Long-term ozone exposures and cause-specific mortality in a US Medicare cohort

We examined the association of long-term, daily 1-h maximum O3 (ozone) exposures on cause-specific mortality for 22.2 million US Medicare beneficiaries between 2000–2008. We modeled the association between O3 and mortality using age-gender-race stratified log-linear regression models, adjusted for state of residence. We examined confounding by (1) adjusting for PM2.5 (particles with aerodynamic diameters <2.5 μm) and NO2 (nitrogen dioxide) exposures, temperature, and neighborhood-level characteristics and behaviors, and (2) decomposing O3 into its temporal and spatio-temporal components and comparing estimated risk ratios. We also examined sensitivity of our results to alternate exposure measures based on warm-season 8-h daily maximum and 24-h average exposures. We found increased risks from long-term O3 exposures to be strongest and most consistent for mortality from respiratory disease (1.030, 95% CI: 1.027, 1.034) (including COPD (chronic obstructive pulmonary disease)), CHF (congestive heart failure), and lung cancer (1.015, 95% CI: 1.010, 1.020), with no evidence of confounding by PM2.5, NO2, and temperature and with results similar across O3 exposure measures. While significant, associations between long-term O3 exposures and CVD (cardiovascular)-related mortality (1.005, 95% CI: 1.003, 1.007) were confounded by PM2.5 and varied with the exposure measure, with associations no longer significantly positive when warm-season 8-h maximum or 24-h average O3 was used to assess exposures. In this large study, we provide strong evidence that O3 exposure is associated with mortality from respiratory-related causes and for the first-time, lung cancer, but raise questions regarding O3-related impacts on CVD mortality. Our findings demonstrate the need to further identify potential confounders.


Introduction
A small but growing number of studies have examined associations between long-term exposure to O 3 and causespecific mortality [1][2][3][4], with mixed results. In an analysis of the ACS (American Cancer Society) cohort, for example, Jerrett et al. [1] found significant associations between warm-season averages of 1-h daily maximum O 3 exposures and increased CVD, IHD (ischemic heart disease), respiratory and cardiopulmonary mortality, but not all-cause mortality. However, with the exception of respiratory mortality, these associations disappeared after adjustment for PM 2.5 , suggesting confounding by PM 2.5 of O 3 -mortality associations for cardiovascular-related causes [1].
Correspondingly, in a follow-up study of the ACS [2], associations of year round and warm-season averages of 8-h daily maximum O 3 were significantly and positively associated with all-cause, circulatory, cardiovascular, and respiratory mortality. Unlike the earlier study, however, these associations, including that for CVD mortality, were not confounded by PM 2.5 . Findings from the CanCHEC study (Canadian Census Health and Environment Cohort) [3], on the other hand, showed significant and positive associations between warm-season 8-h daily maximum O 3 and non-accidental, cardiometabolic disease, and cardiovascular mortality, but not respiratory, CBV (cerebrovascular disease), and lung cancer. Significant associations remained significant but were attenuated after adjustment for PM 2.5 and NO 2 . In contrast, warm-season averages of 24-h averaged O 3 exposure were not associated with all-cause, cardiovascular, CBV, non-malignant respiratory, and lung cancer mortality in the California Teacher's Study [4]. While these studies suggest a relationship between average peak O 3 exposures and mortality, their associations for specific causes of mortality are inconsistent, perhaps due to their use of different measures of long-term O 3 exposure, which may not only affect O 3 's impacts on health but also confounding of O 3 -mortality associations by PM 2.5 and other pollutants.
Yet, few studies have examined how different O 3 exposure measures impact associations with specific causes of death and how these impacts may affect control for potential confounders, such as PM 2.5 and other air pollutants. Our ability to understand these impacts may be furthered by methods that assess unmeasured confounding in air-pollution-health effect models, such as that developed by Greven et al. [5]. We used this method to characterize unmeasured confounding in our recent studies of PM 2.5 [6] and NO 2 -associated [7] mortality risks. We showed significant and positive associations between both PM 2.5 and NO 2 associations and several causes of death, even after adjustment for other air pollutants and neighborhood-level SES (socio-economic status) and behavioral factors; however, unmeasured confounding of these risks remained. No such method has yet been applied to O 3 and cause-specific mortality.
