Chronic Fine and Coarse Particulate Exposure, Mortality, and Coronary Heart Disease in the Nurses’ Health Study

Background The relationship of fine particulate matter < 2.5 μm in diameter (PM2.5) air pollution with mortality and cardiovascular disease is well established, with more recent long-term studies reporting larger effect sizes than earlier long-term studies. Some studies have suggested the coarse fraction, particles between 2.5 and 10 μm (PM10–2.5), may also be important. With respect to mortality and cardiovascular events, questions remain regarding the relative strength of effect sizes for chronic exposure to fine and coarse particles. Objectives We examined the relationship of chronic PM2.5 and PM10–2.5 exposures with all-cause mortality and fatal and nonfatal incident coronary heart disease (CHD), adjusting for time-varying covariates. Methods The current study included women from the Nurses’ Health Study living in metropolitan areas of the northeastern and midwestern United States. Follow-up was from 1992 to 2002. We used geographic information systems–based spatial smoothing models to estimate monthly exposures at each participant’s residence. Results We found increased risk of all-cause mortality [hazard ratio (HR), 1.26; 95% confidence interval (CI), 1.02–1.54] and fatal CHD (HR = 2.02; 95% CI, 1.07–3.78) associated with each 10-μg/m3 increase in annual PM2.5 exposure. The association between fatal CHD and PM10–2.5 was weaker. Conclusions Our findings contribute to growing evidence that chronic PM2.5 exposure is associated with risk of all-cause and cardiovascular mortality.

A substantial body of literature has shown asso ciations of particulate air pollution with mor tality and specifically cardiovascular disease. Recent studies have focused on the fine frac tion < 2.5 µm in diameter (PM 2.5 ) (Eftim et al. 2008;Laden et al. 2006;Miller et al. 2007). However, some studies have suggested that the coarse fraction, particles between 10 and 2.5 µm (PM 10-2.5 ), may be important as well (Brunekreef and Forsberg 2005;Host et al. 2008;Lipsett et al. 2006). A recent study in the Women's Health Initiative (WHI) reported a 24% increase in the risk of a cardio vascular event [hazard ratio (HR) = 1.24; 95% con fidence interval (CI), 1.09-1.41] and a 76% increase in the risk of death from cardio vascular disease (HR = 1.76; 95% CI, 1.25-2.47) for each 10µg/m 3 change in PM 2.5 levels meas ured in 2000 (Miller et al. 2007). The mag nitude of these estimates is higher than those reported in most longterm followup studies (Laden et al. 2006;Pope et al. 2002). Neither Pope et al. (2002) nor Miller and coauthors (2007) observed evidence of a positive associa tion with PM 10-2.5 .
We previously observed a positive asso ciation of chronic PM 10 (particles < 10 µm in aerodynamic diameter) exposures and all cause mortality and fatal coronary heart disease (CHD) in the Nurses' Health Study (NHS), a prospective cohort study of U.S. women ). In the current study, we extend that study to look specific ally at exposures to PM 2.5 and PM 10-2.5 in the same population. This longstanding cohort provides a unique opportunity with biennial updated assessment of covariates to examine these associations. With the geocoding of the nurses' biennially updated residential addresses and the recent development of geographic information sys tems (GIS)-based spatial smoothing models, this study uses time and spaceresolved PM 2.5 and PM 10-2.5 exposures at the monthly level. This individualspecific exposure assessment approach has not been possible in many previ ous studies of chronic air pollutant effects.

