Racial/ethnic disparities in ${\text{PM}}_{2.5}$-attributable cardiovascular mortality burden in the United States

Abstract

Average fine particulate matter (${\text{PM}}_{2.5}$) concentrations have decreased in the U.S. in recent years, but the health benefits of this improvement among different racial/ethnic groups are unknown. We estimate the associations between long-term exposure to ambient ${\text{PM}}_{2.5}$ and cause-specific cardiovascular disease (CVD) mortality rate and assess the ${\text{PM}}_{2.5}$-attributable CVD deaths by race/ethnicity across 3,103 U.S. counties during 2001–2016 (n = 595,776 county-months). A 1 μg/m³ increase in ${\text{PM}}_{2.5}$ concentration was associated with increases of 7.16 (95% confidence interval [CI]: 3.81, 10.51) CVD deaths per 1,000,000 Black people per month, significantly higher than the estimates for non-Hispanic White people (5.40 [95% CI: 2.03, 8.77]; P = 0.001). No significant difference in this association was observed between Hispanic (2.66 [95% CI: −0.03, 5.35]) and non-Hispanic White people (0.90 [95% CI: −1.81, 3.61]; P = 0.523). From 2001 to 2016, the absolute disparity in ${\text{PM}}_{2.5}$-attributable CVD mortality burden reduced by 44.04% between non-Hispanic Black and White people and by 2.61% between Hispanic and non-Hispanic White people. However, in 2016, the burden remained 3.47 times higher for non-Hispanic Black people and 0.45 times higher for Hispanic people than for non-Hispanic people. We call for policies that aim to reduce both exposure and vulnerability to ${\text{PM}}_{2.5}$ for racial/ethnic minorities.

Introduction

Ambient fine particulate matter (${\text{PM}}_{2.5}$) air pollution is a leading risk factor for mortality in the United States (U.S.), especially for mortality of cardiovascular diseases (CVD)¹. The detrimental effect of ${\text{PM}}_{2.5}$ on cardiovascular mortality persists even at concentrations below the current air quality standard^2–5. Over recent decades, ${\text{PM}}_{2.5}$ levels have fallen considerably in the contiguous U.S., as a result of the national Clean Air Act and other regional efforts^6,7. This improvement in air quality has yielded substantial benefits to human health⁸. However, it is unclear whether the health benefits of this decreasing trend of ${\text{PM}}_{2.5}$, especially to cardiovascular health, distributed equitably across different racial/ethnic groups. To tackle this question, in theory, we need information on two potential mechanisms: whether the ${\text{PM}}_{2.5}$ exposure is unevenly distributed across racial/ethnic groups (differential exposure) and whether the health effect of ${\text{PM}}_{2.5}$ exposure, influenced by the capacity of response by coping and adaptability, differs across groups (differential vulnerability)^9,10.

Most previous studies examined this question focusing only on the differential ${\text{PM}}_{2.5}$ exposure. Increasing evidence showed an unequal distribution of ${\text{PM}}_{2.5}$ exposure across communities, with racial/ethnic minorities being exposed to disproportionately high levels of ${\text{PM}}_{2.5}$^11–16. Despite the improvement in air quality during the past years, the disparity in ${\text{PM}}_{2.5}$ exposure remains^11,13,15,16. In the U.S., physical environment and neighborhoods tend to exhibit the symptoms of, and are shaped by, environmental racism, most widely understood as the disproportionate exposure of communities of color to environmental hazards^17,18. For example, racial/ethnic minorities in the U.S., such as Black and Hispanic people, tend to live in neighborhoods where they have higher exposure to air pollution^16,19, close proximity to waste facilities²⁰, and low community green space (e.g. parks)^21,22. Such factors can all greatly increase the levels of ${\text{PM}}_{2.5}$ exposure.

However, the racial/ethnic disparities in ${\text{PM}}_{2.5}$-attributable mortality burden cannot be fully explained by the disparities in ${\text{PM}}_{2.5}$ exposure²³. Differential vulnerability, measured by different ${\text{PM}}_{2.5}$-mortality exposure-response functions (ERFs) across racial/ethnic groups, also needs to be considered. A previous study focusing on Medicare beneficiaries found that the ${\text{PM}}_{2.5}$-related all-cause mortality risk for Black people was three times as high as that for the overall population²³. Therefore, it is important to use race/ethnicity-specific ERF in health burden assessments²⁴.

Utilizing the nationwide ${\text{PM}}_{2.5}$ and cardiovascular mortality data, this study aimed to first estimate the associations of long-term exposure to ${\text{PM}}_{2.5}$ with county-level cardiovascular mortality rates by race/ethnicity, and then assess the average annual ${\text{PM}}_{2.5}$-attributable CVD mortality burden across racial/ethnic groups and how the racial/ethnic disparity in this burden changed over time between 2001 and 2016. Interactive fixed effects (IFE) models were applied to estimate the ${\text{PM}}_{2.5}$-mortality exposure-response functions. In addition to the total CVD mortality, we also analyzed mortality from major categories of CVD.

