Objectives: We investigated whether ozone was associated with survival in four cohorts of persons with specific diseases in 105 United States cities, treating ozone as a time varying exposure.

Methods: We used Medicare data (1985–2006), and constructed cohorts of persons hospitalized with chronic conditions that might predispose to ozone effects: chronic obstructive pulmonary disease, diabetes, congestive heart failure, and myocardial infarction. Yearly warm-season average ozone was merged to the individual follow-up in each city. We applied Cox proportional hazard model for each cohort within each city, adjusting for individual risk factors, temperature, and city-specific long-term trends.

Measurements and Main Results: We found significant associations with a hazard ratio for mortality of 1.06 (95% confidence interval [CI], 1.03–1.08) per 5-ppb increase in summer average ozone for persons with congestive heart failure; of 1.09 (95% CI, 1.06–1.12) with myocardial infarction; of 1.07 (95% CI, 1.04–1.09) with chronic obstructive pulmonary disease; and of 1.07 (95% CI, 1.05–1.10) for diabetics. We also found that the effect varied by region, but that this was mostly explained by mean temperature, which is likely a surrogate of air conditioning use, and hence exposure.

A recent study by Anenberg and coauthors (1) estimated the global burden of mortality caused by ozone and particles with an aerodynamic diameter of 2.5 μm or less (PM_2.5) from anthropogenic emissions, and found that ozone was associated with about 0.7 ± 0.3 million respiratory deaths and 6.3 ± 3 million “years of life lost” annually.

Epidemiologic studies have reported associations between short-term changes in ozone and short-term changes in deaths, especially during the warm weather months (2–5). Three recent meta-analyses reviewed the state of the ozone-mortality time-series literature (6–8). Other studies have shown higher mortality risk in more vulnerable populations, such as the elderly, persons with atrial fibrillation, and blacks (9, 10), and demonstrated the acute effects of ozone are not caused by short-term mortality displacement (11). Another study examined the question of adaptation in the acute mortality risk associated with ambient ozone (12). In contrast, only one study has examined the effects of chronic exposure to ozone in a cohort (13).

The previously mentioned epidemiologic literature is supported by studies of the biologic mechanism of ozone toxicity. There is evidence that short-term ozone exposure is associated with decrements in lung functions, increased respiratory symptoms, and lung inflammation (14–25). The effect of short- and long-term ozone exposure was associated with biomarkers of oxidative stress (26), and with reduced heart rate variability (27). Finally, a review of toxicologic studies found decreased heart rate, metabolism, blood pressure, and cardiac output when rats are exposed to typical concentrations of ozone (28).

All observational epidemiology studies are subject to potential confounding, but different study designs are usually subject to different types of confounding, or confounding by different sets of covariates. Studies of air pollution and mortality fall into two broad categories. Time-series analyses examine the association between day-to-day changes in daily death counts and day-to-day changes in air pollution concentrations. These studies are usually done separately in each geographic area studied, eliminating confounding by factors that vary across geography.

Cohort analyses of long-term exposure to air pollution and survival are very different. In these, long-term average air pollution concentrations in a particular geographic place are assigned to subjects in that place, and their mortality experience across place is compared with the air pollution variation across place. Here season and other short-term fluctuations are not a source of confounding, because the exposure contrast is long-term exposure across place. Potential confounders, therefore, are those that likewise vary across place. These include such factors as diet and socioeconomic position. These are precisely the factors that are not confounders in the time-series studies.

Both types of studies have reported associations of particulate air pollution with mortality risk. However, only one cohort study has examined the association between mortality and warm season average ozone concentrations (13), and the dearth of confirmatory reports prompted the US Environmental Protection Agency's Science Advisory Board to recommend not using that study to estimate the health benefits of reducing ozone concentrations.

Recently, we introduced a variant of the cohort study that seeks to capture some of the advantages of both time series and cross-sectional cohort studies, and that is subject to a different and, we argued, smaller set of potential confounders. The approach focuses on year-to-year differences in exposure from their long-term trend within city. In this approach a time varying Cox proportionate hazard model is used with annual follow-ups, and year-to-year variations in exposure are used as risk factors for the year-to-year survival of the cohort, controlling for long-term trend within city. This analysis can be repeated in multiple cities and the results combined across cities in a second stage. Using that approach, we have demonstrated that particle concentrations were risk factors for reduced survival in persons discharged alive after hospitalization for either myocardial infarctions (MIs) (29) or chronic obstructive pulmonary disease (COPD) (30). We here apply that approach to the question of whether longer-term exposure to ozone is associated with reduced survival in cohorts of putatively susceptible populations, including persons with heart disease, chronic lung disease, or diabetes, using data from 105 cities.

