Short-term exposure to fine particulate pollution and elderly mortality in Chile

Abstract

Exposure to fine particulate matter (PM_2.5) is known to cause adverse health outcomes. Most of the evidence has been derived from developed countries, with lower pollution levels and different demographics and comorbidities from the rest of the world. Here we leverage new satellite-based measurements of PM_2.5, combined with comprehensive public records in Chile, to study the effect of PM_2.5 pollution on elderly mortality. We find that a 10 μg/m³ monthly increase in PM_2.5 exposure is associated with a 1.7% increase (95% C.I.: 1.1–2.4%) in all-cause mortality for individuals aged 75+. Satellite-based measurements allow us to comprehensively investigate heterogeneous effects. We find remarkably similar effect sizes across baseline exposure, rural and urban areas, income, and over time, demonstrating consistency in the evidence on mortality effects of PM_2.5 exposure. The most notable source of heterogeneity is geographical, with effects closer to 5% in the center-south and in the metropolitan area.

Introduction

Chile, a relatively new member of the Organization for Economic Co-operation and Development (OECD), has many environmental challenges due to its recent industrialization. Air pollution levels are far higher than in other wealthy nations, such as the United States or European countries¹; many cities in Chile have annual PM_2.5 concentrations above 20 μg/m³, exceeding the World Health Organization (WHO) guidelines of 5 μg/m^3 2. Multiple studies have linked both chronic and acute PM_2.5 exposure to adverse health outcomes including cardiovascular and respiratory diseases, as well as an increase in all-cause mortality^{3,4,5,6,7,8,9,10,11,12}. Older populations and infants are known to be particularly vulnerable to PM_2.5 exposures^13,14,15,16. Historically, most research has been conducted in developed countries due to the availability of reliable public health statistics, comprehensive air pollution monitoring, and data on other variables that affect both air pollution and mortality. These countries typically have PM_2.5 concentrations below 20 μg/m^3 12. Developed countries also differ markedly in many dimensions, including population demographics, comorbidities, and healthcare infrastructure, from the rest of the world.

The recent emergence of new data measurement techniques, in particular, remote sensing¹⁷, opens the possibility of measuring air pollution exposure with complete spatial coverage and over a much longer time scale. Such measurements have been used to monitor air pollution globally and over time^18,19, and to study its impact on mortality outside the developed world. Recent work has studied the empirical relationship between satellite-based PM_2.5 exposures and mortality in Africa^16,20, Indonesia²¹, Brazil²², and in cities in Latin America²³. For Chile, earlier studies on the short-term association between PM_2.5 and health-related outcomes have used data from ground monitoring stations and focused on the capital, Santiago, and other cities, covering only a fraction of the country and population^24,25,26,27. More recent work that uses satellite-based estimates similarly only focuses on cities²³. In general, evidence from rural areas is scarce worldwide^28,29,30,31, and in particular in Latin America.

Here we study the short-term effect of PM_2.5 on elderly 75+ mortality in Chile, with complete geographical coverage of 327 communes from 2002 to 2019. We focus on the 75+ age group since the elderly population has a higher susceptibility to PM_2.5 pollution^15,32,33. In addition, with global populations aging rapidly, low- and middle-income countries are expected to be disproportionately affected by air pollution^18,19,34; Latin America is projected to see the largest absolute growth in the share of older individuals between 2015 and 2050³⁵, making it even more pressing to study the health of these populations. We first validate satellite-based PM_2.5 estimates³⁶ using available ground-based measurements (see “Methods” section and Supplementary Information (SI), section on “Accuracy of satellite-based PM_2.5 estimates”), and then combine these with a recent national population census to estimate monthly population-weighted exposures to PM_2.5. These monthly exposures are estimated for each commune, the third-level administrative division (the smallest administrative division in Chile), taking into account the geographical distribution of the population within each commune. We construct monthly commune-level mortality statistics using data from publicly available individual death certificates and official yearly population estimates^37,38.

With an assembled panel of commune-level monthly data from 2002 to 2019, we use fixed effects models to answer the question, “What would mortality among older adults look like in a commune during a month with higher than usual fine particulate pollution, compared to an average month in the quarter in the same commune?” Fixed effects for each commune control for commune-specific factors and avoid cross-sectional comparisons. We flexibly control for both time trends and seasonal patterns using quarter-year fixed effects. In other words, our estimates compare mortality for a month in a given commune that has higher levels of pollution than is expected for that quarter-year, to a month in the same commune with pollution that is equal to expectations for that quarter-year. Our estimates represent a short-term effect rather than a chronic one, which is partly captured by the fixed effects. We additionally account for temperature, an important confounding variable that affects both PM_2.5 (e.g., due to the practice of wood burning for residential heat³⁹) and mortality⁴⁰. After accounting for these factors, any remaining variation in PM_2.5 levels is plausibly random, allowing us to estimate the causal effect of PM_2.5 on mortality.