Results

Description of ${\text{PM}}_{2.5}$ exposure and CVD mortality

A total of 13,289,147 CVD deaths was included in the study. From 2001 to 2016, the mean age-adjusted CVD mortality rate was 232.67 per 1,000,000 people per month in the 3,103 counties in the contiguous U.S. (standard deviation [SD]: 140.97), of which over 70% were from ischemic heart disease (IHD), stroke, and hypertensive disease (HD). The mean age-adjusted CVD mortality rate was higher in males (278.43, SD: 230.46) than in females (195.85, SD: 166.79; P value < 0.001), and was higher in non-Hispanic Black people (294.75, SD: 425.72; P value < 0.001) compared to non-Hispanic White (230.19, SD: 158.33; P value < 0.001) and Hispanic people (134.11, SD: 399.58; P value < 0.001). The mean 12-month moving average of ${\text{PM}}_{2.5}$ concentration was 9.12 μg/m³ (SD: 3.13) (Table 1). The average ${\text{PM}}_{2.5}$ concentrations were in general higher in the eastern U.S. (Fig. 1a) and have greatly decreased in most counties in our study period (Fig. 1b). Among people of different racial/ethnic groups who died from cardiovascular diseases from 2001 to 2016, although the long-term ${\text{PM}}_{2.5}$ exposure reduced in all three groups (Supplementary Fig. 1), the median concentration of long-term ${\text{PM}}_{2.5}$ exposure was the highest in the non-Hispanic Black group (Fig. 1c).

Table 1:

Monthly descriptive statistics for all 3,103 contiguous U.S. counties from 2001 to 2016^a

	Mean (SDb)	Min	Median (IQRc)	Max
Environmental factors
12-month moving average of ${\text{PM}}_{2.5}$ (μg/m3)	9.12 (3.13)	1.54	9.23 (4.33)	18.79
Air temperature (°C)	12.70 (9.90)	−20.58	13.57 (15.61)	35.41
Age-adjusted mortality rate (per 1,000,000 individuals)
Cardiovascular disease	232.67 (140.97)	0.00	217.52 (140.42)	15,508.56
Ischemic heart disease	111.37 (97.71)	0.00	97.88 (95.61)	15,508.56
Myocardial infarction	47.70 (65.91)	0.00	30.91 (66.48)	4,169.58
Stroke	40.01 (53.16)	0.00	30.48 (56.68)	3,877.14
Hypertensive disease	14.32 (31.09)	0.00	0.00 (18.90)	2,584.76
Hypertensive heart disease	6.89 (21.80)	0.00	0.00 (3.98)	2,584.76
Age-adjusted CVDd mortality rate by race/ethnicity (per 1,000,000 individuals)
Non-Hispanic White	230.19 (158.33)	0.00	213.54 (144.14)	15,508.56
Non-Hispanic Black	294.75 (425.72)	0.00	234.17 (407.62)	44,840.80
Hispanic	134.11 (399.58)	0.00	0.00 (141.02)	22,011.77
Age-adjusted CVD mortality rate by sex (per 1,000,000 individuals)
Male	278.43 (230.46)	0.00	250.13 (213.63)	15,508.56
Female	195.85 (166.79)	0.00	175.77 (151.65)	15,508.56

^aSome neighboring counties were combined due to historical boundary changes

^bSD: standard deviation

^cIQR: interquartile range

^dCVD: cardiovascular disease

We performed two-sided Wilcoxon Signed Rank Test to compare the age-adjusted CVD mortality rate among demographic groups (n = 595,776). This non-parametric test does not require normality assumption.

Non-Hispanic Black vs. Non-Hispanic White: V = 4.11 × 10¹⁰, P value < 0.001

Non-Hispanic Black vs. Hispanic: V = 2.19 × 10¹⁰, P value < 0.001

Male vs. Female: V = 1.14 × 10¹¹, P value < 0.001

Fig. 1: Distribution of ${\text{PM}}_{2.5}$ concentration in all counties in the contiguous United States.

a, Average monthly ${\text{PM}}_{2.5}$ concentration from 2000 to 2016 in each of the 3,103 counties (μg/m³). b, Percent change in annual mean ${\text{PM}}_{2.5}$ concentration from 2000 to 2016 in each of the 3,103 counties (%). The shapefile in these maps was obtained from the U.S. Census Bureau. c, The distribution of long-term ${\text{PM}}_{2.5}$ exposure (12-month moving average of ${\text{PM}}_{2.5}$ concentration, μg/m³) among people of different racial/ethnic groups who died from cardiovascular diseases from 2001 to 2016 (n = 1,074,613 non-Hispanic White people, 1,568,837 non-Hispanic Black people, and 657,725 Hispanic people). Each violin shows the ${\text{PM}}_{2.5}$ distribution for each racial/ethnic group. The width of the violin describes the frequency of ${\text{PM}}_{2.5}$ concentration at each level. The grey box plot inside each violin summarizes the minimum, first quartile, median, third quartile, maximum, and outliers of ${\text{PM}}_{2.5}$ concentration in each group.

Race/ethnicity-specific exposure-response function

We found that a 1 μg/m³ increase in the 12-month moving average of ${\text{PM}}_{2.5}$ concentration was significantly associated with 2.01 additional CVD deaths per 1,000,000 people (95% confidence interval [CI]: 1.67, 2.35) per month. Among racial/ethnic groups, each 1 μg/m³ increase in the 12-month moving average of ${\text{PM}}_{2.5}$ concentration was associated with 7.16 (95% CI: 3.81, 10.51) additional CVD deaths per 1,000,000 non-Hispanic Black people per month, 5.40 (95% CI: 2.03, 8.77) more deaths compared to non-Hispanic White people (1.76 [95% CI: 1.37, 2.15] CVD deaths per 1,000,000 non-Hispanic White people per month). For Hispanic people, a 1 μg/m³ increase in ${\text{PM}}_{2.5}$ concentrations was associated with 2.66 (95% CI: −0.03, 5.35) additional CVD deaths per 1,000,000 Hispanic people per month (Table 2). We observed no statistically significant difference in this association between Hispanic and non-Hispanic White people (0.90 [95% CI: −1.81, 3.61]). No statistically significant difference was observed between males and females (0.34 [95% CI: −0.38, 1.06]; Supplementary Table 1).

