1 Introduction

In response to COVID-19, many U.S. states issued SAH orders mandating the closure of nonessential businesses and requiring residents to remain indoors except for essential activities between mid-March and late May 2020.Footnote 1 Despite their controversial and adverse socioeconomic impacts (Baek et al. 2021; Baron et al. 2020; Bullinger et al. 2021; Forsythe et al. 2020; Leslie and Wilson 2020; Silverio-Murillo et al. 2021), the SAH orders were found to be effective in reducing social interactions and in slowing the spread of the virus (Chernozhukov et al. 2021; Medline et al. 2020).

This paper aims to evaluate an unintended consequence of the SAH orders—improvement in air quality. With businesses closed and people working from home, there was a sharp drop in road traffic and, therefore, motor vehicle emissions. We estimate the impacts of the SAH orders on two criteria air pollutants in the northeastern USA—nitrogen dioxide (\({\textrm{NO}}_2\)) and carbon monoxide (CO). Motor vehicles are the dominant source of these emissions, responsible for over 50% of CO and 34% of \({\textrm{NO}}_2\) emissions in the USA (Currie and Walker 2011). We focus on six northeastern US states: Connecticut (CT), Massachusetts (MA), New Jersey (NJ), New York (NY), Pennsylvania (PA) and Rhode Island (RI). We use panel data from the Environmental Protection Agency (EPA) to estimate the changes in the concentrations of \({\textrm{NO}}_2\) and CO during the SAH orders. The panel data consist of daily concentrations of these two air pollutants and cover the period from January 1, 2015, to December 31, 2020. Baltagi (2008) and Hsiao (2014) have documented that panel data can help reduce estimation bias by controlling for unobserved heterogeneity.

We find significant reductions in the concentrations of \({\textrm{NO}}_2\) and CO during the SAH orders: \({\textrm{NO}}_2\) decreased by about \(24\%\) and CO by about \(13\%\). Next, we consider the heterogeneity of the estimated effects with regard to population density and to the proximity of the monitor to major roads. We find that the reductions were larger for areas of high population density and for areas near major roads. Finally, we use an event study to examine how these impacts varied over time in each state. For most states, we see a general trend of reductions beginning 1−3 weeks before the orders took effect. This trend began to reverse 6−8 weeks after the orders were lifted, suggesting that the air quality improvements were temporary.

This paper is closely related to the growing literature on the impacts of COVID-19 responses on air quality. While most of the recent studies have found that the SAH orders led to significant improvements in air quality, the effects varied across locations. For example, Berman and Ebisu (2020) find that \({\textrm{NO}}_2\) decreased by \(25.5\%\) in the USA. Dang and Trinh (2021) estimate that global concentrations of \({\textrm{PM}}_{2.5}\) and \({\textrm{NO}}_2\) decreased by 4−5%. Zhang et al. (2021) show that \({\textrm{PM}}_{2.5}\) declined in rural areas of New York state but remained unchanged in urban areas. On the other hand, Jia et al. (2020) find no significant reductions in air pollution in the Memphis metropolitan area. Our paper is also related to the literature on the relationship between air pollution and COVID-19 outcomes. Some recent studies have found that air pollution is associated with a more rapid spread of COVID-19 and a higher mortality rate (Kasioumi and Stengos 2022; Kim and Bell 2021; Travaglio et al. 2021; Wu et al. 2020). The findings from these studies, together with those of our study, suggest an indirect channel through which the SAH orders slowed the spread of virus and saved lives.

This paper makes three contributions to the literature. First, our results provide a more comprehensive picture of the impacts of the SAH orders by focusing on the Northeast, which was among the areas hardest hit in early 2020. Similar to most of the previous studies, ours finds that air quality improved during the SAH orders. It is important to note that because we use monitor-level data, our findings can better speak to the areas in the vicinity of the monitors than to the areas far from the monitors. Second, while much of the previous work uses air pollution data from the early pandemic period, our study includes data through the end of 2020. This allows us to examine how the air quality changed weeks after the SAH orders were lifted and whether the air quality improvements were temporary. Third, we extend our analysis and document the heterogeneous effects of the SAH orders with regard to population density and to the proximity of the air pollution monitors to major roads. Because of data limitation, we are not able to disentangle the effects of all the possible factors and mechanisms contributing to air quality improvements. However, by focusing on the major pollutants from motor vehicles and examining the heterogeneous effects of population density and proximity to major roads, we can provide suggestive evidence on how changes in human mobility affect air quality.

The remainder of the paper is organized as follows: Section 2 describes the data and provides an overview of the timeline of the statewide SAH orders. Section 3 describes our empirical approach. Section 4 presents the estimation results. Section 5 concludes.