Next, we demonstrate the additional insight that can be derived from data with complete and granular spatial coverage by investigating heterogeneity in the estimated effect over multiple dimensions, including region, urbanicity, baseline PM_2.5 levels, population shares of 75+ individuals, and income. These dimensions of heterogeneity have not been fully addressed in previous work, which most commonly focuses on differences due to age and sex. Studying the entire country provides more variation in these characteristics compared to that in limited geographical areas such as cities; in addition, our commune-level data allow us to characterize heterogeneities within cities where they exist. (For instance, Santiago, a city that Gouveia et al.²³ considers in its entirety, contains 52 different communes.) Additionally, by using a longer time period (18 years, compared to up to 10 years in studies covering the same geographical region^{23,24,25,26,27,41}), we investigate changes in effects over time, hinting at potential adaptations to air pollution. We conduct a range of sensitivity analyses to ensure the robustness of our results to different modeling assumptions, outcome variables, and data restrictions. Finally, we estimate the number of avoided deaths under the hypothetical scenario where communes in Chile had reduced pollution to the levels stipulated in WHO guidelines, illustrating the considerable toll that air pollution has on elderly populations, and the importance and urgency of public health interventions to mitigate this impact.

Results

Commune-level monthly population-weighted PM_2.5 exposures range from 3 to 90 μg/m³, with a mean of 22.2 μg/m³. There is substantial spatial and temporal variation (Fig. 1). The populations in the least and most polluted communes were exposed to an average PM_2.5 concentration of 10 and 36 μg/m³, respectively. The southern part of Chile has higher PM_2.5 exposure and lower land temperatures (Fig. 1A and SI Fig. S2). We observe distinct seasonal variation, with higher pollution levels in the winter, consistent with the use of wood burning for residential heating⁴² and meteorological conditions trapping polluted air^43,44,45 (Fig. 1B). Satellite-based land temperature estimates display similarly strong seasonality, with values ranging below 0 °C to above 40 °C, with a population-weighted mean of 24.2 °C (SI Fig. S3). Population-weighted PM_2.5 exposure estimates display a peak in January (summer) 2017 (Fig. 1B), consistent with a historically large wildfire that affected the central-south part of Chile⁴⁶. A histogram of PM_2.5 exposures reveals a large proportion of commune-months that exceed the national annual standard of 20 μg/m³; virtually all commune-months do not meet the annual WHO guideline of 5 μg/m³ (Fig. 2).

Fig. 1: Monthly population-weighted PM_2.5 exposures and 75+ mortality rates.

A Mean monthly population-weighted PM_2.5 exposures (red) and 75+ mortality rates (purple) in 2019 for 327 communes in Chile. For 75+ mortality rates, five communes were omitted from the figure as their high mortality rates distorted the color scheme. These five communes have 75+ populations below 250 and mortality rates above 90 deaths per 1000. The five omitted communes and mortality rates per 1000 in parentheses are Tortel (143), Primavera (200), Ollague (100), Guaitecas (102), and Cisnes (92). (Inset) The enlarged area shows part of the Metropolitan region (region M), including the capital city of Santiago. The black dots show the 59 available PM_2.5 monitor stations that fail to cover the entire country. B Monthly PM_2.5 and 75+ mortality rates, 2002–2019, for each commune. Each line represents one of 327 communes. Representative communes across the country, ordered from north to south, are highlighted in different colors.

Fig. 2: Short-term effect of PM_2.5 on monthly 75+ all-cause mortality.

The figure shows the estimated effect of a 1.7% increase (95% C.I.: 1.1–2.4%) in 75+ mortality rate for a 10 μg/m³ increase in PM_2.5 exposure, centered at the average exposure and mortality rate. The shaded band is the 95% confidence interval. The histogram shows the PM_2.5 exposure distribution for all commune-years in the study. All-cause encompasses all deaths. Vertical lines indicate the Chilean and World Health Organization (WHO) annual PM_2.5 standards.

Estimates of monthly 75+ mortality rate (mortality here includes all causes, which are deaths from any disease as well as external causes; see “Methods” section) display a more uniform spatial and temporal distribution (Fig. 1). There is some evidence of seasonality, with higher mortality rates in the winter. Throughout the entire study period, the average national 75+ population was 691,913. A total of 862,818 deaths of individuals 75 and older were registered, or 3995 a month on average. The average monthly 75+ mortality rate is 5.77 per 1000, and 7.7% of commune-months registered zero deaths.

Effect of PM_2.5 on mortality

We find that a short-term 10 μg/m³ increase in PM_2.5 exposure is associated with a 1.7% increase (95% C.I.: 1.1–2.4%) in monthly all-cause mortality for individuals aged 75 and over (Fig. 2).

