COVID-19 lockdown only partially alleviates health impacts of air pollution in Northern Italy

Evaluating the reduction in pollution caused by a sudden change in emissions is complicated by the confounding effect of weather variations. We propose an approach based on machine learning to build counterfactual scenarios that address the effect of weather and apply it to the COVID-19 lockdown of Lombardy, Italy. We show that the lockdown reduced background concentrations of PM2.5 by 3.84 µg m−3 (16%) and NO2 by 10.85 µg m−3 (33%). Improvement in air quality saved at least 11% of the years of life lost and 19% of the premature deaths attributable to COVID-19 in the region during the same period. The analysis highlights the benefits of improving air quality and the need for an integrated policy response addressing the full diversity of emission sources.


Introduction
Exposure to airborne pollutants is detrimental to human health. Fine particulate matter (PM 2.5 ) increases mortality rates and hospitalizations due to respiratory and cardiovascular disease (Pope and Dockery 2006, Ebenstein et al 2017, Deryugina et al 2019. Additionally, it leads to a decline in physical and cognitive productivity (Graff Zivin and Neidell 2012, Ebenstein et al 2016, Zhang et al 2018, He et al 2019, Kahn and Li 2020. Similarly, exposure to nitrogen dioxide (NO 2 ) leads to an increase in hospital admissions and premature mortality (Mills et al 2015, Amini et al 2019, Duan et al 2019.
The design of effective pollution abatement policies requires a comprehensive understanding of the relationship between reductions of emissions and concentrations. However, the processes of formation, transport, and dispersion of pollutants are complex phenomena, introducing considerable uncertainty on the effect of policies on air quality. Moreover, impact assessments need to address the confounding effect of annual and daily weather variations, a significant driver of pollutant concentrations.
This paper provides novel evidence on the change in concentrations of PM 2.5 following a composite reduction in emissions across different sources.
Specifically, we exploit the dramatic decrease in Italy's mobility and economic activity in response to the COVID-19 outbreak from late February to early May. We provide causal estimates of the change in PM 2.5 and NO 2 over more than two months for Lombardy, one of the most polluted regions among Organisation for Economic Co-operation and Development countries, and one of the first areas outside China that imposed a strict lockdown.
Using a machine-learning algorithm, we address the confounding effect of weather and build a counterfactual scenario of the pollution concentrations that would have occurred if the COVID-19 pandemic had not broken out and no lockdown had been implemented. Finally, we compute the years of life saved (YLS) and the number of premature deaths avoided by the improvement in air quality. We compare these numbers against the years of life lost (YLL) and premature deaths due to COVID-19 in the region over the same period.
Ex-post studies can provide valuable estimates of the sensitivity of concentrations to emissions. However, a host of confounding factors can seriously hinder policy evaluation. In particular, the concentration of airborne pollutants is highly dependent on atmospheric conditions. Formation, transport, dispersion, and even emission of pollutants are directly or indirectly affected by the weather (Kroll et al 2020). For instance, severe haze events in Beijing follow periodic cycles governed by meteorological conditions, especially wind patterns (Guo et al 2014). Unless the confounding impact of weather is accounted for, the estimated change in concentrations following intervention will be biased.
A common approach to impact evaluation of pollution control policies is comparing areas that were affected by a policy and areas that were not (e.g. He et al (2020) and Cole et al (2020) for the case of COVID-19 lockdowns). However, even when differences in weather have been accounted for, unaffected and comparable areas may not always exist. For the problem at hand, a precise separation between affected and unaffected regions is not possible, considering the ubiquitous adoption of measures to control the spreading of COVID-19.
We turn the complex correlation of weather and pollution to our favor, predicting concentrations as a function of weather variables and season with machine learning. We follow a simple strategy, similar to Petetin et al (2020), that does not require the availability of comparable but unaffected regions. For each air pollution monitoring station in Lombardy, we train an extreme gradient boosting regressor (Friedman 2001), a tree-based machine learning algorithm, over daily concentrations from 2012 to 2019 and predict concentrations for the first four months of 2020. We show in supplementary information (available online at stacks.iop.org/ERL/16/035012/mmedia) that this approach is more reliable than linear regression models. To account for any constant error in our prediction, including inter-annual trends (Silver et al 2020), we adopt a difference-in-differences strategy. We identify the average impact of the lockdown on air pollution concentrations as the difference between the prediction error before and during the lockdown.
We find that, despite the unprecedented halt in mobility and economic activity, the concentrations of major pollutants only partially decreased as a consequence of the lockdown. Background concentrations of PM 2.5 and NO 2 decreased by 3.84 µg m −3 (16%) and 10.85 µg m −3 (33%), respectively. Nonetheless, the improvement in air quality saved at least 11% of the YLL and 19% of the premature deaths attributable to COVID-19 in the region during the same period.
This paper contributes to several active strands of literature in air pollution research. First, it speaks to works on the assessment of pollution control policies, and in particular, to the growing corpus of research employing machine learning and fine-grained data. The paper illustrates an innovative procedure to quantify the implications of a change in emissions on outdoor concentrations of pollutants, isolating the effect of weather variability. While existing studies applying a similar approach restrict the analysis to no more than a few days, we show the conditions under which the procedure can be applied to longer time windows, the length of weeks or months. We illustrate the approach through a specific event-the lockdown of Lombardy, in Northern Italy-but it can be generalized wherever spatially and temporally detailed data on air pollution concentrations and atmospheric conditions are available.
Second, this paper is relevant to pollution control policies in the domain of study. Lombardy is a highincome, densely populated region, home to approximately 10 million people, and one of the most polluted in OECD countries. The European Commission has repeatedly referred Italy to the Court of Justice of the European Union over persistently high levels of NO 2 and PM 10 , mainly in Lombardy and the rest of the Po Valley (European Commission v. Italian Republic 2012Republic , 2019Republic , 2020. This study sheds light on the sectoral contributions to emissions of PM 2.5 and NO 2 , offering tools to regulators and policymakers. Finally, our study relates to the literature on source apportionment to different sectors, particularly agriculture, a topic of increasing relevance (Lelieveld et al 2015). During the study period, agricultural production continued unaffected, and on average 11.6 µg m −3 (39%) of PM 10 in Milan, the largest city, were attributable to agriculture. We acknowledge that missing sufficient data on 2020 sectoral emissions and on the composition of PM 2.5 , source apportionment to different sectors remains elusive. Were the data available, our machine learning approach could be used to exactly estimate changes in the composition of PM 2.5 .