Table 2:

Deaths per 1,000,000 individuals associated with a 1-μg/m³ increase in 12-month moving average of ${\text{PM}}_{2.5}$ per month by race/ethnicity

Cause of death	Race/ethnicity	Coefficient (95% CI)a	Difference in coefficient (95% CI)b
Cardiovascular disease	Whole population	2.01 (1.67, 2.35)	-
	Non-Hispanic White	1.76 (1.37, 2.15)	reference
	Non-Hispanic Black	7.16 (3.81, 10.51)	5.40 (2.03, 8.77)
	Hispanic	2.66 (−0.03, 5.35)	0.90 (−1.81, 3.61)
Ischemic heart disease	Whole population	1.67 (1.43, 1.91)	-
	Non-Hispanic White	1.64 (1.37, 1.91)	reference
	Non-Hispanic Black	4.09 (1.80, 6.38)	2.45 (0.14, 4.76)
	Hispanic	1.52 (−0.21, 3.25)	−0.12 (−1.88, 1.64)
Myocardial infarction	Whole population	0.45 (0.29, 0.61)	-
	Non-Hispanic White	0.45 (0.26, 0.63)	reference
	Non-Hispanic Black	1.97 (0.68, 3.26)	1.52 (0.22, 2.83)
	Hispanic	0.45 (−0.58, 1.48)	0.00 (−1.05, 1.05)
Stroke	Whole population	0.22 (0.08, 0.36)	-
	Non-Hispanic White	0.11 (−0.04, 0.26)	reference
	Non-Hispanic Black	1.76 (0.67, 2.85)	1.65 (0.55, 2.75)
	Hispanic	0.05 (−1.01, 1.11)	−0.06 (−1.13, 1.01)
Hypertensive disease	Whole population	0.11 (0.03, 0.19)	-
	Non-Hispanic White	0.05 (−0.04, 0.14)	reference
	Non-Hispanic Black	0.65 (−0.01, 1.31)	0.61 (−0.05, 1.27)
	Hispanic	0.15 (−0.31, 0.61)	0.10 (−0.37, 0.57)
Hypertensive heart disease	Whole population	0.03 (−0.02, 0.09)	-
	Non-Hispanic White	0.03 (−0.03, 0.09)	reference
	Non-Hispanic Black	0.04 (−0.33, 0.41)	0.01 (−0.36, 0.38)
	Hispanic	−0.15 (−0.41, 0.11)	−0.18 (−0.44, 0.09)

^aCoefficient: number of deaths per 1,000,000 individuals associated with a 1 μg/m³ increase in 12-month moving average of ${\text{PM}}_{2.5}$ per month by race/ethnicity; CI: confidence interval

^bWe calculated the differences in coefficient between non-Hispanic Black and non-Hispanic White people and between Hispanic and non-Hispanic White people. CIs were calculated based on the coefficients and standard errors of different groups.

For major CVD categories, 1.67 (95% CI: 1.43, 1.91) IHD deaths, 0.22 (95% CI: 0.08, 0.36) stroke deaths, and 0.11 (95% CI: 0.11, 0.03, 0.19) HD deaths per 1,000,000 people were linked to a 1 μg/m³ increase in ${\text{PM}}_{2.5}$. A 1 μg/m³ increase in ${\text{PM}}_{2.5}$ was also associated with 0.45 additional myocardial infarction (MI) deaths (95% CI: 0.29, 0.61). For different racial/ethnic groups, a 1 μg/m³ increase in ${\text{PM}}_{2.5}$ concentrations was associated with 1.64 (95% CI: 1.37, 1.91), 4.09 (95% CI: 1.80, 6.38), and 1.52 (95% CI: −0.21, 3.25) additional IHD deaths, including 0.45 (95% CI: 0.26, 0.63), 1.97 (95% CI: 0.68, 3.26), and 0.45 (95% CI: −0.58, 1.48) additional MI deaths, per 1,000,000 non-Hispanic White, non-Hispanic Black, and Hispanic people, respectively. In addition, the association between long-term ${\text{PM}}_{2.5}$ exposure and stroke mortality rate was only significantly positive in the non-Hispanic Black group (1.76 [95% CI: 0.67, 2.85], Table 2). The between-group differences in the association of long-term ${\text{PM}}_{2.5}$ exposure with cause-specific CVD mortality were also tested by an interaction analysis (Supplementary Table 2), and the findings were consistent with those presented in Table 2.

Attributable burden by race/ethnicity

The average annual ${\text{PM}}_{2.5}$-attributable CVD mortality burden was in general higher in eastern counties, especially in the Ohio Valley region and the Southeast region, which is consistent with the distribution of ${\text{PM}}_{2.5}$ concentrations (Fig. 2a). In total, approximately 69,675 (95% CI: 57,785, 81,565) CVD deaths were attributable to long-term ${\text{PM}}_{2.5}$ exposure each year from 2001 to 2016, which contributed to about 7.84% of total CVD deaths in the contiguous U.S. (Fig. 2b). Among specific causes, approximately 57,889 IHD deaths (including 15,634 MI deaths), 7,487 stroke deaths, and 3,882 HD deaths (including 1,192 hypertensive heart disease [HHD] deaths) were attributable to ${\text{PM}}_{2.5}$ each year (Fig. 2b; Supplementary Table 3).

