Source-Specific PM_2.5 and Atherosclerotic Cardiovascular Disease Mortality

Abstract

BACKGROUND

Fine particulate matter (PM_2.5) exposure is adversely linked to atherosclerotic cardiovascular disease (ASCVD). However, most studies focused on PM_2.5 mass rather than its chemical composition and specific sources. Particulate pollution sources can have distinct, cumulative, and potentially synergistic health impacts. We investigated the associations of source-specific PM_2.5 exposure with ASCVD mortality in the United States, considering the combined associations and regional variations.

METHODS

We used data from the Centers for Medicare & Medicaid Services (including data from 65,838,403 participants) from 2000 to 2016. We estimated PM_2.5 exposure using machine-learning models and attributed components to five source categories. We used Poisson survival models to assess the associations with the source categories.

RESULTS

Higher ASCVD mortality rate (rate ratio [95% confidence interval (CI)] per interquartile range increase) was associated with oil combustion (1.051 [1.049 to 1.052]), industrial pollution (1.054 [1.052 to 1.056]), coal and biomass burning (1.065 [1.062 to 1.067]), and motor vehicle pollution (1.044 [1.042 to 1.046]). These associations persisted even after limiting our sample to ZIP code–years with PM_2.5<9 μg/m³ — the current National Ambient Air Quality Standard. In these areas the observed rate ratio for a one-unit increase in PM_2.5 mass was 1.028 (95% CI, 1.026 to 1.029).

CONCLUSIONS

We found higher ASCVD mortality rate associated with PM_2.5, with differential effects across sources. These data highlight the importance of considering local population characteristics and exposure patterns when assessing health risks associated with PM_2.5.

Introduction

Lifestyle factors such as physical activity, weight loss, smoking, diet, and stress,¹ and environmental factors such as greenness, temperature, and air pollution,² play a critical role in developing cardiovascular diseases (CVD). Particulate matter in the atmosphere with a mass median aerodynamic diameter less than or equal to 2.5 μm (PM_2.5) has been conclusively linked to the development of CVD,³ particularly atherosclerotic cardiovascular disease (ASCVD).⁴

Air pollution results from complex chemical reactions involving various emissions. Although research often focuses on the mass of PM_2.5, its chemical composition is crucial. The composition of PM_2.5 is complex, and the physical structures, chemical compositions, and sources of its components are diverse.⁵ Therefore, the individual chemical components of PM_2.5 may have distinct, cumulative, and potentially synergistic health impacts. A few studies have documented the links between some components of PM_2.5 and various cardiovascular health–related outcomes.^6–10 However, these results have been mixed,¹¹ and evidence for the combined effects of components is very limited.¹²

In the United States, the distribution of PM_2.5, its chemical components, and its sources are heterogeneous, revealing patterns of exposure that vary from one community to another.¹³ The variability has implications, as studies have demonstrated that subgroups within the population, such as racial or ethnic minorities, the older adult population, and individuals with low socioeconomic status, are disproportionately affected by adverse CVD outcomes.¹⁴ Differences between communities can also be present due to differences in customs, traditions, and lifestyles — often a combination of personal, sociodemographic identities and geographic contexts affect CVD outcomes.

Investigating the combined effects of PM_2.5 chemical components poses two challenges. First, the analysis requires high-quality, component-specific data — with high resolution and low uncertainty — to assign exposures with minimal misclassification. Additionally, data must be available over large areas to allow for comparisons between communities. Second, the analysis of these data requires statistical methods specifically designed for handling the highly correlated structure of component data and extracting meaningful estimates. This study leverages PM_2.5 and component-specific spatial datasets with high resolution and low uncertainly as well as statistical methods capable of handling the complexities of the data to examine the relationships between PM_2.5, its sources, and ASCVD mortality, considering their combined effects, as well as the variations among regions across the United States.

Methods

STUDY POPULATION AND MORTALITY DATA

We obtained two nationwide databases from the Centers for Medicare & Medicaid Services (CMS) for the period 2000 through 2016, including the Master Beneficiary Summary File (MBSF) and the Medicare National Death Index (NDI) segment. Medicare covers over 95% of the population 65 years of age and above in the United States and enrolls new members annually. The outcome tested in this study (extracted from the single underlying cause of death) was ASCVD mortality, defined as either ischemic heart disease or cerebral infarction mortality (International Statistical Classification of Diseases and Related Health Problems, Tenth Revision [ICD-10] codes I20, I21, I22, I23, I24, I25, I63, I65, I66). Individuals included in this study entered the cohort on January 1 of the year after they enrolled in Medicare and were followed for each calendar year until death, end of enrollment in Medicare, or end of follow-up, whichever came first.

