Long-Term Exposure to PM2.5 Species and All-Cause Mortality among Medicare Patients Using Mixtures Analyses

Abstract

Background:

The relationship between long-term exposure to PM_2.5 and mortality is well-established; however, the role of individual species is less understood.

Objectives:

In this study, we assess the overall effect of long-term exposure to PM_2.5 as a mixture of species and identify the most harmful of those species while controlling for the others.

Methods:

We looked at changes in mortality among Medicare participants 65 years of age or older from 2000 to 2018 in response to changes in annual levels of 15 PM_2.5 components, namely: organic carbon, elemental carbon, nickel, lead, zinc, sulfate, potassium, vanadium, nitrate, silicon, copper, iron, ammonium, calcium, and bromine. Data on exposure were derived from high-resolution, spatio-temporal models which were then aggregated to ZIP code. We used the rate of deaths in each ZIP code per year as the outcome of interest. Covariates included demographic, temperature, socioeconomic, and access-to-care variables. We used a mixtures approach, a weighted quantile sum, to analyze the joint effects of PM_2.5 species on mortality. We further looked at the effects of the components when PM_2.5 mass levels were at concentrations below 8 μg/m³, and effect modification by sex, race, Medicaid status, and Census division.

Results:

We found that for each decile increase in the levels of the PM_2.5 mixture, the rate of all-cause mortality increased by 1.4% (95% CI: 1.3%–1.4%), the rate of cardiovascular mortality increased by 2.1% (95% CI: 2.0%– 2.2%), and the rate of respiratory mortality increased by 1.7% (95% CI: 1.5%–1.9%). These effects estimates remained significant and slightly higher when we restricted to lower concentrations. The highest weights for harmful effects were due to organic carbon, nickel, zinc, sulfate, and vanadium.

Conclusions:

Long-term exposure to PM_2.5 species, as a mixture, increased the risk of all-cause, cardiovascular, and respiratory mortality.

1. Introduction

Air pollution has long been linked to mortality (Beelen et al., 2014; Danesh Yazdi et al., 2021a; Di et al., 2017; Dockery et al., 1993; Lepeule et al., 2012; Pope et al., 1995; Pope III et al., 2002; Pun et al., 2017; Shi et al., 2021; Vodonos et al., 2018; Wei et al., 2020). The United States Environmental Protection Agency (EPA) has concluded the association between PM_2.5 and mortality is causal based on the available scientific evidence for both short-term and long-term exposure (US EPA, 2022). However, PM2.5 itself is a mixture of a variety of compounds. It is important to understand, from both a scientific and a policy perspective, to what extent each specific constituent poses a threat to the health of the public. Examining the specific toxicity of PM_2.5 species will strengthen our understanding of which sources of air pollution we need to address and how PM_2.5 causes its health effects.

Previous studies looking at the association between PM_2.5 species and mortality have largely focused on short-term exposure to these pollutants(Franklin et al., 2008; Heo et al., 2014; Kim et al., 2022; Lingzhen et al., 2014), where effects might be lower than long-term exposures(Beverland et al., 2012; Kloog et al., 2012). Moreover, the long-term studies that have been conducted thus far have generally focused on one pollutant at a time due to the high correlation of some of the PM_2.5 species, or they have not used mixtures methods to account for this issue(Bart et al., 2015; Hao et al., 2023; Hvidtfeldt et al., 2019; Krewski et al., 2005; Wang et al., 2022). Furthermore, the exposures in most of the studies previously described were assigned using either monitored data or data from exposure models with lower resolutions than are now available.

In this study, we examined the relationship between long-term exposure to PM_2.5 species, as a mixture, and all-cause mortality using data from the Medicare population from 2000 to 2018. The Medicare dataset is a large administrative dataset which includes the vast majority of elderly individuals in the United States. This data has previously been used to link long-term exposure to air pollutants such as PM_2.5, NO₂, O₃ to health outcomes including but not limited to mortality, cardiovascular disease, respiratory disease, and neurological disease(Abu Awad et al., 2019; Danesh Yazdi et al., 2022, 2021b; Di et al., 2017; Shi et al., 2020). These studies, however, did not use such high-resolution and comprehensive PM_2.5 species exposure data and did not use mixtures methods. We used high-resolution exposure assessment models and a weighted quantile sum (WQS) mixtures approach to conduct the analyses. We further looked at cause-specific mortality and vulnerable sub-groups in stratified analyses. We repeated all analyses amongst ZIP code-years with total PM_2.5 less than 8 μg/m³.

