key: cord-1031617-ey0de2m2
authors: López-Feldman, Alejandro; Heres, David; Marquez-Padilla, Fernanda
title: Air pollution exposure and COVID-19: A look at mortality in Mexico City using individual-level data
date: 2020-11-26
journal: Sci Total Environ
DOI: 10.1016/j.scitotenv.2020.143929
sha: f9d6cfda2e788b4e071a9877939408fabf9cff61
doc_id: 1031617
cord_uid: ey0de2m2

We use individual-level data to estimate the effects of long- and short-term exposure to air pollution (PM2.5) on the probability of dying from COVID-19. To the best of our knowledge, our study is the first to look at this relationship using individual-level data. We find that for Mexico City there is evidence of a positive relationship between pollution and mortality that significantly grows with age and that appears to be mostly driven by long- rather than short-term exposure. By using a rich set of individual- and municipal-level covariates we are able to isolate the effect of exposure to pollution from other crucial factors, thus alleviating endogeneity concerns related to selection. Our results provide yet another reason for the need to implement environmental strategies that will reduce the exposure to air pollution: it is a key element to improve the general population’s health. In addition, and considering that at this moment we do not know when the pandemic will stop or if SARS-CoV-2 will become a recurrent threat, the relationship that we uncovered suggests that financial resources should be allocated to improve medical services in those areas where PM2.5 concentrations tend to be high.

The COVID-19 pandemic has had devastating effects both in terms of human lives and health, and of economic and social costs. Given this challenging context, understanding the factors that might increase mortality risks from COVID-19 is essential so that policy makers can design and implement effective policies to protect individuals with high-risk characteristics and living in contexts that might enhance their vulnerability. While the global pandemic has spread across all regions, the death toll it has brought along has varied enormously between countries, plausibly due to a combination of population characteristics (i.e. age structure, comorbidity prevalence), public policies, and environmental factors. In this study, we focus on the impact of one important environmental factor, air pollution, which has been explored recently in academic studies as a factor contributing to COVID-19 mortality (Conticini et al., 2020; Copat et al., 2020) . We analyze the case of Mexico which, as of October 28 th , ranks fourth in the world in terms of total deaths (>90,000) and is among the 10 countries with the highest number of deaths per 100 thousand inhabitants (>70) (JH-CSSE, 2020). COVID-19 deaths are largely concentrated in the Mexico City Metropolitan Area (MCMA), where air pollution levels, in spite of no longer being at the historically high levels observed during the end of the last century, remain very high.

Polluted air gains access to the body through the respiratory tract but it has systemic effects that can damage many other organs (Schraufnagel et al. 2019) . Therefore, air pollution is linked to many health problems and can lead to premature death in children and adults (Brunekreef and Holgate 2002; Cohen et al. 2017; Burnett et al. 2018) . In particular, air pollutants can cause insecurity"), and percentage of jobs that can be done at home. All the information comes directly from official sources (Coneval, 2017; INEGI, 2015) except the percentage of jobs that can be done from home which was estimated based on information from (INEGI, 2015) following Dingel and Neiman (2020) .

We rely on two sources of information for air pollution, one for long-term exposure and the other for short-term exposure. The average annual concentrations of particulate matter finer than 2.5 micrograms per cubic meter (PM 2.5 µg/m 3 ) were obtained from Hammer et al (2020) .

Their global dataset, estimated from NASA's satellite products combined with ground level observations from the WHO, has been widely used to analyze the impact of air pollution on health outcomes. Wu et al. (2020) and Knittel and Ozaltun (2020) use their data to study the relationship between PM 2.5 exposure and COVID-19 deaths in the US. Our analogous municipality-averaged measure was obtained after averaging 1.1x1.1 km 2 grid-cells (0.01x0.01 degrees) within each municipality each year. The mean exposure between 2000-2018 is shown in Figure 1 ; although all 76 municipalities comprised in the MCMA (which encompasses Mexico City) exceed the WHO's standard of annual 10µg/m 3 , variability in exposure across the territory is comparable to other recent studies focusing in small countries or specific regions within a country (e.g., Cole et al, 2020 and Coker et al 2020) . In the MCMA, the highest exposure is observed in central and northeastern municipalities close to the boundaries between Mexico City and the State of Mexico where most economic activity occurs, and most people inhabit. J o u r n a l P r e -p r o o f Our contemporaneous PM 2.5 exposure variables are estimated using information published by Mexico City's Secretary of Environment from ground-based air quality monitors (SEDEMA, 2020) . Their Automatic Network of Atmospheric Monitoring comprises 34 monitors that report hourly measures of concentrations of different pollutants throughout the MCMA, but that are more densely distributed within Mexico City and bordering municipalities from the State of Mexico. In order to assign more accurately pollution concentrations to each municipality, we restrict our analysis to those municipalities whose centroid is at most 7 kilometers away from a monitoring station. Therefore, given data availability and the geographic location of the air quality monitors, we can measure short-term exposure for 14 out of 16 of the MCMA municipalities in Mexico City using 18 out of the 21 stations that measure PM 2.5 . 2 With the information available, we calculate the weighted average of daily concentrations of PM 2.5 based on the readings from all monitors located within a 7 kilometers radius from each municipality centroid (the average municipality had 2 monitoring stations located within this radius). 3 Weights are given by the inverse of the distance between the municipality centroid and each monitor. Using this information, we calculate alternative versions of our short-term exposure measure that we use as robustness checks.