2 Data

2.1 Timeline of SAH orders

We obtain the dates of each state’s COVID-19 policy responses from The New York Times and local government websites.Footnote 2 The timeline is summarized in Table 1. Most states declared a state of emergency in early March, and all of the states in our sample had announced an SAH order by the end of March, with the effective date in the same or the following week. Each state’s reopening followed a multi-phase, data-driven plan, and the restrictions on businesses and economic activities were relaxed gradually.Footnote 3 The “Date Lifted” column in Table 1 lists the earliest reopening dates for counties that met the key metrics in each state.Footnote 4

Table 1 Timings of COVID-19-related statewide policy responses
Full size table

2.2 Air pollution data

We collect data on the daily concentrations of the criteria air pollutants \({\textrm{NO}}_2\) and CO from the EPA’s Air Quality System (AQS), which contains ambient air pollution data measured by a network of air quality monitoring stations across the USA.Footnote 5 We focus on six northeastern states: Connecticut, Massachusetts, New Jersey, New York, Pennsylvania and Rhode Island. The Northeast is the most economically developed and densely populated region of the four U.S. Census regions. During the early pandemic period (March−April 2020), 40% of northeastern counties (representing 84% of the region’s population) were identified as COVID-19 hotspots (Oster et al. 2020).

Our dataset covers the period from January 1, 2015, to December 31, 2020. The observation is at the monitor level. We only consider monitors that were active for at least two years during our sample period and that had data readings for 2020. This leads to 51 \({\textrm{NO}}_2\) monitors in 39 counties and 33 CO monitors in 23 counties. Panel A of Table 2 presents the summary statistics of the data on air pollutants. Daily \({\textrm{NO}}_2\) denotes the maximum one-hour average concentration and is measured in parts per billion (ppb). Daily CO denotes the maximum eight-hour average concentration and is measured in parts per million (ppm).Footnote 6

Figure 1 plots the monthly average concentrations of the two air pollutants from 2015 to 2020. They are computed as the average daily concentrations in one month across the six northeastern states. We can see that air pollution has distinct seasonal and monthly variations. The concentrations of \({\textrm{NO}}_2\) and CO are higher during colder seasons and lower in warmer seasons. Table 3 displays the average air pollution levels in March, April, and May during the years 2015−2019 (without SAH orders) and the year 2020 (with SAH orders). Columns (1) and (2) present the average concentrations of the air pollutants in each month during the pre-pandemic period (2015−2019) and during the pandemic (2020), respectively. Both the time-series plot in Fig. 1 and a simple mean comparison in Table 3 suggest that the air pollution levels in March, April, and May of 2020 were lower than the levels in those same months averaged over 2015−2019. Note that the analysis here only suggests a correlation, not a causation. Without further study, it is not clear whether the reductions in air pollution were due to the SAH orders or to other factors because air pollution typically is affected by various factors that could vary from year to year and location to location. This motivates our panel data approach in Sect. 3.

Table 2 Summary statistics of data variables
Full size table
Fig. 1
figure 1

Monthly average air pollutant concentrations, 2015–2020

Full size image
Table 3 Mean concentrations of air pollutants (2015–2019 vs. 2020)
Full size table
Fig. 2
figure 2

\({\textrm{NO}}_2\) and CO monitoring stations

Full size image

Notes: The orange icons in Panel A denote the \({\textrm{NO}}_2\) monitors, and the purple icons in Panel B denote the CO monitors.

Figure 2 presents the distributions of the \({\textrm{NO}}_2\) and CO monitors. We observe that the monitors are not evenly distributed. About 67% of CO monitors are located in counties that also have \({\textrm{NO}}_2\) monitors. The number of monitors in a given core-based statistical area is determined by a variety of factors, such as local population and through-traffic. Typically, the literature treats the siting of monitors as exogenous and the resulting sample data as representative. Grainger and Schreiber (2019) find that new monitors tend to be placed in relatively less polluted areas and that race may play a role in local siting decisions. We believe this is not a major concern in our analysis because all of the monitors in our sample were sited at least two years before the COVID-19 pandemic. Therefore, it is reasonable to assume that the SAH orders were not correlated with the location of the monitors.

2.3 City-level population density

We collect data from the U.S. Census Bureau on city-level population density (based on the 2020 Census). Panel B of Table 2 presents the summary statistics of population density.

2.4 Distance to major roads

We use Google Maps to measure the distance between the monitoring station and the nearest major road.Footnote 7 Panel C of Table 2 presents the summary statistics of distance to major roads. Figure 3 illustrates how we measure the distance using Google Maps.