Table 1 shows characteristics of the study population for all the cities together for each of the four cohorts. Overall, our cohorts consisted of 3,210,511 persons with COPD, 1,561,819 with CHF, 2,935,647 with diabetes, and 1,186,496 with MI. Of these, 45% of the COPD population, 55.4% of the CHF population, 38.6% of the diabetes population, and 38.1% of the MI population died before the end of follow-up.

The average duration of the follow-up was 5.1 years for CHF, 5.6 years for COPD, 5.6 years for diabetes, and 6.1 years for MI. The range of survival times in all the cohorts varied from 1–21 years.

The four cohorts differed little in terms of population characteristics with a higher percentage of female, and more whites. The CHF cohort was the older with a mean age of 79 years, whereas the diabetes cohort was younger with a mean age of 76 years.

Table 2 presents the results of the analyses across all cities for ozone averages during summer and transitional seasons. The average interquartile range of ozone over all the cities was 5.5 for the summer season and 4.3 for the transitional season. Therefore, the results are expressed as hazard ratio for 5 ppb of seasonal average ozone. For summertime ozone we found significant associations in the four survival analyses with a hazard ratio of 1.06 (95% confidence interval [CI], 1.03–1.08) for the heart failure cohort; of 1.09 (95% CI, 1.06–1.12) for the MI cohort; of 1.07 (95% CI, 1.05–1.10) for the diabetes cohort; and 1.07 (95% CI, 1.04–1.09) for the COPD cohort. The associations using ozone during the transitional season are lower but still significant.

We then performed a sensitivity analysis by expanding the length of the ozone season by 2 months. That is, we used the average ozone for the months April until October. Person time at risk was therefore April to December. These results (not shown) are similar to the May to September estimates.

In a previous paper, Franklin and coauthors (33) found that the association between PM_2.5 mass and mortality was modified by the seasonal average temperature, showing an inverted U-shaped relationship between the PM_2.5-mortality effect estimates and temperature. In that study temperature was used as a surrogate to explain ventilation of ambient air to the indoor environment (34).

We therefore used a meta-regression between the city-specific effect estimates for ozone and city average temperature (and temperature squared) and we also found an inverted U-shaped relationship. Temperature is a proxy measure for ventilation and therefore also of air conditioning use and may play a role in explaining differences among regions. We then added dummy variables for region to the meta-regression to determine whether there was any remaining regional variation in ozone effect not explained by temperature.

Figure 2 shows the estimated deviation from the overall effect of ozone in each region, as predicted by temperature. The regional effects, then, are the difference of that region from expected, given temperature. In general, there is no significant variation remaining, except for a lower effect of ozone in Region 6 (West Coast), which is significant for the MI cohort and marginal for the others.

A key feature of this study is the nature of the exposure contrast. City-specific regressions eliminated potential confounding by factors that vary across city, which is a common concern in most air pollution cohort studies. Rather, we examined whether within-city year-to-year fluctuations in ozone concentrations about their long-term trend were associated with within-city fluctuations of survival rates about their long-term trend. This study design avoids both confounding by cross-sectional factors that vary by city and the short-term factors that confound daily times series, but are not present in annual analyses. It also costs considerable amounts of power, because much of the exposure variation (short-term within city, and long-term across city) is thrown away. However, the large cohorts available using Medicare data allow such an approach. A key advantage is that fluctuations from summer to summer in ozone concentrations around the long-term level and trend in a specific city is unlikely to be correlated with most other predictors of mortality risk, except for temperature, which is controlled in the regression. For example, smoking rates may have fallen differently in different cities, but we cannot see a mechanism that would induce a correlation whereby year-to-year fluctuations around that trend in smoking would be correlated with year-to-year fluctuations in ozone around its trend. Some factors, such as unemployment rates, may fluctuate with emissions of air pollution. However, our cohorts are retired populations unlikely to be effected by unemployment, and ozone is a secondary pollutant whose concentrations vary with emissions in upwind communities, not the local one. However, a key limitation of our study was the inability to control for PM_2.5, because it was not available in these cities until 1999. This may have resulted in confounding.

The study also differs from most other air pollution cohort studies in that the entire population over 65 was part of the initial sampling frame, rather than a convenience sample. This avoids selection bias, and confounding by possible selection of more or less healthy samples of the population in towns with more or less air pollution.

The finding of an inverted U-shaped relationship between the ozone effect in a city and its mean temperature, control for which renders regional differences insignificant, suggests that those regional differences are mostly related to temperature. The inverted U shape suggests that temperature is mostly capturing exposure differences. Effects are highest in mild temperatures, when windows are most likely to be open. Other possibilities for regional differences also exist, such as differences in medication, coexposures, and so forth. Future studies should address this issue.