2. Materials and Methods

2.1. Study Population

The study population consisted of Medicare beneficiaries, aged 65 years or older, enrolled in the program between 2000 and 2018 who lived in the contiguous United States. Demographic data on the study population was derived from the Medicare denominator file which contained information on beneficiary age, race, sex, Medicaid-eligibility, ZIP code of residence, and date of death. This data was updated annually. The data were aggregated to ZIP code-year for the purposes of this study. We restricted analyses to ZIP code-years with at least 100 Medicare residents.

2.2. Exposure Measurement

Our exposures of interest were annual average levels of 15 PM_2.5 species, namely: elemental carbon (EC), organic carbon (OC), sulfate (SO₄²⁻), nitrate (NO₃⁻), ammonium (NH₄⁺), bromine (Br), calcium (Ca), copper (Cu), iron (Fe), potassium (K), nickel (Ni), lead (Pb), silicon (Si), vanadium (V), and zinc (Zn) from 2000 to 2018. We estimated values for each component using hyperlocal spatio-temporal models that estimated annual levels of PM_2.5 species at a 50 m spatial scale in urban areas and a 1 km scale in non-urban areas. Approximately 80% of the U.S. population lives in urban areas(United States Census Bureau, n.d.). These spatio-temporal models were developed using machine learning methods, measured values from approximately 1,000 monitors including the Environmental Protection Agency’s (EPA) speciation network and the National Park Service’s IMPROVE network, and over a hundred predictor variables predictors included variables such as time, location, satellite data, climate variables, emission sources, etc. For non-urban areas, five machine learners were used to generate initial predictions: stochastic Gradient Boosting (GBM), Cubist regression (Cubist), Random Forest (RF), K-nearest Neighbors (KNN), and eXtreme Gradient Boosting (XGB). The predictions from these five learners were then ensembled using seven approaches, namely Support Vector Machines with Polynomial Kernel (SVM), Cubist, GBM, KNN, RF, XGB, and a generalized additive model (GAM). For the urban areas, due to the computational burden, only three machine learners were run in the first stage of model development: GBM, RF, and XGB. Predictions from these models were ensembled using a GAM and the following super learners: SVM, GBM, RF, and XGB. The ensemble model with the lowest root mean squared error (RMSE) was chosen as the best performer. The R² values derived from ten-fold cross validation in the test sets ranged from 0.79 to 0.96(Amini et al., 2023b, 2023a). The values from these models were aggregated from grid-cell to ZIP code for analysis.

2.3. Outcome Definition

Our primary outcome was all-cause mortality. This data was obtained from the Medicare denominator file. Our secondary outcomes were all-cardiovascular mortality and all-respiratory mortality. Causes of death were extracted from the National Death Index (NDI) data, which was available for 2000 to 2016. We defined cardiovascular disease as ICD-10 codes beginning with “I” and respiratory diseases as ICD-10 codes beginning with “J”. Counts of the outcome were aggregated by zip code and year for the analysis.

2.4. Covariate Data

Covariate data included demographic, socioeconomic, and meteorological variables. We obtained demographic data on the beneficiaries from the Medicare denominator file. We aggregated this data to ZIP code-years and included: proportion of beneficiaries who are male, proportion who identify as Black, proportion who identify as race other than White or Black, proportion who were aged 65-74 years old, proportion who were aged 75-84 years old, and proportion who were also eligible for Medicaid.

