Long-Term Nitrogen Dioxide Exposure and Cause-Specific Mortality in the U.S. Medicare Population

Abstract

Background:

Since 1971, the annual National Ambient Air Quality Standard (NAAQS) for nitrogen dioxide (NO₂) has remained at 53 ppb, the impact of long-term NO₂ exposure on mortality is poorly understood.

Objectives:

We examined associations between long-term NO₂ exposure (12-month moving average of NO₂) below the annual NAAQS and cause-specific mortality among the older adults in the U.S.

Methods:

Cox proportional-hazard models were used to estimate Hazard Ratio (HR) for cause-specific mortality associated with long-term NO₂ exposures among about 50 million Medicare beneficiaries living within the conterminous U.S. from 2001–2008.

Results:

A 10 ppb increase in NO₂ was associated with increased mortality from all-cause (HR: 1.06; 95% CI: 1.05–1.06), cardiovascular (HR: 1.10; 95% CI: 1.10–1.11), respiratory disease (HR: 1.09; 95% CI: 1.08–1.11), and cancer (HR: 1.01; 95% CI: 1.00–1.02) adjusting for age, sex, race, ZIP code as strata ZIP code- and state-level socio-economic status (SES) as covariates, and PM_2.5 exposure using a 2-stage approach. NO₂ was also associated with elevated mortality from ischemic heart disease, cerebrovascular disease, congestive heart failure, chronic obstructive pulmonary disease, pneumonia, and lung cancer. We found no evidence of a threshold, with positive and significant HRs across the range of NO₂ exposures for all causes of death examined. Exposure-response curves were linear for all-cause, supra-linear for cardiovascular-, and sub-linear for respiratory-related mortality. HRs were highest consistently among Black beneficiaries.

Conclusions:

Long-term NO₂ exposure is associated with elevated risks of death by multiple causes, without evidence of a threshold response. Our findings raise concerns about the sufficiency of the annual NAAQS for NO₂.

1. Introduction

Since 1971, the annual National Ambient Air Quality Standard (NAAQS) for nitrogen dioxide (NO₂) has remained at 53 ppb. During this period, ambient NO₂ concentrations within the US have decreased substantially, with all areas in the US in attainment (U.S. Environmental Protection Agency, 2019). Despite this, several studies have demonstrated associations with mortality at currently observed low annual NO₂ levels. For example, in our earlier paper of >14 million Medicare beneficiaries living near EPA monitoring sites (Eum et al. 2019), we showed NO₂ exposures below the annual NAAQS to be associated with increased mortality risks, consistent with findings from the Cancer Prevention II (Turner et al. 2016) and the Canadian Census Health and Environment Cohort studies (Crouse et al. 2015). These findings raise questions regarding the sufficiency of the annual NAAQS NO₂ standard.

Their findings, however, are limited by their geographic and demographic scope. For example, our recent study of NO₂ on mortality was based on a largely urban cohort (89%) (Eum et al. 2019), while the Cancer Prevention Study II included almost entirely White adults (94.6%) (Turner et al. 2016). As such, the generalizability of these findings to other, less studied populations, is not known. Further, prior studies leave unanswered questions regarding whether a threshold level exists below which NO₂ exposures pose no harm.

In this paper, we assess the association between long-term NO₂ exposure and mortality among a near-complete sample of US Medicare beneficiaries.

2. Materials and methods

2.1. Data Source and Study Population

From the Centers for Medicare and Medicaid Services Medicare Enrollment file, we obtained beneficiary data for ~50 million enrollees living in the conterminous US between 2001 and 2008. For each enrollee, we compiled information on age, sex, race, date of death, and ZIP code of residence. We also obtained cause of death from the National Death Index (NDI) (http://U.S..resdac.org/resconnect/articles/117#cause-of-death). Using International Classification of Disease (ICD–10), we identified deaths from cardiovascular (CVD) and respiratory disease and cancer, which together account for ~74 % of all deaths (Table S1). We also identified deaths from specific subcategories, including ischemic heart disease (IHD), cerebrovascular disease (CBV), congestive heart failure (CHF), chronic obstructive pulmonary disease (COPD), pneumonia, lung cancer, and from aggregate non-accidental causes of mortality. We also classified accidental mortality as a negative control.