Fig. 2: Average annual cardiovascular deaths attributable to ${\text{PM}}_{2.5}$ (2001–2016).

a, Average annual ${\text{PM}}_{2.5}$-attributable cardiovascular deaths per 1,000,000 individuals from 2001 to 2016 in each county. The shapefile in the map was obtained from the U.S. Census Bureau. b, Average annual ${\text{PM}}_{2.5}$-attributable deaths from cardiovascular disease (CVD), ischemic heart disease (IHD), myocardial infarction (MI), stroke, hypertensive disease (HD), and hypertensive heart disease (HHD) from 2001 to 2016. c. Average annual ${\text{PM}}_{2.5}$-attributable cardiovascular deaths per 1,000,000 individuals from 2001 to 2016 in each racial/ethnic group (n = 595,776 county-months).

Among racial/ethnic groups, the average annual ${\text{PM}}_{2.5}$-attributable CVD mortality burden was 40,315 (95% CI: 31,426, 49,205) deaths in non-Hispanic White people, 35,136 (95% CI: 18,689, 51,583) deaths in non-Hispanic Black people, and 13,335 (95% CI: −126, 26,795) deaths in Hispanic people. Although the absolute number of ${\text{PM}}_{2.5}$-attributable deaths was the highest in the non-Hispanic White group, non-Hispanic Black people still shouldered the highest relative burden: in average, long-term ${\text{PM}}_{2.5}$ exposure contributed to approximately 905.68 (95% CI: 481.73, 1,329.63) CVD deaths per 1,000,000 non-Hispanic Black people annually, over 3 times higher than the estimated rate in non-Hispanic White people (202.70, 95% CI: 158.01, 247.40). In addition, long-term ${\text{PM}}_{2.5}$ exposure contributed to approximately 279.24 (95% CI: −2.65, 561.13) CVD deaths per 1,000,000 Hispanic people (Fig. 2c).

Changes in racial/ethnic disparity in attributable burden

From 2001 to 2016, the CVD mortality burden attributable to long-term ${\text{PM}}_{2.5}$ exposure has greatly decreased in most U.S. counties, especially the eastern U.S., with the greatest decline in West Virginia and Kentucky (Fig. 3a). In total, ${\text{PM}}_{2.5}$-attributable CVD deaths have decreased by 34.19%, from 82,047 (95% CI: 68,046, 96,048) in 2001 to 53,996 (95% CI: 44,782, 63,210) in 2016 (Fig. 3b; Supplementary Table 4). Despite of the overall decreasing trend, the magnitudes of decrease varied across different racial/ethnic groups (Fig. 4a). The non-Hispanic Black group experienced the greatest reduction in ${\text{PM}}_{2.5}$-attributable CVD mortality burden, from 1,162.76 deaths (95% CI: 618.47, 1,707.05) in 2001 to 655.81 deaths (95% CI: 348.92, 962.79) in 2016, per 1,000,000 Black people. The ${\text{PM}}_{2.5}$-attributable CVD mortality burden also reduced among Hispanic people, from 347.92 deaths (95% CI: −3.30, 699.14) in 2001 to 213.49 deaths (95% CI: −2.02, 429.01) in 2016, per 1,000,000 Hispanic people.

Fig. 3: Changes in annual cardiovascular deaths attributable to ${\text{PM}}_{2.5}$ by county and by specific cause from 2001 to 2016.

a, Percent change in annual ${\text{PM}}_{2.5}$-attributable cardiovascular deaths from 2001 to 2016 in each county. The shapefile in the map was obtained from the U.S. Census Bureau. b, Annual ${\text{PM}}_{2.5}$-attributable deaths from cardiovascular disease (CVD), including myocardial infarction (MI) and other ischemic heart disease (IHD), stroke, hypertensive heart disease (HHD) and other hypertensive disease (HD), and other CVD, from 2001 to 2016 (n = 595,776 county-months).

Fig. 4: Racial/ethnic disparity in annual cardiovascular mortality rates attributable to ${\text{PM}}_{2.5}$ from 2001 to 2016.

a, Annual ${\text{PM}}_{2.5}$-attributable cardiovascular disease (CVD) deaths per 1,000,000 people by race/ethnicity from 2001 to 2016. b, Absolute racial/ethnic disparity in ${\text{PM}}_{2.5}$-attributable CVD deaths per 1,000,000 people in 2001 and 2016, calculated by the difference in annual ${\text{PM}}_{2.5}$-attributable CVD deaths per 1,000,000 people between non-Hispanic Black or Hispanic group and non-Hispanic White group. c, Relative racial/ethnic disparity in ${\text{PM}}_{2.5}$-attributable CVD deaths per 1,000,000 people in 2001 and 2016, calculated by the ratio of annual ${\text{PM}}_{2.5}$-attributable CVD deaths per 1,000,000 people among non-Hispanic Black or Hispanic group to the attributable burden in non-Hispanic White group.

