key: cord-0761339-sisnr09k
authors: Becchetti, Leonardo; Beccari, Gabriele; Conzo, Gianluigi; Conzo, Pierluigi; Santis, Davide De; Salustri, Francesco
title: Particulate matter and COVID-19 excess deaths: Decomposing long-term exposure and short-term effects
date: 2022-01-07
journal: Ecol Econ
DOI: 10.1016/j.ecolecon.2022.107340
sha: 92ea6dad9e4b9123d2ee310d436b78c8b8be2b5b
doc_id: 761339
cord_uid: sisnr09k

We investigate the time-varying effect of particulate matter (PM) on COVID-19 deaths in Italian municipalities. We find that the lagged moving averages of PM2.5 and PM10 are significantly related to higher excess deceases during the first wave (end February-end May) of the disease, after controlling, among other factors, for time-varying mobility, regional and municipality fixed effects, the nonlinear contagion trend, and lockdown effects. Our findings are confirmed after accounting for potential endogeneity, heterogeneous pandemic dynamics, and spatial correlation through pooled and fixed-effect instrumental variable estimates using municipal and provincial data. In addition, we decompose the overall PM effect and find evidence both that pre-COVID long-term exposure and short-term variation during the pandemic matter. In terms of magnitude, we observe that a 1 μg/m3 increase in PM2.5 leads to 20% more deaths in Italian municipalities, which is equivalent to a 5.9% increase in mortality rate.

The impact of the COVID-19 pandemic in relation to contagion and deaths in the first half of 2020 is markedly heterogeneous from a geographical perspective. Several studies have tried to identify the causal factors of this puzzling outcome. Epidemiological literature have identified the frequency of human physical interactions as a leading causal factor of contagions. However, even after controlling for them, a significant part of the observed variability of COVID-19-related outcomes remains unexplained. This paper aims to shed light on this issue by investigating the role of particulate matter (PM) in the pandemic's high mortality rate.

The theoretical background for our research hypothesis can be summarized into two main literature strands. The first strand deems long-term exposure to PM as a contributing factor to COVID-19-related deaths. The research hypothesis relies on the maintained assumption that PM inhalation induces inflammation and oxidative stress, thereby reducing lung efficiency and contributing to respiratory and pulmonary diseases (see Pope and Dockery, 2006) . Over the years, several empirical papers have estimated the relationship between long-term exposure to PM and total mortality, and mortality from cardiovascular and respiratory diseases (Kim et al., 2015; Pelucchi et al., 2009; Pinault et al., 2017; Faustini et al., 2011; Anderson, 2020; Ciencewicki and Jaspers, 2007; Sedlmaier et al., 2009) and the effect of fine PM as a factor in cardiovascular and respiratory morbidity and mortality (McGuinn et al., 2017; Jeong et al., 2017; Yin et al., 2017; Cakmak et al., 2018) .

Based on this literature strand, several researchers have tried to ascertain if the effect of COVID-19 on respiratory and pulmonary diseases can be enhanced by PM exposure. Wu et al. (2020) found that an increase in exposure to PM 2.5 is associated with increased COVID-19 fatality in the US, Cole et al. (2020) find similar results for the Netherlands. 1 Focusing on Italy, Cartenì et al. (2020) find that PM is significant after measuring the number of days in 2019 in which the national PM 10 exceeded the 50 μg/m 3 daily limit, Perone (2020) find that the case fatality rate is affected by ozone and nitrogen dioxide beyond PM, while Becchetti et al. (2020a) used mortality data at province level to confirm the positive relationship 7 To test the impact of particulate matter controlling for potential concurring factors, we estimate the following equation:

Excess Deaths tm = ß 0 + ß 1 Pollution(MA) tm + ß 2 t +ß 3 t 2 + ß 4 t 3 + ß 5 Days_Since_Lockdown t + + ß 6 Population m + ß 7 Density m + ß 8 Over65 m + ß 9 Income m + ß 10 Employees m + + ß 11 Essential_Employees m + ß 12 Temperature(MA) tm + ß 13 Mobility tp + Σ r γ r DRegion rm + u tm (1) where our dependent variable (Excess Deaths tm ) is the difference between total deaths in 20202 in municipality m on day t and the 2015-19 total average deaths in the corresponding municipality and day of the year. The main independent variable of interest is Pollution(MA), calculated as a moving average from day t -10 to day t of PM 10 or PM 2.5 , measured in municipality m on day t. 5 We introduce linear, quadratic, and cubic time trends (t, t 2 , t 3 ) starting with the disease outbreak, which is conventionally fixed as February 24, 2020 (the beginning of our sample period) among control variables. These trends capture part of the deterministic evolution of the pandemic consistently with standard epidemiological modeling approaches (further robustness checks for heterogeneous pandemic dynamics are presented and discussed in section 6). 6 Among other controls, Days_Since_Lockdown counts the days since the national lockdown, taking into account the three government decisions that progressively introduced mobility restrictions in Italian municipalities. 7 Population is the number of residents in municipality m from the last Italian census (2011) (per 1,000 inhabitants); Density is the population density in municipality m (per 1,000 inhabitants); Over65 is the proportion of people aged 65 or above and living in municipality m (per 1,000 inhabitants); Income is the total before-tax income in municipality m (in billion euros); Employees and Essential-Employees are the number of employees operating in all sectors and in essential sectors only (per 1,000 inhabitants) at 8 the municipal level. The essential sectors are those on a list of activities that the Italian government allowed to operate during the lockdown. 8 These last two variables capture lockdown-induced local differences in job commuting flows due to the different incidences of essential and nonessential sectors in each municipality. Temperature(MA) is the 11-day (from t -10 to t) moving average of daily air temperature in each municipality. Last, we control for time-varying human interactions with a variable (Mobility) measuring transit in the subways, bus and train stations, seaorts, taxi stands, highway rest stops, and car rental agencies in Italian provinces. The variable is calculated in first differences, that is, as a change in the number of people in the above-mentioned transit areas compared to the baseline of the median value, for the corresponding day of the week, during the previous 5-week period. We also add region dummies (DRegion) to control for time-invariant features of the regions, such as urbanization rate or health system characteristics. In fact, health policies in Italy are run at the regional level, thereby making health capital endowments highly heterogeneous across regions. 9 Standard errors are clustered at the municipal level. A detailed description of variables and their sources is given in Table A1 in the Appendix. Table 1 presents descriptive findings on the variables used in our econometric specifications. As expected, the moving averages used in the estimates smoothen extreme values of pollution and atmospheric indicators, with maxima of moving averages of particulate concentration reaching 45.71 and 61.72 µg/m 3 , respectively. Nonetheless, the mean value of the PM 2.5 moving average (14.02 µg/m 3 ) during the sample period is above the average yearly threshold suggested by the World Health Organization (10 µg/m 3 ) 10 . Our sample period covered the end of winter and spring, therefore, we did not observe extreme hot temperature events (the single-day maximum is 28.86, while the moving average maximum 24.47). 

Before running our estimates, we perform panel stationarity tests and find that all our series are stationary. More specifically, we perform the Lein-Lin-Chu (2020) test for unit roots in panel datasets and find that the null of non-stationarity is rejected in all cases (p< 0.001 for all series).

In Table 2 , we present the results of our main econometric specification. Columns 1 and 2 display (unweighted) pooled OLS estimates of the effects of PM 2.5 and PM 10 , respectively. In columns 3 and 4, observations are weighted for the inverse of the distance from the centroid to give more importance to municipality centroids that lie closer to the geographical point of our meteorological observation of PM.

Our empirical findings show that the high mortality in 2020 is significantly and positively related to both air pollution measures. In terms of magnitude, the effect of PM 2.5 is larger than that of PM 10 , with results from weighted and unweighted estimates for the same pollution variable being quite similar. The estimated pollution effect in column 3 implies that of additional PM 2.5 concentration creates an approximately 10 percent increase in the average value of the dependent variable, that is 0.113 extra deaths per day per 100,000 inhabitants, which corresponds to a 3.32 percent increase in mortality rate. The total effect over the 94 days of the pandemic considered in our sample is 1.07 extra excess deaths per 100,000 inhabitants. This implies that the effect over the entire Italian population is about 647.96 extra deaths per . Based on our coefficient magnitude, we estimate that a difference of about 19 between the municipalities with the highest and lowest PM 2.5 average concentration in the sample would generate a difference of 1,231.13 more deaths in the overall sample period.