Sectoral emissions during lockdown
The timing and nature of the lockdown of Lombardy and Italy are discussed in detail in the supplementary information. We highlight here two key moments. On 21 February 2020, the first outbreak of COVID-19 in Italy was identified in the south of Lombardy. Within 24 h, 11 municipalities in the region went under strict lockdown: schools were closed, all non-essential economic activities had to stop, and a stay-at-home order was in place. Teaching activities in the rest of Lombardy also were suspended. On 8 March, authorities extended the lockdown to the rest of Lombardy; and to the rest of Italy on the following day. Lockdown measures were kept in place almost unaltered until 4 May.
The progressive spreading of the virus in Northern Italy and the tightening of containment measures have substantially reduced mobility and economic activity. As mobile phone data reveals, the movement of individuals in Lombardy has followed a two-step response, following the first outbreak of COVID-19 cases in lower Lombardy (21 February) and the lockdown of the entire country (9 March) ( figure 1(a)). By mid-March, mobility dropped by three-fourths, according to data compiled by Google and Apple (Apple 2020, Google 2020). Under lockdown, all non-essential industrial production halted. As a consequence, energy demand in Northern Italy steadily decreased since 9 March, as businesses shut down, bottoming to 50% of pre-lockdown levels after two weeks (figure 1(b)).
However, not all major sources of emissions, especially those releasing precursors of PM 2.5 , have been affected by restrictions. The lockdown forced most people to home isolation; it is sensible to hypothesize that emissions from residential buildings increased as a consequence. On the other hand, emissions from non-residential buildings might have decreased. Although data to confirm this is lacking, it is plausible that emissions from heating systems have not been affected substantially.
During the transition between winter and spring, agriculture becomes an important source of secondary PM 2.5 in Lombardy (INEMAR 2017). The dispersal of animal liquids on open fields is a common (though regulated) practice that releases ammonia in the atmosphere, a precursor to secondary PM 2.5 . Public authorities have not restricted agricultural activities during lockdown in the interest of securing food supplies. These practices have continued virtually unchanged compared to previous years (personal exchange with public officials at the regional office for agriculture).
The agricultural sector is responsible for almost all emissions of ammonia (NH 3 ) in the region (INEMAR 2017), a precursor to particulate matter as it combines into ammonium nitrates and ammonium sulfates. Data on the decomposition of background PM 10 in Milan shows that ammonium nitrates and ammonium sulfates accounted for almost 40% of PM 10 concentrations during the lockdown (see figure D.2 in supplementary information). This corroborates the evidence that restrictive measures did not meaningfully alter agricultural emissions.