We also calculate an alternative version of the long-term exposure variable based on groundbased pollution concentrations during 2019. Temporally, this variable could be considered a more accurate measure of the actual exposure of the average inhabitant of a given municipality if there is a concern of people changing residential locations from one municipality to another between 2018 and 2020. Spatially, however, the satellite-based measures are constructed from finer-grid cells within a municipality and thus could be considered to more accurately represent exposure of inhabitants in a given municipality.

It is worth noting that, in addition to the substantial geographic variation in PM 2.5 concentrations within the MCMA that is shown in Figure 1 , we also observe ample variation in the exposure to air pollution across individuals with different characteristics and comorbidities.

In particular, Figure 2 shows variation in PM 2.5 long-term exposure across age, diagnoses for diabetes, hypertension, obesity, and smoking status for individuals living in the MCMA. These characteristics and comorbidities have been found to be important modifiers of the risk of dying from COVID-19; it is therefore particularly important to have variation in exposure to air pollution across them.

As Table 1 shows, the MCMA and the Mexico City sub-sample are very similar in terms of individual characteristics, with slightly less than half of the cases being female, a mean age of 45, and comorbidity rates between 10% and 18%. In terms of pollution exposure and socioeconomic characteristics, the Mexico City sub-sample lives in slightly denser and more polluted areas but has larger access to hospital beds, health care, food security, and telework. Importantly, the death rate among confirmed cases is respectively 10% and 11%. Worth mentioning is that in Mexico the number of total COVID-19 tests performed per capita is very low (Roser et al., 2020) . This implies that the number of COVID-19 deaths is likely to be underestimated. The low level of testing is the result of an explicit decision made by the federal J o u r n a l P r e -p r o o f government to not perform widespread testing. This policy is being applied in the whole country and there is no evidence that suggests that testing levels vary inside the MCMA. 4 Therefore, we argue that underestimation is not correlated to municipal pollution levels and hence should not bias our estimations in any meaningful way. 5 As is the case in many other countries, there are instances where deaths are not immediately attributed to COVID-19 due to lack of testing. These deaths remain classified as suspicious until a committee decides if they should be attributed to COVID-19. Our estimations use the most up-to-date information, which includes deaths that were originally classified as suspicious and later on were attributed to COVID-19. There is no evidence to suggest that in the MCMA the number of deaths classified as suspicious varies with pollution levels.

We specify the following probit model to estimate the effect that pollution has on the probability of dying from COVID-19:

(1)

Where is the standard normal cumulative distribution function, Death im is a dummy variable equal to one if individual i from municipality m who has been infected by COVID-19 dies and zero otherwise, pollution m is a measure of air pollution at municipality m, and X im is a vector of individual-and municipal-level variables that control for potential confounding factors. 6 Based on the literature on air pollution and health, and on the results available for the relationship between exposure to PM 2.5 and COVID-19, we expect  to be positive. In addition to testing this hypothesis, we are interested in estimating the effect that an increase of 1 g/m 3 in PM 2.5 has on the probability that an individual will die after having contracted COVID-19. To do so, we take the derivative of the right-hand side of equation (1) We are also interested in the effects that short-term exposure might have on the probability of dying from COVID-19. We use the air pollution data that is available for Mexico City to analyze this. This allows us to estimate a modified version of Equation (1) where, in addition to longterm exposure, we include additional measures of exposure to PM 2.5 . In particular, in addition to our long-term exposure variable, we include a variable that captures the cumulative average concentration of PM 2.5 during the 14 days before the onset of symptoms (PM 2.5 -14).

Additionally, in some of the models we use the variable that measures average concentrations in Mexico City during 2019 (PM 2.5 -2019).