The linear, quadratic, and cubic trends are strongly significant among the control variables and with the expected sign, displaying non-linear pandemic dynamics during the first phase. The share of employees in essential sectors is positive and significant and likely to capture the positive effect of high death rates on economic activity in industries that could not stop their operations during the lockdown. Time-varying mobility is, as expected, positive and significant as an increase in people in transit stations has a positive and significant effect on excess deaths. The negative sign of the density variable can be explained by the fact that, with population among the regressors, the variable captures the positive effect of municipality surface on excess deaths, likely to be explained by how far inhabitants are from institutions and less accessible health services. The positive and significant effect on excess deaths of the share of the elder population at the municipal level is also expected.

We implement an instrumental variable approach from omitted variables and reverse causality to mitigate a possible estimation bias deriving from measurement error in the dependent variable (Table 3) . We instrument the PM moving averages in eq. 1 with the four-day lagged corresponding 11-day moving average of daily rainfalls controlling in our estimates for the mobility variable. We can confidently argue that the chosen instrument is relevant since rainfalls have a strongly significant and negative effect on PM concentrations in the first-stage estimation. 11 The exclusion restriction is also likely to be satisfied in our case since-apart from its direct effects on pollution-it is implausible that four-day lagged rainfall moving averages significantly affect the difference in deaths between 2020 and the previous years on a given day. Rainfall may discourage mobility and reduce contagion or increase car vs. public-transport mobility (again reducing contagion), thereby potentially invalidating the exclusion restriction. However, these potential threats to the exclusion restriction can be excluded since we condition for mobility in all estimates. Furthermore, most of the mobility decisions made during the lockdown that cover most of our sample period were forced, with little impact on atmospheric conditions. The pairwise correlation between rain and mobility during lockdown is 0.03, which supports our hypothesis. This is a positive (yet low in magnitude) and non-statistically significant correlation, which goes against the prediction of a potential negative association between the two variables. Furthermore, the instrument is not statistically significant if we introduce it in our baseline noninstrumented specification (eq. 1), further supporting its validity hypothesis. Our main findings remain unchanged and the coefficient magnitude is remarkably close to the non-instrumented estimates.

In Table 4 , we re-estimate our benchmark specification through OLS panel fixed effects. This allows for capturing unobservable time invariant idiosyncratic factors at the finest geographical unit, e.g. the quality of local majors or local health governance at the municipality level. The significance of the PM 2.5 and PM 10 variables is also confirmed in this model. In Table 5 , we instrument the PM moving average as the fixed-effect estimates with the instrument used in Table 3 . Our results are again confirmed. (1) and (2) do not weight observations, while columns (3) and (4) 

PM 2.5 0.00123*** 0.00145*** (7.60e-05) (0.000144) PM 10 0.000644*** 0.000713*** (3.80e-05) (5.87e-05) T (linear day trend) 0.00295*** 0.00298*** 0.00302*** 0.00307*** (0.000131) (0.000131) (0.000195) (0.000196) T 2 (quadratic day trend) -5.59e-05*** -7.10e-05*** -5.33e-05*** -7.35e-05*** (5.52e-06) (5.28e-06) (8.85e-06) (9.41e-06) T 3 (Cubic day trend) 3.54e-07*** 4.28e-07*** 3.33e-07*** 4.34e-07*** (3.17e-08) (3.07e-08) (5.01e-08) (5. J o u r n a l P r e -p r o o f 

The 11-day moving average used so far mainly captures the time-varying effect of PM on excess deaths. However, it is reasonable to assume that this measure is also influenced by a long-term component capturing long-term, pre-COVID exposure to PM. This component is regarded as the crucial factor affecting the negative consequences of COVID-19 infection, according to the first strand of the literature described in the introduction.

To disentangle the effects deriving from these two-long-term structural and short-term time -varying-components, we propose the following decomposition. First, since the time-varying component can be correlated with historical levels of PM concentration, we regress the PM 11-day moving average, that is, Pollution(MA), on (time-invariant) average PM concentration in the two years before the pandemic, i.e., PM(2018 PM( -2019 .

