Air pollution below US regulatory standards and cardiovascular diseases using a double negative control approach

Long-term exposure to ambient air pollution is recognized as a modifiable risk factor for cardiovascular diseases, but specific cardiovascular outcomes such as stroke, heart failure (HF), and atrial fibrillation and flutter (AF) have been less studied than overall cardiovascular mortality and morbidity. Prior work suggests long-term PM2.5 is associated with stroke incidence, hospitalization, and mortality, and that HF and AF risks may also increase with chronic pollution exposure, though evidence is scarce and sometimes inconsistent, particularly for NO2 and O3. Importantly, most studies examine the full exposure range; far fewer focus on health effects at concentrations below regulatory standards. Several large cohort studies in the US and Europe indicate curvilinear exposure-response (E-R) relationships for PM2.5 with no clear threshold and possibly stronger relative risks at lower concentrations. This has major implications for settings like the US where average exposures are relatively low. Additionally, unmeasured confounding remains a concern in observational studies, and methods such as propensity scores cannot address unmeasured variables. Negative control strategies, including using future exposure or negative outcomes, offer a way to detect and reduce residual confounding. To fill these gaps, this study applies a double negative control approach to evaluate associations of long-term low-level PM2.5, NO2, and warm-season O3 with hospitalizations for stroke, HF, and AF among US Medicare beneficiaries, and assesses potential effect modification by demographic subgroups.

Multiple studies report positive associations between long-term PM2.5 and cardiovascular outcomes, often without an apparent threshold, with some indicating stronger relative risks at low concentrations. Regulatory comparisons note US EPA annual limits (9 µg/m3 for PM2.5; 53 ppb for NO2) and EU guidelines (10 µg/m3 for PM2.5; 20 µg/m3 for NO2), while WHO recommends even lower targets (5 µg/m3 PM2.5; 10 µg/m3 NO2). Evidence for NO2 and O3 in relation to specific cardiovascular morbidities is less comprehensive, and correlations between pollutants complicate interpretation; recent works emphasize multi-pollutant modeling. Regarding causal inference tools, negative control exposures (e.g., future pollution) and negative control outcomes have been used in air pollution research, and double negative control techniques have recently been employed to mitigate unmeasured confounding in both short- and long-term studies. The literature also highlights potential susceptibility in disadvantaged populations and varying E-R shapes at low exposures, supporting focused analyses in low-pollution ranges and attention to effect modification.

Study design and population: A national cohort of fee-for-service Medicare beneficiaries aged ≥65 across the contiguous US was followed from 2000–2016. Follow-up started January 1 of the year after Medicare enrollment and continued until first hospitalization for the outcome, death, censoring, or end of study. Analyses were restricted to ZIP codes with >100 beneficiaries and consistently low exposure across 2000–2016, defining separate datasets per pollutant: PM2.5 < 9 µg/m3, NO2 < 75.2 µg/m3 (40 ppb), and warm-season O3 < 88.2 µg/m3 (45 ppb). Outcomes: Primary discharge diagnoses for stroke (ICD-9 430–438; ICD-10 I60–I69), heart failure (ICD-9 428; ICD-10 I50), and atrial fibrillation and flutter (ICD-9 427.3; ICD-10 I48) were identified in the MEDPAR database. ZIP code–level annual counts were computed; all hospitalizations after enrollment were counted. Exposure assessment: Daily PM2.5, NO2, and O3 at 1 km × 1 km resolution were obtained from ensemble machine-learning prediction models incorporating meteorology, chemical transport models, land-use and satellite data, validated by 10-fold cross-validation. PM2.5 and NO2 were aggregated to annual averages; O3 was averaged over warm season (Apr 1–Sep 30). ZIP code–level exposures were computed by averaging grid cells within ZIP polygons, or nearest cell for non-polygons, and linked by residential ZIP code and year. Covariates: ZIP code/ZCTA-level SES (percent Black, percent Hispanic, ≥65 in poverty, population density, percent ≥65 without high school diploma, median home value, median household income, percent owner-occupied housing), lung cancer hospitalization rates (surrogate for smoking) from MEDPAR, county-level ever smokers and mean BMI from BRFSS, access-to-care metrics from the Dartmouth Atlas (HbA1c testing, lipid panel, eye exam, ambulatory visits, mammography), distance to nearest hospital, and meteorology (summer and winter average temperature and relative humidity from gridMET). Missing area-level covariates were imputed by linear interpolation/extrapolation; <1% remaining missingness was excluded. Statistical analysis: The primary approach was a double negative control strategy using a quasi-Poisson outcome model to address unmeasured confounding. Negative exposure control (Z) was exposure in the subsequent year; negative outcome control (W) was the preceding year’s outcome counts. Under assumptions that unmeasured confounders (U) are linearly correlated with A and Z and/or that correlations with A and Z are of equal magnitude (βuz=βua), bias can be removed by including predicted W (from regressing W on A and Z) or by differencing coefficients of A and Z. Models adjusted for SES, behavioral, access-to-care, meteorological covariates, and year indicators. Both single- and three-pollutant models were fit; the main results emphasized three-pollutant double negative control models. As secondary analyses, generalized linear models (GLM) without negative controls were run, and E-R curves were explored using natural splines (df=3) in GLMs adjusted for co-pollutants. Effect modification was examined via stratified analyses by age (65–74, 75–84, 85+), sex, race (White, Black), and Medicaid eligibility, with between-strata comparisons using normal approximation for differences in coefficients. Effects were reported as percent change in hospitalization rate per unit increase in exposure (µg/m3 for PM2.5; units reported in results for NO2 and O3). Analyses were performed in R 4.2.3.