The CMS approved this study under the data use agreement (#RSCH-2020–55,733). The Institutional Review Board of Emory University (#STUDY00000316), and the Institutional Review Board of Mount Sinai (STUDY 20–01344) also approved this study and granted a waiver of informed consent. The Medicare dataset was stored and analyzed in the Rollins High-Performance Computing Cluster at Emory University in compliance with the Health Insurance Portability and Accountability Act (HIPAA).

EXPOSURE DATA

We obtained annual mean predictions for 14 PM_2.5 components, including sulfate (SO₄²⁻), nitrate (NO₃⁻), organic carbon (OC), elemental carbon (EC), zinc (Zn), vanadium (V), potassium (K), silicon (Si), lead (Pb), nickel (Ni), iron (Fe), copper (Cu), calcium (Ca), and bromine (Br), at a 50-m spatial resolution for urban areas and 1-km resolution for nonurban areas across the contiguous United States (2000 through 2016) from super-learning approach models.^15,16 The monitoring data used for developing the models had a precision of 0.1 μg/m³ for PM_2.5 mass and major components, and a precision of 0.1 ng/m³ for trace elements. Additional details about the PM_2.5 components’ exposure data can be found elsewhere,^15,16 and the data are publicly accessible at https://sedac.ciesin.columbia.edu/.^17,18 We calculated population-weighted annual averages for each pollutant for each ZIP code (i.e., the finest spatial resolution in Medicare data), considering that the population is not distributed evenly within the ZIP code. Specifically, we regridded the modeled exposure dataset at a 1-km resolution into a common grid consistent with the finest resolution (30 arc-seconds, ~1 km at the equator) of the population density dataset from NASA’s Gridded Population of the World. The exposure data within each ZIP code polygon were aggregated by population-weighted average for each year over the study period (2000 through 2016) and assigned to each Medicare beneficiary based on the calendar year and ZIP code of residences, which are tracked and updated annually.

COVARIATES

We obtained individual-level demographics (age, sex, and race) and Medicaid insurance status from the Medicare MBSF. We created a 5-year age-at-entry group variable based on individual age at entry into the Medicare cohort. To further adjust our models we obtained ZIP code–level socioeconomic characteristics¹⁹ (population density, percentage of the Black population, percentage of the Black Hispanic population, percentage of the population 65 years of age or above living below the Census Bureau’s poverty threshold, percentage of the population receiving public assistance, percentage of the population living in rental houses or apartments, the median household income, and the percentage of the population with less than a high school education), meteorological variables²⁰ (annual mean summer and winter temperatures), land-use variables²¹ (normalized difference vegetation index [NDVI] values), county-level health care capacity indicators²² (number of hospitals and active medical doctors per 1000 people), and a geographic region variable categorized as five U.S. regions²³ (West, Midwest, Northeast, Southeast, and Southwest, Fig. 1). Details and data sources of the covariates are described in Ma et al. (2022).²⁴

Figure 1. Map of Geographical Regions (West, Midwest, Northeast, Southeast, Southwest) in the Contiguous United States.

STATISTICAL METHODS

Descriptive Analysis

Descriptive statistics were used for all included variables for the whole of the United States and stratified for each U.S. region. Pearson correlation was used to examine the correlations for total PM_2.5 and PM_2.5 components.

Single-Pollutant PM_2.5 Analysis

We used Poisson survival analyses using the Anderson–Gill formulation with adjusted individual-level covariates (sex, race, Medicaid eligibility, 5-year categories of age at study entry), as well as the area-level covariates (socioeconomic, meteorological, land use, and health care use variables). We generated an artificial Poisson observation for each person in the risk set at each noncensored year with an outcome of zero for all person-years except those experiencing the event in that calendar year. Person-years (with the same sex, race, Medicaid eligibility, and 5-year age categories at study entry) in the same ZIP code and calendar year were aggregated and treated as interchangeable in the analysis. We allowed a piecewise constant hazard for each year using indicator variables for each calendar year. We then repeated the models within each U.S. region separately, to identify potential differential effects by geographical region (West, Midwest, Northeast, Southeast, and Southwest) and overall. Results are presented in rate ratio with 95% confidence interval (CI) per interquartile range increase for PM_2.5 concentrations.