This overall estimated effect size, sometimes referred to as relative risk in the literature, masks substantial spatial heterogeneity (Fig. 3A). Of the 16 regions (first-level administrative divisions) in Chile, eight in the center-south of Chile have significant positive estimated coefficients. The remaining regions have estimates that are not significantly different from zero, with the exception of the southernmost region XII. This region has a significant negative estimate, but a poor correspondence between satellite and ground-monitor measurements (see SI “Accuracy of satellite-based PM_2.5 estimates” for more details).

Fig. 3: Heterogeneous effects of PM_2.5 on 75+ mortality rate.

A Estimates for each of the 16 regions (first-level administrative division) in Chile, annotated using Roman numerals. Dots are point estimates and bars are 95% confidence intervals. A red dot indicates a significant effect, different from zero, at a 5% level. The red dashed line shows the main effect estimate (Fig. 2) and the red shaded band is the confidence interval. The map shows the PM_2.5 exposure for each region (population-weighted average across communes) in 2019. Note that the accuracy of satellite-based estimates is poorer in regions XV, III, XI and XII (see “Methods” section and SI Fig. S1). Effects for B communes in the metropolitan region versus in the rest of the country, C urban vs. rural communes (above 30% of all-age population in rural areas), D baseline PM_2.5 exposures (monthly mean over the entire study period) above and below 20 μg/m³, and E communes with 75+ population share above and below the median of 4.5%. The effects are centered at the average exposure and mortality rate and the shaded band is the 95% confidence interval. Histograms show the monthly PM_2.5 exposure distribution for each commune in the subgroup over the entire period (2002–2019).

The effects in the center-south of the country tend to be higher, with an average estimated effect of 4.7%. The Metropolitan Region (region M), an important region that contains the capital of Chile where 40% of population lives⁴⁷, has a larger estimated effect of 5.3% (4.2–6.5%) per 10 μg/m³ increase in PM_2.5 (Fig. 3B). Regions in the center-south of Chile have been identified by the Chilean government to be problematic because of high pollution levels, and we observe this clearly in our satellite-based estimates (Fig. 1A). The higher pollution levels can largely be attributed to mobile and industrial sources (center) and wood burning for residential heating (south)³⁹. The government has targeted decontamination programs⁴⁸ and placement of ground monitors in these regions (Fig. 1A). Our analysis reveals that not only are levels of pollution higher in regions in the center-south, their impact on elderly mortality is also larger. SI Fig. S4 plots regional estimates against mean PM_2.5 exposure, highlighting this observation.

The preceding results suggest two potential dimensions of heterogeneity that are of interest and could explain the larger effect sizes in the center-south regions: rural versus urban areas and baseline levels of pollution. Fitting separate models for urban and rural communes (see “Methods” section; Fig. 3C), and by baseline pollution levels (mean PM_2.5 above versus below 20 μg/m³; Fig. 3D), we find that neither of these modify the relationship between PM_2.5 and 75+ mortality. An extended analysis comparing the effect of PM_2.5 exposure between urban and rural areas across specific mortality causes (SI Fig. S5) similarly does not reveal any significant differences; urban areas have a slightly larger effect for respiratory causes (2.6% with 95% C.I.: 0.4–4.8% for urban vs. 1.2% with 95% C.I.: −0.9 to 3.3% for rural).

We find that communes with population shares of 75+ individuals above the median have effects of a larger magnitude of 2.3% (95% C.I.: 1.6–3.0%), compared to 1.4% (95% C.I.: 0.4–2.4%) for those below the median (Fig. 3E). Although the difference in effect sizes is not statistically significant, differences in 75+ population shares might partially explain the larger effect sizes in the center-south, which tend to have higher shares of elderly population (SI Fig. S2; SI Fig. S6 shows that at a regional level, all regions with 75+ population share above 4.5% are in the center-south, except for one). The larger effect in communes with higher elderly population shares may be due to the limited capacity of healthcare systems, increased circulation of respiratory viruses, or different age structures within the 75+ group⁴⁹. Additionally, people in communes with a lower share of the elderly population might be less susceptible to death, due to a potential survivorship bias, where individuals that are most “resilient” to air pollution constitute the majority that have survived to age 75^50,51. Given the global aging of populations, the effect of increasing elderly population shares on mortality should be explored in more detail in future work¹⁹.

Changes over time and income

With documented harms to both population health and the economy due to climate change, a critical question is to what extent populations have adapted to these changes⁵². Fully investigating this question in the context of fine particulate pollution would require data on a decadal scale over multiple decades. Our data span close to two decades, and we estimate effects over 4–5 year periods in our panel (2002–2019). While this does not cover a sufficiently long time period to provide definitive evidence on adaptation, it is considerably longer than other recent work in Chile that covers up to a 10-year period²⁴ and provides a first approximation to potential adaptations over time. We do not find convincing evidence of changes over time; coefficient estimates do not display an obvious trend (Fig. 4A; SI Fig. S7 displays estimated effects for single years). This could be due to two opposing effects: an expected adaptation over time which should decrease the effect size over time, and an aging population (see SI Fig. S8) which might increase the effect size (Fig. 3E).