Machine learning
To identify the causal effect of the lockdown on concentrations without directly observing emissions, we build a synthetic counterfactual. We train a machine learning algorithm that can reproduce pollution concentrations on a business-as-usual scenario, and then predict concentrations during the lockdown. The difference between observed concentrations and the counterfactual, or prediction error, is the effect of the intervention. To account for potential systemic bias in the counterfactual, we adopt a difference-indifferences strategy. We identify the average impact of the lockdown on concentrations as the difference between the average prediction error before and during the lockdown. This approach does not require identifying comparable regions whose concentrations follow a business-as-usual trend.
We first assemble a dataset of air pollution, atmospheric conditions, and calendar variables for the period 2012-2020 for the Italian region of Lombardy. Pollution concentrations are measured at 83 monitoring stations. Data on daily minimum and maximum temperature, average wind speed and wind direction, average relative humidity, daily cumulative precipitation, and atmospheric soundings come from 227 weather stations.
For every monitoring station, we build the counterfactual using an extreme gradient boosting regressor, a tree-based model (Friedman 2001) 4 . Next, monitor by monitor, we train the algorithm on data from 2012 through 2019 and predict concentrations of PM 2.5 and NO 2 in 2020. We use the pre-lockdown period from 1 January to 22 February, which was not included in the training set, to assess the validity of the counterfactual.
As our ultimate goal is a reliable prediction of pollutant concentrations from January through early May 2020, cross-validation is performed over four folds, each one consisting of the months from January to April for 2016-2019. The more common crossvalidation on random subsamples, or folds, gives equal weight to all seasons. However, with such validation strategy it cannot be ruled out that an algorithm make good average predictions, while over-predicting in one season and under-predicting in the opposite one. Suppose, for instance, that the predictions of a learner are positively biased in spring, negatively biased in fall, and unbiased in winter and summer. In this case, testing predictions on the pre-lockdown period (in wintertime) does not give correct estimates of the bias during the lockdown (in springtime). For this reason, we perform cross-validation over the months for which we want predictions to be reliable. Model parameters are selected to maximize the crossvalidated root mean square error (RMSE).
The identification strategy relies on two assumptions. First, input variables should not be themselves affected by the intervention; otherwise, estimated effects will be biased toward zero. To this end, we exploit the sensitivity of concentrations to meteorological conditions and build the counterfactual as a function of weather and season. While emissions are affected by weather (e.g. lower emissions from heating systems on warmer days), our identification assumption is not violated as the weather is not affected by emissions. On the other hand, the 4 We use the python package xgboost (Chen and Guestrin 2016). algorithm implicitly learns the patterns of emissions as the weather varies and seasons pass.
Second, emissions that would have materialized absent the lockdown, and once weather has been accounted for, should be equal to emissions in the training period. One might be concerned that differences in technology (such as upgrading of the vehicle fleet) or economic activity between the training and prediction sample violate this assumption (Silver et al 2020). We address this concern adopting a difference-in-differences strategy that excludes any constant prediction bias from the estimated effects of the lockdown. As long as the variation of observed values around the true counterfactual mean is well reproduced, estimates will be valid. Furthermore, the learner is cross-validated on data from 2016 through 2019; thus, recent years are given more weight.
We estimate the average effect of the lockdown with the following equation: where y it is concentration measured at monitor i on day t, y it is the predicted value, and Lockdown is a dummy equal to 1 during the lockdown and 0 prior to it. α captures any time-invariant bias of the predictor; β is the parameter of interest; and ε it is a random term. The preferred specification then distinguishes treatment effects by type of monitoring station 5 . Since concentrations are consequential to the extent that they reflect exposure, we weight observations by population within 20 km from monitors 6 . We leave estimates of unweighted regressions, which yield qualitatively similar results, to the supplementary information. To our knowledge, there is little guidance in the literature on how to estimate standard errors in this context properly. Thus, where reasonable, we cluster standard errors by monitor; where the number of clusters is small, we use robust standard errors.

