The Effect of Fine and Coarse Particulate Air Pollution on Mortality: A National Analysis

Background Although many studies have examined the effects of air pollution on mortality, data limitations have resulted in fewer studies of both particulate matter with an aerodynamic diameter of ≤ 2.5 μm (PM2.5; fine particles) and of coarse particles (particles with an aerodynamic diameter > 2.5 and < 10 μm; PM coarse). We conducted a national, multicity time-series study of the acute effect of PM2.5 and PM coarse on the increased risk of death for all causes, cardiovascular disease (CVD), myocardial infarction (MI), stroke, and respiratory mortality for the years 1999–2005. Method We applied a city- and season-specific Poisson regression in 112 U.S. cities to examine the association of mean (day of death and previous day) PM2.5 and PM coarse with daily deaths. We combined the city-specific estimates using a random effects approach, in total, by season and by region. Results We found a 0.98% increase [95% confidence interval (CI), 0.75–1.22] in total mortality, a 0.85% increase (95% CI, 0.46–1.24) in CVD, a 1.18% increase (95% CI, 0.48–1.89) in MI, a 1.78% increase (95% CI, 0.96–2.62) in stroke, and a 1.68% increase (95% CI, 1.04–2.33) in respiratory deaths for a 10-μg/m3 increase in 2-day averaged PM2.5. The effects were higher in spring. For PM coarse, we found significant but smaller increases for all causes analyzed. Conclusions We conclude that our analysis showed an increased risk of mortality for all and specific causes associated with PM2.5, and the risks are higher than what was previously observed for PM10. In addition, coarse particles are also associated with more deaths.


Research
Many multicity studies have shown that ambi ent particulate air pollution, generally meas ured as particulate matter with aerodynamic diameter ≤ 10 µm (PM 10 ), is associated with increased risk of death for broadly defined cardiovascular or respiratory causes, using time series analysis Katsouyanni et al. 1997;Schwartz 1994Schwartz , 2000 or case-crossover analysis (Schwartz 2004;Zeka et al. 2005).
It is generally thought that fine particles (those with an aerodynamic diameter of ≤ 2.5 µm; PM 2.5 ) are more harmful to health than larger particles (PM 10 ) (Cifuentes et al. 2000;Schwartz et al. 1996), although some studies have shown stronger effects with coarse particles (particles 2.5-10 µm in aerodynamic diameter) (Ostro et al. 2000).
The literature on the association between fine particles (PM 2.5 ) and mortality is rela tively sparse, because of two main issues: The U.S. Environmental Protection Agency (EPA) began PM 2.5 monitoring in 1999, and mortality data from the National Center for Health Statistics (NCHS) were not available nationwide after the year 2000. Nevertheless, several multicity studies have been performed. Using early PM 2.5 monitoring data from the Harvard Six City Study, Schwartz et al. (1996) reported a strong association between 2day average PM 2.5 and daily deaths, but little association with coarse particles. A study of eight Canadian cities similarly found asso ciations with fine but not coarse particles (Burnett et al. 2000). Dominici et al. (2007) examined PM 2.5 -mortality associations using national data in the United States, but only for the years 1999-2000.
Franklin and coauthors in two papers (Franklin et al. 2007(Franklin et al. , 2008 addressed this issue a different way. They used mortal ity data up to year 2000 from the NCHS, whereas they obtained mortality data directly from state health departments for the years 2001-2005 to examine the mortality effects of PM 2.5 for an extended period of time. They were limited by their ability to obtain mortal ity state data, however. Ostro et al. (2006) examined the associa tions between PM 2.5 and daily mortality for the years 1999-2002 in nine California coun ties. They also found significant associations with PM 2.5 .
With a common effort of the U.S. EPAfunded PM research centers and the help of the U.S. EPA, we were able to obtain mor tality data from 2001 through 2005 from each state in the country (except Hawaii and Idaho) through NCHS. As a result, national data for an extended period are available for the first time.
Based on the associations found in the previous studies and the new mortality data available, we hypothesized that PM 2.5 is associated with increased risk of deaths in a national study. Because coarse particles are not currently regulated by the U.S. EPA, we also test the hypothesis that PM coarse (PM with aerodynamic diameter > 2.5 and < 10 µm) are associated with mortality. We therefore conducted a multicity time series study of the acute effect of PM 2.5 and PM coarse on the increased risk of death for all causes, all cardio vascular disease (CVD), myocardial infarction (MI), stroke, and respiratory mortality.

