask mandate. This interaction term allows for different slopes on the running variable across the mask border – a standard assumption in RD designs. ${\tau }_c$ is a county fixed effect absorbing time-invariant factors in a given county, such as geography, population density, or industrial structure, that could affect COVID-19 outcomes. ${\tau }_t$ is a time fixed effect absorbing nation-wide factors affecting the pandemic. $X_{c,s,t}$ includes other controls, including lagged mobility in the baseline,8 but whose inclusion is not crucial for the overall effects. Other public health and containment measures and county-level mandates are added in robustness checks. State-level public health and containment measures are collected by the Oxford COVID-19 Government Response Tracker, and cover restrictions on (i) school closures, (ii) workplace closures, (iii) public events, (iv) size of gatherings, (v) public transportation, (vi) stay at home requirements, (vii) restrictions on movement, (viii) public information campaigns, (ix) testing policies, (x) contact tracing, and (xii) facial coverings.9 We also include controls for county level mask mandates from Wright et al. (2020). Standard errors are clustered at specific pairwise state borders yielding a total of 363 clusters.10

In the Online Appendix B we show that the specification in Eq. (1) for new weekly COVID-19 cases identifies the effect of mask mandates on the contact rate of a Susceptibles-Infected-Recovered (SIR) model up to a proportion. This is because active infections are continuous around mask borders, as shown there. The tight relationship between our estimate and the contact rate of the SIR model allows for a structural interpretation of our results and underpin our subsequent counter-factual policy analysis.

3. Results

This section reports and discusses our estimates of the effect of mask mandates on COVID-19 cases, hospital admissions, and deaths, including on how effects vary across political leaning.

We find that state mask mandates significantly reduced new weekly COVID-19 cases, hospital admissions, and deaths, see Table 1(a). Specifically, statewide mask mandates reduced new weekly COVID-19 cases by 54.95 cases per 100,000 inhabitants (column 1), COVID-19 hospital admissions by 11.44 persons per 100,000 inhabitants (column 2), and new COVID-19 deaths by 0.73 by 100,000 inhabitants (column 3). To put these numbers in perspective, average new cases, hospital admissions and new deaths in our sample are 166.44, 23.57 and 2.64.

Table 1.

Results per 100,000 inhabitants .

(a) Unconditional results
	(1)	(2)	(3)
	Cases	Admissions	Deaths
State mask mandate	−54.95***	−11.44***	−0.73**
	[17.71]	[4.24]	[0.34]
Observations	45 577	22 042	41 034
R2	0.487	0.417	0.215
Mean of dep. variable	166.44	23.57	2.64
Linear terms	Yes	Yes	Yes
Time & county FE	Yes	Yes	Yes
Lagged mobility	Yes	Yes	Yes
(b) Results conditional on political leaning
	(1)	(2)	(3)
	Cases	Admissions	Deaths
State mask mandate	−121.10***	−16.65	−2.88***
	[38.60]	[10.93]	[0.83]
Rep. vote share × State mask	107.44**	8.50	3.53***
Mandate	[49.38]	[16.29]	[1.18]
Observations	45 577	22 042	41 034
R2	0.488	0.417	0.216
Mean of dep. variable	166.44	23.57	2.64
Linear terms	Yes	Yes	Yes
Time & county FE	Yes	Yes	Yes
Lagged mobility	Yes	Yes	Yes

Notes: The estimates are based on data for 3107 counties over 48 weeks. Counties within 150 miles of a state border over which mask mandates vary are included. We lead cases(hospital admissions)[deaths] by 1(2)[4] weeks with respect to the state mask mandate dummy. “Republican vote share” denotes the share of votes obtained by the Republican presidential ticket in 2020 at the county level. All specifications include time and county fixed effects. Standard errors are clustered at specific pairwise state borders. ***, **, and * indicate statistical significance at 1, 5, and 10 percent, respectively.

Authors’ calculations.

The covariates we use do not exhibit a discontinuity at the mask border. Importantly for the validity of our design (see Lee and Lemieux (2010)), Table C1 in the Online Appendix shows that mobility, Republican vote share, and other containment and health policies at the state-level do not jump at the mask border (except weakly for work place closures, and negatively for stay-at-home orders). We also check for jumps in the outcome variables across the mask border in the two weeks before statewide mask mandates were imposed. If such jumps exist and are negative, it could lead us to erroneously conclude there is an effect of mask mandates, when instead the differences in outcomes pre-date the introduction of mandates. Table C2 shows that this is not the case. For cases it is actually the opposite. In the two weeks before a mandate is imposed, cases are significantly higher in counties that eventually get mask mandates. There is no significant difference for deaths. Hence, our baseline estimate for cases is a lower bound on the true effects of mandates, meaning mandates likely prevented even more cases than we estimate.