Methods
Study population. The NHS began in 1976 with 121,700 female registered nurses who lived in 11 states (California, Texas, Florida, Massachusetts, Pennsylvania, Ohio, New York, New Jersey, Michigan, Connecticut, and Maryland), were born between 1921 and 1946, completed a baseline questionnaire, and provided informed consent. The Brigham and Women's Hospital Institutional Review Board approved all aspects of this study. Participants have been mailed biennial questionnaires to their residential address to obtain information on risk factors and health outcomes since the study's inception. Among nurses available for followup, about 6% did not respond to cur rent questionnaires. For the current study we included participants who were living from 1992 until 2002 in metropolitan statistical areas (MSAs) of 13 contiguous states in the northeast and midwest United States (Maine, Vermont, New Hampshire, Ohio, Pennsylvania, Maryland, Michigan, Massachusetts, Connecticut, Rhode Island, New York, New Jersey, Delaware). We chose to limit the study population to women residing in MSAs (about 87% of participants in this geographic region) to allow for comparisons with results from pre vious studies that also have focused on met ropolitan areas (Eftim et al. 2008;Pope et al. 1995Pope et al. , 2002 and because the distributions of air pollution monitors and nurses were more sparse outside MSAs. Women were excluded for any time period of followup during which they resided outside this geographic region. Nonfatal myocardial infarctions (MIs) were assessed through biennial questionnaires and confirmed through medical record review by physicians blinded to study parti cipants' exposure status. Deaths were obtained through next of kin, postal authority reports, death cer tificates, or the National Death Index. Fatal CHD was confirmed by death certificate, hos pital records, or autopsy. Additional details regarding the assessment and confirmation volume 117 | number 11 | November 2009 • Environmental Health Perspectives of nonfatal MI, firstincident nonfatal or fatal CHD, fatal CHD, and allcause mortality for the current study are described elsewhere ). Only cases indicated as definitely or probably confirmed were counted. Women reporting cancers prior to 1992 (other than nonmelanoma skin cancer) were excluded at the beginning of followup. Accidental deaths were excluded from the allcause mortality analysis. Women with nonfatal MIs prior to baseline were excluded from fatal and nonfatal incident CHD cases for the current study.
Exposure assessment. Separate spatio temporal models for PM 10 and PM 2.5 were developed and validated, with coarse particle levels estimated by subtraction of predicted PM 2.5 from predicted PM 10 Yanosky et al. 2008Yanosky et al. , 2009. The PM 10 model is a GISbased spatial smoothing model that predicts monthly outdoor concentrations specific to each participant's bien nially updated residence. This generalized addi tive mixed model [GAMM, detailed elsewhere Yanosky et al. 2008)] used monitoring data from sites in the U.S. Environmental Protection Agency's Air Quality System (U.S. EPA 2007), the Interagency Monitoring of Protected Visual Environments (IMPROVE) network (National Aeronautics and Space Administration 2009), Clean Air Status and Trends Network (CASTNet, U.S. EPA 2009) data, and Harvard research stud ies to estimate monthly smooth spatial terms and penalized regression terms of GISbased and meteorologic covariates. These covariates included population density; distance to near est road by Census Feature Class Code A13; elevation; urban land use; point and areasource PM 10 emissions; wind speed; and precipitation.
We followed a similar process to pre dict monthly outdoor PM 2.5 concentrations Yanosky et al. 2009). Briefly, because of the lack of PM 2.5 moni tor data before 1999, we constructed separate models for 1988-1998 and 1999-2002. The post1999 PM 2.5 model was of similar form to the PM 10 model, and with similar covariates but used pointsource PM 2.5 emissions. The pre1999 model used a simpler spatiotemporal structure to estimate the PM 2.5 to PM 10 ratio seasonally and used estimated extinction coef ficients and covariates described previously. We estimated PM 10-2.5 exposures by subtracting the modeled PM 2.5 estimates from the PM 10 modeled estimates for each month at each location. The PM 10 model and post1999 and pre1999 PM 2.5 models were validated using crossvalidation. This procedure involved divid ing the monitoring locations into 10 sets, and fitting the model with each set held out. Then, we calculated the squared correlation between heldout observations and predictions from the model with each set removed. We used sets 1-9 to assess model performance, whereas we reserved set 10 to evaluate model over fitting. Each of these models performed well using crossvalidation, exhibiting little bias and high precision Yanosky et al. 2008Yanosky et al. , 2009. In comparison, predicted PM 10-2.5 levels exhibited little bias but were less precise (crossvalidation results are detailed elsewhere) ).
Evaluation of confounders and effect modifiers. Data from the biennial questionnaires were used to assess potential confounding and effect modification by covariates, including hypertension (yes, no); physiciandiagnosed dia betes (yes, no); hypercholesterolemia (yes, no); physical activity (< 3, 3 to < 9, 9 to < 18, 18 to < 27, or ≥ 27 metabolic equivalent (MET) hr per week); body mass index (BMI) (con tinuous); smoking status (never, former, or cur rent); and smoking packyears. Family history of MI (yes, no) was included based on answers to the 1976 and 1984 questionnaires. Census 2000 data were used to assign two grouplevel socioeconomic status variables, median house hold income and median household value at the census tract level. Confounding was assessed through adjustment for each of these covari ates in individual and multivariate Cox models.
Effect modification was evaluated through strat ification and the use of interaction terms.