2.2. Ambient NO₂, PM_2.5, Black Carbon Exposures

We estimated 12-month moving average NO₂ exposure for beneficiaries based upon their ZIP code of residence using NO₂ estimates from Bechle et al. (Bechle et al. 2015), who used land-use regression (LUR) and spatially varying temporal scaling factors to estimate monthly NO₂ concentrations on a 100-meter grid across the conterminous U.S. from 2000 to 2008 (Bechle et al. 2015). In brief, monthly average NO₂ concentrations were calculated for each of 370 EPA regulatory monitoring stations which met the reliability criterion of ≥75% valid hourly values. These values were scaled using 2006 LUR estimates and through interpolation were used to estimate monthly NO₂ along a 100-meter grid. Estimates had high validity and low error, explaining 81% of the spatial (R² = 0.81), 73% of temporal (R² = 0.73), and 84% of the spatiotemporal variation (R² = 0.84) in monthly mean NO₂ concentrations from 2000 to 2010, with an absolute average bias of 2.4 ppb. Although the mean error was similar, the model performance was lower in rural (R² = 0.69) as compared to urban (R² = 0.80) areas, suggesting greater exposure error for rural ZIP codes.

We adjusted for potential confounding by fine particulate matter (PM_2.5) using daily PM_2.5 concentrations estimated on a 6×6 km grid across the conterminous U.S. from a set of well-validated spatio-temporal smoothing models (Yanosky et al. 2014). The models predicted PM_2.5 concentrations from measured PM_2.5 data from the US EPA Air Quality System, meteorological data, geographical factors, and traffic-related PM_2.5 estimated from a line-source Gaussian dispersion model. Model performance was strong, with a cross-validation R² for daily PM_2.5 concentrations of 0.76. We averaged monthly NO₂ and daily PM_2.5 concentrations for each beneficiary to obtain 12-month moving average exposure estimates for each pollutant. We linked estimated PM_2.5 concentrations to aggregated beneficiary mortality data by month and ZIP code (using the PM_2.5 concentration estimated at the grid point closest to the centroid of each beneficiary’s residential ZIP code), accounting for residential moves.

We adjusted for potential confounding by BC with an aerodynamic diameter less than 2.5 μm using annual BC concentrations for each ZIP code and year from 2000 to 2008 estimated using a combined Geoscience-Statistical Method by van Donkelaar et al. (van Donkelaar et al., 2019). We linked BC estimates to aggregated beneficiary mortality data by calendar year and ZIP code, again accounting for residential moves.

2.3. Urbanicity, Region, and SES assessment

We classified ZIP codes as urban, micropolitan, and rural using the Rural Health Research Center (RUCA) Categorization B. Specifically, ZIP codes were characterized as ‘urban’ (codes: 1.0, 1.1, 2.0, 2.1, 3.0, 4.1, 5.1, 7.1, 8.1, 10.1), ‘micropolitan,’ (codes: 4.0, 4.2, 5.0, 5.2, 6.0, 6.1), or ‘rural,’ (codes: 7.0, 7.2, 7.3, 7.4, 8.0, 8.2, 8.3, 8.4, 9.0, 9.1, 9.2, 10.0, 10.2, 10.3, 10.4, 10.5, 10.6) (http://depts.washington.edu/uwruca/ruca-uses.php). We classified states into four U.S. Census regions: ‘West’, ‘Midwest’, ‘Northeast’, and ‘South’ (https://www.census.gov/prod/1/gen/95statab/preface.pdf). To estimate ZIP code- and state-level SES, we obtained data from the U.S. Internal Revenue Service (IRS) on the annual mean gross adjusted income for each state and ZIP code. The data were based upon the individual tax returns filed with the IRS (Internal Revenue Service).

2.4. Statistical Analyses

For each month between 2001 and 2008, we computed the number of beneficiaries and deaths for each ZIP code, sex, race, age. For each month, we calculated the number of deaths for each age interval, as year from 65–89 years and as one interval for individuals 90 years and older. Since our Cox proportional hazards model stepped through time every month, we aggregated the number of deaths (by cause) for each month, allowing us to calculate the changes in both the number at risk and number of deaths for each month. This discretized version of the Cox PH model essentially resulted in a coarse grid of times with ties that were broken by adding an extremely small amount of noise to each death time. We used a 12-month moving average of NO₂ as our exposure measure, which changes monthly, reflecting the average exposure over the previous 12 months. As our base model, we examined the impact of a 10 ppb increase in 12-month moving average NO₂ exposure on cause-specific mortality in age, sex, race, ZIP code-stratified Cox proportional-hazard models, adjusting for ZIP code- and state-level SES. We used the calendar month as a time scale.

We also fit PM_2.5-adjusted models to assess potential confounding of our results by PM_2.5. To do so, we used a two-stage approach, given the strong correlation between NO₂ and PM_2.5 concentrations (r= 0.59; Table S2). In the first stage, we regressed 12-month moving average NO₂ on a 12-month moving average PM_2.5. In the second stage, we used the residuals from this regression as the exposure measure in Cox proportional-hazard models, with the resulting HRs representing the NO₂-associated mortality risk, adjusting for PM_2.5. In sensitivity analyses, we also fit two-pollutant models that included either NO₂ and PM_2.5 or NO₂ and BC together in the same model.