NONNEGATIVE MATRIX FACTORIZATION OF 14 PM_2.5 COMPONENTS

To evaluate the independent associations of PM_2.5 sources, we used nonnegative matrix factorization (NMF) to group PM_2.5 components into source categories. This method identifies PM_2.5 source categories containing a mixture of pollutants associated with higher ASCVD mortality rate.¹³ We tested for four to six possible factors and conducted 100 multiple iterations to obtain the best factorization fit. We selected five factors based on the inflection point on the residuals sum of squares curve, a visual inspection of the mixture coefficient matrix, and an assessment of the correlation between the factors. We then labeled the source categories based on the high-loading components, known as tracers, of PM_2.5 sources.

ASSOCIATIONS BETWEEN PM_2.5 SOURCES AND ASCVD MORTALITY

Since the correlation among the individual factors was modest, from −0.39 to 0.10, we considered the derived five NMF loading factors as coexposures and fitted the outcome model as a multivariable Poisson regression. We used the same model construction, covariates, and stratification structure as the single-pollutant models.

ZIP CODE–YEARS BELOW THE ANNUAL NAAQS PM_2.5 THRESHOLD OF 9 μG/M³

We subset our dataset to ZIP code–years with an estimated annual exposure below 9 μg/m³ PM_2.5 to evaluate the associations between PM_2.5, the source categories, and ASCVD in exposure levels lower than the new annual National Ambient Air Quality Standards (NAAQS) for PM_2.5. First, we repeated the PM_2.5 and ASVCD model. Then, we repeated the multivariable regression evaluating the associations of the five source categories with ASCVD. In this analysis, we report the associations per one-unit increase to allow for direct comparisons among datasets.

All computational analyses were completed with the support of the Rollins High-Performance Computing cluster at Emory University and conducted using R (version 4.0.2) with the mgcv R package.

Results

DESCRIPTIVE STATISTICS

We included a total of 65,838,403 persons amounting to 559,591,306 person-years of follow-up; 7.2% of these people died of ASCVD, with mortality rates per 1000 person-years in the Northeast of 9.80 and Midwest of 8.45 compared with the West (7.71), Southwest (8.01), or Southeast (7.82). The mean age at entry into the dataset was 71.0 years; about 56% of individuals included in the study were female; 84.7% were White, and 10.4% had Medicaid insurance (Table 1). Table 2 shows the national and regional summary statistics of the exposures to PM_2.5, its 14 chemical components, and the NMF source categories. Overall, the average PM_2.5 mass level was 9.9 μg/m³.

Table 1.

Descriptive Characteristics of the Study Population and Area-Level Covariates among the Cohort, by U.S. Region and Overall.^*