Additional socioeconomic information was obtained from the Decennial Census and the American Community Survey: proportion of adults living below the poverty line and median household income(Manson et al., 2021). We also used proxy variables for access-to-care from the Dartmouth Atlas of Health Care: proportion of Medicare beneficiaries with at least one ambulatory care visit to a primary care clinician, proportion of diabetic Medicare beneficiaries aged 65-75 who received an annual eye exam, proportion of diabetic beneficiaries aged 65-75 who received an annual blood lipid tests, and proportion of female Medicare beneficiaries aged 67-69 receiving at least one mammogram during a two-year period(Dartmouth Atlas of Healthcare Data, 2022). We included data on population density derived from NASA’s Socioeconomic Data and Applications Center (SEDAC)(Center for International Earth Science Information Network, 2015). We further calculated the distance from the centroid of each ZIP code to the nearest hospital using ESRI data from 2010(ESRI Data and Maps; United States Geological Survey, 2010). We used linear interpolation to fill in missing values.

Meteorological variables included seasonal ambient temperature. We obtained temperature data from the gridMET dataset which estimated maximum and minimum daily temperature values from 2000-2018 on an approximately 4 km spatial scale(Abatzoglou, 2013). Values were aggregated to ZIP codes. Average daily temperature was defined as the mean of the maximum and minimum temperature. Cold-season temperature was defined as the average of the daily temperatures in January, February, March, October, November, and December of the same calendar year. Warm-season temperature was defined as the average of the daily temperatures in April-September in the same calendar year. These temperature values were weighed by population.

2.5. Statistical Analysis

We used a weighted quantile sum (WQS) approach to look at the relationship between exposure to PM_2.5 species, jointly and individually, and mortality in the study population. A WQS approach quantizes the multiple exposures and creates a weighted index that serves as the variable representing the mixture in the outcome regression. In order to estimate the weights, the data is divided into a training set and a validation set. The training data set is bootstrapped, and the regression model is run in each bootstrapped sample to determine the average weight for each component which optimized the algorithm used to estimate parameter effects. These weights are constrained to sum up to one, have values between zero and one, and be either non-negative or non-positive. After the weights have been determined, the final outcome regression is run in the validation dataset to determine the effect estimate of the weighted index (Carrico et al., 2015). The coefficient for this term represents the change in outcome due to a quantile change in the overall mixture. Individual toxicities can then be compared using the weights derived from the bootstrap approach in the training dataset. The equation for the model, using a quasi-Poisson distribution, can be seen below.

$$\log (CountsofOutcome)={\beta }_0+{\beta }_{1}WQS+{\beta }_{2}Covariates+s(year,3)+offset(\log (population))$$

Where

$$WQS=\sum _{i=1}^n{w}_i{q}_i$$

Here $i$ is a component of the mixture, $n$ is the total number of components in the mixture, $w_i$ is the weight of the $i\text{th}$ component, $q_i$ is the quantile of the $i\text{th}$ component, and $s$ is a natural cubic spline term for year with three degrees of freedom. A non-linear time trend was included to account for secular changes in pollution levels over time. This model assumes unidirectionality, namely that the relationship between all the components of the mixture and the outcome is in a single direction or zero. As it is unlikely that any constituent of PM_2.5 has a beneficial effect, this assumption seems reasonable in this context. We calculated 95% confidence intervals using robust standard errors to account for heteroskedasticity by zip code.

2.6. Sensitivity and Subgroup Analyses

We conducted several sensitivity analyses to check the robustness of our results and subgroup analyses to identify vulnerable populations. First, we looked at the association between one-year lags of exposure and the outcome. Then, we created a two-year average exposure for each component and the mixture. Next, we looked at the exposure outcome relationship in analyses stratified by sex, race, Medicaid eligibility status, and region. We did this by stratifying individual data by these characteristics, aggregating the ZIP code-years, and re-running the analyses. Finally, we conducted all the analyses again while restricting the dataset to ZIP code-years with total PM_2.5 exposure levels below 8 μg/m³. For these analyses, total PM_2.5 was defined as the sum of the mass of all fifteen components.