We fit PM_2.5-adjusted Cox models for the conterminous U.S. as well as for each mutually exclusive U.S. region. We examined effect modification using interaction terms for sex, age (age <75, age ≥75), race (White, Black, Asian, Hispanic), urbanicity (urban, micropolitan, rural), and NO₂ level (low <7 ppb, medium <12 ppb, high >12 ppb). Given that minority beneficiaries lived predominantly in urban locations, we also examined effect modification by race only for urban beneficiaries. We examined the exposure-response curve using restricted cubic splines with 3 knots, which demonstrated superior model performance compared to 4 and 5 knot models based on BIC criterion (Figure S1). We also assessed the exposure-response curve stratified by the potential effect modifiers. We implemented all Cox models in Java, incorporating linkage, grouping and other data mining techniques to reduce memory needs and improve computational efficiency (Wang et al. 2020). For comparison with previous studies, we rescaled results from prior studies to reflect HRs per 10 ppb increase in NO₂ exposure.

3. Results

Our study population included almost 50 million Medicare beneficiaries (65−120 years), with 13.2 million deaths: 96.7% from non-accidental and 2.4% from accidental causes. CVD accounted for 39.9% of all non-accidental mortality, followed by cancer (22.6%), and respiratory mortality (11.2%; Table 1, Table S1). Approximately 78.0%, 11.1%, and 10.9% of beneficiaries lived in urban, micropolitan, and rural areas, respectively. Minority populations lived predominantly in urban areas, with 74.2%, 92.5%, and 85.2% of all Black, Asian, and Hispanic beneficiaries, respectively, while 55.8% of all White beneficiaries lived in urban areas. The annual mean NO₂, PM_2.5, and BC concentrations were 10.9 ppb, 9.6 μg/m³, and 0.78 μg/m³, respectively, with higher mean concentrations in urban (12.8 ppb, 10.4 μg/m³, 0.87 μg/m³) as compared to micropolitan (6.8 ppb, 8.5 μg/m³, 0.67 μg/m³), and rural areas (5.4 ppb, 7.8 μg/m³, 0.61 μg/m³). The NO₂ concentration were higher in the West (14.45 ppb) and Northeast (14.38 ppb) regions than Midwest (9.55 ppb) and South (7.98 ppb) regions (Table 1, Figure S2). Mean 12-month NO₂ exposure also differed by race, with Asians (14.3±6.8 ppb), Hispanics (13.6±7.4 ppb), and Black (11.1±6.4 ppb) beneficiaries having higher exposures as compared to White (9.5±5.7 ppb). The interquartile range (IQR) for 1-year moving average NO₂ concentrations equaled 7.86 ppb for our entire study population, with IQR values of 8.12 ppb, 3.12 ppb, and 2.46 ppb for participants living in urban, micropolitan, and rural areas, respectively.

Table 1.

General Characteristics of the Study Population

Population	N (%) or Mean ± SD
Persons	49,712,702
Deaths	13,213,500
Person-months	3,119,343,105
Pollution concentration	Mean ± SD	Median (25th, 75th)	Range
NO2, ppb	10.90 ± 6.44	9.14 (6.13, 13.98)	1.03 – 47.04
PM2.5, μg/m3	9.59 ± 3.12	9.44 (7.56, 11.45)	0.70 – 25.79
Black Carbon, μg/m3	0.78 ± 0.38	0.70 (0.53,0.95)	0.00 – 3.23
Individual characteristics	N (%)	NO2, ppb Mean ± SD	PM2.5, μg/m3 Mean ± SD	Black Carbon, μg/m3 Mean ± SD
Sex
Male	22,032,670 (44.32)	10.84 ± 6.43	9.55 ± 3.13	0.77 ± 0.38
Female	27,680,032 (55.68)	10.94 ± 6.45	9.62 ± 3.12	0.78 ± 0.38
Race
White	43,220,223 (86.94)	9.51 ± 5.65	9.17 ± 3.00	0.72 ± 0.33
Black	4,518,885 (9.9)	11.13 ± 6.40	10.61 ± 2.74	0.88 ± 0.35
Asian	512,041 (1.03)	14.33 ± 6.75	10.54 ± 3.33	0.94 ± 0.48
Hispanic	805,346 (1.62)	13.61 ± 7.36	9.85 ± 3.55	0.85 ± 0.50
Native American	656,208 (1.32)	9.30 ± 5.79	7.75 ± 3.30	0.61 ± 0.35
Neighborhood characteristics
Urban	38,775,908 (78.00)	12.8 ± 6.5	10.4 ± 3.1	0.87 ± 0.41
Micropolitan	5,518,110 (11.10)	6.8 ± 2.3	8.5 ± 2.6	0.67 ± 0.27
Rural	5,418,685 (10.90)	5.4 ± 1.9	7.8 ± 2.5	0.61 ± 0.27
Geological regions
West	9,296,275 (18.70)	14.45 ± 7.48	8.39 ± 4.53	0.81 ± 0.61
Midwest	10,489,380 (21.10)	9.55 ± 5.25	9.72 ± 2.83	0.66 ± 0.23
South	11,985,732 (24.11)	7.98 ± 3.75	10.07 ± 2.29	0.80 ± 0.28
Northeast	17,936,343 (36.08)	14.38 ± 7.18	9.79 ± 2.71	0.88 ± 0.33

