We included 265,223 deaths that occurred among adults 45 years of age or older during 2007–2008 in the contiguous United States. Each death record had a U.S. county identifier in the restricted mortality data file, which allowed us to summarize death counts by county and to link them with other county-level data sources. CLRD deaths included those with underlying cause coded as J40–J47 (ICD-10 codes). We derived county-level average daily ozone and PM_2.5 by aggregating 2001–2008 census-tract-level 8-hour maximum ozone and 24-hour average PM_2.5 concentration, generated by the U.S. Environmental Protection Agency for the Tracking Network based on monitored data and output from the Community Multi-scale Air Quality modeling system (16). County-level lifetime smoking prevalence (percentage of adults who were current or former smokers) and obesity prevalence (percentage of obese adults with body mass index [the ratio of height to weight] ≥30) were derived from the Behavioral Risk Factor Surveillance System (2007–2008) (17), using the method suggested by Zhang and colleagues (18). The county-level percentage of adults at least 65 years of age and poverty levels (percentage of adults below the federal poverty line) were obtained from 2007–2008 census data (19). Extremely hot days were defined as the average annual number of days with maximum temperature equal to or greater than 90 degrees Fahrenheit (°F). County-level daily maximum temperatures during 2001–2008 were obtained from the Tracking Network, originally from the North American Land Data Assimilation System (14).

We fit Bayesian hierarchical spatial Poisson models using CLRD death counts as the outcome and county-level variables as predictors. Five models were explored with different random effect specifications: model 1, state unstructured random effects only; model 2, state unstructured and county unstructured random effects; model 3, state unstructured and county spatially structured random effects; model 4, state unstructured, county unstructured, and county spatially structured random effects; model 5, model 4 with a mixture parameter embedded between county unstructured and county spatially structured random effects. Epidemiologically, state unstructured random effects specify state-level contextual effects on mortality; county unstructured random effects specify county-level heterogeneous contextual effects whereas county spatially structured random effects capture possible spatial dependence (i.e., spatial autocorrelation between adjacent counties). The mixture parameter allows the balance of county-level heterogeneity and spatial dependence. Our log-link Poisson regression model is log[y_i] = log[E_i] + α + X_iβ + ST_j[i] + ρ_iU_i + (1 – ρ_i)S_i, where y_i is the number of deaths for county i (i = 1, …, 3,109), E_i is the population (≥45 yr), α is the intercept, X_i is the vector of seven predictors (X_1,i, …, X_7,i), β_{[1, …, 7]} is the corresponding regression coefficient, ST_j[i] (j = 1, …, 49) is state unstructured random effects, U_i is county unstructured random effects, S_i is county spatially structured random effects, and ρ is the mixture parameter (0 ≤ ρ ≤1). County spatially structured random effects are formulated as Si|Sj~Normal(S¯i,τi2)(i≠j) (20), where S¯i=(1/∑jwij)∑jSjwij,τi2=τS2/∑jwij, w_ij = 1, if i, j are adjacent counties, otherwise w_ij = 0. The state unstructured and county unstructured random effects are formulated as STj[i]~Normal(0,τST2) and Ui~Normal(0,τu2). τST2,τs2, and τu2 are the variance parameters of ST_j[i], S_i, and U_i. In full Bayesian analyses, prior distribution must be specified for these three variance parameters. We assigned diffusive/noninformative gamma distributions for these three parameters, as suggested by Bernardinelli and colleagues (21). We implemented these five models in WinBUGS1.4.3 and used the deviance information criterion (DIC) to compare model fit (15, 22).

Table 1 shows the mean, range, and quartiles of ozone, PM_2.5, and five selected demographic, socioeconomic, behavioral, and meteorological characteristics. Ozone exposure ranged from 27.8 to 52.0 ppb (median, 41.2 ppb), PM_2.5 exposure ranged from 4.8 to 16.8 µg/m³ (median, 10.9 µg/m³), percentage of adults 65 years of age or older ranged from 3.1 to 39.6% (median, 15.1%), percentage of adults below the federal poverty line ranged from 2.7 to 49.5% (median, 12.4%), lifetime smoking prevalence ranged from 24.6 to 68.9% (median, 51.5%), obesity prevalence ranged from 16.6 to 50.2% (median, 30.0%), and extremely hot days ranged from 0 to 197 (median, 46).