Data sources
We assemble a dataset of air pollution, atmospheric conditions and calendar variables for the period 2012 to 2020 for the Italian region of Lombardy. The region is the home to about 10 million people and is the first contributor to national GDP by size. Its natural geography is conducive to low winds and stable air masses throughout the cold season. Mountain ranges to the North, West and South effectively block transboundary air streams extending wintertime thermal inversions and aggravating pollution events. For exceeding recommended air quality thresholds, Italy has been fined and subject to infringement procedures by the 5 Namely background, industrial, and traffic monitoring stations. 6 Territory within 20 km of two or more monitors is assigned to the closest monitor. The construction of population weights is described in more detail in supplementary information.
European Commission. We describe the data sources and pollution trends in Lombardy.

Air pollution
Data for air pollution is collected, checked, and published by ARPA Lombardia, the regional environmental agency 7 . We obtain readings for NO 2 and total PM 2.5 for background, traffic, and industrial stations as available. Hourly readings are averaged to daily readings. We exclude all monitoring stations that are not functioning during the lockdown or have been set up after 2015. Background stations account for about 60% of pollution monitors, traffic stations for about 30%, and the remaining 10% is located in industrial areas.
Average yearly concentrations of PM 2.5 in Milan, the region's capital, are systematically above the safety levels established by the WHO (10 µg m −3 ); from December to the end of February, daily concentrations average above 40 µg m −3 . Average levels of NO 2 during the period are also well above WHO safety standards.

Weather data
Data on weather conditions at weather stations throughout the region are also elaborated and made available by ARPA Lombardia. We retrieve the daily minimum and maximum temperature; average wind speed and wind direction; average relative humidity; and daily cumulative precipitation. We further include a host of atmospheric sounding indices measured at Milano Linate airport and made available by the University of Wyoming, namely Showalter index, Lifted index, SWEAT index, K index, and Cross Totals, and Vertical Totals indices. All atmospheric variables enter as predictors in the form of contemporaneous and lagged values. Although monitor data and atmospheric soundings have gone through quality checks at the source, we winsorize all atmospheric predictors at 1 and 99 percentiles to bound the influence of extreme values.

Additional predictors
The ratio of PM 2.5 to PM 10 in Lombardy is typically altered in presence of pollution transported from long distances. For instance, a mass of dust from the Caspian Sea reached Northern Italy in late March, substantially altering the ratio. We assume the PM 2.5 to PM 10 ratio is independent of the lockdown and include it among predictors as the concentration of PM 2.5 is affected by such shocks. Additional predictors are calendar variables to capture trends over time and seasons. We include year, month, week of the year, day of the month, day of the week in the form of continuous variables as well as dummy variables. 7 Both air pollution and weather data are publicly available at www.dati.lombardia.it/stories/s/auv9-c2sj.
We further include sine functions of time to mimic seasonality.

Population weights
Population weights for monitoring stations reflect the population within 20 km of monitors (figure D.4 in supplementary information). Population data on a 1 km by 1 km grid comes from the Italian National Statistical Office (ISTAT) 8 . Grid cells within less than 20 km from two or more monitors are assigned to the closest one.

Health impact assessment
To compute the number of avoided deaths and years of life saved by the reduction in PM2.5, we follow Fowlie et al (2019) and take all-cause mortality relative risk (RR) ratios for PM 2.5 from two influential studies, Krewski et al (2009) and Lepeule et al (2012). In addition, we use the RR ratio recommended by the WHO (Henschel et al 2013) and adopted by the European Environment Agency (European Environment Agency 2019). For NO 2 , we only use the WHO recommendations. The calculation of avoided deaths and years of life saved from concentrationresponse functions is described in supplementary information A.
The more conservative estimates are based on Krewski et al (2009), who report an hazard ratio 1.056 for an increase of 10 µg m −3 of PM 2.5 . Lepeule et al (2012) estimate instead a larger hazard ratio of 1.14 for the same change in concentrations. The WHO recommends estimating the long-term impact of exposure to PM 2.5 in adult populations using an RR of 1.062 for 10 µg m −3 ; it recommends an RR of 1.055 for 10 µg m −3 of NO 2 above 20 µg m −3 in adult populations.