Addressing endogeneity concerns is crucial in order to establish a causal relationship between exposure to pollution and health. In particular, the evidence suggests that environmental 6 In some of the estimations for Mexico City we take advantage of the fact that for that sub-sample we have information at the locality level (a smaller administrative area than a municipality) and hence use pollution concentrations at that level.

J o u r n a l P r e -p r o o f quality is an important dimension for choosing residential location, which is likely to lead to a correlation between higher pollution levels and both observable and unobservable characteristics correlated with health-potentially leading to omitted variable bias.

Additionally, avoidance behavior may similarly lead to nonrandom assignment of pollution exposure (Currie et al., 2014) . In order to partially address these concerns, we include a rich set of municipal-and individual-level covariates that we use to control for factors that may be correlated to pollution exposure and simultaneously affect mortality. In our preferred specification we include population size, population density, hospital beds per capita, poverty measures ("health-service access deficiency" and "food insecurity"), and percentage of jobs that can be done at home, all measured at the municipal level, as covariates. We also include individual indicators for sex, obesity, smoking status, diabetes, hypertension, age, age squared, and the day in which symptoms started to control for time effects. The inclusion of these covariates is likely to alleviate endogeneity concerns-Cole et al (2020) and Wu et al (2020) point out that individual-level data, which they cannot include in their analyses since their COVID-19 data is at the county or municipal levels, is crucial to establish a rigorous statistical link between long-term exposure to pollution and COVID-19. Our standard errors are clustered at the municipality level in order to correct for spatial correlation of the individuals' unobserved variation that may remain after including covariates.

To begin the econometric analysis, we use the full sample for the MCMA and estimate a version of Equation (1) that only includes long-term exposure to PM 2.5 as the explanatory variable. Our J o u r n a l P r e -p r o o f results (Model 1, Table 2) show that there is a positive and statistically significant relationship between exposure and the probability of dying from COVID-19. The relationship remains valid when we include municipal level covariates in Model 2. The last column (Model 3) shows the results of including individual level covariates in our analysis and is our preferred specification.

The positive relationship between long-term exposure to PM 2.5 and the probability of dying from COVID-19 is robust to the inclusion of covariates at both the individual and municipal levels. 7

In order to better interpret our results, we estimate marginal effects for the three versions of the model, which are shown in Figure 3 . For the version without covariates (Model 1, Table 2) we estimate an average marginal effect of 0.0076, which implies that for an increase of 1 g/m 3 in PM 2.5 the probability of dying from COVID-19 increases on average by 0.76 percentage points.

For the model with municipal-level covariates (Model 2) the marginal effect is 0.0129. Our preferred specification is the one that takes into account age, obesity, diabetes, hypertension, smoking, and municipal level covariates, as in Model 3. For that model we estimate that on average the probability of dying from COVID-19 increases by 0.77 percentage points with an increase of 1 g/m 3 in PM 2.5 . Given that the average mortality rate for our sample is of approximately 10.4%, our estimates suggest that an increase of 1 g/m 3 in PM 2.5 may increase mortality risk by approximately 7.4%. Figure 4 presents the average marginal effects of pollution across age. It shows that the effects of long-term exposure on the probability of dying are heterogeneous and increasing with age. 7 The relationship is also robust to the way in which the probability function is modeled; our results remain valid when we estimate a linear probability model instead of a probit model (Table A1 of Table 3 shows the results of using the data from air pollution monitors that is available for Mexico City. With these data we are able to, in addition to our long-term exposure variable, include short-term PM 2.5 cumulative average concentration during the 14 days before the onset of symptoms. Model 1 shows that the coefficient of short-term pollution is significant only at the 10% level. In Model 2, we use PM 2.5 concentrations during 2019 instead of using the longterm variable. We only find weak evidence to support that there is an effect of short-term exposure on the probability of dying (Model 2, Table 3 ). The effect of long-term exposure seems to be partially captured with the variable PM 2.5 -2019, which is weakly statistically significant. This is reassuring; both variables (PM 2.5 -2019 and PM 2.5 ) seem to be in fact capturing the same phenomenon. As a way to test that this is in fact the case, the last column shows the results of estimating the model with the three pollution variables simultaneously (Model 3).