Fig. 3
figure 3

Illustration of using Google Maps to measure the distance between monitoring station and the nearest major road

Full size image

3 Empirical approach

Our baseline econometric model is the following:

$$\begin{aligned} \log (y_{\textrm{icst}})=\alpha _i+\alpha _c+ \alpha _{sy}+\textbf{X}_t + \beta _1 {\textrm{SAH}}_{{\textrm{icst}}}+\beta _2 {\textrm{PreSAH}}_{{\textrm{icst}}}+\beta _3 {\textrm{PostSAH}}_{{\textrm{icst}}}+\epsilon _{{\textrm{icst}}},\nonumber \\ \end{aligned}$$
(1)

where \(\beta _1\) is the parameter of interest; \(y_{{\textrm{icst}}}\) is the daily concentration of the air pollutant \({\textrm{NO}}_2\) or CO at monitor i, county c, state s and time t; \(\alpha _i\), \(\alpha _c\) and \(\alpha _{sy}\) are the monitor, county and state-by-year fixed effects, respectively; \(\textbf{X}_t\) is a collection of time-varying controls that captures any time-varying shocks to air pollution common across counties—these include month fixed effects, week fixed effects, day-of-week fixed effects and indicators for holidays;Footnote 8\({\textrm{SAH}}_{{\textrm{icst}}}\) is an indicator equal to one if the county c where monitor i is located was under a SAH order at time t; \({\textrm{PreSAH}}_{{\textrm{icst}}}\) is an indicator equal to one if the county where monitor i is located at time t was in the preSAH orders period, which is defined as the period from the state of emergency declaration date to the day before the SAH order effective date; \({\textrm{PostSAH}}_{{\textrm{icst}}}\) is an indicator equal to one if the county c where monitor i is located at time t was in the postSAH orders period; and \(\epsilon _{{\textrm{icst}}}\) is an error term.

The monitor and county fixed effects control the time-invariant monitor-specific and county-specific characteristics, respectively. The state-by-year fixed effects capture the differences across states (e.g., state-level economic environment and policies) that vary from year to year. We include the holiday and day-of-week fixed effects to account for the fact that people may exhibit different travel behavior during holidays and weekends. The month fixed effects account for seasonality in air pollution. The PreSAH indicator accounts for the fact that some people started to work from home and reduced travelling before the SAH orders took effect. The PostSAH indicator accounts for possible lagged effects. For example, some people continued to work from home, and businesses only opened at limited capacity. Therefore, there still may have been less travel and reduced economic activity than during the same period over the years 2015–2019.

Our identification strategy exploits the exogenous variation in the timing of the SAH orders, i.e., \(E({\textrm{SAH}}_{{\textrm{icst}}} \cdot \epsilon _{{\textrm{icst}}}|\alpha _i, \alpha _c,\alpha _{sy},\textbf{X}_t)=0\). More intuitively, the identifying assumption is that the changes in air pollution across the days of the week during the weeks without SAH orders provide a good counterfactual for the changes that would have been observed on the lockdown days in the absence of the SAH orders. We cluster the standard errors by state-year to address the possible serial and cross-sectional correlation of error terms. This allows for the error term \(\epsilon _{{\textrm{icst}}}\) to be correlated across all observations within the same state and year.

4 Estimation results

The main estimation results of Eq. (1) are shown in Table 4. Columns (1) and (2) use the full sample, and column (2) includes a third-order time trend. Although we focus our subsequent discussion on the richer specification corresponding to column (2), the estimates vary little across specifications. As expected, the SAH orders led to significant reductions in the concentrations of the air pollutants \({\textrm{NO}}_2\) and CO. As shown in column (2), \({\textrm{NO}}_2\) and CO decreased by approximately 24% and 13%, respectively. In addition, there was approximately a 15% reduction in \({\textrm{NO}}_2\) and a 10% reduction in CO during the preSAH orders period. We also find that \({\textrm{NO}}_2\) was approximately \(8\%\) lower during the postSAH orders period. However, we find no significant reductions for CO during the postSAH orders period. Columns (3) and (4) report the results of using the sub-samples 2016–2020 and 2017–2020, respectively.Footnote 9 We use the two sub-samples as a robustness check to see if changing the baseline period affects our estimates. Comparing columns (2)–(4) shows that, overall, the estimation results using different samples are quite similar.

Next, we explore the potential heterogeneity of the impacts of the SAH orders with regard to population density. Specifically, we use the same empirical strategy as in Sect. 3, but we separately estimate Eq. (1) for “low population density areas” and “high population density areas”. The “low population density areas” are defined as areas located in cities with population densities below the median, and the “high population density areas” are defined as those with population densities above the median.Footnote 10 The results are presented in Table 5. We observe that the estimated reductions in \({\textrm{NO}}_2\) are larger for the high population density areas; the reductions in CO are similar for the two areas.

We also evaluate the potential heterogeneity with regard to proximity to major roads. Similar to what we show in Table 5, we split the samples based on the distance between the monitor and the nearest major road with the cutoff at the median, and we estimate Eq. (1) for areas “near major roads” and areas “far from major roads”, respectively.Footnote 11 The results are shown in Table 6. We observe that there are significant estimated reductions in air pollution for both areas near and far from major roads. In addition, in general, the reductions are larger for the areas near major roads than for those far from major roads. Our findings indicate that with people working from home and driving less, the decrease in road transportation led to reductions in emissions and, therefore, improvement in air quality. This analysis provides suggestive evidence that a drop in motor vehicle use is a possible channel for air quality improvement.