- In three-pollutant double negative control models restricted to low exposure areas: • PM2.5 (<9 µg/m3): per 1 µg/m3 increase was associated with +1.82% (95% CI: 1.44%, 2.19%) stroke hospitalization rate; +2.83% (95% CI: 2.36%, 3.30%) HF; +0.13% (95% CI: -0.39%, 0.65%) AF. • NO2 (<75.2 µg/m3): per 1 µg/m3 increase was associated with +0.01% (95% CI: -0.002%, 0.03%) stroke; +0.18% (95% CI: 0.16%, 0.19%) HF; +0.09% (95% CI: 0.07%, 0.10%) AF. • Warm-season O3 (<88.2 µg/m3): per 1 µg/m3 increase was associated with +0.32% (95% CI: 0.27%, 0.38%) stroke; +0.05% (95% CI: -0.01%, 0.12%) HF; +0.12% (95% CI: 0.04%, 0.20%) AF. - Co-pollutant adjustment generally strengthened PM2.5 associations; NO2 and O3 estimates were similar across modeling approaches. GLMs without negative controls yielded comparable but slightly different effect sizes. - Hospitalization rates in low-pollution areas were approximately 0.87% (stroke), 0.84% (HF), and 0.41% (AF) annually in areas with concurrently low PM2.5, NO2, and O3; rates varied modestly by pollutant-specific low areas, with higher rates in low NO2 areas (where PM2.5 was more typical). - Subgroup analyses: • PM2.5: Larger effects among Black and Medicaid-eligible beneficiaries for all or multiple outcomes; HF effects stronger in ages 65–74; no consistent sex modification. • NO2: Greater risks among age ≥85 and non–Medicaid-eligible for stroke (and similar patterns for HF and AF); males had higher NO2-associated HF risk; White beneficiaries had higher NO2-associated HF and AF risks than Black beneficiaries. • O3: Greater risks in ages 65–74 for all outcomes; Black individuals more susceptible for stroke and HF; females more susceptible for HF and AF; Medicaid-eligible had higher HF risk. - Exposure-response: For PM2.5, positive associations for stroke and HF down to the lowest concentrations; AF showed positive association above ~5 µg/m3 and negative below. For NO2, nearly linear positive curve for HF with steeper slope below 20 µg/m3; stroke and AF curves negative at very low levels then increasing around ~25 µg/m3. For O3, linear positive association with stroke; non-linear patterns for HF and AF with increases toward higher concentrations.

Findings indicate that even at concentrations below current US EPA standards, long-term exposure to PM2.5, NO2, and warm-season O3 is associated with increased hospitalization rates for stroke, HF, and AF among older adults. PM2.5 and NO2 showed the strongest associations with HF, while O3 effects were most pronounced for stroke. The double negative control approach likely reduced bias from unmeasured confounding, and consistency with GLM results suggests any remaining confounding is small (negative for PM2.5 and positive for NO2). Multi-pollutant modeling helped address confounding by co-pollutants. The E-R curves showed no clear thresholds for PM2.5 (stroke and HF) and steeper NO2–HF risks at very low levels, implying health benefits from further reductions even within low ranges. Mechanistically, systemic inflammation, oxidative stress, autonomic dysfunction, endothelial dysfunction, and thrombogenicity provide plausible pathways linking air pollutants to cardiovascular events. Environmental justice patterns emerged: Black and Medicaid-eligible beneficiaries exhibited greater susceptibility to PM2.5 and O3, and younger-old and females appeared more susceptible to O3. In contrast, NO2-related risks showed different modification patterns (e.g., higher risks in White, non–Medicaid-eligible, and very old groups), potentially reflecting urbanicity/commercial activity proxies and differences in baseline risk. Overall, results support tightening air quality guidelines and targeted protections for vulnerable subgroups.

Using a double negative control approach in a large nationwide Medicare cohort, the study found positive associations of long-term low-level PM2.5, NO2, and warm-season O3 with hospitalizations for stroke, HF, and AF, with evidence of susceptibility among certain demographic and socioeconomic groups. The absence of clear thresholds and observed risks at low concentrations suggest current National Ambient Air Quality Standards for PM2.5 and NO2 may be insufficient, and that consideration of guidelines for warm-season O3 is warranted. Future research should refine causal inference in multi-pollutant contexts, investigate mechanistic pathways and exposure-response shapes at very low levels, and identify and protect vulnerable populations.

- Generalizability: Results may not extend to younger populations or highly polluted regions. - Residual confounding: Despite extensive adjustment and the double negative control framework, unmeasured confounding could persist if assumptions (linearity of U with A and Z or equal magnitude correlations) are violated. - Collinearity/over-control: Moderate correlations between annual PM2.5 and NO2 may introduce collinearity and potential over-adjustment in multi-pollutant models. - Exposure misclassification: ZIP code–level ambient estimates may not fully capture personal exposures, indoor sources, or time-activity patterns; however, likely Berkson-type error reduces bias. - Mobility and exposure history: Low-exposure selection across 17 years reduces but does not eliminate potential misclassification from migration/travel. - Outcome measurement: Hospital discharge data may miss milder cases; potential SES-related differences in healthcare access could yield differential classification. - Unit reporting: Differences in unit conventions (µg/m3 vs ppb for gases) and model frameworks could complicate direct comparisons across studies.