The two long-term exposure variables appear to be substitutes of each other, while once again we only find a slightly statistically significant effect for short-term exposure. 9 As a robustness check, we also estimate all the models presented in Table 3 using cumulative average concentrations during 2, 7, 5 and 10 days instead of 14 days (results are presented in Table A4 of the Appendix). While the results for long-term exposure are similar to those presented in Table 3 , results for short-term exposure are not statistically significant for any of the alternative exposure windows. Finally, as shown in Figure 5 , the marginal effects for changes in long-term PM2.5 exposure estimated for Mexico City are very similar to those estimated for the MCMA (Figure 3 ). Marginal effects for short-term exposure are very small and only weakly significant (90% CI: 0.00006, 0.00099).

Given the magnitude of the health crisis caused by SARS-CoV-2 and the amount of research that it has spurred across many fields, many researchers have looked at the relationship between exposure to particulate matter and morbi-mortality by COVID-19, perhaps unsurprisingly. Many of these studies have analyzed the relationship between exposure and infection and have shown that there is in fact a positive association between high levels of PM 10 and PM 2.5 and confirmed cases (Coccia, 2020; Zhu et al., 2020) . Consistent with results for the SARS 2003 outbreak (Cui et al. 2003) , the research that looks at the effects on mortality has found, as we do, that there is a positive relationship between exposure to PM 2.5 and COVID-19 mortality.

Our results cannot be directly compared to other results for COVID-19 as ours is, to the best of our knowledge, the first study to estimate the effects of air pollution on the probability of dying included the average temperature that the individual experienced in the 14 days before the onset of symptoms as an additional variable in all the models presented in Table 3 . Results for PM 2.5 and PM 2.5 -2019 remain virtually identical, while PM 2.5 -14 ceases to be statistically significant (Table A3 of Using data at the city-level they show, like we do, that the association between short-and longterm exposure and fatality is positive.

Contrary to all these studies that use data at an aggregate level, by using individual-level information we are able to estimate how exposure to PM 2.5 can have a differential effect on the probability of dying according to the age of the infected individual. We show that the effect of PM 2.5 is monotonically increasing with age and reaches a maximum at around 80 years of age, a finding that is consistent with the available evidence that clearly shows that COVID-19 fatality rates vary with age (Roser et al., 2020) . Furthermore, although the effects are small, according to our results exposure to PM 2.5 can actually increase the probability that infants and small children die from COVID-19. This is consistent with evidence showing that exposure to PM 2.5 can lead to systemic inflammation and suppress early immune responses to infections even in young people (Conticini et al., 2020; Wu et al 2020) .

This study is subject to some limitations. First, although we include a rich set of municipal-and individual-level covariates, available information does not allow us to directly control for individuals' residential choice or avoidance behavior. Both factors could lead to nonrandom assignment of pollution exposure and hence represent a threat to establishing a causal relationship. Second, we are implicitly assuming that air pollution is evenly distributed within each municipality (locality). If that is not the case and the distribution of PM 2.5 inside a municipality (locality) is correlated to some individual characteristics that also affect the probability of dying, then our results could be biased. Third, although we know the residence of J o u r n a l P r e -p r o o f all the individuals in our sample, we do not know for how long they have been living there or how much of their time they actually spend at home. In our estimations, individuals' long-term exposure to PM 2.5 is based on residence. Nonetheless, if an individual recently moved to a municipality, their true exposure to air pollution would be different to the one that we are attributing. Although we partially control for this by using exposure during 2019 as an alternative measure to exposure during 2000-2018, it is still possible for our results to be biased because of this. Fourth, we do not find robust evidence of an effect of short-term PM2.5 exposure on the probability of dying from COVID-19. We had expected to find a statistically significant relationship because pollution exposure could be expected to generate an acute deterioration of the immunological system and to facilitate transportation of the virus.

However, due to heterogeneous changes in daily activities and mobility across the population during the pandemic, measured PM2.5 concentrations may be less indicative of actual exposure and thus become a bad predictor of the probability of dying from COVID-19. In spite of these limitations, our study makes relevant contributions to the literature.

We find evidence of a positive relationship between PM 2.5 air pollution and the probability that an individual will die after contracting COVID-19; this relationship increases with age, especially for individuals that are 40 years old or more. Although our results suggest that this is mainly driven by long-term exposure, we cannot rule out that short-term exposure also has an effect.

These results are robust to model specification and inclusion of confounders at the municipaland individual-level. The last is the most important contribution of this paper as, in contrast to J o u r n a l P r e -p r o o f the previous literature, we are able to explicitly include in our estimations not only the age and gender of the person affected by COVID-19 but also if she has underlying health conditions. By doing this, we reduce the likelihood that unobservables could explain away the effects that we report. In spite of these efforts, our robust statistical evidence cannot be fully interpreted as causal as we do not have experimental or quasi-experimental data. Furthermore, although in our estimations we use cluster-robust standard errors, there might be spatial autocorrelation issues that cannot be fully accounted for by our empirical strategy that is based on individuallevel data.