As a result, the absolute racial/ethnic disparity became smaller: comparing to non-Hispanic White people, the ${\text{PM}}_{2.5}$-attributable CVD mortality rate in non-Hispanic Black people was 909.58 higher in 2001 and 509.02 higher in 2016, which reduced by 44.04%; this difference also reduced by 29.59% for the Hispanic group, from 94.75 in 2001 to 66.71 in 2016 (Fig. 4b). However, the changes in the relative disparity were relatively small. The ratio of ${\text{PM}}_{2.5}$-attributable CVD mortality rate in non-Hispanic Black people to that in non-Hispanic White people was 4.59 in 2001, and this ratio slightly decreased to 4.47 in 2016. For Hispanic people, this ratio was 1.37 and 1.45 in 2001 and 2016, respectively.

Effect modification by urban vs. rural counties

We observed a higher coefficient of ${\text{PM}}_{2.5}$ in rural counties compared to urban counties. In rural counties, a 1 μg/m³ increase in long-term ${\text{PM}}_{2.5}$ exposure was significantly associated with 2.05 CVD deaths per 1,000,000 people (95% CI: 1.54, 2.56) per month, while in urban counties, the effect estimate of ${\text{PM}}_{2.5}$ became smaller (1.00, 95% CI: 0.62, 1.37; P = 0.001). There was no difference in the association between ${\text{PM}}_{2.5}$ exposure and CVD mortality rate in non-Hispanic Black people compared to non-Hispanic White people (urban: P = 0.143; rural: P = 0.058), nor between Hispanic and non-Hispanic White people (urban: P = 0.414; rural: P = 0.133) (Supplementary Table 5).

Sensitivity analyses and placebo tests

Sensitivity analyses showed that our results generally remained robust when excluding counties with population size smaller than 5,000 or greater than 200,000, using 12-month moving average of CVD mortality rate as the outcome, applying traditional two-way fixed effects (TWFE) models, additionally adjusting for ${\text{NO}}_2$, ${\mathrm{\text{O}}}_3$, or dew point temperature in the model, using alternative numbers of degree of freedom for air temperature, using an alternative way to control for the long-term trend, and setting the 10^th percentile of ${\text{PM}}_{2.5}$ distribution as the reference concentration in the calculation of attributable deaths (Supplementary Table 6). The time-series plot of the monthly CVD mortality rates in counties with population size smaller than 5,000 indicated that the outcome measure was stable even in relatively small counties (Supplementary Fig. 2). When extending the exposure period of ${\text{PM}}_{2.5}$ from 12 months to 24, 36, 48, and 60 months, the model coefficients became slightly higher, indicating the lag effect of ${\text{PM}}_{2.5}$ on cardiovascular mortality (Supplementary Fig. 3).

The IFE model we applied can potentially account for both unmeasured spatial and temporal confounders, as well as those unmeasured time-varying county-specific factors. The estimated common time-varying effects, time-invariant county effects, and unmeasured time-varying county effects in the main IFE model were visualized in Supplementary Fig. 4. One unmeasured factor that displayed different temporal variations across counties was estimated, but this factor may not be directly interpretable because it is, at best, a linear transformation of the true factors (see more details in Methods).

A spatial randomization test and a temporal randomization test were used to show that our main results were not spurious due to model misspecification. As shown in Supplementary Fig. 5, the distributions of the model coefficients when the ${\text{PM}}_{2.5}$ concentrations were randomly assigned 2,000 times spatially or temporally were centered at zero and the coefficient estimate from our main model fell substantially outside this distribution, indicating that the estimated ${\text{PM}}_{2.5}$-mortality association in our study was unlikely driven by spatial or temporal dependence due to a misspecification of model.

Discussion

This study demonstrates substantial racial disparities in both vulnerability to and burden of ${\text{PM}}_{2.5}$-related CVD mortality. First, the association between ${\text{PM}}_{2.5}$ exposure and CVD mortality rate was stronger for non-Hispanic Black people than non-Hispanic White people in the study. No significant difference in this association was observed between Hispanic and non-Hispanic White people. Second, non-Hispanic Black people also had the highest average annual ${\text{PM}}_{2.5}$-attributable CVD mortality burden, followed by Hispanic people. Finally, the absolute racial disparities in this burden narrowed over time from 2001 to 2016, but the relative disparity between the minorities and non-Hispanic White people persisted.

We found that long-term exposure to ambient ${\text{PM}}_{2.5}$ was associated with a greater increase in CVD mortality rate among non-Hispanic Black people compared to non-Hispanic White people, though no significant difference was observed between Hispanic and non-Hispanic White people. This finding from our study for all ages is consistent with the finding from a previous study focusing on people 65 years of age or older.²³ A cohort study based on all Medicare beneficiaries reported that Black people had a higher estimated risk of all-cause mortality in association with ${\text{PM}}_{2.5}$ exposure than the general population²³. This high vulnerability of Black people may be explained by structural racism. Structural racism refers to “the totality of ways in which societies foster racial discrimination through mutually reinforcing systems of housing, education, employment, earnings, benefits, credit, media, health care and criminal justice, that in turn reinforce discriminatory beliefs, values and distribution of resources”^25,26. Due to structural racism and its downstream social determinants, including lack of access to equitable healthcare resources²⁶ and chronic racial stressors²⁷, Black people may suffer from greater CVD mortality consequences of air pollution than White people, even when they live in the same physical environment²⁸.