Statistical analysis. Timevarying Cox pro portional hazards models were used to assess the relationship of allcause mortality and CHD outcomes with PM 2.5 and PM 10-2.5 . These models were based on a monthly time scale and were used to estimate HRs and 95% CIs. Personmonths of followup time were cal culated from baseline (30 June 1992) until the end of followup (30 June 2002), censoring at death or loss to followup. Persontime spent living outside the selected region was excluded, as were nurses with cancers or outcomes of inter est (e.g., nonfatal MI) reported prior to baseline. Incidence rates were estimated as the number of new cases divided by personmonths of fol lowup. We focused on the average exposure to PM 2.5 and PM 10-2.5 in the 12 months prior to the outcome of interest because a previous study has shown that to be the most relevant exposure . However, in separate models, we also considered other time windows of exposure, including average exposure in the 1, 3, 24, 36, and 48 months prior to the event. We assessed PM 2.5 and PM 10-2.5 in single and two pollutant models. All Cox models were strati fied by age in months and adjusted for state of residence (indicator variables), year (linear term), and season (indicator variables). By including state variables, the model adjusts for largescale spatial patterns in mortality that might be caused by factors other than pollution, thereby esti mating the effect of pollution based on within state variation (Dominici et al. 2006;Pope et al. 2002). All statistical analyses used SAS version 9.1 (SAS Institute Inc., Cary, NC, USA).

Results
The study population consisted of 66,250 women who lived in MSAs in the north eastern and midwestern United States in 1992 (Table 1)  . Their mean age was 62.4 years. During the followup period, most were never or former smokers and 42% had a BMI under 25. Mean (± SD) levels of PM 2.5 and PM 10-2.5 exposures in the previ ous 12 months were 13.9 ± 2.4 and 7.7 ± 2.6 µg/m 3 , respectively (further details on PM described elsewhere) Yanosky et al. 2009). There were 3,785 deaths; 1,348 incident CHD events; 379 fatal CHDs; and 854 nonfatal MIs.
HRs and 95% CIs for allcause mortality and other outcomes of interest for a 10µg/m 3 unit change in PM 2.5 and PM 10-2.5 averaged over the previous 12 months are presented in Table 2. In models adjusted for age, calen dar time, and state of residence, PM 2.5 was significantly associated with allcause mortal ity (HR = 1.45; 95% CI, 1.19-1.78). Results also suggest PM 10-2.5 may be associated with increased mortality risk (HR = 1.13; 95% CI, 0.98-1.30). The HRs for both size fractions of PM and incident CHD were also elevated. Risks associated with fatal CHD were larger for PM 2.5 (HR = 2.29; 95% CI, 1.26-4.18) than for PM 10-2.5 (HR = 1.28; 95% CI, 0.82-1.98) in crude models adjusted for age, calendar time, and state of residence. In sensitivity analyses using average annual PM 2.5 exposure estimates from the nearest U.S. EPA AQS monitor in 2000, as opposed to PM 2.5 estimates from the timevarying geospatial model, the risks of allcause mortality (HR = 1.35; 95% CI, 1.08-1.69) and fatal CHD (HR = 1.47; 95% CI, 0.73-2.99) were attenuated but elevated. Fully adjusted models included hyperten sion, family history of MI, hypercholesterolemia, BMI, physical activity, smoking, diabetes, median house value, and household income for census tract of residence, season, and state of residence and were stratified by age in months ( Table 2). Confounders were not highly cor related. The HRs were attenuated compared to models with limited control for confounding. PM 2.5 was associated with allcause mortality (HR = 1.26; 95% CI, 1.02-1.54) and fatal CHD (HR = 2.02; 95% CI, 1.07-3.78). Effect estimates for PM 2.5 and PM 10-2.5 were generally stable in two pollutant models, although esti mates with allcause mortality and fatal CHD were strengthened for PM 2.5 and attenuated for PM 10-2.5 . Overall, for nonfatal MI, the effect strengthened for PM 10-2.5 .
We assessed the sensitivity of our results to different time periods of exposure: 1, 3, 24, 36, and 48 months before the event. In sin glepollutant, fully adjusted models of PM 2.5 exposure, the associations with each outcome (except nonfatal MI) were stronger with times greater than 3 months and similar among time periods 12-48 months. In equivalent mod els for PM 10-2.5 , there were no apparent dif ferences among exposure windows (data not shown). Results for different periods of exposure were similar for multipollutant fully adjusted models (Figure 1). Table 3 shows relationships of PM 2.5 exposures in the previous 12 months with all cause mortality and fatal CHD adjusting for each potential confounder one at a time (after adjusting for state of residence, year, and sea son, and stratifying by age). For both fatal CHD and allcause mortality, median house value for the census tract of residence elevated the risk, whereas physical activity attenuated the risk associated with PM 2.5 exposures.