To address this issue, we examined the association between long-term O 3 exposures and all-cause and cause-specific mortality in a cohort of >22 million Medicare beneficiaries living across the conterminous US (United States) and assessed the consistency of these associations across different measures of exposure, control for confounders, and as modified by region of residence and urbanicity.

Materials and methods
This study was approved by the Institutional Review Board of Tufts and Northeastern Universities.

Air pollution data
We obtained hourly O 3 concentrations from 1151 air quality monitors from the US EPA (Environmental Protection Agency) AQS (Air Quality System) for 2000 through 2008. These monitors, 80% of which were located in urban areas, were selected based on their having hourly measurements for 4 + calendar years, with each year having data for at least five warm-season months (April-September), with each month having at least 75% valid daily measurements. Daily measurements were determined to be valid if 75% of its hourly measurements were available. For each monitor and day, we calculated the daily maximum hourly O 3 concentration as our main exposure measure and averaged daily values from April through September to obtain warmseason average, 1-h daily maximum O 3 concentrations (hereafter referred to as the warm-season 1-h maximum O 3 ) for each monitor and calendar year, as used in Jerrett et al. [1]. In addition, we calculated two secondary exposure measures, including the (1) warm-season average of daily 8h maximum and (2) warm-season average of 24-h average.
To examine potential confounding, we also estimated (1) 1-year moving average PM 2.5 concentrations using well validated GIS-based (Geographical Information System) spatio-temporal models that estimated daily PM 2.5 exposures on a 6 km grid covering the conterminous US [8], (2) 1-year moving average NO 2 concentrations using land use regression models developed by Bechle et al. [9] that estimated monthly NO 2 exposure for census blocks, and (3) 3day moving average temperature using spatio-temporal GAMs that estimated daily temperature on a 6 km grid [8]. PM 2.5 , NO 2 , and temperature values were linked to corresponding O 3 exposures using the grid point or in the case of NO 2 , the census block closest to each of our study's O 3 monitors.

Medicare beneficiaries
We compiled enrollment data from the Centers for Medicare and Medicaid Services for 52.9 million Medicare beneficiaries aged 65 to 120 between 2000-2008. We identified those beneficiaries that resided in ZIP (Zone Improvement Plan) codes with a geographic centroid within a 10 km of an eligible O 3 monitor between 2000-2008. For each enrollee, we obtained information on date of birth, gender, race/ethnicity, ZIP code of residence, and survival. Using the ICD-9 (International Classification of Disease 9th version) codes from the National Death Index, we extracted mortality from non-accidental and accidental causes, as well as from all CVD (cardiovascular) and all respiratory-related causes, and specific subcategories, including IHD (ischemic heart disease), CBV (cerebrovascular) disease, CHF, COPD and pneumonia. Altogether, CVD and respiratory-related causes accounted for over 50% of all-cause mortality ( Table 1). We computed the number of Medicare beneficiaries and cause-specific deaths for each 5-year age interval-gender-race (white/non-white) group, monitor, and study month. To avoid excessive zero counts, we collapsed ages 90 and above into one interval.

Statistical analyses
We linked demographic and mortality data for our Medicare beneficiary cohort to O 3 concentrations measured at the monitor closest to the centroid of the ZIP code of residence.
As our main analysis, we used log-linear regression models to examine the association between warm-season averages of 1-h maximum O 3 exposure and cause-specific mortality, stratified on age, gender, and race (White/non-White) and adjusted for state of residence in both single and two-pollutant models controlling for 1-year moving average PM 2.5 (Supplementary Appendix 1, SI (Supplementary Information)). All results are expressed as the RR (risk ratio) of dying in a given month per 10 ppb (part per billion) increase in O 3 exposure and associated 95% CI (confidence interval).
