COVID-19 Lockdowns and Air Quality in the United COVID-19 Lockdowns and Air Quality in the United States States

,


Introduction and Background
It is well known that both long-term and short-term air pollution exposure lead to negative health outcomes.Of the many industries that contribute to the degradation of air quality, the transportation sector is a major source of emissions.The COVID-19 pandemic and subsequent lockdowns caused an unprecedented halt in human mobility.For this reason, the lockdowns serve as a natural experiment to investigate the link between transportation and air pollution.
Media outlets were quick to attribute observed reductions in air pollution to lockdowns (Popovich 2020;Duncan 2020;Watts & Kommenda 2020).However, to my knowledge, there are no existing studies that definitively link U.S. lockdowns to lower levels of pollution.
Empirical evidence from China confirms this phenomenon to a certain extent, highlighting heterogenous effects across different pollutants (Almond et al. 2020;Cole et al. 2020).If this were also the case in the U.S., policymakers would benefit from knowing what types of restrictions on human activity effectively reduce air pollution levels.
There are several mechanisms through which lockdowns may cause reductions in pollution.Because the main purpose of lockdowns is to restrict mobility, transportation-related pollutants should see dramatic improvements.For example, NO21 is emitted through fuel burning, so reductions in human mobility in the form of bus, truck, and airplane miles could drive air quality improvements (U.S. Environmental Protection Agency 2016).Furthermore, observed reductions in gasoline sales during the lockdowns (Fraser 2020) indicate a fall in personal vehicle travel which could lead to reductions in both PM2.5 and PM10 considering particulate matter levels are positively associated with personal vehicle travel (Lu et al. 2017).
For these reasons, I focus on changes in PM2.5, PM10, and NO2 in this analysis, as they are most likely to respond to lockdown-induced decreases in mobility.On the other hand, pollutants tied to coal burning such as CO and SO2 may remain unaffected.I also examine Air Quality Index (AQI) in my analysis to capture air quality in the general sense.This paper exploits the incidence of COVID-19 to examine the link between reduced human mobility and various air pollutants in the U.S. Taking advantage of the fact that some states had stringent lockdowns while others had lenient lockdowns, I investigate whether there were differential air pollution outcomes between the two groups.I utilize a difference-indifferences approach and empirical data on lockdown stringency and air pollution levels to answer this question.Contrary to expectations, I find no clear relationship between lockdowns and AQI, PM2.5, PM10, or NO2 levels.I also narrow my study area to only metropolitan counties to test if strict lockdowns are more influential in urban areas but find similar results to the main analysis.Although the findings of this paper are unexpected, they make an important contribution to the underdeveloped literature surrounding the environmental implications of the U.S. COVID-19 lockdowns.To my knowledge, this paper is the first to use empirical data to investigate a causal link between lockdowns and air pollution levels.
This paper proceeds as follows: Section 2 reviews existing literature on air pollution and COVID-19 lockdowns, Section 3 presents the data, Section 4 lays out the empirical methodology, Section 5 reports the results, Section 6 discusses the results, and Section 7 concludes.

Lockdowns and Air Quality in China
Most of the existing papers on this topic utilize either difference-in-differences or regression discontinuity design to examine air quality variations following the lockdowns in China (Almond et al. 2020;Cole et al. 2020;Liu et al. 2020;Ming et al. 2020).This is due to the fact that COVID-19 first appeared in China, and data availability on most other countries was limited until recently.Whether or not air quality improved in China depends on the examined pollutants.For example, Almond et al. (2020) finds reductions in PM2.5 and NO2 resulting from the lockdowns, with O3 worsening and SO2 seeing only modest improvements.It hypothesizes that reductions in China's NO2 levels are attributable to reductions in transportation sector activity, which is consistent with the mechanisms laid out in Section 1. Furthermore, Almond et al. (2020)'s results are consistent across the China literature; Cole et al. (2020) finds NO2 and PM10 reductions, but no improvements in SO2 or CO.Ming et al. (2020) finds reductions in AQI and PM2.5.These findings further reinforce the theory that PM10, PM2.5, and NO2 levels should decrease as a result of the lockdowns whereas other criteria pollutants such as SO2, CO, and O3 may remain ambiguously affected.