Variables	Whole United States	West	Midwest	Northeast	Southeast	Southwest
Person-years	559,591,306	100,702,965	130,554,793	123,441,954	146,829,687	58,061,907
Population	65,838,403 (100)	12,469,527 (100)	15,761,151 (100)	14,943,121 (100)	18,365,091 (100)	7,421,795 (100)
Cause of death
ASCVD	4,708,030 (7.2)	776,773 (6.2)	1,103,724 (7.0)	1,209,118 (8.1)	1,148,885 (6.3)	469,530 (6.3)
Other all-cause	20,944,933 (31.8)	3,506,867 (28.1)	5,112,426 (32.4)	4,525,940 (30.3)	5,650,701 (30.8)	2,148,999 (29.0)
Age at entry, years	71.0±7.4	70.8±7.1	71.3±7.5	71.5±7.6	70.8±7.2	70.5±7.0
Sex
Male	28,783,373 (43.7)	5,602,097 (44.9)	6,846,587 (43.4)	6,374,429 (42.7)	7,976,529 (43.4)	3,295,903 (44.4)
Female	37,055,030 (56.3)	6,867,430 (55.1)	8,914,564 (56.6)	8,568,692 (57.3)	10,388,562 (56.6)	4,125,892 (55.6)
Race
White	55,753,469 (84.7)	10,107,533 (81.1)	14,256,366 (90.5)	12,624,785 (84.5)	15,196,569 (82.7)	6,286,933 (84.7)
Black	5,358,398 (8.1)	469,124 (3.8)	977,175 (6.2)	1,300,423 (8.7)	2,385,373 (13.0)	451,382 (6.1)
Other	4,726,536 (7.2)	1,892,870 (15.2)	527,610 (3.3)	1,017,913 (6.8)	783,149 (4.3)	683,480 (9.2)
Medicaid eligibility
Eligible	6,853,852 (10.4)	1,733,607 (13.9)	1,070,680 (6.8)	1,503,191 (10.1)	2,110,990 (11.5)	824,486 (11.1)
Area-level covariates
Population density, people per km2	1114.8±2930.5	1498.5±2099.3	582.6±1216.8	2413.9±5606.2	522.8±973.4	604.9±817.8
Black, %	11.8±18.6	4.5±7.4	8.0±17.5	11.0±18.3	21.6±22.4	8.6±12.9
Black Hispanic, %	0.3±0.6	0.2±0.2	0.1±0.2	0.5±1.0	0.3±0.4	0.2±0.2
Below the poverty level, %	14.4±9.2	14.3±8.5	12.7±8.6	11.3±8.8	16.9±9.0	17.7±9.9
Receiving public assistance, %	22.8±10.8	22.1±10.8	23.7±11.0	23.0±10.8	22.9±10.8	22.1±10.6
Renting a house or apartment, %	32.1±17.9	39.4±18.1	27.5±15.7	33.0±21.3	30.2±15.3	32.6±16.9
Median household income, US$1,000	50.7±22.5	56.1±23.5	48.8±17.6	61.1±27.2	43.5±18.5	45.0±19.2
Not graduated from high school, %	13.2±7.5	12.7±8.3	10.9±5.7	10.9±6.5	15.7±7.4	15.9±8.4
Number of hospitals	11.3±20.4	24.9±36.4	8.1±15.1	7.8±6.4	5.2±6.8	17.1±21.8
Number of active medical doctors	2878.0±5565.4	6609.8±9613.8	2076.1±4876.3	2880.7±3693.4	1136.1±1881.0	2783.8±4129.4
NDVI	0.29±0.07	0.22±0.07	0.28±0.05	0.30±0.07	0.34±0.04	0.26±0.08
Summer mean temperature, °C	23.7±3.4	21.6±3.7	22.3±2.0	21.8±2.1	25.9±2.1	28.1±2.7
Winter mean temperature, °C	3.9±7.0	7.5±6.1	−3.2±3.9	−0.7±3.1	8.3±5.7	8.9±4.1

^*Summary statistics for the study population are presented as number (%) mean±SD. “Other” in the race group includes Asian, Hispanic, American Indian or Alaskan Native, and unknown. ASCVD denotes atherosclerotic cardiovascular disease mortality; NDVI, normalized difference vegetation index; PM_2.5 fine particulate matter (with a mass median aerodynamic diameter less than or equal to 2.5 μm); and SD, standard deviation.

Table 2.

Summary Statistics Presented as Mean±SD, (Interquartile Range) of PM_2.5 Particle Component Concentrations and NMF Loading Factors over the Study Period, by U.S. Region and Overall.^*