3.1. Associations of Long-term NO₂ and Specific Causes of Death

In base models, we found 12-month moving average NO₂ to be significantly associated with increased mortality for all examined causes of death, with increased risks ranging from 4% for lung cancer (HR: 1.04; 95% CI: 1.02–1.07) to 33% for pneumonia (HR: 1.33; 95% CI 1.29–1.37) (Table 2). While attenuated, HRs remained significant and positive for all causes of death after adjusting for PM_2.5. PM_2.5-adjusted HRs were highest for CVD-related mortality (HR: 1.10; 95% CI: 1.10–1.11), with 1.12 (95% CI: 1.11–1.13), 1.08 (95% CI: 1.06–1.10), and 1.10 (95% CI: 1.06–1.13) times the risk of death for IHD, CBV, and CHF, respectively (Table 2). For all respiratory mortality, we observed a HR of 1.09 (95% CI: 1.08–1.11); risks were higher for pneumonia (HR: 1.23; 95% CI: 1.19–1.27) and lower for COPD (HR: 1.03; 95% CI: 1.01–1.05). While lower, NO₂-associated HRs (HR: 1.03; 95% CI: 1.01–1.05) were positive and significant for lung cancer mortality and were marginally significant for all cancer mortality (HR: 1.01; 95% CI: 1.00–1.02). Associations between NO₂ and accidental mortality were null for all examined models, supporting the validity of the analysis. When we additionally adjusted for year and month (Table S4), we found similar patterns of association. We found greater associations between NO₂ and cause-specific mortality in models that did not adjust for ZIP- and state-SES (Table S5), or that used the two-pollutant models (versus our two-stage approach) to control for confounding by PM_2.5 or BC exposure (Table 2).

Table 2.

Mortality Hazard Ratios (95% CI) Associated with a 10 ppb Increase in 12-month Moving Average NO₂ Exposure: by Cause of Death in Base and PM_2.5- and BC-Adjusted Models

Cause of Death	HR a	PM2.5-Adjusted HR		BC-Adjusted HR d
		2-stageb	2-pollutant c
All-Cause	1.14 (1.13,1.15)	1.06 (1.05,1.06)	1.14 (1.13,1.15)	1.14 (1.13,1.15)
Non-Accidental	1.14 (1.14,1.15)	1.06 (1.06,1.07)	1.15 (1.14,1.15)	1.15 (1.14,1.15)
Accidental	0.99 (0.95,1,03)	0.98 (0.95,1.02)	0.98 (0.94,1,02)	0.98 (0.94,1,02)
All cardiovascular	1.23 (1.22,1.24)	1.10 (1.10,1.11)	1.24 (1.23,1.25)	1.24 (1.23,1.26)
IHD	1.26 (1.24,1.27)	1.12 (1.11,1.13)	1.27 (1.25,1.29)	1.28 (1.27,1.30)
CBV	1.25 (1.22,1.28)	1.08 (1.06,1.10)	1.24 (1.21,1.27)	1.28 (1.25,1.30)
CHF	1.16 (1.12,1.21)	1.10 (1.06,1.13)	1.19 (1.15,1.24)	1.19 (1.14,1.23)
All respiratory	1.17 (1.15,1.19)	1.09 (1.08,1.11)	1.19 (1.17,1.21)	1.17 (1.15,1.19)
COPD	1.06 (1.03,1.08)	1.03 (1.01,1.05)	1.06 (1.04,1.09)	1.06 (1.03,1.08)
Pneumonia	1.33 (1.29,1.37)	1.23 (1.19,1.27)	1.41 (1.36,1.47)	1.36 (1.31,1.40)
All Cancer	1.04 (1.03,1.06)	1.01 (1.00,1.02)	1.04 (1.03,1.05)	1.04 (1.03,1.06)
Lung Cancer	1.04 (1.02,1.07)	1.03 (1.01,1.05)	1.05 (1.03,1.08)	1.04 (1.02,1.07)

^aBase models with strata for sex (male, female), race (White, non-White), age (1 year age categories with 90+ years old as one category), ZIP code and adjusted for ZIP code- and state-level SES by including the two SES variables as covariates.