Lockdowns and Air Quality in the United States
The literature surrounding the U.S. COVID-19 lockdowns is sparse.Among the only papers to examine the relationship between the lockdowns and air quality in the U.S. is Newbold et al. (2020), which utilizes a modified SIR model to estimate the number of avoided deaths resulting from air quality improvements.However, it simply assumes air quality improvements are a causal outcome of the U.S. lockdowns rather than proving this phenomenon with empirical data.Similarly, Cicala et al. (2020) estimates air quality improvements in the U.S. through variations in electricity use, meaning no empirical air pollution data is used.With that being said, is it safe to assume lockdowns had a causal effect on air quality in the U.S.?And if so, which components of air quality drove these changes?My paper addresses these unanswered questions through the use of empirical data on various pollutants, similar to the aforementioned China studies.

Air Pollution Data
The EPA maintains monitors throughout the U.S. that collect daily summary data on various pollutants, which are made available through their Pre-Generated Data Files 2 .I obtain daily air quality data over the period from January 1, 2019 to October 31, 2020.This data is at the county level, though only 1,043 of 3,007 U.S. counties are observed 3 .Large gaps and inconsistent observational intervals make an analysis with the daily data problematic.Therefore, I instead calculate monthly averages of air pollution.
Overall air quality is reflected through AQI, which ranges from 0 to 500.Air quality classifications are characterized by 50-point intervals.For example, 0 to 50 is "Good" and 51 to 100 is "Moderate."However, any measurement in excess of 300 is considered hazardous.AQI is the maximum of six sub-indices, which are calculated by individually plugging raw O3, PM2.5, PM10, CO, SO2, and NO2 concentrations into a piecewise linear function 4 (Office of Air Quality 2 Explanation of the EPA's air pollution data collection process is documented at https://aqs.epa.gov/aqsweb/airdata/FileFormats.html#_monitors 3 Only about a third of counties are observed due to monitor constraints.Not every county in the U.S. contains at least one monitor.An interactive map of EPA air quality monitor locations is found at https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=&cad=rja&uact=8&ved=2ahUKEwjjouS_korxAhVVqp4KHTgRB7EQFjAPegQIFBAE&url=https%3A%2F%2Fwww.arcgis.com%2Fapps%2Fwebappviewer%2Findex.html%3Fid%3D5f239fd3e72f424f98ef3d5def547eb5%26extent%3D-146.2334%2C13.1913%2C-46.3896%2C56.5319&usg=AOvVaw24rvygp_sLud1zyAeAkDnH  4 Formulas used in the calculation of AQI are found at https://www.airnow.gov/sites/default/files/2018-05/aqitechnical-assistance-document-may2016.pdfPlanning and Standards 2016).This means that on any given day, AQI will reflect only one of the six criteria pollutants.The second column of Table 1 shows U.S. AQI averages in each month of 2019.Throughout the year, AQI stayed in the 30 to 40 range-comfortably within the "Good" air quality classification.Figure 1 shows the change in average AQI from 2019 to 2020 by month.Negative values, which are synonymous with reductions in pollution averages between the two years, indicate air quality improvements.In most months average AQI improved from the year before, with larger decreases of about 6 and 8 points observed in March and October respectively.
In addition to AQI, I obtain data on the three criteria pollutants that are most likely to be affected by lockdowns-PM2.5, PM10, and NO2.Monitor-level daily arithmetic means are available for each of these pollutants from the EPA.Because some counties contain multiple monitors, I take the simple average of all monitors within each county.I then calculate monthly averages for compatibility with the rest of the data.Units of measurement and 2019 U.S. averages for each pollutant are shown in Columns 3 -5 of Table 1.Changes in the averages between 2019 and 2020 are shown in Figure 1.
Improvements in nitrogen dioxide are observed in most months, though they are not exceptionally large when compared to the average levels in Table 1.Most notable are the changes in average PM10 and PM2.5, which are positive in June, August, September, and October.
In September and October, these changes are quite large in magnitude compared to average levels in 2019.While averages are not particularly compelling, these results are contrary to findings in the literature.It is possible that the widespread wildfires that occurred in the Summer of 2020 influenced these results, but further regression analysis is necessary to differentiate these effects.