Table 2 shows that model 3 produced the lowest DIC. The difference between the DIC for this model (21,474.7) and the DICs for models 4 and 5 (21,475.1 and 21,479.1, respectively) is admittedly small (<5), indicating that any of them could be the best model for describing the data (22). Still, model 3, with state unstructured and county spatially structured random effects, is preferred because it contains fewer parameters (23). In contrast, the difference is substantial between the DICs for models 3, 4, and 5 (21,474.7, 21,475.1, and 21,479.1, respectively) and the DICs for models 1 and 2 (21,606.4 and 21,607.4, respectively) (>5). It is evident that county spatially structured random effects dominate spatial dependence between neighboring counties, reflecting the effects of unobserved, spatially structured covariates.

Bayesian inference is based on posterior means (analogous to means) and credible intervals (CIs, analogous to confidence intervals). Table 3 presents adjusted rate ratios (RRs) and 95% CIs from model 3 (the preferred model with state unstructured and county spatially structured random effects), measured per five-unit increment for all variables. All predictors were positively associated with CLRD deaths. Specifically, the RR was 1.05 (95% CI, 1.01–1.09) per 5-ppb increase in ozone exposure.

Ozone and PM_2.5 were associated with a 5% (per 5-ppb increase in average ozone) and a 7% (per 5-µg/m³ increase in average PM_2.5) increase in CLRD mortality, respectively, although the association between PM_2.5 and CLRD mortality was not statistically significant. Together, ozone and PM_2.5 explained about 3% of the total variation in log RRs (Table 4). Table 4 also shows that all predictors combined explained about 35% of the total variation with lifetime smoking, age (adults aged ≥65 yr), and poverty explaining most, whereas other unobserved covariates at state (5%) and county (60%) levels explained about 65%.

Our principal finding is that after controlling for selected demographic, socioeconomic, behavioral, and environmental risk factors, and other spatially unstructured and structured contextual influences, ozone is associated with increased CLRD mortality rates across U.S. counties. A few cohort studies have observed similar results for ozone, but they were limited in terms of demographic or geographical coverage (8, 11, 24). This nationwide study explored the linkage between county-level concentration levels and aggregated deaths across the contiguous United States. Our analyses differ fundamentally from traditional ecologic analyses in that we used Bayesian hierarchical spatial modeling. Bayesian modeling allows for direct control of known (i.e., percent adults aged ≥65 yr, lifetime smoking, poverty, obesity, and temperature) and unknown (unstructured and spatially structured) risk factors. These analyses showed that CLRD mortality was significantly associated with ozone exposure. Such correlation might reflect an amalgam of complex pathophysiological pathways through which ozone could induce or accelerate pulmonary inflammation leading to CLRD mortality (25, 26).

Our results for the long-term effects of ozone on CLRD mortality are generally consistent with the findings from the ACS cohort (448,850 participants in 86 U.S. metropolitan areas during 1977–2000) (8) and Medicare subpopulations (3,210,511 persons with COPD in 105 U.S. cities during 1985–2006) (11). Our adjusted RR estimate of 1.05 per 5-ppb increase in ozone—which is equivalent to 1.10 per 10-ppb increment— is lower than the estimate from Medicare participants hospitalized with COPD (RR, 1.07 [95% CI, 1.05–1.10] per 5-ppb increment), but it is higher than that from the ACS cohort (RR, 1.04 [95% CI, 1.01–1.07] per 10-ppb increment). The study of a California ACS cohort reported a positive albeit insignificant association (RR, 1.02 [95% CI, 0.90–1.15] per 10-ppb increment), which might be due to the small number of participants (n = 73,711). The difference in RR estimates could be due in part to the difference in participants or study areas included in these studies. Participants hospitalized with COPD in the Medicare study might have had higher risk of CLRD death due to ozone exposure than did people without preexisting COPD as well as the general U.S. population. Participants in the ACS study were mainly white with relatively high educational attainment (27). Thus, subjects in the ACS study might have been healthier, and have had lower risk of CLRD death due to ozone exposure, than the U.S. population generally.