Fig. 4: Changes over time and for different income levels.

A Effects are estimated separately for each group of years, from 2002 to 2019. B Communes are divided into quintiles based on their mean per capita annual income level from a 2017 socioeconomic survey. Dots are point estimates and bars are 95% confidence intervals. A red dot indicates a significant effect, different from zero, at a 5% level. The red dashed line shows the main effect estimate (Fig. 2) and the red shaded band is the confidence interval.

Wealth is a commonly hypothesized factor that enables populations to minimize the health impacts of environmental shocks⁵³, and this is especially relevant given that low and middle-income countries are exposed to higher levels of air pollution¹⁸. In terms of health outcomes, wealth can act through multiple pathways, such as better healthcare access, dietary habits, and physical well-being. Stratifying the analysis by income level (see “Methods” section), we observe the largest effects in communes in the poorest income quintile, but an increasing risk moving from the second to fourth quintile followed by a decline in the highest income quintile (Fig. 4B). There is no obvious trend or significant differences between income quintiles, suggesting that income is not the determining factor in explaining mortality impacts of PM_2.5 in Chile. Instead, the previously documented factors of location and population shares of 75+ may be more important.

Other sensitivity analyses

Our results are robust to a range of sensitivity checks (Fig. 5). First, we consider 75+ mortality of males and females separately, finding that overall estimated effects are virtually identical even though males have a higher baseline mortality rate. To check the sensitivity of the results, we remove communes with small 75+ populations and regions with a poor linear correlation between satellite-based and ground-monitor PM_2.5 measurements (regions XV, III, XI, and XII; see SI “Accuracy of satellite-based PM_2.5 estimates”). We find that estimates are unchanged.

Fig. 5: Effect of PM_2.5 for the full sample and for various subgroups.

Each row shows a model fitted for a different subgroup, with information on the number of commune-month observations (n), the average monthly mortality rate (monthly MR, deaths per 1000 inhabitants), monthly 75+ population, mean PM_2.5 (μg/m³, population-weighted), and the overall effect (black and red dots; red means statistically significant) with bars representing 95% CIs.

To check the robustness of our results to modeling assumptions, we consider a wide range of alternative model specifications. First, we use Poisson and ordinary least squares models instead of the negative binomial model that was presented in the main results (SI Fig. S9). Instead of quarter-year fixed effects, we consider the more restrictive month-year fixed effect, and other variations of month-fixed effects (SI Fig. S10). While these more restrictive fixed effects improve the identification of the causal effect, they absorb most of the variation in PM_2.5, thus diminishing its observable effect on mortality. While estimates remain positive, they are not statistically significant. We also consider fixed effects that vary by region and more flexible functional forms for the temperature variable. Our main results use a linear temperature term, as is used in other studies (e.g.,^16,54), while in robustness checks we consider polynomial and spline terms, producing very similar estimates to the main result (SI Fig. S10). Of particular note is that removing temperature, a confounding variable, increases the effect size of PM_2.5 to around 6%. This larger estimate is consistent with the hypothesized effect of higher temperatures both reducing PM_2.5 levels and mortality rates (at the lower range of the well-established U-shaped temperature-mortality curve, where warmer temperatures decrease death rates⁵⁵). We repeat the analysis using mortality rates for the 65+ age group instead of 75+, which has been used in the previous literature^13,14,26; our estimates are not sensitive to this change (SI Fig. S11). Finally, we considered more flexible functional forms for the PM_2.5 variable (SI Fig. S12), finding roughly the same shaped response function as simple linear models, particularly in regions where the majority of the data lie.

In terms of specific causes of death, we find the largest effects for deaths from respiratory causes (Fig. 5). This is consistent with one of the main known acute effects of PM_2.5 pollution, the other being cardiovascular causes^12,56,57. As expected, we also found positive effects for cardiovascular and cardiorespiratory causes. In addition, there is a positive effect for mortality excluding cardiorespiratory causes, indicating other pathways in which PM_2.5 affects mortality, such as other diseases or indirect effects like reduced healthcare capacity.

Discussion

Our study uses satellite-based data to estimate population-weighted PM_2.5 exposures in Chile with complete geographical coverage and uses these to estimate the effect of PM_2.5 exposure on elderly mortality. The inclusion of rural areas and the use of granular commune-level data allow us to study dimensions of heterogeneity with a level of detail that has not been explored in previous work. Our work contributes more broadly to the understanding of the relationship between fine particulate pollution and mortality in countries outside the developed world, and to the consistency of the current body of evidence in relation to mortality due to short-term exposure to particulate matter in older people. The approach we develop is applicable to other countries and contexts.