Variable	Whole United States	West	Midwest	Northeast	Southeast	Southwest
PM2.5 major components
SO42− (μg/m3)	2.3±1.2 (1.9)	1.1±0.7 (0.6)	2.3±1.0 (1.4)	2.6±1.1 (1.8)	2.9±1.1 (2.0)	2.0±0.9 (1.6)
NO3− (μg/m3)	1.1±0.6 (0.9)	1.3±0.9 (1.3)	1.6±0.6 (0.8)	1.1±0.5 (0.6)	0.7±0.3 (0.3)	0.7±0.3 (0.4)
OC (μg/m3)	1.9±0.7 (0.8)	2.3±1.0 (1.4)	1.5±0.4 (0.5)	1.7±0.4 (0.5)	2.0±0.5 (0.8)	1.9±0.6 (0.7)
EC (μg/m3)	0.5±0.3 (0.4)	0.6±0.4 (0.6)	0.5±0.2 (0.3)	0.6±0.3 (0.4)	0.5±0.2 (0.2)	0.5±0.2 (0.3)
PM2.5 trace elements
Zn (ng/m3)	7.8±4.2 (4.8)	7.1±3.9 (5.1)	9.0±4.7 (5.3)	1.0±4.7 (5.2)	6.2±2.5 (2.9)	6.6±3.3 (4.2)
V (ng/m3)	1.0±0.9 (0.9)	1.0±0.9 (1.1)	0.6±0.4 (0.5)	1.3±1.2 (1.6)	1.1±0.9 (1.0)	0.9±0.7 (0.8)
K (ng/m3)	58.9±14.8 (17.2)	60.6±17.2 (17.3)	53.7±9.4 (11.0)	48.6±7.6 (8.7)	63.6±10.9 (13.9)	73.0±19.0 (19.2)
Si (ng/m3)	104.8±55.7 (69.9)	116.4±59.0 (72.7)	85.5±33.7 (41.5)	60.6±15.9 (20.9)	107.7±29.5 (46.8)	193.6±64.0 (57.2)
Pb (ng/m3)	2.2±1.3 (1.6)	1.9±1.2 (1.7)	2.6±1.5 (2.0)	2.7±1.4 (2.2)	1.9±0.9 (1.1)	1.8±0.9 (0.9)
Ni (ng/m3)	0.6±0.6 (0.5)	0.6±0.5 (0.8)	0.4±0.3 (0.5)	1.1 ± 1.1 (1.0)	0.5±0.3 (0.4)	0.4±0.3 (0.4)
Fe (ng/m3)	61.7±29.0 (34.6)	76.0±36.8 (55.8)	53.5±24.2 (28.1)	55.4±26.8 (38.9)	56.1±17.8 (23.8)	79.4±32.7 (36.0)
Cu (ng/m3)	2.5±1.7 (2.3)	3.4±2.3 (3.8)	1.8±1.3 (1.8)	2.6±1.5 (2.2)	2.2±1.3 (1.8)	2.6±2.0 (2.7)
Ca (ng/m3)	46.0±25.9 (28.1)	51.8±30.9 (31.0)	48.4±18.5 (25.1)	28.0±9.2 (11.6)	38.3±12.9 (16.7)	82.7±30.6 (34.8)
Br (ng/m3)	2.8±0.7 (0.8)	2.7±1.0 (1.5)	2.6±0.7 (0.9)	2.8±0.6 (0.8)	2.9±0.4 (0.6)	2.7±0.6 (0.6)
Total PM2.5 mass
PM2.5 (μg/m3)	9.9±3.2 (4.0)	8.9±4.5 (5.6)	10.5±2.8 (3.7)	9.9±2.8 (4.2)	10.4±2.6 (3.9)	9.2±2.7 (3.3)
NMF loadings
Oil combustion	7.4±4.1 (4.1)	8.2±3.6 (4.2)	7.5±2.7 (3.2)	10.3±5.9 (5.4)	5.3±2.7 (3.3)	6.2±2.4 (3.2)
Soil and dust	7.5±4.2 (5.4)	7.9±4.3 (4.9)	6.6±2.9 (4.1)	3.7±1.4 (1.5)	8.1±2.3 (3.4)	14.0±4.5 (4.6)
Industrial pollution	6.1±3.1 (4.1)	3.7±2.2 (2.0)	9.3±2.7 (3.5)	6.8±2.4 (2.8)	5.2±2.3 (3.1)	4.9±2.1 (3.0)
Coal and biomass burning	9.7±5.4 (7.6)	8.4±4.2 (6.3)	5.2±3.3 (4.6)	11.0±5.0 (7.0)	13.7±4.8 (6.7)	8.8±4.9 (6.9)
Motor vehicle pollution	9.3±5.9 (7.5)	14.6±7.4 (11.7)	10.2±4.8 (6.5)	8.1±4.8 (6.2)	6.7±3.8 (4.5)	7.8±5.5 (7.1)

^*Summary statistics for air pollution concentrations are presented as mean±SD (interquartile range). Br denotes bromine; Ca, calcium; Cu, copper; EC, elemental carbon; Fe, iron; K, potassium; Pb, lead; Ni, nickel; NMF, nonnegative matrix factorization; NO³⁻, nitrate; OC, organic carbon; PM_2.5 fine particulate matter (with a mass median aerodynamic diameter less than or equal to 2.5 μm); SD, standard deviation; Si, silicon; SO₄²⁻, sulfate; V, vanadium; and Zn, zinc.