More specifically, we estimate the following model:

Then, we compute the time-varying residuals ̂ , which can be interpreted as a "cleaner" measure of the time-varying effect, i.e., the variation Pollution(MA) tm that is not explained by the variation in the long-term PM component.

We, therefore, run our benchmark model as in eq. 1 by replacing Pollution(MA) with its time-varying residual component ̂ , and the two-year (time-invariant) average of PM concentration. The estimating model reads as:

Excess Deaths tm = ß 0 + ß 1 ̂ + ß 2' PM(2017-2018 ) m + ß 3 t +ß 4 t 2 + ß 5 t 3 + +ß 6 Days_Since_Lockdown t + ß 7 Population m + ß 8 Density m + ß 9 Over65 m + ß 10 Income m + +ß 11 Employees m + ß 12 Essential_Employees m + ß 13 Temperature(MA) tm + ß 14 Mobility tp + + Σ r γ r DRegion rm + u tm

The results from the OLS pooled estimates of eq. 3 show that the coefficients of both components (ß 1 and ß 2 ) are positive and statistically significant (columns 1 and 2 of Tables 6 -7 for PM 2.5 and PM 10 , respectively). Our interpretation is that long-term exposure and time-varying effect significantly predict excess mortality.

We also re-estimate eq. 3 through an OLS fixed-effects model. The results are in columns 3 and 4 of Tables 6-7 for PM 2.5 and PM 10 , respectively. Given the nature of this regression model, the effect of pre-COVID time-invariant exposure to PM is now absorbed by municipality fixed effects. The rationale of this last estimate is to check whether the time-varying PM component is statistically significant when local unobserved timeinvariant characteristics are accounted for. Our findings confirm that this is the case. J o u r n a l P r e -p r o o f Columns (1) and (2) pooled estimates, columns (3) and (4) 

The first robustness check we perform features the use of an alternative instrument calculated as the residual from the following regression:

Pollution(MA) tm = γ 0 + γ 1 Mobility tp + γ 2 Excess Deaths tm t + γ 3 Rain(MA) + η tm

The residual η tm is, by construction, exogenous when used as an instrument in our benchmark estimate in eq. (1). The advantage of this instrument is that through eq. (4), we control for the complex pattern of relationships through which rain and mobility can affect the relationship between pollution and excess deaths.

The new IV findings confirm that this instrument is also relevant since first-stage regression coefficients are significant. Moreover, the falsification exercise of introducing the instrument in non-instrumented estimates confirms that the former has no significant direct impact on the dependent variable. In terms of magnitude, we note, however, that the coefficient size of the instrumented variable is much higher in the new IV estimates than in the non-IV ones.

We also test whether the short term effect estimated in our decomposition exercise presented in Tables 6 and 7 remains significant when instrumented under our two different IV approaches. We find this to be the case (Panel 8.5, columns 1 and 2).

There are two additional potential concerns in our estimates: (i) heterogeneity of the pandemic dynamics at the municipal level; and (ii) spatial dependence of the pandemic.

With regard to the first concern, we take two approaches. First, we estimate the Pesaran (1995) mean group estimator model where slope coefficients are separately calculated for each municipality and averaged across all municipalities. Our main variables of interest remain strongly significant. However, this approach corrects more for heterogeneity of PM impact than of the virus spread net of the PM effect. We, therefore, estimate this model with a mean group estimator specification allowing for province-specific trends. Again, our main results are unchanged (Table  8 , panel 8.5, column 4). Second, we test whether our findings are confirmed when data are aggregated at the province level as the problem of heterogeneous infection dynamics is particularly severe at the municipal level, but less so at the province level. Our main findings are confirmed in non-instrumented and instrumented specifications with province-level data (Table 9 ). Finally, we check for the contemporaneous presence of PM J o u r n a l P r e -p r o o f 19 between and within effects to test whether particulate matter has an impact through both effects at the municipal level. This is another way to address the heterogeneity of pandemic dynamics problem since PM between-effects cannot be affected by such a problem. To this purpose, we estimate hybrid models that split the effect of particulate matter into within-and between-municipality effects (Schunck 2013; Schunck and Perales 2017) using a Mundlak (1978) random-effects approach. The estimated findings show that both between and a within municipality variation in PM 2.5 and PM 10 significantly matter in explaining variation in excess deaths. The within-effect, however, has more power since it accounts for three-fourth of the overall effect in the decomposition estimated in the hybrid model (Table 8 , panel 8.5, column 3). Note that this decomposition allows us to disentangle contemporary between and within effects; this is a different approach from that proposed in section 4 eq. 3, where the between effect is long term, lagged, and aims to capture previous long-term exposure to particulate matter.