Our findings indicate that a 10 μg/m³ increase in monthly PM_2.5 is associated with an increase in all-cause mortality of approximately 1.7% (95% C.I.: 1.1–2.4%), for individuals over 75 years old in Chile in the same month. The estimated effects should be interpreted as the short-term mortality induced by monthly deviations in PM_2.5 concentrations. We do not capture the chronic effect of exposure to PM_2.5 pollution, which has been established in multiple other studies^6,7,12,58, even for Chile⁴¹, with generally much larger effect sizes (in the range of 2–25%¹²). Daily, acute effects have also been documented in other studies, with similar effect sizes (0.8–4.2%), despite important differences in methodologies and contexts^{9,13,24,25,26,27,28,59}. We provide a slightly different perspective that is focused on monthly exposures, which do not capture immediate daily responses to increases in PM_2.5. Monthly effects capture the effect of sustained weeks-long increases in PM_2.5, such as those produced by wildfires, as well as the effect of illnesses with a longer window between onset and death. Such monthly effects are less well-studied and are not subject to regulatory standards, but could become increasingly important due to more frequent moderate exposure events such as wildfires and droughts⁶⁰. Consistent with findings in the US⁶⁰ and in Latin American cities²³, we find that the magnitude of monthly effects is similar to or exceeds that for daily acute effects, but is smaller than chronic effects. A limitation of monthly aggregated data is that a fraction of deaths occur before the peak PM_2.5 exposure in the same month. This temporal mismatch could limit the causal interpretation of our estimates and attribution of the effects to known biological pathways. In particular, the main known acute effects of PM_2.5 pollution are respiratory and cardiovascular illnesses leading to increased mortality. These same factors would naturally explain the increased monthly mortality that we observe, but this relies on the assumption that mortality increases after PM_2.5 exposure.

Overall, the estimated effects are remarkably consistent across many dimensions, such as baseline pollution levels, urban versus rural, men and women, 65+ versus 75+, income, and over time. Small differences exist due to population shares of 75+, with larger effects in areas with larger population shares of 75+. Additionally, urban areas have a slightly larger effect on respiratory causes than rural areas. To our knowledge, these sources of heterogeneity have not been fully explored in previous work, which focuses on age, sex, individual risk factors^6,7,14, other aggregate socioeconomic characteristics, and meteorological conditions^41,61,62. Differences between urban and rural areas have been understudied in the literature, with one study finding a similar result as we do in China, where PM_2.5 exposure had larger effects on respiratory causes of death in urban areas²⁸. This result may be due to differences in the source and composition of PM_2.5; in the case of Chile, in rural areas, pollution comes mainly from natural sources and wood burning⁴², while in cities, it is due to traffic and industrial activity³⁹. Another potential explanation might be more circulation of respiratory viruses in areas with higher population density.

The most notable source of heterogeneity that we find is geographical, where the effect is closer to 5% in the center-south of Chile and the metropolitan area. The effect size for the latter is similar to a study that uses daily mortality data from 1988 to 1996 in Santiago city and finds an effect of 4.2% on all-cause mortality (all ages) excluding accidental and trauma deaths²⁵. The regional effects we find are not explained by differences between rural and urban areas or baseline levels of PM_2.5, but there is suggestive evidence that population shares of 75+ may have some effect. Another potential explanatory factor for the regional heterogeneity is the chemical composition of PM_2.5, which could affect multiple health outcomes⁶³. There is substantial regional heterogeneity in the sources of PM_2.5 emissions in Chile: in the north emissions come mainly from mobile sources and industrial smelters for minerals like copper; in the center from mobile and industrial sources; and in the south from mobile sources and residential wood combustion³⁹. A full characterization of emissions sources and their impacts on particulate matter composition is beyond the scope of this study, but this is a key area for future work. A final potential factor explaining regional differences is that elderly people are likely to spend more time indoors in colder regions in the south, increasing their exposure to indoor sources of PM_2.5, such as wood combustion⁴². Our study is limited in its ability to study indoor air pollution, as we discuss in more detail below.

To understand the implications of fine particulate pollution on 75+ mortality, we calculate the avoided deaths under the hypothetical scenario that all communes in Chile were in compliance with either the PM_2.5 annual national standard (20 μg/m³) or the WHO guidelines (5 μg/m³) during the period of our analysis (2002–2019) (see “Methods” section). Under the national standard achievement scenario, we estimate an average of 27 avoided deaths (95% C.I.: 16–38) due to air pollution per month, and under the WHO guidelines scenario, we estimate 126 avoided deaths (95% C.I.: 76–176) per month. For the entire study period (2002–2019), the latter represents over 25,000 avoided deaths due to air pollution in Chile, which represents 3.2% of the total 75+ deaths. For the metropolitan region alone, with its higher estimated effect size, a total of 178 (95% C.I.: 140–217) deaths per month could have been avoided under the WHO guidelines scenario; this represents 11.6% of total 75+ deaths. These numbers underscore the importance of reducing short-term PM_2.5 exposures in Chile and help to quantify the benefits of potential policy interventions, such as a comprehensive wildfire management strategy.