ANALYSIS BASED ON OVERALL PM_2.5 CONCENTRATIONS

We found an association between ASCVD mortality and overall PM_2.5 mass exposure (rate ratio 1.012, 95% CI, 1.011 to 1.012 per one μg/m³ increase; and rate ratio 1.048, 95% CI, 1.046 to 1.050 per interquartile range, i.e., 4.0 μg/m³, increase). Regional models showed different associations across U.S. regions, and these are shown in Table S1.

NONNEGATIVE MATRIX FACTORIZATION (NMF) OF 14 PM_2.5 COMPONENTS

Based on the NMF results, we labeled the first category as “oil combustion” based on high loadings of Ni and V; the second category as “soil and dust” based on high loadings of Si and Ca; the third category as “industrial pollution” based on high loadings of SO₄²⁻, NO₃⁻, Z, and Pb; the fourth category as “coal and biomass burning” based on high loadings of OC and Br; and the fifth category as “motor vehicle pollution” based on high loadings of EC, Cu, and Fe (Fig. S1). Exposure levels of the different source categories also differed by region, with notable exposures to oil combustion in the Northeast, to soil and dust in the Southwest, to industrial pollution in the Midwest, to coal and biomass burning in the Southeast, and to motor vehicle pollution derived PM_2.5 in the West, as shown in Table 2.

The correlation among the 14 PM_2.5 components varied greatly, with correlations ≥0.75 for Cu and Fe (correlation coefficient [r]=0.82), Cu and EC (r=0.78), Fe and EC (r=0.76), Si and Ca (r=0.80), and Z and Pb (r=0.75). Overall PM_2.5 mass had a correlation≥0.60 with SO₄²⁻ (r=0.75), OC (r=0.62), and Br (r=0.62) (Fig. S2). The correlation between the source categories ranged from r=0.10 (for oil combustion and motor vehicle pollution) to r=−0.39 (for the oil combustion and soil and dust sources). Among source categories, industrial sources had a correlation with PM_2.5 mass of r=0.50; all correlation data are shown in Figure S3.

ASSOCIATIONS BETWEEN PM_2.5 SOURCES AND ASCVD MORTALITY

In the overall national results, ASCVD mortality rate (rate ratio per interquartile range increase) was adversely associated with oil combustion (rate ratio 1.051, 95% CI, 1.049 to 1.052), industrial pollution (rate ratio 1.054, 95% CI, 1.052 to 1.056), coal and biomass burning (rate ratio 1.065, 95% CI, 1.062 to 1.067), and motor vehicle pollution sources (rate ratio 1.044, 95% CI, 1.042 to 1.046). Stratified by regions, exposure to the oil combustion source category was associated with increased ASCVD mortality across most of the United States, except for the West (rate ratio 1.095, 95% CI, 1.089 to 1.101 in the Northeast; rate ratio 1.088, 95% CI, 1.084 to 1.092 in the Southeast; rate ratio 1.042, 95% CI, 1.036 to 1.049 in the Southwest; and rate ratio 1.079, 95% CI, 1.074 to 1.083 in the Midwest). Similarly, industrial pollution was associated with increased ASCVD mortality in most U.S. regions except for the Northeast (rate ratio 1.017, 95% CI, 1.014 to 1.020 in the West; rate ratio 1.022, 95% CI, 1.017 to 1.027 in the Midwest; rate ratio 1.099, 95% CI, 1.093 to 1.104 in the Southeast; rate ratio 1.056, 95% CI, 1.047 to 1.064 in the Southwest). The PM_2.5 attributed to soil and dust sources was adversely associated with ASCVD rate only in the eastern regions and the Midwest (rate ratio 1.033, 95% CI, 1.027 to 1.039 in the Northeast; rate ratio 1.074, 95% CI, 1.069 to 1.080 in the Southeast; rate ratio 1.071, 95% CI, 1.065 to 1.078 in the Midwest). ASCVD mortality associations with coal and biomass burning PM_2.5 were present in the western regions (rate ratio 1.087, 95% CI, 1.079 to 1.094 in the West; rate ratio 1.043, 95% CI, 1.038 to 1.048 in the Midwest; rate ratio 1.071, 95% CI, 1.062 to 1.079 in the Southwest). Finally, ASCVD mortality associations with the motor vehicle source category were present across all U.S. regions except for the Southeast (rate ratio 1.054, 95% CI, 1.047 to 1.062 in the West; rate ratio 1.036, 95% CI, 1.030 to 1.041 in the Midwest; rate ratio 1.042, 95% CI, 1.037 to 1.047 in the Northeast; rate ratio 1.044, 95% CI, 1.035 to 1.052 in the Southwest) (Fig. 2 and Table S1).