For the second concern, that is, spatial correlation, we run a spatial Durbin model for our panel with the province level data following the approach proposed by Belotti et al. (2017) .

Furthermore, to account for other possible endogeneity issues, we build a spatial panel IV model. First, we run the fixed-effects quasi-maximum likelihood estimator on the endogenous regressor against both the instruments and the exogenous covariates of the main model. Then, after getting the control function, ie., the prediction of the overall error component from this regression, we run the full spatial model again, controlling for this component. This allows us to further mitigate the remaining endogeneity of the PM variables (Table 9 , panel 9.5, columns 1-3).

To check whether our findings are robust to a more flexible control for the aggregate pandemic dynamics that do not assume any particular functional form, we repeat our estimates by introducing day fixed effects (Table 8 , panels 8.1-8.4). Our main results remain significant and the coefficient magnitude do not vary significantly.

In an additional robustness check, we calculate COVID-19 non-synchronous regional trends by assuming independent regional pandemic dynamics. To this purpose, we set the regional contagions at n=100 and use this conventional number as the starting point of the pandemic trends in each region. This approach allows us to account for unobserved time-varying region-level characteristics. Our main findings do not change after attributing a specific regional trend to each municipal (Table 10 , panels 3-5,8-10, 13 and 16). 12 We further refine our main instrument by ruling out episodes of extreme rainfalls from the sample. More specifically, we eliminate observations where the instrument (rainfall moving average) is above the 95 th centile and can be suspected to directly affect excess deaths (Table 10, panels 2,4-5,7,9-10, 12 and 15).

12 Table 10 presents a summary of the empirical findings from these final robustness checks. Full findings are in the online Appendix.

To test whether our findings are robust for "super-spreader" events during the pandemic, we consider the UEFA Champions League match between Atalanta and Valencia that took place February 19, 2020, when around 40,000 Atalanta supporters gathered in the San Siro stadium in Milan for the match 13 . To do so, we repeat our estimates by removing data for the Bergamo and Milano provinces. Again, the results remain unchanged in terms of magnitude (the pooled estimate coefficient changes only at the fifth decimal digit) and statistical significance (Table 10, panels 1,4-5,6,9-10, 11 and 14) .

To compare the magnitude of our results with those of the existing literature, we calculate what our coefficients imply in terms of the impact of 1 µg/m 3 of PM on mortality. For the magnitude of the PM effects, the estimated PM coefficients vary between different estimates that look at different sources of variability. However, presenting all of them at least as a robustness check is important to evaluate the robustness and extension of the significance of our findings. For example, the fixed-effect coefficient compared with its pooled estimated counterpart captures only the within-effect controlling for unobserved time-invariant municipality effects. The IV effect depends in turn on the quality of the instrument and corrects for endogeneity problems.

Based on all our different estimates, we conclude that the overall non-instrumented PM 2.5 effect can be reasonably estimated in a range between 0.001 and 0.002. The highest coefficient is that of the fixed effect estimates augmented for day fixed effects. The same numbers for the PM 10 effect are between 0.0006 and 0.001, also when considering estimates of provincial data in Table 8 and robustness checks in Table 9 . Given the average daily mortality rate in Italy in the last four years, the effect implies that one additional µg/m 3 of PM 2.5 is associated with an increase in mortality rate by 2.9 to 5.59 percent. This effect is in the range of findings made in other studies, slightly above that estimated in the Netherland (Cole et al. 2020 ) and below that obtained in the US and Northern Italy (Cocker et al. 2020) (see introduction).