We evaluated the sensitivity of our findings to the exposure measure, replacing warm-season 1-h maximum O 3 concentrations with (1) warm-season average of daily 8-h maximum or (2) warm-season average of 24-h average O 3 concentrations in our health models in single and twopollutant models controlling for PM 2.5 . We also examined whether associations differed by region of residence (Northeast, Midwest, South, West) as classified using US Census Bureau classifications (Fig. S1, SI), and by urbanicity (urban vs. non-urban) using USGS (United States Geological Survey) Land Cover Trends classifications by stratifying our data into corresponding groups.
To further assess potential confounding of an association between O 3 and mortality, we fit models that additionally adjusted for NO 2 and 3-day moving average temperature. We also fit multi-variable models for beneficiaries living near the 836 of the 1151 monitors for which we had demographic covariates from the 2000 U.S. Census [10]. We assessed the association of O 3 and mortality adjusting for PM 2.5 with and without controlling for the proportion of the population ≥25 years old with at least a high school degree, median income, and proportion of the population living in urban area within the census tract corresponding to the ZIP code centroid. The spatial distribution of the monitors with Census data was similar to that for all monitors, although with fewer monitors in the Midwest (18% vs. 24% overall) and more monitors in the West (31% vs. 26% overall) (Table S1, SI).
We also examined potential confounding by behavioral covariates measured in Selected Metropolitan/Micropolitan Area Risk Trends of the BRFSS (Behavioral Risk Factor Surveillance System) [11] for the subset of beneficiaries living near monitors (586 of 1151 monitors) located in a county with BRFSS data. For these beneficiaries, we assessed the association of O 3 and mortality, adjusting for PM 2.5 , and with and without adjusting for potential confounding by monthly county-level prevalence of current smokers, diabetics, heavy drinkers (i.e., >two drinks per day), asthma, average median income, and body mass index. Note that the spatial distribution of the monitors with Warm-season average of daily 1-h maximum ozone concentrations BRFSS data was similar to that for all monitors, although with fewer monitors in the Midwest (16% vs. 24% overall) and more monitors in the West (31% vs. 26% overall) (Table S1, SI). The spatial distribution of the monitors with Census and BRFSS data was similar by region.
To examine the extent to which our findings remain affected by confounding, we followed the method suggested by Greven et al. [5] and decomposed O 3 into two orthogonal, component measures, "temporal" and "spatiotemporal" O 3 , and estimated their associations with causespecific mortality in base, PM 2.5 -adjusted, and BRFSSadjusted models (Supplementary Appendix 1, SI). "Temporal" O 3 , which represents the national temporal trends in monthly O 3 concentrations, is O 3 centered by the average concentrations for all monitors and across the study period; "spatio-temporal" O 3 is a measure of the monitor-specific temporal trends in monthly O 3 concentrations compared to the national "temporal" trends. In the absence of residual confounding by long-term time trends of mortality, estimates of the "temporal" and "spatio-temporal" O 3 should be similar.

Results
Our study population included 22.2 million Medicare enrollees residing in 260 metropolitan areas across the US (Table 1), comprising 42% of all Medicare enrollees. During the study period, 5.8 million deaths were reported; 97.8% from non-accidental causes. CVD accounted for 40.5% of all mortality, followed by cancer (22.2%) and respiratory mortality (11%). More than 50% of all CVD mortality was caused by IHD. Fifty-two percent of respiratory deaths were attributed to COPD and 27.6% to pneumonia, while 27.4% of cancer deaths were attributed to lung cancer. The warm-season average of 1-h maximum O 3 concentration across all monitors and years was 56 ppb (9 ppb) and exhibited no consistent temporal trend over the study period (Fig. S2, SI). Warm average concentrations varied by region, with concentrations highest in the West (Fig. S2, SI). Warm-season averages of 1-h maximum O 3 were strongly correlated with corresponding averages of 8-h maximum (r = 0.98) and 24-h average (r = 0.74) levels. The correlation coefficient between the warm-season average of 1-h maximum O 3 and 1-year average PM 2.5 , 1-year average NO 2 , and 3-day average temperature equaled 0.24, 0.26, and 0.04, respectively.