Accuracy of predictions
To assess the accuracy of predictions, we test the counterfactual against observed values during the pre-lockdown period from 1 January to 22 February, which has not been used for training. Table 1 reports mean values of Pearson's correlation coefficient (Corr), mean bias (MB), normalized mean bias (nMB), and RMSE. As we ultimately compute the difference-in-differences between observed values and the counterfactual, we also report the centered RMSE (cRMSE) and the normalized centered RMSE (ncRMSE) 9 . For completeness, the table also includes statistics for the training set. The correlation between observed and predicted values in the pre-lockdown period is 0.87 and 0.88 for PM 2.5 and NO 2 , respectively. The counterfactual overestimates observed values by 1.34 µg m −3 (PM 2.5 ) and 4.7 µg m −3 (NO 2 ), thus motivating the use of a difference-in-differences strategy. The cRMSE is 30% (PM 2.5 ) and 27% (NO 2 ) of mean observed concentrations. A graphical summary of model predictive performance, Taylor diagrams, can be found in supplementary information.
In air pollution forecasting, machine learning techniques are typically used to predict concentrations an hour to few days ahead, and studies that can be used as benchmark are scarce. To the best of our knowledge, Petetin et al (2020) is the only work whose methodology and length of forecast are comparable. They use machine learning to build a counterfactual for NO 2 concentrations in Spain during the COVID-19 lockdown. They report a nMB of 2%-7%, depending on the type of station, a correlation coefficient of 0.71-0.75, and normalized RMSE of 28%-32%. Compared to their study, our algorithm better mimics variation around the mean, than the mean itself. However, in our estimation strategy, any constant bias is captured by the constant in equation (1).

Effect of the lockdown on air pollution
Following the lockdown, air quality in Lombardy improved only partially. Figure 2 plots the population-weighted observed and counterfactual values for PM 2.5 (figure 2(a)) and NO 2 ( figure 2(b)). NO 2 at background stations reached levels below the yearly limit set by the WHO Air Quality Guidelines. However, background concentrations of PM 2.5 still exceeded the daily limit of 25 µg m −3 every one in four days.
The counterfactual well mimics observed values in the pre-lockdown period, corroborating the validity of the statistical approach. In contrast, a gap between observed and counterfactual values is evident as restrictions are tightened. We show in supplementary information that the method outperforms a linear regression.
Suggestive evidence of the effect of the lockdown on concentrations of NO 2 , which in Lombardy largely originate from motor vehicles, is visible from the week of 25 February, consistent with the reduction in mobility documented in figure 1(a). The effect on PM 2.5 only appears as non-essential economic activities are halted in Lombardy and the rest of Italy, and is smaller in magnitude.
The lockdown may have affected PM 2.5 concentrations mainly through two channels: the reduction of primary PM 2.5 emissions, such as black and organic carbon, and reduction of precursors of secondary PM 2.5 . We remark that NO 2 is a precursor of secondary PM 2.5 ; a reduction in NO 2 may, therefore, lead to a decline in PM 2.5 . However, as data on PM 2.5 composition is insufficient, we cannot quantify the contribution of NO 2 to the reductions in PM 2.5 concentrations 10 . Therefore we treat both pollutants independently.
We estimate a population-weighted version of equation (1) in Methods and report results in table 2. Results of unweighted regressions are qualitatively similar and can be found in supplementary information. From 22 February to 4 May, the lockdown has on average reduced daily concentrations of PM 2.5 and NO 2 by 5.32 and 13.56 µg m −3 . That is a reduction of 21.8% and 35.6%, respectively, from the average levels that would have been observed had not the epidemic broken out.
Next, our preferred specification distinguishes effects of the lockdown by type of monitor. Background monitors are located where concentrations are representative of the ambient exposure of the general population; industrial monitors are located in the proximity of industrial sites or industrial sources; traffic monitors are located near a major road.
Population-weighted average background concentrations of PM 2.5 decreased by 3.84 µg m −3 from 24.42 µg m −3 (table 3) 11 . The reduction was almost twice as large in monitored industrial sites and near major roads. Background concentrations of NO 2 dropped by 10.85 µg m −3 from 33.22 µg m −3 , by 10.66 µg m −3 near monitored industrial sites and by 15.85 µg m −3 more at major roads. 10 At the time of writing, data on composition of PM2.5 has not been released. Data on composition of PM10 is available only for three monitoring station. 11 The very low number of monitors by type makes clustered standard errors inappropriate. We thus use robust standard errors.  Notes: Regression weighted by population within 20 km of a monitoring station. Territory within less than 20 km from two or more monitors is assigned to the closest one. The dependent variable is the difference between the observed values and the counterfactual. Lockdown is a dummy variable equal to 0 from 1 January 2020 to 22 February, and equal to 1 after 22 February 2020. Average baseline concentration is the population-weighted average of counterfactual values during the lockdown, less the constant in case the latter is statistically significant at 10%. Standard errors, in brackets, are clustered by monitor. * p < 0.1, * * p < 0.05, * * * p < 0.01.