Although no interaction terms were statis tically significant, stratified analyses for each covariate, adjusting for all other co variates, showed some differences in the association between allcause mortality and chronic PM 2.5 exposure (Table 4). Women with hyper cholesterolemia or in the lowest category of physical activity were at higher risk. Risks were greatest for nonsmokers and least for current smokers. There was no evidence of effect modi fication for the relationship between allcause mortality and PM 10-2.5 (data not shown).
Women with a family history of MI were at significantly higher risk of fatal CHD associated with PM 2.5 exposure (Table 4). Stratified analyses also suggested greater risks for women with hypertension, hypercholester olemia, larger BMIs, and living in census tracts in the lowest quartile of median house value or the lowest two quartiles of median family income. Never smokers showed the highest risk and current smokers, the least. Similar stratified differences by BMI and smoking were evident for PM 10-2.5 (data not shown).

Discussion
In this study among women in the northeastern and midwestern region of the United States, we found each 10µg/m 3 elevation of PM 2.5 expo sure in the previous 12 months was associated  with an increased risk of allcause mortality (HR = 1.26; 95% CI, 1.02-1.54) and fatal CHD (HR = 2.02; 95% CI, 1.07-3.78) after controlling for known risk factors. Although we found evidence of a positive association between PM 10-2.5 exposure and allcause mortality, there was no association after adjustment for covari ates. An association between fatal CHD and PM 10-2.5 exposures was also evident but weaker in fully adjusted models. Finally, there was little evidence of an association between incident MI and PM. The relationship between PM 2.5 and fatal CHD was modified by family history of MI, and nonsmokers were at greatest risk, sug gesting the strong impact of smoking exposures conceals the effects of air pollution. However, CIs were wide. The attenuation of risk of allcause mortality and fatal CHD by physi cal activity as well as the increased risk of all cause mortality for women recording the least activity raises questions about the biological mechanism underlying these relationships. Few previous studies have examined the influence of physical activity. These results are consistent with those observed in the growing body of literature on chronic air pollution and health effects. In the extended followup of the American Cancer Society (ACS) Study, a 10µg/m 3 change in PM 2.5 was associated with an HR of 1.06 (95% CI, 1.02-1.11) for allcause mortality (Pope et al. 2002). The equivalent HR for a 10µg/m 3 change in the updated Harvard Six Cities Study was 1.14 (95% CI, 1.06-1.22) (Laden et al. 2006). Recently, Eftim et al. (2008) replicated these analyses among Medicare patients resid ing in the same counties included in these two studies. Their results were more consistent with those observed in our cohort (ACS: HR for a 10µg/m 3 change = 1.11; 95% CI, 1.09-1.13; Six Cities: HR = 1.21; 95% CI, 1.15-1.27). Specific associations with CHD also have been consistently observed. The Harvard Six Cities (Laden et al. 2006) and ACS (Pope et al. 2002) studies observed associations of 1.28 (95% CI, 1.13-1.44) for cardiovascular mortality and 1.09 (95% CI, 1.03-1.16) for cardiopulmo nary mortality, respectively. The recent WHI study reported overall risks of 2.21 (95% CI, 1.17-4.16) for cardiovascular mortality and 1.76 (95% CI, 1.25-2.47) for incident CHD (Miller et al. 2007). Two additional studies of men and women have found greater suscepti bility among women for cardiovascular events associated with particulate matter exposures (Chen et al. 2005;Rosenlund et al. 2006).
Neither the original ACS study (Pope et al. 1995) nor the WHI study (Miller et al. 2007) observed an association between allcause mortality and chronic exposure to PM 10-2.5 . However, the Harvard Six Cities study found an elevated relative risk associated with exposure to PM 15-2.5 (HR = 1.19; 95% CI, 0.91-1.55) (U.S. EPA 1996). McDonnell et al. (2000) observed an HR of 1.05 (95% CI, 0.92-1.20) for an IQR increase (9.7 µg/m 3 ) in PM 10-2.5 among men living near an airport in a cohort of Seventhday Adventists. In an acute exposure study of fine and coarse particles in Shanghai, China, Kan et al. (2007) found no significant effect of PM 10-2.5 on mortality, but increases of total and cardiovascular mortality were reported. Other studies of acute exposure to coarse particulate matter have also suggested a relationship with cardiovascular outcomes (Host et al. 2008;Lipsett et al. 2006;Peng et al. 2008;Tolbert et al. 2000), although Peng et al. (2008) reported the association between daily CVD hospital admissions and coarse particu late matter was no longer statistically significant when adjusted for PM 2.5 . We found a stronger association between coarse particulate matter and fatal CHD than with allcause mortality in fully adjusted singlepollutant models and with nonfa tal MI in multipollutant models. In multipollut ant models, the WHI also found an association of PM 2.5 (HR = 1.53; 95% CI, 1.21-1.94) and PM 10-2.5 (HR = 1.10; 95% CI, 0.97-1.23) with first cardiovascular event (Miller et al. 2007).