All data were processed and analyzed using R Statistical Software(R Core Team, 2023). Analyses were conducted using the “gwqs” package(Renzetti et al., 2021). This study was approved by the institutional review boards of Emory and the Harvard School of Public Health.

Results

The characteristics of the main dataset can be seen in Table 1. There were about 500,000 observations from over 28,000 ZIP codes. Across ZIP code-years, the average proportion of males was about 44%. The vast majority of beneficiaries were White and under the age of 85.

Table 1.

Demographic Characteristics of Data (2000-2018)

Total Observations (N)	506,897
ZIP Codes	28,631
Demographic Characteristic	Average Percentage Across ZIP Code-Years
Sex: Male	43.8 %
Race: Black	7.1%
Race: White	88.4%
Medicaid-Eligible	12.8%
Age Group: 65 ≤ Age < 75	50.3%
Age Group: 75 ≤ Age < 85	34.7%

Table 2 shows the distribution of the fifteen PM_2.5 components which were the exposures of interest in this study. Five main species contribute most to the mass of PM_2.5: EC, NH₄, NO₃, OC, and SO₄. The trace elements comprise a much smaller proportion. However, even among the major components, the concentrations are relatively low, a reflection of measures implemented through the years to control air pollution.

Table 2.

Distribution of PM2.5 Components Across ZIP Code-Years in μg/m³

	Minimum	10th Percentile	25th Percentile	Mean	50th Percentile	75th Percentile	90th Percentile	Maximum
EC	0.034	0.214	0.298	0.461	0.418	0.576	0.759	2.567
NH4	0.023	0.278	0.448	0.849	0.778	1.188	1.558	2.806
NO3	0.054	0.412	0.561	1.013	0.852	1.355	1.893	4.917
OC	0.316	1.060	1.296	1.718	1.621	2.038	2.494	6.763
SO4	0.122	0.771	1.160	2.166	1.928	3.047	4.032	6.824
Br	0.000	0.002	0.002	0.003	0.003	0.003	0.003	0.008
Ca	0.007	0.021	0.027	0.044	0.037	0.054	0.075	0.325
Cu	0.000	0.001	0.001	0.002	0.002	0.003	0.004	0.015
Fe	0.005	0.026	0.036	0.053	0.047	0.064	0.085	0.292
K	0.011	0.041	0.048	0.057	0.055	0.065	0.074	0.218
Ni	0.000	0.000	0.000	0.000	0.000	0.001	0.001	0.006
Pb	0.000	0.001	0.001	0.002	0.002	0.003	0.004	0.010
Si	0.012	0.048	0.061	0.100	0.089	0.128	0.164	0.593
V	0.000	0.000	0.000	0.001	0.001	0.001	0.002	0.009
Zn	0.001	0.003	0.005	0.007	0.006	0.009	0.012	0.046

Figure 1 shows the correlation matrix between the fifteen components across ZIP code-years. The correlations ranged from very weak to fairly strong. The largest correlation was between NH₄ and SO₄ at 0.87. This is expected as much of the NH₄ is in the form of ammonium sulfate and bisulfate(Chen et al., 2018; Seinfeld, 1989).

Figure 1.

Correlation matrix for PM2.5 species.

The main results of the study are summarized in Table 3. Each decile increase in the levels of PM_2.5 species increased the rate of all-cause mortality by 1.4% (95% CI: 1.3%-1.4%), CVD mortality by 2.1% (95% CI: 2.0%-2.2%), and respiratory mortality by 1.7% (95% CI: 1.5%-1.9%). These results held when the data used for analysis was restricted to ZIP code-years with total PM_2.5 levels less than 8 μg/m³, with effect estimates slightly higher in the restricted analyses (Table 3). For better comparison of the all-cause and cause-specific mortality analyses, we also ran an all-cause mortality restricted to 2000-2016. The results from this analysis were very similar to the analysis including all years.

Table 3.