Human health benefits
As the reduction in road transport and the slowing of economic activity reduced toxic emissions, the burden of pollutants on human health eased. For calculations, we use the estimated change in concentrations at background stations. Avoided deaths and YLS should be considered a lower-bound estimate of total health benefits avoided deaths. The reduction in PM 2.5 prevented 10.2-24.8 premature deaths per 100 000 individuals and saved 72.1-175.9 years of life per 100 000 individuals, depending on the concentration-response function (table 4). The reduction in NO 2 prevented 28.8 premature deaths and saved 203.7 years of life per 100 000 individuals. Given the high correlation between concentrations of PM 2.5 and NO 2 , the concentration-response function of these pollutants are interdependent. It is recommended that avoided deaths and YLS be not aggregated across pollutants, lest incurring in partial double counting.
As a comparison, in Italy in 2016 for every 100 000 individuals, there have been 96.6 premature deaths attributable to PM2.5 and 24.1 attributable to NO 2 , or 23.8 and 5.9 premature deaths in three months, respectively (European Environment Agency 2019). Since most of the premature deaths happen in the more polluted North of Italy, including Lombardy, the lockdown has temporarily reduced the cost of pollution by a substantial amount.
We compare the results against the number of deaths and the YLL related to COVID-19 in Lombardy during the same period, computed from Notes: Regression weighted by population within 20 km of a monitoring station. Territory within less than 20 km from two or more monitors is assigned to the closest one. The dependent variable is the difference between the observed values and the counterfactual.
Lockdown is a dummy variable equal to 0 from 1 January 2020 to 22 February, and equal to 1 after 22 February 2020. Average baseline concentration is the population-weighted average of counterfactual values during the lockdown, less the constant in case the latter is statistically significant at 10%. Robust standard errors are in brackets. * p < 0.1, * * p < 0.05, * * * p < 0.01. In Lombardy, from 22 February to 3 May 2020, every 100 000 people 155 died after testing positive for COVID-19 and 1891 years of life have been directly lost to the virus. The hazard ratio is the ratio of two concentration-response functions, or hazard rates, between a high and a low concentration differing by 10 µg m −3 . Avoided premature deaths are calculated using the population-weighted change in concentrations at background stations. patient-level data 12 . In Lombardy, from 22 February to 3 May 2020, every 100 000 people 155 died after testing positive for COVID-19 and 1891 years of life have been directly lost to the virus. Avoided deaths from the reduction in PM 2.5 are 6.5%-16% of COVID-19 deaths; YLS are 3.8%-9.3% of YLL to COVID-19. Avoided deaths from the reduction in NO 2 are 18.6% of COVID-19 deaths; YLS are 10.8% of YLL to COVID-19.

Conclusions
The dramatic reduction in emissions of airborne pollutants that has come with the response to COVID-19 provides a unique natural experiment to assess the sensitivity of pollutants concentrations and health to emissions. We estimate a substantial yet partial improvement in air quality in Lombardy following the outbreak, and suggest that the improvement originates primarily from the reduction 12 Data on the individual COVID-19 patients has been shared by regional health officers under an institutional agreement. of road transport; and to a lesser degree from the reduction in industrial activity. Important sources of emissions as heating systems and agriculture have not been substantially affected by the outbreak. The methodology used to build the counterfactual does not require identifying comparable but unaffected regions, but relies on the assumption of emissions absent the lockdown following historical variation around the mean. The approach is not limited to this case study, but can be applied in a variety of settings due to the increasing and reliable availability of pollution and weather data.
Finally, we are nowhere near suggesting the pandemic has been beneficial for the affected communities, yet the health benefits from improved air quality are noticeable. While global pandemics are rare phenomena, exposure to unhealthy levels of toxic air pollutants is the rule, including in affluent regions of the world such as the one considered here. This paper has emphasized some of the health benefits of cleaner air, but also highlighted the variety of emissions sources and the need for a broader policy response to solve Europe's biggest environmental health risk.