The effects also remain unchanged when controlling for other public policies. We control for the vast array of state-level containment and health policies collected by the Oxford COVID-19 Government Response Tracker in Tables C3–C5 in the Online Appendix.11 Our results are robust to controlling for all policies, an index of containment and health measures and to each individual measure in the index. Most measures are insignificant except for international travel restrictions for cases and deaths, but even in that case our coefficient of interest remains unchanged. Note that “facial coverings” are insignificant too. That variable is coded differently than ours, containing information on whether a state has significant sub-state mask mandates. In that sense, by controlling for it we show that such sub-level mandates are not obviously affecting our results.

We also investigate whether the presence of county-level mandates affect our results. We do this using data on county-level mandates from Wright et al. (2020). Conceptually, such mandates could introduce a downward bias in our baseline results since counties without statewide mandates may have county-level mandates. This confounding factor would make it harder to find an effect of statewide mandates. In practice, the results in Table C6 show that estimates are essentially unchanged when controlling for county-level mandates. Note that this is distinct from investigating whether county-level mandates are themselves effective as that would necessitate a regression discontinuity design around the borders of counties implementing mandates. We do not pursue that line of inquiry as our research question in on the effectiveness of statewide mandates.12

Our estimated coefficients, and their significance, are also robust to a range of other checks. The design controls for linear terms on the distance to the mask border that may be different on either side of said border. Table C7 in the Online Appendix shows that restricting the linear coefficient to be the same or using a second order polynomial on minimum distance to the mask border does not affect the coefficients of interest significantly, except in the most flexible specification for deaths. However, none of the polynomial coefficients are significant in those specifications, pointing to over-fitting. In general, we find the coefficients on these polynomials on distance to be weak, with the only robustly significant coefficient being the effect at the border. The baseline controls for county and time fixed effects and county-level prior mobility lagged by one week. However, neither of these controls affect the significance or overall level of the estimates, see Table C8 in the Online Appendix. The coefficient on lagged mobility is negative for all outcomes, likely reflecting reverse causality and slow moving outcomes and mobility — i.e. concerns about worsening COVID-19 outcomes likely lead to lower mobility, all else equal. As we are not interested in the causal effects of mobility, we do not explore this correlation further but retain lagged mobility as a control in all main regressions to avoid biased estimates. However, as mentioned this inclusion is not crucial. Table C9 in the Online Appendix shows that estimates remain significant and broadly of the same magnitude across different bandwidths, i.e. the maximum distance that defines mask borders. An exception is deaths where wider bandwidths render the coefficient on mask mandates insignificant. In any case, the RD design is well motivated for reasonably close counties, and we think 165 $+$ miles is too wide of a bandwidth (for reference, that is close to the “crow” distance between New York City and Baltimore, which crosses several state lines). The choice of a bandwidth of 150 miles is well motivated as it makes sure all counties in state borders are included. Below that value, regressions exclude border counties, above that value regressions start including a lot of counties far from the border and even potentially not even in a neighboring state.13 In Table C10 in the Online Appendix, we also vary the leads for cases, deaths, and hospital admissions from 1–3, 2–4, and 3–5 weeks, respectively. Cases remains significant across all choices of leads. Hospital admissions and deaths are significant for leads of 2–3 and 4–5 weeks, respectively, consistent with hospital admissions preceding deaths, with the latter calling for the longest leads.14 As further robustness, we also estimate models using a contiguous border county-pair approach (Online Appendix F) and a difference in discontinuity approach (Online Appendix G). Our main findings are robust across these alternative specifications, although the exact coefficients vary across the approaches, partly due to differences in the samples generated by the different methodologies.