Note that the severe lockdown measures adopted at the beginning of March 2020 significantly contributed to air quality. The lockdown, therefore, reduced the short term effect of PM on high mortality rate during the pandemic. To understand to what extent this occurred, we calculated the difference between the average daily PM concentration during the pandemic's first wave (February 2020 to May 2020) and during the corresponding days in the previous two years (2018-2019 average). If we limit our analysis to the Northern regions, 14 the difference is above 1µg. Hence, if we can interpret estimates in section 4 as causal, we may conclude that lockdown measures saved between 1-2 extra deaths per 100,000 inhabitants.

Extracting the PM concentration differential using fixed-effects estimates do not change the significance and magnitude, thereby confirming our previous analysis.

This study has a number of caveats and limitations. First, our instruments are relevant, but we cannot test their validity and have to prove that on logical grounds. However, the instruments' lack of significance when included as explanatory variables in the main specification, the control for time-varying mobility, the robustness in the sensitivity analysis when excluding extreme rainfalls supports our exclusion restriction.

Second, data at the municipal level are rarely available and, when available, come from the last census in 2011. Consequently, while our analysis controls for many variables like the share of population aged above 65 and the number of employees, we cannot control for other possible factors influencing COVID-19 contagion. For instance, we cannot control for the number of doctors in a given municipality. However, this characteristic is likely captured by municipality fixed effects at the finest level of geographical disaggregation. In our robustness checks described in section 5, we account for municipal heterogeneous pandemic dynamics looking at between effects, using province trends in mean group estimators, and aggregating data at the province level. Note that the two above-mentioned and all other unobserved components were invariant during our 3month sample period but updated in time to the 2011 census. Third, similarly to other papers (see Cocker et al. 2020), our dependent variable measures total deaths and does not discriminate between COVID-19 deaths and deaths caused by other diseases. This is because of the heterogeneity of COVID-19-related deaths registration, both over time and across regions, that we have explained when motivating the choice of our dependent variable.

A final consideration relates to the interpretation of our findings on the decomposition between the two-year average and the time-varying component of PM. The first fixed component captures the ex-ante long-term exposure effect, while the second the effect of changes in PM during the pandemic. We do not explicitly and exclusively identify this last component in the "short term effect." Therefore, it can be questioned whether the time-varying effect derives from the PM capacity to increase survival outside the human body (short term effect) or it may further weaken the capacity of lungs and alveoli to resist respiratory and pulmonary diseases on top of the long-term exposure. While further research could clarify this point, this paper is the first, to the best of our knowledge, to show that historical pre-COVID and contemporary time-varying effects matter.

The two research hypotheses in the literature on the potential impact of particulate matter on COVID-19 contagion and deaths argue that prolonged pre-COVID exposure and contemporary levels of particulate matter can play a positive and significant role.

To test these two hypotheses, we evaluate the impact of particulate matter concentration in Italian municipalities on daily deaths between the first COVID-19 outbreak in Italy and the previous five years. The specific contribution of this study to the literature hinges on the use of the geographically finest controls for concurring factors through municipality fixed-effects, instrumental variable estimates to tackle endogeneity issues within a model taking spatial correlation into account, and the decomposition of the two effects, that is, long-term pre-COVID exposure and time-varying effect.

Our findings show that the impact of both components is positive and significant. Our estimates control for standard time-trend components accounting for the non-linear deterministic evolution of the pandemic, the effects of lockdown measures, and several other controls, such as time-varying mobility. The results from the province level data accounting for spatial correlation and instrumental variable estimates addressing endogeneity problems are further checks on the robustness of our findings. More specifically, taking an average PM 2.5 effect estimated across all our different models, we find that particulate matter concentration predicts more than 1,231 more deaths if we consider the difference between the municipalities with the highest and lowest average PM 2.5 concentration during the first wave of the pandemic.

Indeed, findings from our instrumental variables are inevitably subject to discussion and limitations. However, if they can be interpreted as causal, our empirical results have relevant policy implications. They highlight an additional and important reason to contrast particulate matter beyond those already known. For example, beyond the impact of atmospheric phenomena, the sources of PM depend on around 90 percent of human choices such as domestic heating systems, mobility, agriculture, and industrial production processes. Therefore, urgent steps should be taken to accelerate the transition to frontier technology, reducing each source's contribution to PM. Second instrument: instrument built as explained in eq. (4) section 5. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 (1) and (2) do not weight observations, while columns (3) and (4) (1) and (2) presents pooled estimates, while columns (3) and (4) fixed effect estimates. 