^bBase models with additional adjustment for 12-month PM_2.5 using a two-stage approach.

^cBase models with additional adjustment for PM_2.5 using 2-pollutant modeling approach.

^dBase models with additional adjustment for annual BC using a 2-pollutant modeling approach.

The shape of the exposure-response curves varied by cause of death (Figure 1). Associations of long-term NO₂ exposure and non-accidental and lung cancer mortality were linear, while those for CVD, IHD, CVB, and pneumonia mortality were supra-linear, with higher HRs when NO₂ exposures were lower as compared to higher than ~12 ppb. In contrast, exposure-response curves for respiratory mortality were sub-linear in shape, with lower risks when NO₂ exposures were less than ~12 ppb.

Figure 1.

Non-linear Relationship of Mortality Hazard Ratio (95% CI) for Cause-Specific Mortality and 10 ppb Increase in 12-month Moving Average NO₂. All analyses were conducted using Cox Proportional Hazard models with strata for sex (male, female), race (White, non-White), age (1 year age categories with 90+ years old as one category), and ZIP code, and including ZIP code- and state-level SES as model covariates. Models were also adjusted for 12-month moving average PM_2.5 using a 2-stage approach. The bold solid line represents the non-linear relationship estimated using restricted cubic spline (3 knots); the dotted line represents the 95% C.I.s for the non-linear association, and the muted solid line represents the linear association for (A) All-cause and non-accidental mortality, (B) Cardiovascular-related mortality, (C) Respiratory-related mortality, (D) Cancer and lung cancer mortality

3.2. Effect Modification

Of the examined effect modifiers, race had the greatest impact on NO₂-associated HRs (Figure 2, Table S6). PM_2.5-adjusted HRs for Black, White, and Asian beneficiaries were positive and significant for most causes of death. Black beneficiaries had the highest NO₂-associated risks of death, with HRs 1.3 to 1.9 times higher for CVD-related diseases, 3.7 to 12 times higher for respiratory-related disease, and 7 times higher for cancer and lung cancer, as compared to White beneficiaries (p-values for differences < 0.05). Likewise, Asian and Hispanic beneficiaries also had higher risks of respiratory and COPD mortality as compared to White beneficiaries (p-values for differences < 0.05). NO₂-associated HRs for Black, Asian, and Hispanic, but not White beneficiaries, were higher for respiratory- as compared to CVD-related mortality. NO₂-associated risks of lung cancer were highest for Black and Asian, and while lower, were also significant and positive for White beneficiaries. NO₂-associated risks of lung cancer for Hispanic beneficiaries were null. When we limited our analyses to beneficiaries living in urban ZIP codes, we found similar pattern of associations by race (Table S7).

Figure 2.

Mortality Hazard Ratio (95% CI) Associated with a 10 ppb Increase in 12-month Moving Average NO₂ Exposure by Beneficiary Characteristics. Analyses were based on Cox proportional hazard models with strata for sex (male, female), race (White, non-White), age (1 year age categories with 90+ years old as one category), and ZIP code, and including ZIP code- and state-level SES as model covariates. Models were also adjusted for 12-month moving average PM_2.5 using a 2-stage approach.

The exposure-response curves for CVD, IHD, and CBV mortality were again supra-linear for all racial groups (Figure S3, S3a). However, the curves for Hispanic and Asian beneficiaries were positive and significant only until ~66^th percentile of exposure (12 ppb), after which the associations were non-significant. The associations for lung cancer mortality were largely linear and positive for Asian and Black beneficiaries, but as in the linear analysis, were null for Hispanics. Conversely, the curves for all races for CHF, respiratory, and COPD mortality were sub-linear, with lower associations at lower levels of exposure. For cancer, NO₂ exposures were significantly associated with mortality in all races other than Hispanics, for whom associations were null. (Table S6, Figure 2, S3, S3a).

HRs also varied by age, sex, and urbanicity. Associations for CVD, IHD, CHF, respiratory, COPD, and lung cancer were higher for older as compared to younger beneficiaries (Figure 2, Table S6). The curves for each age group largely mirrored those for the whole population (Figure S4, S4a). Associations for both sexes were significant and positive, except for COPD, cancer, and lung cancer mortality in men. HRs for men and women were comparable for all-CVD, CBV, and CHF, and all respiratory mortality. For other causes of death, HRs were higher for women as compared to men (Table S6, Figure 2). Exposure-response curves were similar to those we observed by age (Figure S5, S5a).