Change in Mortality per Decile Increase in PM_2.5 Species in Full Dataset and Restricted to ZIP Code-Years with Low PM_2.5 Concentrations

Main results in full dataset
Model	Rate Ratio	Lower 95% CI	Upper 95% CI
All-cause Mortality	1.014	1.013	1.014
All-cause Mortality 2000-2016	1.014	1.013	1.014
CVD Mortality*	1.021	1.020	1.022
Respiratory Mortality*	1.017	1.015	1.019
Main results for ZIP code-years with total annual PM2.5 <8 μg/m3
All-cause Mortality	1.015	1.014	1.016
All-cause Mortality 2000-2016	1.015	1.014	1.016
CVD Mortality*	1.023	1.021	1.024
Respiratory Mortality*	1.020	1.019	1.022

^*Data availability limited to 2000-2016

In the model for all-cause mortality, the largest contributors to the mixture’s effects were OC, zinc, nickel, sulfate, and vanadium, respectively. OC can be a primary or secondary pollutant from various sources including traffic, combustion, biomass burning, and reactions of volatile organic chemicals(Amini et al., 2022). The considerably higher correlation of OC with EC than with SO₄ suggests that in this analysis, most of the OC is from traffic or biomass. Zinc can come from non-tail-pipe traffic emissions (particularly tire wear) and the metals industry. Again, the high correlation with EC suggests a primarily traffic source. Nickel and vanadium are typically from heavy oil use and metals industry, but can also be traffic related, as seen in their high correlation with EC, and sulfate is a byproduct of coal combustion(Thurston et al., 2011). For cardiovascular mortality, the largest contributors were nitrate, sulfate, ammonium, nickel, and vanadium, respectively. Nitrate is typically an indicator of traffic pollution and secondary pollution, and ammonium comes primarily from chemical reactions in the air between ammonia and nitrogen oxides and sulfur dioxide gases (Amini et al., 2022; Thurston et al., 2011). For respiratory mortality, the largest contributors were OC and nitrate, followed by relatively similar weights for the other species. In the low concentration analyses, similar trends persisted for all-cause mortality and respiratory mortality. For CVD mortality, the results were also largely similar, except that OC was one of the top five contributors to the mixture (Figure 2).

Figure 2.

A) Weights from main models at all concentration levels B) Weights from main models at ZIP code-years with total PM2.5 levels less than 8 μg/m3.

In the sensitivity analyses, using the two-year calendar average for each species and using the previous year’s levels as the exposures of interested resulted in fairly similar effect estimates, both in the main analyses and in the low-exposure analyses (Figure 3). The highest weight contributors were also similar to the same-year exposure with the same five elements represented in the highest weight categories (Tables S1 & S2).

Figure 3.

A) Sensitivity analyses at all concentrations. B) Sensitivity analyses at zip code-years with total PM2.5 levels less than 8 μg/m3.

In the stratified analyses, we found that those who identified as Black had an increased rate of mortality due to air pollution exposure as compared to those who identified as White. We further found that men had an increased rate of death as compared to women. Amongst the US Census divisions, the highest rates of mortality were found in the East and West North Central Division states and the East and West South Central Division States and the lowest rates were found in the Middle Atlantic Division states (Figure 4). This was largely consistent in the lower concentration analyses as well.

Figure 4.

A) Stratified analyses at all concentrations. B) Stratified analyses at zip code-years with total PM2.5 levels less than 8 μg/m3.

In terms of the weights of each species in the stratified analysis, we found ammonium, zinc and copper to have higher weight for those who identify as Black and for OC, sulfate and nickel to have higher weight for those who identify as White. Copper is an indicator of non-tail-pipe traffic emissions, mostly from brake wear(Schauer et al., 2006). Men and women had similar high-weight species, though weights were higher for sulfate and nickel were among men than women. For those who were Medicaid-eligible, copper, potassium, iron, and vanadium had the highest weights while OC, zinc, and vanadium represented the highest weights for those who were not Medicaid-eligible. Iron is an indicator of pollution of the steel industry, but also comes from traffic, and had a high correlation with EC and Cu(Thurston et al., 2011). The weights varied largely by region. For example, in the East North Central region, the highest weights were seen for copper and sulfate, while in the East South Central region the highest weights were for copper and nitrate. In the Middle Atlantic region, the weights seemed fairly evenly distributed among the species (Figure S1).