In our study, the burden of ${\text{PM}}_{2.5}$-attributable CVD mortality in non-Hispanic Black people was found to be over 3 times greater than the burden in non-Hispanic White people. Hispanic people also had a high ${\text{PM}}_{2.5}$-attributable CVD mortality burden, but this did not significantly differ from people of other race/ethnicity. This finding reflects both disproportionally higher ${\text{PM}}_{2.5}$ exposure and the aforementioned higher vulnerability of non-Hispanic Black people in the U.S. Multiple studies consistently demonstrated that people of color are typically exposed to higher levels of air pollution than non-Hispanic White people in the U.S.^11–14, which is a pervasive and persistent consequence of redlining, inequitable siting of emission sources (e.g., highways and industrial facilities), and other policies and practices^29,30. In addition, consistent with a previous study²⁴, our results showed that using a uniform ERF for all racial/ethnic groups could yield a smaller estimate of total ${\text{PM}}_{2.5}$-attributable CVD mortality burden (approximately 69,675 deaths per year) compared to the sum of estimated attributable deaths across different groups (approximately 88,786 deaths per year), indicating that ignoring the differential vulnerability to ${\text{PM}}_{2.5}$ of different groups may underestimate the overall magnitude of ${\text{PM}}_{2.5}$-related mortality burden.

Benefited by the improvement of air quality in recent decades in the U.S., the overall ${\text{PM}}_{2.5}$-attributable CVD mortality burden was greatly reduced. Among racial/ethnic groups, this burden decreased more among non-Hispanic Black and Hispanic people, thus narrowed the absolute racial disparity, particularly between non-Hispanic Black and White groups. However, the reduced disparities found in our study are consistent with the explanation that they were mainly driven by the high vulnerability of non-Hispanic Black and Hispanic people, not by a more equal distribution of air pollution. If we assume the ERF of non-Hispanic Black people is the same as the one of non-Hispanic White people, the decreasing trends of ${\text{PM}}_{2.5}$-attributable CVD mortality burden in these two groups would become almost parallel (Supplementary Fig. 6). These different results indicate the importance of using race/ethnicity-specific ERF when quantifying the racial/ethnic disparities in the ${\text{PM}}_{2.5}$-attributable mortality burden²⁴.

Our results suggested an effect modification by urbanization level of residence, with stronger associations between ${\text{PM}}_{2.5}$ and CVD mortality rates in rural counties compared to urban counties. A similar pattern was found for the association between ${\text{PM}}_{10}$ and non-accidental mortality in Italy³¹. One possible explanation of this finding is the different mixture compositions of ${\text{PM}}_{2.5}$ in rural and urban counties³², which may have different effects on cardiovascular mortality^31,33.

Revealing the differential public health burden of air pollution across racial groups, our findings could help inform environmental agencies to be more aware of the disproportionate disparity in exposure and vulnerability to air pollution. In addition, there is a need to better design and inform policies that efficiently reduce environmental inequity and protect vulnerable population in the pursuit of improving overall air quality.

We examined the changes in racial/ethnic disparity in ${\text{PM}}_{2.5}$-attributable cardiovascular mortality burden over time using race/ethnicity-specific ERF. In addition, we applied IFE model that allows us to better control for unmeasured time-varying county-level confounders compared to the traditional TWFE model, to estimate the association between long-term ${\text{PM}}_{2.5}$ exposure and mortality from principal CVD types.

However, some limitations should be noted. First, this is a county-level ecological study, which is susceptible to ecological fallacy. We were unable to capture the within-county heterogeneity of ${\text{PM}}_{2.5}$ exposure among different racial/ethnic groups. In an additional analysis, using ${\text{PM}}_{2.5}$ exposure that weighted by race/ethnicity-specific population (Supplementary Fig. 7), we estimated a slightly larger racial disparity in ${\text{PM}}_{2.5}$-attributable CVD burden (Supplementary Table 7), indicating that the estimated disparity in attributable burden in our main analysis could be a conservative estimate. Studies with a finer spatial resolution or using individual-level data are needed in the future. Second, we did not specifically analyze for people of races/ethnicities other than non-Hispanic White, non-Hispanic Black and Hispanic people. Members of other races such as American Indians, Asians, and Pacific Islanders were not included in the race/ethnicity-specific analyses due to lack of statistical power. Third, the potential measurement error in mortality data and misreport in racial/ethnic categories may influence our results. Fourth, although the IFE model relaxed the assumptions of the traditional TWFE model, it still relies on its own assumptions (see more details in Methods) and violation of these assumption may bias the results. There are emerging literatures on relaxing the model assumptions³⁴, and more methodological advancements are needed in the future to further extend the IFE model to fit more complicated circumstances. In addition, we used 12-month moving average of ${\text{PM}}_{2.5}$ concentrations to represent the exposure in the previous one year, but as the health effects of ${\text{PM}}_{2.5}$ can last for more than one year, the estimated attributable CVD mortality burden is likely to be an underestimate of the long-term impact of ${\text{PM}}_{2.5}$. Furthermore, this study assumes the relationship between ${\text{PM}}_{2.5}$ and CVD mortality to be constant over time, which is a simplification of reality. Future studies examining the time-varying vulnerabilities to the long-term ${\text{PM}}_{2.5}$ exposure are warranted. Finally, we did not further explore the racial/ethnic disparity in ${\text{PM}}_{2.5}$ attributable burden of other diseases and the potential mechanisms of this disparity, which can be examined in depth in future studies.