Our study has limitations arising from the data sources used. Our validation of satellite PM_2.5 measurements with ground measurements reveals that satellite-based estimates tend to be less accurate in the low and high ranges. Compared to ground measurements, they overestimate lower concentrations (below 12 μg/m³) and underestimate higher concentrations (above 50 μg/m³) (SI Fig. S1). This might introduce an upward bias in our effect size estimates if the actual variation in PM_2.5 is larger than the satellite-based estimations. To partially address the issue, we remove several regions identified to be particularly problematic in our data validation exercise as a sensitivity check and do not find significant differences. Next, both satellite and monitor PM_2.5 measurements capture outdoor pollution and not indoor pollution, limiting our ability to accurately measure population exposures, which may be due to in-home sources of pollution. In this regard, elderly people living in colder regions in Chile are more likely to spend less time outside and have higher levels of PM_2.5 exposure indoors^25,64,65. Measuring indoor pollution is difficult even in wealthy countries, and has been identified as a critical research priority⁶⁶. Next, importantly, our data do not include information on the source of the PM_2.5 pollution, which may affect overall particulate size^67,68,69 and subsequently affect health through different pathways^70,71,72. Finally, death certificates register the commune of residency at the time of death and do not necessarily reflect the location of the majority of a person’s PM_2.5 exposure, both in the short-term and in the past. For example, if people move to a different commune after retirement (the retirement age in Chile is 65), commune fixed effects might not fully capture their previous air pollution exposure. However, in general, mobility in Chile is low, and more so for the elderly; hence, we do not anticipate this limitation to be a major issue.

Using satellite-based PM_2.5 estimates with comprehensive spatial and temporal coverage, we provide insights into the effect of short-term exposures to PM_2.5 on 75+ mortality in a rapidly developing country. Our findings highlight the urgent need for effective air quality regulations and interventions to reduce short-term exposures to PM_2.5, particularly in the central-south part of Chile. Our approach illustrates the potential of using satellite-based measurements of PM_2.5 to study the effect of air pollution on health outcomes, which could be particularly valuable in developing countries that lack a comprehensive network of ground monitors, and where air pollution is a growing public health concern.

Methods

Study area

The study was conducted in the continental area of Chile, a Latin American country in the southern hemisphere, located between the −17° and −56° latitude and −65° and −76° longitude. Chile has multiple climatic regions due to latitude differences and oceanic and mountain boundaries: tropical in the north, Mediterranean in the center, and Antarctic in the south⁷³. Chile has a population of 17.5 million (2017 estimates), with 0.8 million in the 75+ age group⁴⁷, and a GDP per capita of 15 thousand USD in 2022⁷⁴. The geographical aggregation used for the analysis was the commune level, the third-order (smallest) administrative unit in Chile. There are a total of 346 communes (third-level administrative division) in 56 provinces (second-level administrative division) in 16 regions (first-level administrative region) in Chile.

Air pollution data

For fine particulate matter pollution (PM_2.5) we use satellite-based data produced by the Atmospheric Composition Analysis Group from the Washington University in St. Louis³⁶ and made publicly available at https://sites.wustl.edu/acag/datasets/surface-pm2-5/. These data contain global estimates at a 0.01° ( ~1.11 km) resolution at a monthly level from 2000 to 2021. These PM_2.5 estimates are calculated using satellite measurements of aerosol optical depth and a chemical transport model, and calibrated using ground-based observations. To validate the satellite data, we use available PM_2.5 ground-monitor data (Fig. 1) from Chile’s National Air Quality Information System (SINCA), publicly available at https://sinca.mma.gob.cl/⁷⁵.

Land temperature data

For land temperature, we use the Terra Moderate Resolution Imaging Spectroradiometer (MODIS) Land Surface Temperature/Emissivity Daily (MOD11A1) Version 6.1, publicly available at https://developers.google.com/earth-engine/datasets/catalog/MODIS_061_MOD11A1⁷⁶. These data have worldwide coverage and a spatial resolution of 1 km and are available daily starting from March 2000. We derive monthly measurements by taking the mean over all days in each month. Another commonly used source of temperature data is the 2-m ambient temperature (https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-land.) We chose land temperature for two main reasons. First, the satellite land temperature data has a finer resolution (1 km) than available ambient temperature data (9 km). This allows us to more accurately estimate population exposures within urban centers and minimizes errors in coastal areas where pixels overlap the ocean. Second, land temperature and ambient temperature at 2-m have a high correlation (0.9)⁷⁷, and studies have recommended the use of land temperature datasets^77,78. Temperature is a control variable in our empirical model (see “Empirical strategy” section), and since the two measures are highly correlated, the choice of either should not affect estimates of our coefficients of interest.