(1) (2) (3)(4)

☒ The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

☐The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:

J o u r n a l P r e -p r o o f

As the wind blows: The effects of long-term exposure to air pollution on mortality

COVID-19 mortality and contemporaneous air pollution. No. paper2016. International Center for Public Policy

Understanding the heterogeneity of adverse COVID-19 outcomes: the role of poor quality of air and lockdown decisions

Lagged particulate matter, contagions and deaths: the relationship between quality of air and COVID-19 at European level

Spatial panel-data models using Stata

First data analysis about possible COVID-19 virus airborne diffusion due to air particulate matter (PM): the case of Lombardy (Italy)

Associations between long-term PM2. 5 and ozone exposure and mortality in the Canadian Census Health and Environment Cohort (CANCHEC), by spatial synoptic classification zone

How mobility habits influenced the spread of the COVID-19 pandemic: Results from the Italian case study

Clinical progression of patients with covid-19 in shanghai

Air pollution and respiratory viral infection

The effects of air pollution on COVID-19 related mortality in northern Italy

Air pollution exposure and COVID-19

Air pollution and Covid-19: the role of particulate matter in the spread and increase of Covid-19's morbidity and mortality

Particulate Matter and COVID-19 Disease Diffusion in Emilia-Romagna (Italy). Already a Cold Case

Features of 20 133 uk patients in hospital with covid-19 using the isaric who clinical characterisation protocol: prospective observational cohort study

Pandemic Meets Pollution: Poor Air Quality Increases Deaths by COVID-19

A review on the human health impact of airborne particulate matter

Epidermal growth factor receptor (EGFR)-MAPK-nuclear factor (NF)-κB-IL8: A possible mechanism of particulate matter (PM) 2.5-induced lung toxicity

Tools for improving air quality management. A review of Top-down source apportionment techniques and their application in developing countries

Unit root tests in panel data: Asymptotic and finite-sample properties

Early transmission dynamics in wuhan, china, of novel coronavirus infected pneumonia

The relationship between air pollution and COVID-19-related deaths: an application to three French cities

Fine particulate matter and cardiovascular disease: Comparison of assessment methods for long-term exposure

On the pooling of time series and cross section data

Assessing nitrogen dioxide (NO2) levels as a contributing factor to the coronavirus (COVID-19) fatality rate. Science of The Total Environment

Long-term particulate matter exposure and mortality: a review of European epidemiological studies

The determinants of COVID-19 case fatality rate (CFR) in the Italian regions and provinces: an analysis of environmental, demographic, and healthcare factors

Associations between fine particulate matter and mortality in the 2001 Canadian Census Health and Environment Cohort

Within and between estimates in random-effects models: Advantages and drawbacks of correlated random effects and hybrid models

Within-and between-cluster effects in generalized linear mixed models: A discussion of approaches and the xthybrid command

Generation of avian influenza virus (AIV) contaminated fecal fine particulate matter (PM2. 5): Genome and infectivity detection and calculation of immission

Position Paper Relazione circa l'effetto dell'inquinamento da particolato atmosferico e la diffusione di virus nella popolazione

Links between air pollution and COVID-19 in England

Exposure to air pollution and COVID-19 mortality in the United States

Long-term fine particulate matter exposure and nonaccidental and cause-specific mortality in a large national cohort of Chinese men

Association between short-term exposure to air pollution and COVID-19 infection: Evidence from China

Assessing the relationship between surface levels of PM2. 5 and PM10 particulate matter impact on COVID-19 in

0.00271* **

Excluding super-spreader events Excluding extreme rainfalls Using nonsynchronous regional epidemic trends (1) (2) (3) (4) (5) (6)

Excluding super-spreader events Excluding extreme rainfalls Using nonsynchronous regional epidemic trends 

Excluding super-spreader events Excluding extreme rainfalls Using nonsynchronous regional epidemic trends Second instrument: instrument built as explained in eq. (4) section 5. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 J o u r n a l P r e -p r o o f