In conclusion, our study indicated that non-Hispanic Black people have a higher ${\text{PM}}_{2.5}$-attributable CVD mortality burden compared to non-Hispanic White people. While Hispanic people also had a high ${\text{PM}}_{2.5}$-attributable CVD mortality burden, the estimates were insignificant. Although the absolute racial disparity in this burden was narrowed in recent decades, the gap between non-Hispanic Black and White people, as well as between Hispanic people and non-Hispanic White people, still exists. Policies and practices that aim to reduce both exposure and vulnerability to ${\text{PM}}_{2.5}$ for racial/ethnic minorities, in addition to investigating the effects of structural racism on health, are needed.

Methods

Using anonymized monthly county-level mortality records, this study was approved by the Yale Institutional Review Boards (protocol ID: 2000026808).

Study area and population

In total, this study covered 3,103 counties or county equivalents in the contiguous U.S., with consistent boundaries over the study period. The cartographic boundary for counties in the contiguous U.S. was downloaded from the U.S. Census Bureau’s TIGER/Line geodatabase³⁵. We merged counties with boundary changes from 2001 to 2016 with neighboring counties (Supplementary Methods 1). County-level population data were collected from the Surveillance, Epidemiology, and End Results Program, National Cancer Institute³⁶. We extracted the total population and population estimates by sex, age, race, and Hispanic origin for each county across years 2001–2016.

Mortality rate

Mortality data from 2001 to 2016 were provided by the National Center for Health Statistics, which included the year and month of death, the cause of death (International Statistical Classification of Diseases and Related Health Problems, 10^th Revision [ICD-10] codes), and the sex, age, race, and ethnicity of each deceased person. Information about the race and Hispanic origin of a decedent was based on death certificates. The categorization of race and ethnicity followed the national standards “Revisions to the Standards for the Classification of Federal Data on Race and Ethnicity”³⁷. This study focused on CVD mortality (ICD-10 code: I00-I99) and its major categories, including IHD (I20-I25), stroke (I60-I69), and HD (I10-I15). Deaths from more specific causes, MI (I21-I23) in IHD and HHD (I11) in HD, were also analyzed in this study. No statistical methods were used to pre-determine sample sizes but our sample sizes of CVD mortality are larger than those reported in previous studies in the U.S.^38–40.

We calculated the monthly county-level cause-specific mortality rates for male, female, non-Hispanic White, non-Hispanic Black, and Hispanic population. All mortality rates were age-adjusted by direct standardization using the 2000 U.S. Census population as the standard population. Mortality data in 2000 were also obtained to calculate the 12-month moving average of mortality rates in a sensitivity analysis.

Air pollution and temperature

Daily ${\text{PM}}_{2.5}$ concentration data from 2000 to 2016 for the contiguous U.S. at a resolution of 1 km were obtained from the NASA Socioeconomic Data and Applications Center⁴¹. In brief, the ${\text{PM}}_{2.5}$ concentration was estimated by an ensembled model based on neural network, random forest, and gradient boosting, using satellite data, meteorological data, land-use variables, elevation, chemical transport model predictions, and other variables as predictors⁶. The 10-fold cross-validated coefficients of determination (R²) of this model was 0.86 for daily predictions, indicating a strong predictive performance⁶.

The daily ${\text{PM}}_{2.5}$ data were aggregated to monthly county-level concentrations by area weighting to match to the mortality data. For months from January 2001 to December 2016, we calculated the 12-month moving average of ${\text{PM}}_{2.5}$ concentration for each county to represent the average exposure to ${\text{PM}}_{2.5}$ in the previous one year^23,42. Nitrogen dioxide (${\text{NO}}_2$) and ozone (${\mathrm{\text{O}}}_3$) data were downloaded from the same source and processed in the same way as ${\text{PM}}_{2.5}$^43–46.

Monthly mean air temperature data at 4 km × 4 km were obtained from the PRISM Climate Group⁴⁷. Similar to the air pollution data, we averaged the air temperature data for each county.

Statistical analysis

We applied an IFE model to estimate the associations between long-term exposure to ${\text{PM}}_{2.5}$ and cardiovascular mortality rates. The IFE model is more flexible than the traditional TWFE model⁴⁸. TWFE models can control for both spatial confounders that only vary across counties (time-invariant confounders) and temporal confounders that only vary by time (county-invariant confounders), either measured or unmeasured⁴⁹. However, this approach assumes no unmeasured confounders that display different temporal variations across counties (time-varying county effects), which is often violated⁵⁰. In a preliminary analysis, we detected the existence of such factors that a classical TWFE model failed to control for (P value < 0.001, Supplementary Methods 2). Therefore, we used an extension of the TWFE model, the IFE model, which can potentially account for unmeasured time-varying county-specific confounders by decomposing them into heterogeneous impacts of common trends^48,50,51. The IFE approach requires weaker assumptions than TWFE models: it assumes the unmeasured time-varying county effects to have a factor structure and all regressors and factors to be stationary, and requires a weak serial and cross-sectional correlation⁴⁸. In our study, the model can be expressed as

$${Mortalityrate}_{i,t}=\mu +{\alpha }_i+{\theta }_t+\beta {PM}_{{2.5}_{i,t}}+ns\left({Temperature}_{i,t},df=5\right)+v_{i,t}+{\epsilon }_{i,t},$$