In single pollutant models we found a 10 ppb increase in warm-season 1-h maximum O 3 to be significantly associated with an increased risk of dying from all ( In models adjusting for PM 2.5 , O 3 -associated mortality risks remained significant and positive for all causes of death, except for CBV and pneumonia (Table 2). For several causes of death, PM 2.5 -adjusted associations were attenuated, with RRs decreasing slightly to 1.030 (95% CI: 1.027, 1.034) for respiratory mortality and more substantially to 1.005 (95% CI: 1.003, 1.007) for CVD mortality per 10 ppb increase in long-term O 3 . In contrast, for COPD and CHF mortality, RRs increased after adjustment for PM 2.5 , resulting in a 1.072 (95% CI: 1.067, 1.077) and 1.063 (95% CI: 1.055, 1.071) times increase in risk per 10 ppb increase in O 3 , respectively. In models adjusting for NO 2 , O 3 -associated mortality risks remained unchanged for respiratory and lung cancer, increased slightly for CHF and COPD, and were slightly attenuated for other causes of death (Table S2, SI). Associations were generally robust to adjustment for temperature (Table S3, SI). As shown on Table S4 (SI), RRs of the PM 2.5 -adjusted association between warm-season 1-h maximum O 3 and mortality were essentially unchanged after adjustment for potential confounding by ecologic covariates from the 2000 U.S. Census or county-level behavioral covariates from BRFSS.
When we examined the potential for unmeasured confounding, we found statistically significant, positive RRs for temporal O 3 and negative RRs for spatio-temporal O 3 in base models (Table 3). In two-pollutant models adjusting for PM 2.5 (Table 3) and in multi-variable models adjusted for both PM 2.5 and behavioral covariates (Table S5, SI), the magnitude of "temporal" RRs decreased for all causes, especially in the multi-variate model, but the "spatio-temporal" RRs remained unchanged.
Risks of death were higher for beneficiaries living in non-urban as compared to urban areas for all-cause and CVD mortality, but were similar for respiratory and COPD mortality (Table S6, SI). For all causes of death, O 3 -associated mortality risks were highest in the Northeast and West as compared to the South and Midwest (Table S7, SI). Region-specific RRs were generally lower in PM 2.5 -adjusted as compared to non-adjusted models for all regions except the Northeast, for which RRs were generally comparable.
We examined the robustness of our results to the O 3 exposure measure (Table 4). In PM 2.5 -adjusted models, RRs based on the warm-season average of 8-h maximum O 3 were similar for mortality from all causes, CHF, all respiratory, COPD, and lung cancer, but were attenuated and no longer significant for CVD and IHD mortality [note that associations between warm-season 8-h maximum O 3 and both CVD and IHD mortality were significant and positive in single pollutant models, suggesting confounding of 8-h maximum-but not 1-h maximum associations-by PM 2.5 (Table S8,

Discussion
In our cohort of over 22 million Medicare beneficiaries, we found consistent associations between long-term O 3 exposures and increased mortality. Specifically, we found a 0.4%, 3.0%, 0.5%, and 1.5% increased risk in all-cause, respiratory, cardiovascular and for the first-time, lung cancer mortality, respectively, per 10 ppb increase in warmseason average, 1-h maximum O 3 exposure in PM 2.5adjusted models.