Population data

We use yearly population estimates from 2002 to 2019, released by the National Statistics Institute of Chile (El Instituto Nacional de Estadísticas) at https://www.ine.gob.cl/estadisticas/sociales/demografia-y-vitales. These are derived from the 2002 and 2017 Chilean national censuses and by interpolation using demographic models³⁷. Estimates are available at a commune-year level and are disaggregated by one-year age groups and sex.

In addition, for the 2017 Census, population estimates are available at the level of enumeration area (district zone in urban areas, aggregating multiple blocks of houses and rural entity for rural areas), disaggregated by five-year age groups and sex. We use these more granular estimates to estimate population exposures to PM_2.5 and temperature (see “Estimating population-weighted exposures” section).

Mortality data

Individual death certificates from 1990 to 2020 are released by the Chilean Department of Statistics and Health Information³⁸, and made available at https://deis.minsal.cl/#datosabiertos. Death certificates have information on the commune of residency, sex, age, date, and cause of death of the deceased. The cause of death is defined using the International Classification of Diseases, Tenth Revision (ICD-10) codes.

Data validation

To validate the satellite-based PM_2.5 data, we compare 3858 monthly PM_2.5 measurements with ground-monitor data from 59 different stations from January 2015 to December 2021. Full details are in the SI (Accuracy of satellite-based PM_2.5 estimates). The Pearson correlation coefficient between the satellite data and ground measurements is 0.78, with minor differences across years; the root-mean-square error is 12.76. There are several regions that have noticeably poorer correlations between satellite and ground measurements, largely due to inaccuracies in the low and high ranges. Satellite-based estimates tend to be higher than ground measurements for lower concentrations (below 12 μg/m³); the opposite applies for higher concentrations (above 50 μg/m³), where satellite measurements are lower than monitor measurements. As robustness checks, we exclude the regions in which these issues are most problematic.

Constructing a panel data set

We construct a panel of data for 327 communes over 216 months from January 2002 to December 2019. Each of the 70,632 observations has information on population exposure to PM_2.5 and temperature and 75+ mortality. We discard data from 19 communes that were either outside the continental area of Chile or had an average population aged 75 or older of fewer than 50 people. For communes with a small number of older individuals, mortality estimates tend to be more variable and have a larger number of zeros. As a robustness check, we test the sensitivity of our results to this cutoff. The time period was selected as the intersection of availability of PM_2.5, temperature, death certificates, and population data. We did not include data beyond 2019 to avoid major changes due to the COVID-19 pandemic.

Estimating population-weighted exposures

To estimate population-weighted exposure to air pollution and land temperature at a commune-month level, we use the satellite-based PM_2.5 and temperature data, combined with population estimates at the level of enumeration area, based on the 2017 Census⁴⁷.

For each month, we compute (I) the exposure for each district zone or rural polygon, aggregated from satellite measurements, weighted by the land area intersecting with satellite grids, and then (II) the population-weighted exposure for each commune, aggregated from the district zone and rural polygon exposures, weighted by population in district zones or rural polygon. For (I), we use the spatial intersection of polygons and satellite rasters to calculate the fraction of the total polygon area intersected with satellite rasters. For example, if a district zone falls within two grids of satellite data accounting for 80% and 20% of the district zone area, the PM_2.5 assigned to the district zone is the average of the two satellite measurements, using 80% and 20% as weights. For (II), we take the average exposure among all district zones and rural polygons contained in the commune, using the 2017 population as weights.

Monthly mortality rates

The mortality rate was calculated at the commune-month level for the 75+ age group by dividing the number of deaths in a given month (from death certificates) by the total population in that age group from the yearly population estimates. We compute mortality for all causes, cardiorespiratory causes, cardiovascular causes, respiratory causes, and all causes excluding cardiorespiratory. These categories are based on ICD-10 codes available in the death certificates. All causes encompasses all deaths. Deaths due to cardiovascular issues are categorized under codes I00 to I99 and respiratory causes under codes J00 to J99.

Empirical strategy

To estimate the effect of PM_2.5 exposure on mortality, we use a generalized linear model with two-way fixed effects:

$$g({\mathbb{E}}[{Y}_{imqt}])=\beta P{M}_{imqt}+\gamma {T}_{imqt}+{\alpha }_{i}+{\theta }_{qt},$$

(1)

where Y_imqt corresponds to the 75+ mortality rate in commune i at month m in quarter q and year t, and β and γ represent the effects of PM_2.5 and temperature on mortality, respectively. α_i represents fixed effects for each commune (327 in total), and θ_qt represents time-fixed effects at a quarter-year level (72 in total; 4 quarters each for 18 years). g() represents the link function (in this case, the log transformation; see below for details).