Mexico is still in the midst of the pandemic; health conditions are continuously evolving, and data related to COVID-19 cases and mortality is being updated regularly. In the coming months we should have better data and a cleaner picture of the full extent of the effects of the pandemic both in the MCMA and in the country. Meanwhile, our results provide yet another reason for the need to implement more stringent controls to air pollution. In addition, and considering that at this moment we do not know when the pandemic will stop or if SARS-CoV-2 will become a recurrent threat, the relationship that we uncovered suggests that financial resources should be allocated to improve medical services in those areas where PM 2.5 concentrations tend to be high; long-term exposure has made people living in those areas more vulnerable to the disease. Additionally, policy makers and governments should rethink and reinforce environmental strategies that will reduce levels of air pollution as a key element of public health policies.

In our analysis we focus on the MCMA because of its high levels of air pollution and high rate of COVID-19 mortality. Nonetheless, it is not clear if our results are valid for other big metropolitan areas in the world or even for other parts of Mexico. The fact that our results are in general consistent with other studies on the relation between air pollution and COVID-19 mortality gives us greater confidence that our findings are externally valid-at least to some extent. Finally, our findings suggest that there are groups that, given their exposure to PM 2.5 , could be particularly vulnerable to COVID-19 in developing countries with characteristics similar to Mexico. Notorious among these groups are users of public transportation, street vendors, and households that use solid fuels for all or part of their cooking or heating needs.

Understanding how COVID-19 affects these groups will be essential for the design of public policies that can successfully decrease the toll that this pandemic is exerting on the population.

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. PM2.5 (2000 PM2.5 ( -2018 shown with 95% cluster-robust confidence intervals. The models behind each estimation are those described in Table 2 . Estimates of marginal effects to exposure to PM2.5 (2000 PM2.5 ( -2018 shown with 95% cluster-robust confidence intervals. The model behind the estimation (Model 3 in Table 2 ) includes municipal-level covariates (population density, population, density of hospital beds, percentage of population without access to health care, percentage of population with moderate or severe food insecurity, and percentage of labor force with jobs that can be done from home), and individual-level covariates (gender, age and age squared, obesity, diabetes, hypertension, smoking status, and day in which symptoms started).

The effects of long-term exposure to PM2.5 on probability of COVID-19 death in Mexico City. Estimates of marginal effects to exposure to PM2.5 (2000 PM2.5 ( -2018 shown with 95% cluster-robust confidence intervals. The models behind each estimation are those described in Table 3. J o u r n a l P r e -p r o o f J o u r n a l P r e -p r o o f J o u r n a l P r e -p r o o f Notes: The dependent variable is a dummy equal to one if an individual diagnosed with COVID-19 dies and zero otherwise. Estimations are done using a probit model. Cluster robust standard errors at the municipal-level shown in brackets. * p<0.1, ** p<0.05. Municipal-level covariates are: Population density, population, density of hospital beds, percentage of population without access to health care, percentage of population with moderate or severe food insecurity, and percentage of labor force with jobs that can be done from home. Individual-level covariates are: gender, age and age squared, obesity, diabetes, hypertension, smoking status, and day in which symptoms started. The Mexico City sample only considers the 14 municipalities for which, in addition to long-term exposure to PM 2.5 , we know PM 2.5 for 2019 and 2020. 

Air pollution and health. The Lancet

Global estimates of mortality associated with long-term exposure to outdoor fine particulate matter

Hazardous air pollutant exposure as a contributing factor to COVID-19 mortality in the United States

Coronavirus Pandemic (COVID-19

Air Pollution and Noncommunicable Diseases

Calidad del aire en la Ciudad de México, informe 2016, Dirección general de Gestión de la Calidad del Aire, Dirección de Monitoreo Atmosférico

Bases de datos -Red Automática de Monitoreo Atmosférico (RAMA)

Sistema Nacional de Vigilancia Epidemiológica

SARS-Cov

Casos diarios por municipio. Dirección General de Epidemiología, Secretaría de Salud

Links between air pollution and COVID-19 in England

The World's cities in 2018 -Data Booklet. Department of Economic and Social Affairs, Population Division

Association Between Air Pollution in Lima and the High Incidence of COVID-19: Findings from a Post Hoc Analysis

Exposure to air pollution and COVID-19 mortality in the United States: A nationwide cross-sectional study

Association of particulate matter pollution and case fatality rate of COVID-19 in 49 Chinese cities