NO₂-associated mortality risks were higher among beneficiaries living in urban, as compared to non-urban ZIP codes for non-accidental, CHF, and respiratory-related causes of death (Table S6, Figure 2). In contrast, HRs for CVD, IHD, CHF, cancer, and lung cancer mortality were similar regardless of urbanicity. As with age and sex, exposure-response curves were largely linear for all-cause, lung cancer, and pneumonia, supra-linear for CVD, IHD, and CBV, and sub-linear for CHF, respiratory, and COPD mortality (Figure S6, S6a).

4. Discussion

In our cohort of Medicare beneficiaries, we found that long-term exposure to NO₂ was associated with increased mortality from all causes as well as from specific causes related to CVD, respiratory diseases, and cancer. Significant associations remained after adjusting for PM_2.5 and BC, suggesting the adverse effect of NO₂ is independent of that from PM_2.5 and BC. Risks varied by beneficiary characteristics, including race, age, urbanicity, and to a lesser extent by sex. Importantly, we found no evidence of a threshold in NO₂-associated HRs on mortality from non-accidental, CVD, respiratory and lung cancer mortality, with significant and positive associations across the range of examined NO₂ exposures.

Our findings of increased mortality risk are consistent with those from previous studies (Beelen et al. 2008; Crouse et al. 2015; Cesaroni et al. 2013; Eum et al. 2019; Filleul et al. 2005; Fischer et al. 2015; Gehring et al. 2006; Hart et al. 2011; Heinrich et al. 2013; Katanoda et al. 2011; Krewski et al. 2000; Schikowski et al. 2007; Turner et al. 2016; Zhang et al. 2011), including our earlier study of Medicare beneficiaries living near air pollution monitoring sites (Eum et al. 2019) (Figure S7). As in our prior study, we found an increased risk of non-accidental mortality, CVD, IHD, CBV, cancer, respiratory disease, and pneumonia (Eum et al. 2019; Figure S7), with similar HRs as found previously for non-accidental mortality (HR: 1.06 vs. 1.04), CVD (HR: 1.10 vs. 1.11), CBV (HR: 1.08 vs. 1.05), all cancer (HR: 1.01 vs. 1.02), pneumonia (HR: 1.23 vs. 1.29), and respiratory disease (HR: 1.09 vs. 1.03). In contrast, we found statistically significant increased risk of COPD (HR: 1.03; 95% CI: 1.01–1.05) and lung cancer (HR: 1.03; 95% CI: 1.01–1.05) mortality in our present study, for which associations were null in our prior study, possibly due to its smaller sample size. Similarly, other studies (Figure S7) also found the increased risk of non-accidental mortality (or all cause mortality), CBV, respiratory disease, COPD, or lung cancer (Beelen et al. 2008; Crouse et al. 2015; Cesaroni et al. 2013; Filleul et al. 2005; Fischer et al. 2015; Gehring et al. 2006; Hart et al. 2011; Heinrich et al. 2013; Katanoda et al. 2011; Krewski et al. 2000; Schikowski et al. 2007; Turner et al. 2016; Zhang et al. 2011). We also newly found a significant 10% increased risk of CHF mortality (HR: 1.10; 95% CI: 1.06–1.13), an association which has only rarely been explored in prior literature, given that CHF comprises only a small proportion (3%) of overall deaths. It is also notable that adjustment for PM_2.5 resulted in attenuated effect estimates across all mortality outcomes, and the impact of this adjustment was greater in rural environments than in urban environments (Table S6, S8). This may be due to differences in the correlation of NO₂ and PM_2.5 (Beelen et al. 2014; Monn et al. 1995; Putaud et al. 2010), in PM_2.5 composition (Kulshreshtha et al. 2014), and/or mortality outcome ascertainment (Kulshreshtha et al. 2014) between urban and non-urban environments.

Our findings were insensitive to model specifications, with similar patterns of association in models additionally adjusting for year or month (Table S4), although confounding by weather remains possible, given that both mortality and NO₂ are associated with weather (Marti-Soler and Marques-Vidal 2015; Braga et al. 2002), which warrants further study.

We showed the shape of the exposure-response curve for NO₂ and mortality to vary by aggregate cause of death groupings (e.g. cardiovascular, respiratory), but to be generally consistent within these groupings. For example, we found linear associations for non-accidental and lung cancer mortality, supra-linear associations for all CVD, IHD, and CBV mortality, and sub-linear (but still significant and positive) associations for all respiratory and COPD mortality. These findings add substantially to the scientific literature, as to date only two studies have examined the shapes of the exposure-response curves for NO₂ and mortality, with mixed results. As in our study, Cesaroni et al. (Cesaroni et al. 2013) found associations to be linear for non-accidental and lung cancer mortality and supra-linear for IHD mortality, but showed a linear association for CVD mortality. In contrast, in a Canadian cohort, Crouse et al. (Crouse et al. 2015) reported a supra-linear pattern of association for all-cause mortality, with higher risks below as compared to above approximately 10 ppb. Lack of adjustment for PM_2.5 in the previous studies may explain the different findings (Crouse et al. 2015; Cesaroni et al. 2013). Neither study examined the impact of NO₂ on other causes of death, or the shape of the exposure-response curve for different sub-populations based on demographic characteristics or place of residence.