Figure 2. Rate Ratio of ASCVD Mortality per Interquartile Range Increase in PM_2.5 Mass and Each NMF Loading Factor by U.S. Regions and Overall.

The estimated rate ratios for NMF loading factors were obtained using the multivariable Poisson regression model. Error bars represent the 95% confidence intervals. ASCVD denotes atherosclerotic cardiovascular disease; CI, confidence intervals; NMF, nonnegative matrix factorization; and PM_2.5, fine particulate matter (with a mass median aerodynamic diameter less than or equal to 2.5 μm).

Zip Code–Years with PM_2.5 Exposure below the Annual NAAQS PM_2.5 Threshold of 9 μg/m³

This subset includes 223,548,803 person-years in 36,879 ZIP codes. While most of the ZIP codes were included for at least a year in the subset, the Southeast and bordering areas were included in fewer years than the rest of the United States (Fig. S4). Concentrations of PM_2.5 and most components were lower in the subset, except for Si (Table S2). Similarly, the means of the factors were smaller in this subset, except for soil and dust (Si being a major component). The observed rate ratio for a one-unit increase in PM_2.5 mass was 1.028 (95% CI, 1.026 to 1.029) in the below 9 μg/m³ PM_2.5 dataset and 1.012 (95% CI, 1.011 to 1.014) in the full-range dataset. The regional effect estimates are presented in Figure S5 and Table S3. The observed rate ratio per one unit, increase in coal and biomass burning exposure was 1.013 (95% CI, 1.013 to 1.014) in the below 9 μg/m³ PM_2.5 dataset and 1.008 (95% CI, 1.008 to 1.009) in the full-range dataset. The remaining source-specific associations are presented in Figure S6 and Table S4.

Discussion

In this study, we examined the associations between PM_2.5 and its sources with ASCVD mortality, emphasizing individual and combined effects, as well as regional variations in the United States. We found higher ASCVD rates associated with PM_2.5 mass, as well as PM_2.5 from industrial and combustion sources (i.e., oil combustion, coal or biomass burning, industrial pollution, and motor vehicle sources) but not with soil and dust. The exposure patterns and associations varied regionally.

PM_2.5 AND ASCVD MORTALITY

We found an increase in ASCVD mortality with each interquartile range increase in PM_2.5 mass. This association aligns with the existing literature.³ A 2017 study found a 7.3% increase in all-cause mortality associated with a 10 μg/m³ increase in annual overall PM_2.5 exposure. For comparison, our study shows a 5.3% increase in ASCVD mortality for a 10 μg/m³ increment. The difference in the effect estimates may be related to the difference in the outcome studied, but both studies show the mortality rate associated with long-term PM_2.5 exposure.²⁵ Additionally, a study in the United States demonstrated that long-term exposure to moderate PM_2.5 levels increases the risk of death from cardiovascular diseases.²⁶ Similarly, a meta-analysis conducted in China indicated a rise in cardiovascular mortality with an increase in PM_2.5 levels, although the actual PM_2.5 levels were notably higher in China than in the United States.²⁷

PM_2.5 SOURCES AND ASCVD MORTALITY

Our NMF results indicated that oil combustion, industrial pollution, coal and biomass burning, and motor vehicle pollution sources were associated with increased rate of ASCVD mortality across the United States, while soil and dust exposure was not associated with a higher ASCVD mortality rate. It is challenging to draw direct comparisons with other studies that use factorization methods to group components into sources because different studies use different factors. Yet, multiple studies find adverse cardiovascular outcomes associated with oil combustion, industrial pollution, coal and biomass burning, and motor vehicle sources. In a 2022 study, Kazemiparkouhi et al. categorized 18 components into eight factors and found that cardiovascular mortality had the strongest associations with traffic and coal. However, unlike our study, they also found strong associations with soil sources.¹² Like our study, Henneman et al. highlight the mortality rate associated with coal-sourced PM_2.5 among the Medicare population in the United States.²⁸ Moreover, Thurston et al. found substantial associations between coal combustion and traffic-related sources and ischemic heart disease. However, they did not find associations between soil and biomass combustion and heart disease.²⁹