We found a positive but statistically insignificant association between long-term PM_2.5 exposure and CLRD mortality, with an adjusted RR estimate of 1.07 (95% CI, 0.99–1.14) per 5-µg/m³ increase in PM_2.5 exposure, which is equivalent to 1.14 per 10-µg/m³ increment. The Harvard Six Cities cohort (8,096 white participants) and California ACS cohort (73,711 participants) resulted in similar findings, with adjusted RRs of 1.05 (95% CI, 0.95–1.15) for the California cohort and 1.08 (95% CI, 0.79–1.49) for the Harvard Six Cities cohort for an increase of 10 µg/m³ in PM_2.5 exposure (7, 9). A study of a large ACS cohort (448,850 participants) using a single-pollutant model reported a similar association (RR, 1.03 [95% CI, 0.96–1.11]) but reported an inverse and insignificant association in a two-pollutant model when ozone was included (RR, 0.93 [95% CI, 0.84–1.03]) (8). Further studies are needed to confirm the association between PM_2.5 and CLRD mortality at both the individual and aggregated population levels.

The increased risk of death due to CLRD associated with ozone was small compared with the risk posed by lifetime smoking, older age (≥65 yr), and temperature. Nevertheless, the positive association between air pollution and CLRD mortality persisted after controlling for known and unknown factors at county and state levels. The adjusted RR for ozone was close to that for poverty, a county-level socioeconomic indicator—although lifetime smoking and older age were two known leading contributors to the CLRD mortality variation. It is worth noting that county-level unknown, spatially correlated influences contributed more than half of the CLRD death variations across U.S. counties; these influences were not considered in most previous studies. Also, the inclusion of spatially correlated random components could potentially increase the precision of the RR estimate for the known risk factors in which we were interested.

This study has several limitations. First, our single ecological study could not make any causal inference. Although we adjusted for available county-level covariates and unobserved influences, the possibility of ecologic bias remains. Furthermore, including 8 years of exposure data could be one of the strengths; however, it could also be a limitation because people could move during this time period and disease latency could be longer than the years we included in this study. Similarly, we could not account for any seasonal variation in or trend of exposure during this time period. National annual average ozone and PM_2.5 both showed downward trends from 1980 (for ozone) and 2000 (for PM_2.5), but such trends are not smooth and do show year-to-year influences of weather conditions, which contribute to ozone and PM_2.5 formation in the air (28, 29). We used 8-year (2001–2008) average ozone and PM_2.5 as proxies for their long-term exposure estimates; however, using exposure averaged during the early years (2001–2002) did not meaningfully change county-level RR estimates (see Table E4 in the online supplement). In addition, we did not address the seasonality of CLRD deaths, ozone, and PM_2.5 because of the small sample size of CLRD deaths by season at the county level. Ozone shows a clear seasonal pattern, and linking the seasonal timing of death might strengthen the association found, if the number of deaths by season at the county level were sufficient to allow us to do so. Finally, we could not evaluate the sensitivity of ICD-10 codes (J40–J47) for the diagnosis of CLRD. Potential misclassification of CLRD as an underlying cause of death might introduce additional uncertainties in our findings. Findings of this national study suggest that ozone and PM_2.5 might have contributed to increased CLRD mortality across U.S. counties, although residual confounding cannot be excluded. The association observed between long-term ozone exposure and CLRD mortality across U.S. counties is in line with findings from previous cohort studies, but this study expands the evidence to the U.S. population. The positive association between long-term PM_2.5 exposure and CLRD mortality is consistent with findings from previous studies. This U.S. national study provides additional evidence that ambient air pollutants, particularly ozone, could be important contributing factors in CLRD mortality.