We found both demographic characteristics and place of residence to be important modifiers of the NO₂-mortality association. Of these factors, we found race to have the largest impact on NO₂-associated HRs, with minority, especially Black beneficiaries being particularly susceptible to the impacts of NO₂. Higher HRs for minority may result from their higher NO₂ exposures, with Asians having the highest exposures, followed by Hispanic and Black beneficiaries, and finally White beneficiaries, consistent with previous studies (Clark et al. 2014; Grineski et al. 2007; Jones et al. 2014; Su et al. 2011). For respiratory mortality, higher NO₂ exposures for minority beneficiaries corresponded to higher mortality risks, as evidenced by its sub-linear curve, with higher HRs for NO₂ exposures above 12 ppb. Correspondingly, we found lower HRs for CVD mortality among Asian and Hispanic as compared to White beneficiaries despite their higher overall exposure, which is consistent with our findings of lower mortality risks at higher NO₂ exposures for CVD mortality. Notably, Black beneficiaries had the highest CVD mortality risks of all racial groups, possibly reflecting the fact that their average NO₂ exposures, while higher than that for Whites, were below the inflection point of our exposure-response curves, for which HRs were higher. Alternatively, high CVD mortality risks may also be result from the influence of other, unmeasured factors. While ours is the first study to examine modification of the NO₂-mortality association by race, our findings are indirectly supported by prior studies showing Black persons to have higher PM_2.5-related mortality risks (Di et al. 2017; Wang et al. 2017), as found in our study of PM_2.5 and cause-specific mortality in the Medicare cohort (Wang et al. 2020). We note that our findings of effect modification by race does not control for individual level socioeconomic status, which is closely tied with race/ethnicity in the United States (Tibuakuu et al. 2018). Since individual-level race/ethnicity data were not available in our study, our effect modification analyses controlled for ZIP code-level SES and were restricted to beneficiaries living in urban areas, given that the majority of non-White beneficiaries lived in urban areas. However, given the lack of control for individual level SES, our findings of effect modification by race may reflect the combined impacts of both race/ethnicity and SES, or as suggested by Gwynn and Thurston (Gwynn and Thurston 2001) and Grineski et al. (Grineski et al. 2010), modification by SES instead. Future analyses should control or conduct stratified analyses using individual-level SES data to better understand whether and how NO₂-associated mortality impacts differ separately or jointly by race and SES.

Associations of NO₂ and mortality were also significantly modified by urbanicity. Urban beneficiaries showed generally greater mortality risks than their rural counterparts, consistent with their higher mean NO₂ exposures (Table 1) and higher proportion of minorities (18.75%, 9.22%, 8.4% of Black, Asian and Hispanic among urban population versus 9.09%, 1.03%, 1.62% among the whole study population). There findings are consistent with a previous Minnesota study, which found minorities and lower SES urban populations to be exposed to higher levels of traffic-related air pollution and to be at higher health risk, than those living in rural environments (Want et al. 2017). Together, our findings suggest that the impact of NO₂ on mortality vary by both race and urbanicity, underpinning the importance of estimating risk among representative sub-populations.

While our study adds to the literature showing significant and positive associations between NO₂ and mortality for several causes of death, the biological mechanism through which long- term NO₂ exposures may independently cause mortality is not well understood. It is possible, however, that observed NO₂-associated sub-clinical effects contribute over long time periods to increased mortality. NO₂ exposures, for example, have been shown to cause lipid peroxidation in cell membranes and damage to structural and functional molecules by the release of free radicals (Sandström, 1995). Consistent with this, NO₂ exposures have been associated with pulmonary inflammation, bronchial hyperresponsiveness, and increased risk of respiratory infection in toxicological studies (Koenig, 2000), while epidemiological findings show associations between NO₂ exposure and increased systemic inflammation and blood pressure and decreased pulmonary function (Gao et al. 2020; Lepeule et al. 2014). Importantly, these NO₂-associated sub-clinical impacts have been shown to be independent of PM_2.5.

It is possible that these NO₂-associated adverse effects contribute over long time periods to increased mortality, as has been observed in ours and other epidemiological studies. Importantly, our findings suggest independent impacts of NO₂ on mortality, given that NO₂ remained significantly and positively associated with mortality even after adjustment for PM_2.5 and BC. However, it is possible that some unmeasured confounding by PM_2.5 and/or BC remains.