Given that the states had slightly different timelines of SAH orders, we further conduct an event study analysis to examine how the concentrations of the air pollutants changed over time for each state. Specifically, we average the daily data by week and include indicators for the weeks prior to and after the SAH orders. The model is as follows:

$$\begin{aligned} \log (y_{\textrm{ict}})=\alpha _i+\alpha _c+ \textbf{X}_t + \sum _{j=-6}^{24}\beta _j \textbf{1}\{j \text { weeks since SAH orders}\}+\epsilon _{ict}, \end{aligned}$$
(2)

where \(y_{\textrm{ict}}\) is the weekly concentration of the air pollutant \({\textrm{NO}}_2\) or CO at monitor i, county c and week t; \(\alpha _i\) and \(\alpha _c\) are the monitor and county fixed effects, respectively; \(\textbf{X}_t\) is a collection of time-varying controls that captures any time-varying shocks to air pollution common across counties—these include year fixed effects and month fixed effects; \(\textbf{1}\{j\text {weeks since SAH orders}\}\) is an indicator which takes the value one if the SAH orders have been in effect for j weeks, and \(\epsilon _{\textrm{icst}}\) is an error term. We include six weeks prior to and twenty four weeks after the implementation of the SAH orders.Footnote 12 The event study coefficients allow us to examine the dynamics of air pollution during the pandemic.

In Figs. 4 and 5, we present the event study graphs that show the estimates of the effects of the SAH orders on the air pollutants \({\textrm{NO}}_2\) and CO, respectively. The corresponding 95% confidence intervals are shown in gray. In each subfigure, the solid circles represent the estimated effects at different weekly time points. The vertical yellow and purple dashed lines mark the weeks that the statewide SAH order was implemented and lifted, respectively. Although precise magnitudes are difficult to discern from the figures, the results as a whole reveal some clear patterns. Specifically, the estimates show that for most states, there was a general trend of reduction in air pollution 1−3 weeks before the SAH orders took effect, and that trend started to reverse 6−8 weeks after the orders were lifted. The reductions prior to the SAH order effective date might be due to the fact that some people started to work from home and drove less after the declaration of the state of emergency and the closure of schools. The reversed trend suggests that the air quality improvements were temporary.Footnote 13

Finally, we note that the econometric approach in this paper uses the pandemic-caused SAH orders as exogenous policy shocks to identify the causal effects of the SAH orders on air quality. However, if there are some unobserved pre-COVID time-varying state characteristics correlated with the orders, there will be endogeneity and biased results. For example, some states may already have had better measures in place to protect air quality. These same states may also tend to be more protective of workers. Therefore, such states would be more likely to issue stringent SAH orders. In this case, the effects of the SAH orders will be overestimated.Footnote 14 In addition to the results presented in this paper, we conduct an event study using Model (2), where we include 16 weeks of indicators prior to the implementation of the SAH orders to examine the dynamics of air pollution prior to the pandemic. However, we do not observe a persistent downward trend in the concentrations of air pollution, which provides suggestive evidence in favor of the exogeneity identification assumption.Footnote 15

Table 4 Estimated effects of SAH orders on air pollution
Full size table
Table 5 Effects by population density
Full size table
Table 6 Effects by proximity to the major roads
Full size table
Fig. 4
figure 4

Effects of SAH orders on air pollutant \({\textrm{NO}}_2\) by state

Full size image
Fig. 5
figure 5

Effects of SAH orders on air pollutant CO by state

Full size image

5 Conclusion

In this paper, we study the effects of SAH orders on air quality in the northeastern USA. We find that during the SAH orders, there were significant reductions in the air pollutants \({\textrm{NO}}_2\) and CO. However, the reductions got smaller and the air pollution gradually returned to normal levels after the orders were lifted, indicating that the air quality improvements were temporary. One extension of this paper is to consider the quantile treatment effects of SAH orders.Footnote 16 We leave this as a future research topic.

Our paper contributes to the growing literature on assessing COVID-19 pandemic-related policies and their impacts on the environment as well as on human health and well-being. This paper has two policy implications. First, the reductions in air pollution estimated in this study and their associated health benefits found from other studies (e.g., Kasioumi and Stengos (2022) and Kim and Bell (2021)) highlight the need for governments to invest in sustainable transportation and to take actions to protect our environment. Second, as there has been an ongoing debate over the SAH order policies, our findings provide some evidence of their environmental benefits. This can help policymakers weigh the pros and cons of SAH orders and make more balanced decisions in the case of future pandemics.