Commune fixed effects capture cross-sectional, time-invariant heterogeneity between communes, and allow us to control for confounding variables that do not vary over time, such as health infrastructure and services, income level, and dietary habits. This avoids cross-sectional comparisons, for example, if certain communes have higher levels of PM_2.5 and mortality than others due to a common confounding variable, it would create the appearance of a positive association between PM_2.5 and mortality. Time-fixed effects control for national yearly and seasonal trends in PM_2.5 and elderly mortality. We use quarter-year (e.g., 2010-Q1) instead of month-fixed effects, as most of the variation in PM_2.5 is absorbed by month-fixed effects, thus diminishing its observable effect on mortality. Potential time-varying confounders include hospital saturation, seasonal activities, other air pollutants (e.g., O₃, SO₂), and behavioral changes such as alcohol consumption, diet, or physical activity, which are likely to be captured by quarter-year fixed effects. Temperature is an important confounding variable that varies over space and time and is included in the model. We use a linear temperature term in the main results and polynomial and spline terms in robustness checks.

We use a negative binomial distribution to model the mortality rate outcome, which is appropriate given the over-dispersion and zero-inflation of this outcome variable⁷⁹. The log transformation is the canonical link function for the negative binomial regression. We do robustness checks using other distributional assumptions for the outcome variable (see “Robustness checks” section). Standard errors are clustered at the commune level. Both coefficients for PM_2.5 and temperature (β and γ) can be interpreted as the short-term effect of the respective variables on elderly mortality in Chile, after accounting for spatial heterogeneity and temporal trends.

Heterogeneity

To examine heterogeneity in the relationship between PM_2.5 and 75+ mortality, we fit separate models (Equation (1)) for different subsets of the data. For spatial heterogeneity, we consider communes in each of the 16 regions separately, using the same model as in Equation (1).

For the Metropolitan region versus the rest of the country, we fit two separate models for communes in Region M and in all regions excluding M. For the rural versus urban analysis, we classify communes as rural and urban-based on the proportion of the population (all ages) in each commune that are in rural entities, according to the 2017 Census. Communes that have more than 30% (approximately the median) of the population in rural areas are classified in the rural category.

For baseline PM_2.5 exposures, we calculate the average monthly PM_2.5 for each commune over the whole period; communes with means above and below 20 μg/m³ are analyzed separately. 20 μg/m³ was chosen as it is roughly the median and is also Chile’s national annual standard. For 75+ population shares, we similarly take the average yearly 75+ population share for each commune over the 18-year period. We analyze communes with an average 75+ population share above 4.5% (the median), and below 4.5% separately.

For differences over time, we consider data for each year separately and use quarter-fixed effects instead of quarter-year. Finally, for income, we stratified the communes based on their average annual per capita income, according to a national socioeconomic survey from 2017⁸⁰. We split the 321 communes with available survey data into five income quintiles, with average 2017 per capita income cutoffs of $3,182, $3,577, $4,138, and $4,961 US dollars.

Robustness checks

For male and female mortality rates, we use death certificates and population estimates for the respective sexes to construct separate estimates. To investigate the sensitivity of estimates to the inclusion of communes with fewer older adults, we discard communes with an average population aged 75 or older of fewer than 500 people, instead of 50 in the main analysis.

Next, our validation of satellite-based estimates highlighted four regions, XV, III, XI, and XII, with a poor linear correlation between satellite and ground-monitor measurements (see SI Fig. S1). We exclude these regions as an additional robustness check.

Finally, we test the sensitivity of our estimates to modeling assumptions. We use ordinary least squares regression as well as Poisson regression, instead of the negative binomial model used in the main analysis. We also consider different model specifications, including interactions, month-fixed effects, and various functional forms of the temperature variable (SI Fig. S10). We also replicate the full set of results for 65+ individuals (SI Fig. S11).

Hypothetical avoided deaths

For each commune-year, we first compute the population-weighted monthly PM_2.5 exposure, and the percentage reduction from the recorded mean necessary to meet the annual target of either 5 μg/m³ for WHO guidelines, or 20 μg/m³ for Chile’s national standard. We then apply this percentage reduction to each month in the year, to get the target monthly PM_2.5 for the entire study period. Then, for each month we apply the coefficient estimate for the effect size in the main model to derive the percentage reduction in mortality rate for the reduction in monthly PM_2.5 from the observed to the target exposure. Given the 75+ population in that commune month, we compute the corresponding number of deaths with the target exposure. The difference between actual deaths and hypothetical deaths is the number of avoided deaths. For the metropolitan region, we use the estimated coefficient for that model, along with all the communes in the region.