Lockdown Stringency Data
Data on lockdown stringency is obtained from the University of Oxford's COVID-19 Government Response Tracker.The database contains a lockdown stringency index that is scaled to fall between 0 and 100.It is a simple average of nine sub-indices, which are derived from nine COVID-19 lockdown variables.The variables are measured on an ordinal scale with higher values indicating stricter policies.Each variable also has a binary flag that captures whether the policy was targeted or general in geographic scope (Tatlow & Phillips 2021).The variables include school closures, workplace closures, public event cancellations, gathering restrictions, public transportation closures, stay at home orders, internal movement restrictions, international travel restrictions, and public information campaigns5 .The index aims to reflect restrictive COVID-19 policies (Hale et al. 2021), making it a suitable proxy for COVID-19 lockdowns in my analysis.
Daily lockdown stringency index values are available for all 50 U.S. states over the period from January 1, 2020 to February 9, 2021.Because I utilize data from the year leading up to the pandemic as well, a stringency index of zero is assumed for the entirety of 2019.I also calculate monthly averages for compatibility with the air quality data.Figure 2 shows the overall U.S. average stringency index in each month of 2020.In general, the U.S. saw very lenient lockdown policy in January and February of 20206 before becoming progressively stricter in March and April of the same year.Since then, lockdown stringency has steadily declined.

Control Variables
Additional factors, such as weather, might affect air quality (Cole et al. 2020).According to the National Weather Service (2021), higher temperatures can expedite chemical reactions that produce ground level ozone.Alternatively, rainfall can wash away dissolvable pollutants and

Methods
First, I classify states into "strict" and "lenient" groups using each state's sum of stringency index from January 2020 to October 2020: This formula is applied to the 48 contiguous U.S. states9 .I focus on the January through October timeframe because air pollution data for November and December of 2020 has not yet been made available.Although this paper utilizes a county-level model, this categorization step is performed at the state level because Oxford University's database does not provide county-level information.With that being said, each county within a given state is assumed to have the same lockdown stringency index and therefore the same strict-lenient classification.
Table 2 shows the 15 states with the highest sum of stringency index and the 15 states with the lowest sum of stringency index.I model air pollution with three renditions of strictlenient groups.First, I designate the top 15 states as strict and the bottom 15 states as lenient.
Then, I designate only the top and bottom 10 states as strict and lenient.Lastly, I reduce the number of states in each group to include just the top 5 and bottom 5 states.
Figure 3 shows plots for each of these groups' stringency indices over time.For the 15state specification in Panel a, there is no distinction between strict and lenient states in terms of stringency index.Because of this, I expect to find little difference in air pollution between the two groups.However, the separation becomes more apparent as the number of states in each group is reduced.There is a slight gap between the two groups in the 10-state specification in Panel b.The 5-state specification in panel c shows a clear gap between strict and lenient states' stringency indices that persists throughout 2020.In this case, I would expect to see the most dramatic air pollution discrepancies between the groups.
To test whether air quality improved as a result of the COVID-19 lockdowns, I estimate the following county-level fuzzy difference-in-differences model: is an interchangeable measure of air pollution 10 .Note that yearly differenced values are utilized in this model 11 , so ∆  can be interpreted as the change in air pollution in county  from 2019 to 2020 in any given month.Differencing controls for timeinvariant county characteristics.Furthermore, air pollution levels fluctuate naturally month-tomonth-differencing removes these discrepancies.ℎ  is an indicator variable for months January through October and   is a dummy variable equal to one if the county is in a strict state and equal to zero if the county is in a lenient state.The coefficient of interest,  3 , is the difference-in-differences estimator.∆ℎ  is a matrix containing the differenced weather control variables outlined in Section 3.3 and   is the error term.
January, which was characterized by extremely low lockdown stringency in 2020, serves as the pre-treatment and is omitted from ℎ  in the model.Therefore,  3 is interpreted as the difference in the change in air pollution relative to the pre period between the strict and lenient groups.I expect to see negative  3 's in March through October, as this would indicate strict counties saw larger improvements in air pollution than lenient counties following COVID-19 lockdowns.
The model relies on several assumptions.First, all five OLS assumptions must hold.Of those, most likely to fail is the zero conditional mean assumption, which states that unincluded factors that affect air pollution cannot be correlated with the included variables.If violated, the results will be biased.For example, if a county's political leaning affects air pollution levels but is also correlated with its strict-lenient classification, the model will overstate the magnitude of the coefficient on   if left unaccounted for.The difference-in-differences framework also requires the assumption of parallel trends.That is, absent COVID-19 lockdowns, the difference 10   takes the form of AQI to capture air pollution in the general sense.The model is also estimated using PM2.5, PM10, and NO2 for   to capture any heterogenous effects across pollutants. 11Differences are calculated by subtracting pollution levels in each month of 2019 from pollution levels in the respective month of 2020.Before differencing each pollutant, I drop observations in which the respective pollutant is recorded as zero or negative.These observations are likely indicative of a coding or monitor error.This results in 32, 7, 4, and 3 lost observations for AQI, PM2.5, PM10, and NO2 respectively.in air pollution between the strict and lenient groups remains constant throughout 2020 (Mailman School of Public Health 2013).Violation of this assumption will result in biased estimates as well.