Our study has several limitations. First, while we used ambient NO₂ and BC estimates based upon validated models, exposure misclassification is likely (Bechle et al. 2015; Yanosky et al. 2014). As stated in Bechle et al. (Bechle et al. 2015), exposure error was similar in areas near versus far from EPA monitoring sites and was not dependent on urbanicity, which was associated with mortality in our study and in others (House et al. 2000). Since the vast majority of our cohort lived in urban areas, exposure misclassification in our study was likely non-differential, and any bias in effect sizes in our study would likely be towards the null (Hart et al. 2015a, Zeger et al. 2000). Support for this is provided by previous epidemiological studies that corrected for measurement error using calibration factors that accounted for differences between personal exposures and corresponding ambient concentrations, finding an increase in the association’s magnitude when the exposure measurement error corrections were made (Avery et al. 2010a; Avery et al. 2010b; Hart et al. 2015b). Consistent with this, when we additionally adjusted for urbanicity, associations were similar, supporting the robustness of the findings among the whole population. We note, however, that since the error associated with NO₂ estimates in rural areas was relatively larger, our associations of 1-year NO₂ exposures and cause-specific mortality for rural populations may be more vulnerable to exposure misclassification. Second, while our analyses controlled for individual covariates, including age, sex, race, and residential ZIP code, and ZIP code- and state-level SES, we were unable to adjust for individual-level SES or behavioral confounders, given our reliance on Medicare data, which lack this information. Despite this, confounding by individual-level SES or behavioral characteristics is unlikely to explain our findings, given results from Krewski et al. (Krewski et al. 2009) which found little change in air pollution-associated mortality risk estimates after adjustment for individual-level characteristics. Nonetheless, residual confounding by unmeasured covariates, such as individual-level SES remains a possibility (Tibuakuu et al. 2018). Third, our data included only older adults living in the U.S. between 2000 and 2008, limiting generalizability to younger or non-U.S. populations and to a lesser extent more recent time periods. While the gender, race and age composition of the US older adult population has remained largely unchanged from 2008 to now (U.S Census Bureau, 2008; U.S Census Bureau, 2019), annual NO₂ concentrations have decreased in the US across this same time period. As a result, NO₂-associated mortality risks at the high end of the NO₂ concentration distribution may be less applicable to the current US older adult population, given that fewer US older adults experience long-term NO₂ exposures of this magnitude. Our findings for lower NO₂ exposures, however, are generalizable to today’s US older adults, as evidenced by our exposure-response curves, which show 1-year moving average NO₂ exposures to be associated with higher mortality risks at concentrations below 12 ppm with very low associated uncertainty. Fourth, we used a two-stage model to adjust for PM_2.5 given collinearity between ambient PM_2.5 and NO₂ concentrations, which may overcontrol for PM_2.5 concentrations and underestimate NO₂-associated mortality risks. Support for this theory is provided by our findings from a two-pollutant model that includes both NO₂ and PM_2.5, which showed greater effects for NO₂ as compared to that found in our two-stage model. These findings are consistent with an independent effect of long-term NO₂ exposures on mortality. While our findings may also reflect the ability of NO₂ to serve as a better proxy of combustion-related PM_2.5, as has been found in previous studies (Pope et al. 1995), our finding of similar HRs for NO₂ in single and BC-adjusted models suggests this to not be the case.

These limitations are balanced by our study’s substantial strengths, including its large size and its ZIP code-specific pollution estimates, which provide ample statistical power to estimate associations and exposure-response curves in understudied populations. Our study provides strong evidence of an association between NO₂ and increased all-cause and cause-specific mortality in U.S. older adults, especially for racial minorities and older adults living in urban areas. Black beneficiaries, those living in urban environments, and those >75 years were found to be most susceptible. Given that NO₂ exposures were well below the annual NAAQS, our findings suggest that its current annual NAAQS does not sufficiently protect public health.

Highlights.

• Long-term NO₂ exposure was associated with increased nonaccidental mortality, as well as mortality from cardiovascular disease, respiratory disease, and cancer.

• NO₂-mortality associations were without a lower threshold, suggesting that any increase in long-term NO₂ exposures is associated with elevated mortality risk.

• The increased mortality associated with NO₂ exposure was considerably higher among the Black beneficiaries than other races.

Abbreviations

NAAQS

National Ambient Air Quality Standard

NO₂

Nitrogen dioxide

PM_2.5

Particulate matter with an aerodynamic diameter less than 2.5 μm

HR

Hazard ratio

CI

Confidence interval

ICD

International Classification of Disease

CVD

Cardiovascular disease

IHD

Ischemic heart disease

CBV

Cerebrovascular disease

CHF

Congestive heart failure

COPD

Chronic obstructive pulmonary disease

LUR

land-use regression

RUCA

Rural health research center

IRS

Internal Revenue Service

SD

standard deviation