This is the first study of longer-term exposure to ozone that focused on cohorts of susceptible subjects–persons with chronic conditions (specifically COPD, heart failure, diabetes, and MI) and as expected the hazard ratios were higher in these high-risk cohorts than in the American Cancer Society cohort. We should recognize, however, that there are differences between the exposure measures in the different studies. Time-series studies examine the more immediate risk on a high-pollution day. In contrast, long-term studies, such as Jerrett and coworkers (13), should capture the integrated effects of these acute responses, the effects of semi-chronic effects (response to a high ozone summer), and the additional responses resulting from long-term exposure to ozone over years. Our study focuses on the semi-chronic effect. Hence, comparisons of effect sizes across such studies should recognize that they are examining different things (and in the case of our study, susceptible populations).

In this context, the risks of semi-chronic exposures in this cohort analysis were higher than in studies analyzing the acute effects of ozone on the risk of mortality. For example, in a previous study (11) we reported a 0.32% increase in total mortality for 10 ppb of 8-hour average ozone, and the meta-analyses of Levy and coworkers (7) and Ito and coworkers (8) reported effect size estimates of 0.41% and 0.39% for a 10-ppb increase in maximum hourly ozone. This study report around a 7% (depending on the cohort) increase in mortality per 5 ppb (which is approximately the interquartile range) increase in summer average of 8-hour mean ozone.

Only one cohort study has investigated the effect of warm season average ozone on the risk of death in a large cohort (13). The authors used data from the American Cancer Society Cancer Prevention Study II cohort and found a 1.1% increase in cardiovascular causes, a 1.5% increase in ischemic heart disease, and a 2.9% increase in respiratory causes, which increase when adjusting for PM_2.5. In that study, the authors investigated a younger population that was more educated and socially advantaged than the general population, which may account for some of the difference in effect size compared with our analysis of older, sicker subjects. They also found the effect primarily in respiratory deaths. Susceptibility alone cannot explain the difference. Given that respiratory deaths are a small fraction of all deaths, our results also suggest an effect of ozone on nonrespiratory deaths.

The biologic mechanism by which ozone can affect mortality is still under examination. A review of toxicologic studies found decreased heart rate, metabolism, blood pressure, and cardiac output when rats are exposed to typical concentrations of ozone (28).

A study (22) found a possible ongoing inflammatory response in the lungs of recreational joggers exposed to ozone and associated copollutants during the summer months. A previous paper (23) showed that airway inflammation persists after repeated ozone exposure, despite attenuation of some inflammatory markers in bronchoalveolar lavage fluid and adaptation of lung function.

Long-term exposure to ozone was associated with reduced lung function in a study of freshmen at the University of California, Berkeley (24). The associations were independent of any effects related to PM and NO₂. In the same cohort (26) the effect of short- and long-term ozone exposures on biomarkers of oxidative stress were examined, and the authors found that estimated 2-week before, 1-month before, and lifetime O₃ exposures of young healthy individuals were significantly and independently associated with increased 8-iso-prostaglandin F levels, while controlling for sex and ethnicity.

Another study (25) examined the association between current respiratory health status and long-term ozone exposure histories in 520 Yale University students. The authors found that lung function was significantly lower in the group with high ozone exposures, and chronic phlegm wheeze and respiratory symptom index were increased in the ozone-exposed group. They conclude that living for 4 or more years in regions of the country with high levels of ozone is associated with diminished lung function and more frequent reports of respiratory symptoms. Ozone exposure has also been associated with reduced heart rate variability (27).

Our study presents some major limitations. One is that Medicare does not provide the underlying cause of death. In addition, our definitions of persons with COPD are restrictive. Doubtless there are other persons in the 105 cities, who were covered by Medicare, who had the conditions we were interested in, and who were not included in the cohort either because they were not hospitalized for the condition or because it was misdiagnosed. Although this limits generalizability and moves our focus to a frailer subset of the persons with the conditions, it does not limit internal validity. The main limitation, however, is the absence of information on subject characteristics, such as smoking, body mass index, or medicine use. However, because of our study design, differences in these variables across cities, or long-term trends in smoking within city, cannot confound our estimates. In our model we controlled for all the available personal characteristics, such as age, race, sex, severity of the index admission, and detailed data on previous and secondary diagnosis. Moreover, we conducted a city-specific analysis to remove location-specific differences from the analyses, and removed long-term time trends. Hence, differences across cities in smoking rates cannot confound the association, and neither can long-term time trends in smoking rates within city.