In general, our results are elevated in rela tion to other studies, with the exception of the WHI, another cohort study of women. This disparity could be due in part to the use of different air pollution exposure estimation methods. We modeled monthly exposures for each biennially updated residential address using GISbased spatial smoothing models. For example, the ACS study used mean expo sure in metropolitan areas measured during a few of the years of followup (Pope et al. 1995). A reanalysis restricting to subjects with monitors in their county of residence reported higher risks (Willis et al. 2003). Additionally, a study in Southern California using spatially estimated exposures reported stronger results (Jerrett et al. 2005). To the extent that our exposure modeling accounts for local variation that other studies do not, we might be cap turing different sources of pollution resulting in different effect sizes. Further, our sensitiv ity analyses, using a lessprecise exposure esti mate, showed an attenuation of the effect size. Therefore, it appears that studies using spa tially estimated exposure measures may pro duce higher risk estimates. This has important implications for risk assessment. Because PM 10-2.5 estimates were derived from PM 10 and PM 2.5 estimates, more uncer tainty is associated. This may contribute to the lower effect estimates we observed for PM 10-2.5 , but Yanosky et al. (2009) show that longterm average PM 10-2.5 was reasonably well estimated (crossvalidation R 2 = 0.63 and 0.65 for long term post1999 and pre1999 PM 10-2.5 , respec tively). To improve our exposure modeling, we focused on the northeastern and midwestern United States (63% of the total study popula tion), an area with more uniformly distributed study population and monitors, and results may differ for other U.S. regions. In addition, we were unable to account for nurses who moved between the biennial questionnaires or for lengthy stays away from their residence in another geographic region. We estimated time and spaceresolved exposures with GIS based smoothing models. Although smooth ing reduces variability relative to measured concentrations, Gryparis et al. (2008) show this is a type of Berkson measurement error that should not cause substantial bias toward the null. Additionally, modeling allows us to assign exposures specific to biennially updated residential addresses for the entire period of followup. Thus, compared with using only monitor measurements, fewer participants are lost because of missing exposure data.
A strength of this study is that we have updated information on numerous covariates throughout the followup period. However, there does exist the possibility of residual confounding or confounding by unmeasured covariates and/or by additional pollutants.
Biological mechanisms for the relationship of particulate matter exposure with mortality and CVD have not yet been fully explained. Several mechanisms have been proposed, including changes in autonomic function, oxi dative stress, and systemic inflammation leading to endothelial dysfunction, thrombosis, or ath erosclerosis (Donaldson et al. 2001;Gurgueira et al. 2002;Pope et al. 2004;Pope and Dockery 2006;Utell et al. 2002). Although fine par ticles deposit deeper into the lung (Brunekreef and Forsberg 2005;Venkataraman and Kao 1999), some studies have shown coarse particles have a greater ability to stimulate inflammatory responses and macrophages cytokine produc tion (Becker et al. 2002(Becker et al. , 2005Brunekreef and Forsberg 2005).

Conclusions
Our findings contribute to growing evidence that annual exposure to particles is associated with increases in risk of allcause and cardiovas cular mortality. The extended followup of the Harvard Six Cities study (Laden et al. 2006) found mortality risks associated with exposure in the year of death were similar to those for the entire followup period. In another reanaly sis, Schwartz et al. (2008) reported that the association was with exposure in the previous 2 years. A recent study of a Medicare cohort of MI survivors and PM 10 exposure examined the effect of multiple lags of exposure on survival. Again, the effect of particles on mortality risk seemed to go back only a few years (Zanobetti and Schwartz 2007).
In summary, with chronic coarse and fine particulate exposures estimated on a finer spa tial and temporal scale than in previous cohort studies (Dockery et al. 1993;Pope et al. 1995), we found PM 2.5 was associated with increased risks of allcause mortality and fatal CHD. Coarse particulate matter exposures were not associated with an increase in risk after control for confounders. In addition, our results sug gest that health benefits may be evident within a few years of reductions in particle levels.