Notably, we found increased mortality risks from longterm O 3 exposures to be the strongest and most consistent for respiratory-related diseases (including all respiratory and COPD), CHF, and lung cancer mortality, with no evidence of confounding by PM 2.5 , NO 2 , or temperature or of sensitivity to the exposure measure. O 3 -associated RRs for respiratory-related mortality, for example, were not attenuated in models controlling for PM 2.5 or NO 2 and further were comparable for beneficiaries living in urban and nonurban areas, for which the composition of PM 2.5 differs substantially [12,13]. In contrast, associations between long-term O 3 exposures and CVD-related mortality were less consistent, confounded by PM 2.5 as evidenced by their attenuation in two-pollutant models and their stronger effects in non-urban as compared to urban environments. Further, we found O 3 -associated risks to vary with the exposure measure, with associations no longer significantly positive when warm-season 8-h maximum or 24-h average O 3 was used to assess exposures, suggesting that long-term peak rather than average exposures reduce life expectancy. Altogether, these findings provide strong evidence that long-term peak O 3 exposure is associated with mortality from respiratory-related causes and for the first-time lung cancer, but raise questions regarding O 3 -related impacts on CVD mortality. Our findings were robust to further adjustment by ecological and behavioral covariates for all causes of death. They, however, likely remain affected by some unmeasured confounding, as suggested by the unequal coefficients for temporal and spatio-temporal O 3 . Our finding of unmeasured confounding, even after adjustment for PM 2.5 and behavioral covariates, is not surprising given our reliance on administrative mortality data, which precludes us from controlling for covariates describing individual-specific SES, behaviors, and health histories, as was done in the CanCHEC study [3]. Despite this, findings from our study are consistent with previous studies of ozone and mortality, many of which did adjust for numerous individual-specific covariates. Of these studies, only Jerrett et al. [1] assessed O 3 exposures for 448,850 ACS CPS-II (Cancer Prevention Study II) study participants as warm-season averages of daily 1-h maximum ozone concentrations, finding significant, positive associations for cardiovascular (1.011, 95% CI: 1.003, 1.023), IHD (1.015, 95% CI: 1.003, 1.026), respiratory (1.029, 95% CI: 1.010, 1.048) and cardiopulmonary (1.014, 95% CI: 1.007, 1.022) mortality as in our study, but not for all-cause mortality (1.001, 95% CI: 1.996, 1.007), in contrast to our study. Further, associations in the CPS-II cohort remained significant after adjustment for PM 2.5 for mortality from respiratory but not from cardiovascular-related causes, which like our study suggests confounding by PM 2.5 of O 3 -mortality associations for cardiovascular-related causes [1].
In a follow-up to the CPS-II study [2], the impact of long-term O 3 exposures on cause-specific mortality was reexamined using new estimates of ambient O 3 , PM 2.5 , and NO 2 concentrations, a larger number of participants, and an extended follow-up period. The new O 3 exposure estimates were based on individual-specific measures of yearly and summer-only 8-h maximum O 3 averaged between 2002-2004 concentrations, thus capturing spatial but not temporal variability in O 3 exposures. As in the original CPS-II and in our study, the follow-up study found summer 8-h daily maximum O 3 exposures to be significantly and positively associated with all-cause (1.02, 95% CI: 1.02, 1.03), respiratory (1.10, 95% CI: 1.07, 1.12), and COPD (1.08, 95% CI: 1.05, 1.12) mortality and consistent with the original CPS-II but not our study, reported null associations For mortality from all causes, CVD, IHD, CBV, pneumonia, and all cancers, our non-positive findings for warmseason averages of 24-h O 3 in PM 2.5 -adjusted models are consistent with those from Lipsett et al. [4] in the California Teachers study, which reported null associations between summer averaged 24-h O 3 and all-cause, cardiovascular, non-malignant respiratory, and CBV mortality, and significant associations with increased mortality for IHD in single, but not 2-pollutant models that included PM 2.5 . While Lipsett et al. found null associations between 24-h O 3 and lung cancer, we found these associations to be significant and positive. The insignificant associations for lung cancer and other causes of death reported by Lipsett et al. [4] have been attributed to its relatively small sample size of approximately 125,000 women. Our results suggest that its use of warm-season averages of 24-h O 3 as the exposure measure may also contribute to the observed null findings.