At lower concentrations, the most toxic species were somewhat changed in the stratified analyses. We found higher weights for OC, zinc, potassium, and nitrates for blacks and OC, sulfate, and nickel for Whites. Potassium is an indicator of pollution resulting from biomass burning(Thurston et al., 2011). Once again, men and women had similar high-weight species, most notably OC and zinc. For those who were Medicaid-eligible, copper, potassium, and OC had the highest weights and OC, zinc, and nickel had the highest weights for those who were not Medicaid-eligible. The weights by region varied largely and did not suggest a trend (Figure S2).

Discussion

In this study, we used a high-resolution model for 15 PM_2.5 species and examined its relationship with all-cause mortality in the Medicare population using all species together in a modern mixture model. We found an increased rate of all-cause mortality, cardiovascular mortality, and respiratory mortality associated with long-term exposure to joint PM_2.5 species as a mixture. Blacks were disproportionately affected, with an effect size roughly twice that of Whites, indicating that environmental injustice is more than differential exposure, but also includes differential susceptibility. There was also a higher effect size among those who identified as male, and those who lived in the East and West North Central Divisions. The increased rate of mortality persisted at lower concentrations well below EPA standards. We found OC, nickel, zinc, and sulfate were the main drivers of PM_2.5 toxicity in our main analysis. OC suggests sources such as traffic emissions, industrial pollution, and biomass burning. Nickel is indicative of heavy oil use but also of traffic, and zinc is indicative of non-tail-pipe traffic emissions and those from the metals industry. Sulfate is mainly a marker of coal use. The highest weights for cardiovascular mortality were seen for nitrate which is primarily a marker of traffic pollution. For respiratory disease, OC had the highest toxicity. Taken together these findings indicate that coal burning powerplants, which are sources of sulfate, nitrate, and secondary organic aerosols are a key source of PM-derived mortality and the other key source is traffic. The finding of associations below current and proposed standards in the U.S. indicates that further controls should be focused on these two sources. The finding of higher susceptibility for Blacks indicates this is an important environmental justice issue.

Our study is most directly comparable to another study looking at the relationship between PM_2.5 species and mortality in the Medicare population. That study used two exposure methods to look at PM_2.5 species, one of which was the model we utilized as well. They found an increased risk of mortality with both models. However, they found sulfate and ammonium to have the largest effects while we found OC to be the most influential species but also identified sulfate. Their study analyzed components individually and did not use a mixture model incorporating all components simultaneously. For one exposure, it also did not include the trace elements we included. (Hao et al., 2023). In a previous study looking at long-term exposure to PM_2.5 species and mortality in all age groups in Massachusetts using a WQS approach, we found a non-accidental mortality rate ratio of 1.0303 (95% CI: 1.0265-1.0341) which is higher than our national estimate of 1.014 (95% CI: 1.013-1.014) for all-cause mortality. Similar to this study, we found higher effect estimates for cardiovascular mortality and respiratory mortality in the previous study. We also found OC to be an important component in the effect of the mixture, though we did not find potassium and iron to be as important in this study as it was in Massachusetts(Jin et al., 2022). Another study conducted in the Medicare population from 2000 to 2008 looking at PM_2.5 constituents for those beneficiaries who lived within a 12-mile radius of an EPA monitor found a mortality hazard ratio of 1.062 (95% CI: 1.052, 1.072) for each IQR increase PM_2.5 levels. In their single pollutant models, researchers found that lead, selenium, nitrate, OC, and silicon had the highest association with mortality(Kazemiparkouhi et al., 2022). Similar to this study, we also found OC to be an important species in all-cause mortality. In an analysis of the American Cancer Society Cancer Prevention Study II (ACS CPS II) examining PM_2.5 constituents and mortality, researchers found an increase in the hazard ratio of death after long-term exposure. The components with highest effect estimates for all-cause mortality were arsenic and sulfur(Thurston et al., 2013). We did not study arsenic but we did include sulfate in our models and found it to be an important contributor to all-cause mortality as part of the PM_2.5 mixture.