Main Models
The results from Equation (1) are shown in Tables 3 and 4. 15-state, 10-state, and 5-state specifications are shown from left to right for each pollutant.The coefficients on ℎ  are interpreted as the difference in the change in air pollution between 2019 and 2020 relative to the pre period.For example, all AQI, PM2.5, and NO2 models display negative and statistically significant coefficients on ℎ  , meaning that on average, counties saw lower levels of these pollutants during that month relative to before the lockdowns.However, a lack of negative and significant coefficients on the remaining months suggests negligible improvement from the pre period.In terms of PM2.5 and PM10, it appears that air quality actually worsened in the post period on average, considering several months have positive and significant coefficients.
Coefficients on   represent the difference in the change in air pollution from 2019 to 2020 between the strict and lenient groups.Aside from the 10-state NO2 specification, none of the coefficients are significantly different from zero.For that reason, there is little evidence that strict counties saw lower pollution levels than lenient counties on average.Difference-in-differences coefficient results are visualized in Figure 4.There is little significance in the AQI models.The 15-state and 5-state specifications suggest that the change in AQI was about 8 to 9 points lower in strict states in October compared to the pre period.This is large in magnitude compared to mean changes of around -3 points, but these differences are only observed in October.The NO2 results are slightly more compelling.The 10-state specification shows negative and significant results from March to August, hovering around -2 parts per billion.However, these results largely diminish in the 5-state specification, with significance retained only in March, May, and July.
Results from the PM2.5 models are mixed.In the month of October, coefficients are negative and significant in the 10-state and 5-state specifications with magnitudes of -4.878 and -3.188 / 3 respectively (quite large compared to means of -0.118 and -0.326 / 3 ).
However, the February coefficients are positive and significant in all specifications, with coefficients ranging from about 1.5 to 2.5 / 3 .In terms of PM10, the only significant result is seen in June of the 15-state specification.The magnitude is sizable at -2.906 / 3 compared to a mean of 1.171 / 3 , however significance diminishes in the 10-state and 5-state specifications.

Parallel Trends
If the assumption of parallel trends holds, the difference-in-differences estimator,  3 , should be insignificant in the pre-lockdown period.Because average lockdown stringency was still quite low at 8.667 in February, this month is reasonably within the pre period.Therefore, the difference in the change in air pollution from 2019 to 2020 compared to January between the strict and lenient groups should not be statistically significant in February.If a statistically significant  3 is observed for February, this would signal a systemic difference between the strict and lenient groups even before treatment occurred, resulting in biased results.
Revisiting the main results in Figure 4, I observe statistically significant February  3 s in terms of AQI, PM2.5, and NO2.There are two main takeaways from this.First, the 10-state NO2 specification is the only model with a negative and significant  3 in February, meaning the results from this model are potentially negatively biased.Coincidentally, this is the only specification with negative and significant results in most of the tested months.It is likely that strict counties are systemically prone to lower levels of NO2 compared to lenient counties, and the negative results I find are biased due to parallel trends not holding.Second, aside from the 10-state NO2 specification, all February  3 s are positive.This suggests that if anything, the results from those models would be positively biased.Therefore, if strict lockdown policies do cause reductions in air pollution, the true results might be undetectable due to said bias.