in which

$$v_{i,t}={\sum }_{l=1}^d{\lambda }_{i,l}{f}_{l,t}.$$

The outcome variable ${Mortalityrate}_{i,t}$ is the cause-specific age-adjusted mortality rate in county $i$, month $t$ and the exposure of interest ${PM}_{{2.5}_{i,t}}$ is the moving average of ${\text{PM}}_{2.5}$ concentration of the current and previous 11 months for county $i$, month $t.{\alpha }_i$ refers to time-invariant county effects and ${\theta }_t$ refers to time-varying effects that are common in all counties. The year-month and county fixed effects can help isolate the effect of ${\text{PM}}_{2.5}$ by removing the effects of factors that are common to all counties in the same month and factors that are consistent in time in the same county. Air temperature was controlled by a flexible natural cubic spline with five degrees of freedom (df). $v_{i,t}$ is the unmeasured time-varying county effect, which is decomposed into $d$ common time-varying factors $f_{l,t}$, with corresponding unobserved county-level loading parameters ${\lambda }_{i,l}$^52,53. The criteria for selecting the number of factors ($d$) can be found in Supplementary Methods 3. The distribution of our main model residuals was plotted (Supplementary Fig. 8). We also estimated the potential non-linear association between long-term ${\text{PM}}_{2.5}$ exposure and cardiovascular mortality and found a near-linear trend when ${\text{PM}}_{2.5}$ concentrations were approximately higher than the 10^th percentile of its distribution (4.58 μg/m³), indicating that approximately 90% of the data fell in the near-linear part of the exposure-response curve (Supplementary Fig. 9). Therefore, we used a linear term to model this relationship in the main model. The IFE analyses were conducted with R software (version 4.1.3) using the package phtt⁵³. All statistical tests in this study are two-sided.

Based on the estimated coefficient of ${\text{PM}}_{2.5}$ ($\beta$) for each cause, which can be interpreted as the increase in mortality rate associated with a 1 μg/m³ increase in 12-month moving average of ${\text{PM}}_{2.5}$ per month, we then calculated the ${\text{PM}}_{2.5}$-attributable number of deaths (AN) in each county in each month by ${AN}_{i,t}={\beta PM}_{{2.5}_{i,t}}\times {Population}_{i,t}$, where ${Population}_{i,t}$ is the total population in county $i$, month $t$. The reference ${\text{PM}}_{2.5}$ concentration was set as 0, since recent epidemiological evidence suggests that there is no safe level of ${\text{PM}}_{2.5}$^4,5. In a sensitivity analysis, we alternatively set 4.58 μg/m³ (the 10^th percentile of ${\text{PM}}_{2.5}$ distribution) as the reference concentration based on the findings of the non-linear analysis (Supplementary Fig. 9).

We further estimated the association between long-term exposure to ${\text{PM}}_{2.5}$ and cause-specific cardiovascular mortality rates and the attributable mortality burden by race/ethnicity (non-Hispanic White, non-Hispanic Black, and Hispanic people). We also conducted subgroup analysis by sex (male and female) to test sex differences and stratified analysis by urban or rural counties⁵⁴ to investigate its potential modification effect. We calculated the confidence interval of the difference in effect estimates between different groups by $\left({\widehat{Q}}_1-{\widehat{Q}}_2\right)\pm 1.96\times \sqrt{{\left({S\widehat{E}}_1\right)}^2+{\left({S\widehat{E}}_2\right)}^2}$, where ${\widehat{Q}}_1$ and ${\widehat{Q}}_2$ are the estimates, and ${S\widehat{E}}_1$ and ${S\widehat{E}}_2$ are their respective standard errors⁵⁵. This calculation assumes that the estimates are distributed independently. To relax this assumption, we additionally used an expanded dataset nesting the subgroups, interacted the subgroup variable with all the other variables in the model, and reported the statistical significance of the interaction term between the subgroup variable and ${\text{PM}}_{2.5}$.

Several sensitivity analyses were performed to test the robustness of our results: (a) counties with population size smaller than 5,000 or larger than 200,000 were excluded (approximately the 10^th and 90^th percentiles of county-level population size); (b) 12-month moving average of CVD mortality rates were used as the outcome; (c) traditional TWFE model was applied; (d) we additionally adjusted for ${\text{NO}}_2$, ${\mathrm{\text{O}}}_3$, or dew point temperature in the model; (e) an alternative four or six degrees of freedom was used in the natural cubic spline of air temperature; (f) we used an alternative way to control for the long-term trend by adjusting for year using a natural cubic spline with three dfs; and (g) we used the 10^th percentile of ${\text{PM}}_{2.5}$ distribution as the reference concentration in the calculation of attributable deaths. We tested the lag pattern in the effects of ${\text{PM}}_{2.5}$ by extending the 12-month moving average of ${\text{PM}}_{2.5}$ concentration to 24, 36, 48, and 60 months, representing longer-term ${\text{PM}}_{2.5}$ exposure in the past up to five years.

In addition, we performed a spatial randomization test and a temporal randomization test to evaluate the specification of our main model. In the spatial randomization test, we randomized the ${\text{PM}}_{2.5}$ concentrations for 2,000 times across county while keeping their de facto year-month; in the temporal randomization test, we randomized the ${\text{PM}}_{2.5}$ exposure for 2,000 times across year-month while keeping the corresponding counties. Such placebo tests are commonly used to detect spatial and temporal dependence due to model misspecification in panel models⁵⁶.

Data availability

The monthly county-level cause-specific mortality data can be requested from the National Center for Health Statistics (https://www.cdc.gov/nchs/index.htm). Other data that support the findings of this study are available at https://zenodo.org/record/8121894 (DOI: 10.5281/zenodo.8121894)⁵⁷.