As noted above, the key components of PM driving the mortality associations are traffic, electricity generating plants, and biomass burning, where wildfires are a growing source. The Clean Air Transport Rule and increased competition from natural gas plants have resulted in a large decline in coal based electric power generation in the U.S. in the last 15 years(US Energy Information Administration, 2023). However, some coal plants still do not incorporate scrubbers as sulfur controls(US Energy Information Administration, 2011). Moreover, secondary organic aerosols continue to be of concern, in part because NOx is a key input to their production and is also produced by gas-fired powerplants(Berkemeier et al., 2016; Venkatesh et al., 2012; Xu et al., 2020). Traffic is a major source of primary OC, and large vehicles such as diesel trucks and off-road construction equipment are important contributors(Hu et al., 2010; Lane et al., 2007). These sources will be important targets for further action to lower the key PM species. Moreover, brake and tire wear will continue even with electric vehicles and research on how to limit those emissions will be needed.

Our study had several notable strengths. We used comprehensive data which included all individuals 65 years or older who are enrolled in Medicare. This contrasts with the non-representative samples in many cohort studies. This large nationwide dataset gave us enough power to conduct extensive subgroup analyses. Moreover, we used a mixtures analysis to account for the fact that several of the PM_2.5 species were strongly correlated with one another, without the use of single species models. We adjusted extensively for covariates including demographic, socioeconomic, and temperature variables. Finally, our exposure model had strong prediction metrics on a very fine spatial scale, 50 meters in urban areas and 1 km in non-urban areas. This allowed us to assign exposure to cohort members living across the contiguous US and to capture fine-scale variations within these exposures.

Our study had several limitations. Firstly, although our exposure assignment was based on a highly spatio-temporally resolved modeling approach, and this approach has strong validation metrics, it is possible to have some measurement error, both due to error in the modeling approach and Berkson error resulting from aggregation. However, simulation studies have suggested this error to be minimal and biased towards the null(Wei et al., 2022). There are also other PM_2.5 species that we did not have exposure models for, but these would have made up a very small portion of the PM2.5 mass. Secondly, the assumption of unidirectional effects for all PM_2.5 species was untestable. We believe that this is reasonable given that there is no reason to believe that any of the PM_2.5 species have beneficial effects. Thirdly, our data was derived from an administrative data source which was not collected for purposes of research. We tried to ameliorate this issue by collecting area-level data from several data sources including the Census, ACS, and Dartmouth Health Atlas, though some components of these data sources are prone to measurement error themselves. Finally, as with any epidemiological study, there may be unmeasured or residual confounding. Socioeconomic status is a very complex concept, and it is possible that despite the inclusion of numerous variables, we may have not fully adjusted for this issue.

Conclusion

Exposure to PM_2.5 as a mixture increased the risk of all-cause mortality in the Medicare population. Our analyses suggest several sources of exposure including combustion of fossil fuels, traffic sources, industrial sources, biomass burning, etc. We found variations in toxicity by demographic characteristics and region. The increased rate of mortality persisted when restricting to zip code-years with exposure to lower PM_2.5 concentrations, indicating that current and EPA proposed PM_2.5 standards are inadequate to protect public health.

Supplementary Material

Supplementary Tables. Sensitivity analyses-Weights for PM2.5 species.

Table S1. Sensitivity analyses at all concentrations-Weights for PM_2.5 species.

Table S2. Sensitivity analyses at lower concentrations-Weights for PM_2.5 species.

Figure S2. Stratified analyses at lower concentrations- Weights for PM2.5 species

Figure S2. Stratified analyses at lower concentrations- Weights for PM_2.5 species

Figure S1. Stratified analyses at all concentrations- Weights for PM2.5 species

Figure S1. Stratified analyses at all concentrations- Weights for PM_2.5 species

Data Sharing:

The authors are not permitted under the terms of their Data Use Agreement with the Centers for Medicare and Medicaid Services (CMS) to share their data. Individuals may apply to CMS to obtain the data.