Urban Counties
The main results in Section 5.1 wavered from expectations.It might be that air quality improvements only occurred in metropolitan counties, in which case including non-metropolitan counties would muddy the results.Urban areas are known to have higher congestion levels and poorer air quality (Strosnider et al. 2017).For these reasons, we might see larger effects when rerunning the same models from Section 5.1 on only metropolitan counties in each state 12 .The results from this robustness check are shown in Tables 5 and 6, with coefficient plots shown in Figure 5.
The results from this robustness check are almost identical to those of the main analysis.
Again, the most compelling result is seen in the 10-state NO2 specification with coefficients hovering around -2 parts per billion in March through August.However, the parallel trends violation is still a concern, and significance largely diminishes in the 5-state specification as was the case in the main results.The October coefficient is again consistently negative and significant in the AQI models, but is slightly larger in magnitude this time (estimates cap out at -12.64 points in the 5-state specification).There is again evidence to support lower PM2.5 levels in strict counties in October, with coefficients of -4.872 and -3.745 / 3 for the 10-state and 5state specifications respectively.

Discussion
The results from the main analysis provide little evidence to suggest strict lockdown policies are associated with larger improvements in pollution.Although the 10-state NO2 specification found negative and statistically significant results in March through August, these findings are likely overstated due to a parallel trends violation.It does appear that strict counties saw larger decreases in AQI and PM2.5 than lenient counties in October.These differences were as large as -9.240AQI points and -4.878 / 3 -sizable compared to mean changes in AQI and PM2.5 from 2019 to 2020 of -2.730 AQI points and -0.118 / 3 respectively.However, these differences were only consistently observed in one month, which is not enough evidence to conclude lockdown stringency made more than a negligible difference on air quality.
While the strict and lenient groups may not be significantly different in this sense, one might still expect the incidence of COVID-19 to have had a negative impact on pollution levels in general.However, a lack of consistently significant coefficients on the noninteraction ℎ  terms make it difficult to conclude whether COVID-19 coincided with reductions in pollution that were significantly larger in magnitude relative to the pre period.
The metropolitan county robustness check produced very similar results to those of the main analysis.Therefore, there is no evidence to suggest that pollution in densely populated areas is most affected by strict lockdowns and subsequent decreases in human mobility and transportation.The inclusion of rural counties in the main model does not appear to muddy the results whatsoever.One explanation for this might be animal agriculture pollution.Farms, which COVID-19 LOCKDOWNS AND AIR QUALITY IN THE U.S. 12 are usually located in rural areas, are major emitters of particulate matter (Cambra-Lopex et al. 2009).If lockdown policies disturbed agriculture supply chains causing variations in farmrelated pollutants, I would not necessarily expect differential impacts of lockdown policies between urban and rural areas.
It is important to note that this analysis looks strictly at overall levels of PM2.5, PM10, and NO2.It is possible that pollution decreased in one industry but increased in another, therefore resulting in zero net change as captured in my results.A plausible story is that transportationrelated particulate matter and NO2 decreased more in areas with stricter lockdowns, but home electricity usage increased proportionally with more people staying indoors as a result.We know that NO2 is also emitted from power plants (United States Environmental Protection Agency 2016), so elevated power usage could have offset decreases that happened elsewhere.Another plausible story is that stricter lockdown policies resulted in more businesses taking advantage of relaxed pollution reporting requirements that were granted by the EPA during the pandemic (Friedman 2020).This could also cancel out air quality improvements made elsewhere.While it would be informative to capture heterogenous changes in pollution levels across industries, net changes are arguably most important from a health standpoint.Lastly, it does not appear that moving from the 15-state to 10-state to 5-state specifications resulted in more pronounced results as previously thought based on the stringency plots from Figure 3.The coefficient plots in Figures 4 and 5 show minimal difference in the estimates as the strict-lenient groups were reduced in size.If anything, the most compelling results were observed in the 10-state specifications.

Conclusion
Based on the above results, I cannot conclude that lockdown stringency significantly impacted AQI, PM2.5, PM10, and NO2 in the U.S. My results are contrary to the existing literature, which suggests a negative relationship between China's COVID-19 lockdowns and these pollutants (Almond et al. 2020;Cole et al. 2020;Liu et al. 2020).However, my results contribute to the literature by investigating this link using empirical data from the U.S., which is the first analysis of its kind to my knowledge.
While my unexpected findings may have been due to confounding factors such as elevated home power usage in stringent counties, the failure of parallel trends puts the validity of the estimates into question.With that being said, an extension of this paper could utilize a COVID-19 LOCKDOWNS AND AIR QUALITY IN THE U.S. 13 synthetic control method to ensure a suitable counterfactual.Also, I suspect that my models suffer from omitted variable bias.Controlling for wildfire incidences and county political leanings would add to the robustness of this analysis.2) using the results from step one, find the simple average of all 48 contiguous states' pollution levels for each given month.(    (