Our results consistently pointed toward an effect modification by U.S. region in the sources-centered analysis; we illustrate this in Figure 3. These regional differences require further investigation, as these differences can be attributed to a combination of diverse factors. For example, each U.S. region has a distinct climate. Atmospheric temperature, as well as other climate variables, were found to modify PM_2.5-associated health effects. Furthermore, components of the built environment such as greenness levels³⁰ and the socioeconomic, racial, and ethnic composition of the population^31,32 can also contribute to regional differences. Finally, the uneven distribution of ASCVD incidence and its underlying risk factors should also be considered. For example, the Southeastern region in the United States has sometimes been known as the U.S. “Stroke Belt” due to clustering of stroke incident cases in these states. This heterogeneous incidence distribution can be linked to differences in population vulnerability, socioeconomic status, race, and ethnicity as well as environmental and behavioral factors.³³

Figure 3. Illustrative Summary of the PM_2.5 Components with the Strongest Regional Associations with ASCVD Mortality.

In this figure, the regions outlined in Figure 1 are populated with an illustrative summary of the results of the PM_2.5 components that have the strongest regional associations with ASCVD mortality.ASCVD denotes atherosclerotic cardiovascular disease; and PM_2.5, fine particulate matter (with a mass median aerodynamic diameter less than or equal to 2.5 μm).

ZIP CODE–YEARS BELOW THE ANNUAL NAAQS PM_2.5 THRESHOLD OF 9 μG/M³

Lastly, our results based on ZIP code–year below the current NAAQS of 9 μg/m³ PM_2.5 showed that PM_2.5 is still detrimental to health even when the overall annual exposure to PM_2.5 is <9 μg/m³. These results are consistent with previous studies that found associations between PM_2.5 at <12 μg/m³ and excess mortality,²⁵ as well as excess CVD mortality, in low-exposure regions of North Carolina³⁴ and Canada.³⁵ These results highlight the need to close the gap between the NAAQS and the World Health Organization recommendation — currently set at <5 μg/m³.

STRENGTHS AND LIMITATIONS

We conducted a national analysis including diverse older adults and exposure characteristics. We used tailored statistical methods for mixtures capable of handling the high complexity and correlation structure of the components data. We leveraged high-resolution geospatial models of numerous PM_2.5 components. Our data add evidence to current research largely focused on a limited set of PM_2.5 components due to restricted data availability or limited geographic areas.³⁶

Our study has limitations. First, despite the high performance of our exposure assessment models, the reliance on modeled pollutant concentrations introduces measurement errors, and the use of ZIP code–level exposure might introduce exposure misclassification. Second, while our statistical models accounted for many potential confounders, residual confounding due the lack of certain individual-level risk factors, such as smoking, drug use, diet, and alcohol consumption, which are all linked to ASCVD mortality, might have influenced our risk estimates. However, since the exposure is assigned on a ZIP code–level, we do not expect this to bias our results. Third, our outcome definition includes ICD-10 codes for ischemic heart disease and stroke, excluding other ASCVD diagnoses. However, ischemic heart disease and stroke are the main drivers of ASCVD mortality — as well as overall mortality — in the United States.³⁷ Therefore, we expect to capture a representative portion of the association between PM_2.5 and ASCVD. Fourth, our data are for the region of residence while enrolled in Medicare; we do not have data on each individual patient’s residential history prior to Medicare enrollment. Lastly, the use of factorization into sources provides critical insights about the origins of PM_2.5, potentially informing policy. However, additional research is required to investigate the toxicity of specific PM_2.5 chemical components.

Conclusions

Our study provides an assessment of the associations between PM_2.5 sources and ASCVD mortality, highlighting both individual and combined effects, as well as distinct regional variations across the United States. We established an association between PM_2.5 and ASCVD mortality, identifying specific sources as key drivers of this finding. Furthermore, our findings showcase regional heterogeneity, with stronger associations in certain areas. These associations persisted even after limiting our sample to ZIP code–years with PM_2.5<9 μg/m³ — the new NAAQS. These results emphasize the role of geographical context and the need to consider local population characteristics and exposure patterns when assessing health risks associated with air pollution.