More puzzling are results from Crouse et al. [3], who used data from the CanCHEC to examine the association between warm-season averages of 8-h maximum O 3 and a number of causes of death. The authors reported significant increased O 3 -associated risks of all-cause mortality (1.031, 95% CI: 1.026, 1.036), and null association for CBV (0.981, 95% CI: 0.961, 1.001) as in our study, but null associations with lung cancer (1.006, 95% CI: 0.990, 1.023) mortality and significant positive associations with mortality from CVD (1.037, 95% CI: 1.028, 1.047), and diabetes (1.156, 95% CI: 1.121, 1.190), even after adjustment for PM 2.5 and NO 2 . The O 3 -associated risks of increased CVD mortality-together with observed null associations between PM 2.5 and CVD mortality-caused the authors to hypothesize that the multi-pollutant results were affected by collinearity among the multiple pollutants and/or overlapping spatial patterns in O 3 and PM 2.5 . Support for this hypothesis is provided by our findings of effect modification of the O 3 -mortality relationship by region for each cause of death. One possible explanation for this region-specific variability in O 3 -mortality risks may be corresponding regional variability in the correlation between O 3 and PM 2.5 , which may result in differential confounding by PM 2.5 of regional O 3 effects on mortality. Other factors that may contribute to observed heterogeneity in O 3 -mortality associations may also include regional differences in (1) PM 2.5 composition, which may differentially confound O 3 -mortality associations, (2) air conditioning prevalence, which may contribute to differential exposure error, and (3) unmeasured confounders that were not captured by our control for strata for age, race, gender, and our adjustment for state of residence, PM 2.5 , census, and behavioral variables. Note that certain causes of death, such as pneumonia or CBV, may be more sensitive to the regionspecific effects of confounding, due to their small number of deaths.
Importantly, our positive and consistent associations linking long-term peak O 3 exposures to cardio-respiratory mortality are biologically plausible. Numerous studies have demonstrated the negative permanent effects of long-term exposure to O 3 on pulmonary function and increased lung inflammation [14][15][16][17][18][19][20][21][22], which may lead to tissue injury [14], airway remodeling [15,16], and pulmonary fibrotic changes [17]. Further, exposed animals were shown to exhibit similar or greater morphological and other changes after episodic as compared to continuous O 3 exposures [18], lending support for our findings showing that warm-season averages of 1-h daily maximum O 3 were more relevant to mortality as compared to 8-h daily maximum and 24-h average O 3 exposures.
Our study has several limitations. First, we did not have information on beneficiaries' activity and mobility patterns, which may contribute to exposure misclassification. Further, we used ambient, nearest monitor O 3 concentrations to assess exposures, which are imperfect proxies of personal O 3 exposures. This exposure error has been shown to bias RR estimates towards the null [23], possibly suggesting that the magnitudes of our associations are underestimated. We, however, minimized other sources of exposure error by (1) linking O 3 exposure to beneficiary information by year in order to account for residential moves and ZIP code boundary changes and (2) using exposure values that were geographically close to residences, capturing exposures that result from nearby O 3 emission sources, and thus reducing the potential for exposure misclassification. Second, although we did not have data on personal-level characteristics, we adjusted for county-level behavioral variables, including those related to smoking and comorbidities, with results essentially unchanged, consistent with little, if any, confounding of O 3 -mortality associations by the examined covariates. While our findings showed that adjusting for PM 2.5 and behavioral covariates reduces unmeasured confounding, note that differences in the coefficients for temporal and spatio-temporal O 3 remained, suggesting unmeasured confounding of the O 3 -mortality associations. These findings demonstrate the need for additional studies to examine the impact of unmeasured confounding on estimated risks posed by long-term O 3 on mortality. Lastly, our findings may not be generalizable to younger age groups or to beneficiaries living away from monitors.
These limitations are balanced by the substantial strengths of our study and long time period of study. With more than 22 million Medicare beneficiaries with near 5.8 million deaths over the 9 years of study period, our study is well powered to detect meaningful associations for mortality from specific causes, allowing us to provide valuable, new information on the relationship between O 3 exposures and specific CVD, respiratory and cancer-related deaths in addition to all-cause mortality.

Code availability
All analyses were performed using SAS 9.4 Software (SAS Institute Inc., Cary, North Carolina). SAS code used to generate the results is provided as supplementary information (Supplementary Appendix 2, SI).