Appendix
particulate matter.The National Oceanic and Atmospheric Administration's (NOAA) Climate at a Glance provides monthly data on temperature and precipitation for all counties in the contiguous U.S. plus Alaska dating back to 1895.The temperature data are monthly averages measured in Fahrenheit degrees and the precipitation data are monthly cumulations measured in inches.I include these in my model to control for variations in air pollution resulting from climactic factors.As a robustness check, I narrow my study area to metropolitan counties, which I identify using data from the United States Department of Agriculture's (USDA) 2013 Rural-Urban Continuum Codes 7 .These codes follow the criteria of the Office of Management and Budget (OMB) in assigning metro or nonmetro status to each county.Metropolitan Statistical Areas (MSAs) are defined as having one or more urban area with a population exceeding 50,000 and socially/economically integrated neighboring areas 8 .All counties within these MSAs are classified as "metropolitan" in the data.In 2013, 37 percent of counties fell into this classification-these counties contained about 85 percent of the U.S. population (Office of Management and Budget 2013).

Figure 1 :
Figure 1: Change in U.S. average air pollutants by month, calculated from the Environmental Protection Agency's Pre-Generated Data Files.Changes are calculated in three steps;(1) find each state's average monthly level of pollution from the available counties in the dataset, (2) using the results from step one, find the simple average of all 48 contiguous states' pollution levels (3) subtract the average U.S. pollution levels for each month in 2019 from the average U.S. pollution levels from the same month in 2020.

Figure 2 :
Figure 2: Average lockdown stringency in the U.S. for each month of 2020, calculated from the University of Oxford's Government Response Tracker Data.Average stringency index is the simple average of all 48 contiguous states' index values for each given month.

Figure 3 :
Figure 3: Graph of state stringency index from January 2020 to October 2020.Data is from the University of Oxford's Government Response Tracker.Panel a includes the top 15 and bottom 15 strict/lenient states, Panel b includes the top 10 and bottom 10 strict/lenient states, and Panel c includes the top 5 and bottom 5 strict/lenient states.

Figure 4 :
Figure 4: Plots of the coefficient of interest in Equation (1) for AQI, PM2.5, PM10, and NO2.Each point represents the difference in the change in air pollution relative to January between the strict and lenient groups.Bands show 95 percent confidence intervals.The 15-state, 10-state, and 5-state specifications are shown in Panels a, b, and c respectively.October*Strict estimates for the 10-state and 5-state PM10 model specifications are omitted because there are no strict counties with average PM10 levels available for October 2020 in the data.

Figure 5 :
Figure 5: Plots of the coefficient of interest in Equation (1) for AQI, PM2.5, PM10, and NO2, but with only metropolitan counties included.Each point represents the difference in the change in air pollution relative to January between the strict and lenient groups.Bands show 95 percent confidence intervals.The 15-state, 10-state, and 5-state specifications are shown in Panels a, b, and c respectively.October*Strict estimates for the 10-state and 5-state PM10 model specifications are omitted because there are no strict counties with average PM10 levels available for October 2020 in the data.

Table 1 :
Average levels of AQI, NO2, PM10, and PM2.5 in 2019 in the U.S., calculated from the Environmental Protection Agency's Pre-Generated Data Files.Units of measurement are shown in parentheses.These averages are calculated in two steps: (1) find each state's monthly average level of air pollution from the available counties in the dataset, (

Table 2 :
Top 15 strict and top 15 lenient states, as ranked by the sum of stringency index (calculated from the University of Oxford's Government Response Tracker Data) from January 2020 through October 2020.

Table 6 :
Model results from Equation (1) for PM10 and NO2, but with only metropolitan counties included.January is omitted from the model as the base month.(15),(10),and(5) denote 15-state, 10-state, and 5-state specification results respectively.Outcome variable means are shown in the last row to contextualize the magnitude of the estimates.Standard errors are shown in parentheses with significance levels denoted as follows: *** p<0.01, ** p<0.05, * p<0.1.October*Strict estimates for the 10-state and 5-state PM10 model specifications are omitted because there are no strict counties with average PM10 levels available for October 2020 in the data.