Crucially, we find that the effects of mask mandates vary with political leaning. Given existing evidence of differing views on mask wearing across political leaning (Milosh et al., 2020), we investigate whether the effects found in Table 1(a) depend on political leaning as proxied by the share of Republican votes in the 2020 Presidential election (see Table 1(b)). Cases and deaths indeed seem to be affected by mask mandates differently depending on political leaning at the county level (columns 1 and 3),15 while we do not find an effect for hospital admissions (column 2), potentially reflecting the much smaller sample for that variable. Specifically, mask mandates reduce new weekly COVID-19 cases and deaths by −77.50 and −1.44, respectively, in the median Democratic-leaning county in our sample (those with a Republican vote share of 41%). While the same numbers for the median Republican-leaning county are −44.37 and −0.36, also respectively (Republican vote share of 71%). In these illustrative calculations, median counties across political leaning are defined as the vote share for the Republican ticket among counties whose share is smaller than 50% (Democratic counties) or larger than 50% (Republican counties). However, for Republican voting shares above 65% the estimated effect on deaths becomes insignificant (Fig. 4).16 This underscores the idea that mask mandates can be effective provided attitudes towards them are favorable. Separate regressions exploring an additional interaction with mobility are not consequential — the mobility interaction itself is never significant, while the interaction with political leaning remains significant for cases and deaths. Thus, differences in mobility do not explain our finding that the effect of mandates depends on political leaning.

Fig. 4.

The estimated effects of statewide mask mandates on deaths per 100,000 inhabitants across political leaning . Note: The effect is computed using the estimates in column 3 of Table 1. “Republican vote share” denotes the share of votes obtained by the Republican presidential ticket in the 2020 election.

A key channel through which mask mandates are likely to decrease the spread of COVID-19 is by encouraging and achieving greater mask usage among the population. To study this possible effect, we use the results of a survey on mask usage across counties in early July of 2020, which is around the middle of our sample, and just after several states introduced statewide mandates. Using a regression discontinuity design exploiting this single cross section, we find that statewide mandates increase the share of the population wearing masks (see Online Annex H). Also, consistent with our results we find that this effect decreases with the share of Republican votes across counties. These findings are consistent with existing literature that finds that political leaning shapes the use of masks, even when mask mandates are in place (Milosh et al., 2020).

4. Conclusion

This paper estimates the effect of statewide mask mandates on COVID-19 cases, hospital admissions, and deaths in the United States. To identify such an effect, we rely on variation in the adoption of state-level mandates across time and on how COVID-19 outcomes vary across counties near state mask borders.

Our estimates imply that statewide mask mandates saved 87,000 lives and could have potentially saved 57,000 additional lives up to December 19, 2020,17 $avg\left(mas{k}_{c,t}\right)\times weeks$ is the average duration of county c’s mask mandate in the period of April 18 and December 19, 2020, and $populatio{n}_c$ is county c’s population. Potentially saved lives use $1-avg\left(mas{k}_{c,t}\right)$ instead of $avg\left(mas{k}_{c,t}\right)$ in the above formula. Lives saved are calculated comparing the actual mask mandates in place after April 18, 2020 – when the first states, Maryland and New York, instituted their mandates – with a scenario without any statewide mandates. Lives potentially saved are calculated under a scenario in which a nationwide mask mandate would have been imposed starting April 18, 2020.

Given existing evidence on mask usage being shaped by political leaning, we investigated whether the effect of mask wearing depends on the political leaning at the county level. We found this to be the case, and these conditional effects are important enough that they must be taken into account when conducting counterfactual policy experiments. This paper is the first to look at such conditional effects, thus complementing evidence that political leaning shapes the use of masks, even when mask mandates are in place (Milosh et al., 2020). If we ignored conditional effects (using instead column 3 of Table 1(a)), estimates for lives saved and potentially saved would be 45 and 33 percent lower, respectively. These conditional results highlight that mask mandates are particularly powerful if attitudes towards them are favorable among a large portion of the population.

Saved and potentially saved lives vary across political leaning and geography. Most lives saved are located in California, Texas, Michigan, and New York (Online Appendix Figure D1a). Potentially saved lives are notably concentrated in California and Florida (Online Appendix Figure D1b). These distributions across states also clearly reflect population sizes.

Our findings have important broader implications beyond the year 2020 in the United States. First, not all countries are rolling out vaccines with equal speed and global widespread vaccination is not likely to be achieved in the near term. Second, COVID-19 may become an endemic disease much like the influenza,18 and thus the effectiveness of current vaccines against current and future variants of the virus could be low. At the same time, our results also establish the limitations of mask mandates and their dependence on individual’s attitudes proxied by political leaning. Hence, mask mandates are likely to remain an important policy in the toolkit against COVID-19 and other future pandemics. Our results also underscore the need for complementary policies that underscore the reason and importance of adhering to mandates, as well as adjusting other behaviors, to stem the spread of any airborne virus.

Uncited references

Eichenbaum et al., 2021, Judson and Owen, 1999, Dube et al., 2010, Butts, 2021, Grembi et al., 2016

Appendix A. Supplementary data

The following is the Supplementary material related to this article.