Materials and Methods
Health data. The NCHS provided researchers with national data sets of mortality records, including date of death. Certain locations were excluded from the data, either because the population of the county was low, or because the state was not willing to cooperate with the project. The analysis was conducted at the county level, because this was the smallest resolution available for all mortality data; the name of the major city within each county was used as an identifier rather than the county name. The mortality data used provided non confidential information on decedents includ ing state of death, county of death, age, sex, date of death and primary cause of death.
We excluded those individuals who died in a state different from their state of residence. Only those individuals who died of nonaccidental causes were examined [i.e., International Statistical Classification of Diseases, 10th Revision (ICD10; World Health Organization 2007) codes S00 through U99 were excluded].
For all nonaccidental deaths and for each specific cause, we created daily counts of deaths in each of the examined counties. This work was done under an exemption from Human Subjects Committee of the Harvard School of Public Health.
Environmental data. We obtained data on PM 2.5 and PM 10 from the U.S. EPA Air Quality System Technology Transfer Network (U.S. EPA 2008), which provides daily PM 2.5 concentrations from the U.S. EPA National and State Local Ambient Monitoring stations.
In most cities, the analysis was con ducted on a county level, because the city lies within a single county. However, we used multiple counties for When more than one monitor was avail able in one county, the 24hr integrated mass concentrations were averaged over the county using a method previously described (Schwartz 2000;Zanobetti et al. 2000). Briefly, we first excluded any monitor that was not well cor related with the others (r < 0.8 for two or more monitor pairs within a county), because it likely measured a local pollution source and would not represent the general population exposure over the entire community. We then computed the annual mean for each monitor and year and subtracted that mean from the daily values of that monitor. We then stan dardized these daily deviances by dividing by the standard deviation for that monitor. The daily standardized deviations for each monitor on each day were averaged, producing a daily averaged standardized deviation. We finally multiplied this by the standard deviation of all of the monitor readings for the entire year and added back in the annual average of all of the monitors. This process automatically compensates for missing data in some moni tors on individual days by preventing that missingness from contributing to false varia tions in the daily value. This process has been reported previously (Schwartz 2000) and used extensively in previous publications (O'Neill et al. 2003;Wellenius et al. 2006;Zanobetti and Schwartz 2005;Zanobetti et al. 2000;Zeka et al. 2005).
To be included in our study, we required that at least 265 days of data in at least 1 year be available. We found 112 cities with at least 265 days of monitoring of PM 2.5 per year and at least 300 days of mortality data per year from NCHS. They represented a geographic distribution across the country [ Figure 1; Table 1 in Supplemental Material (available online at http://www.ehponline. org/members/2009/0800108/suppl.pdf)]. PM coarse was estimated by differencing the countywide averages of PM 10 and PM 2.5 . PM coarse was available for fewer locations, because less monitoring of PM 10 is currently being done. We had 47 locations that met our criteria for PM coarse.
We obtained local meteorologic data from the U.S. Surface Airways and Airways Solar Radiation hourly data (National Environmental Satellite, Data and Information 2003).
Because climate may effect exposure to particles (e.g., because ventilation varies by climate) and may also modify particle charac teristics (e.g., because climate affects secondary particle formation), we divided the United States into regions based on the Köppen cli mate classification (Kottek 2006;Kottek et al. 2006), which is one of the most widely used climate classification systems. The Köppen cli mate classification scheme divides the climates into five main groups and several types and subtypes. We used the following classifica tion: region 1: humid subtropical climates and maritime temperate climates, which includes Florida, Louisiana, Texas, Georgia, Alabama, Mississippi, Arkansas, Oklahoma, Kansas, Analytical strategy. We investigated the association between PM 2.5 and PM coarse concentrations averaged over the day of death and day before death and mortality with a time series analysis. The analysis was stratified by season because the composition of par ticles varies seasonally (partly due to different source contributions at different times of the year) and because the penetration of outdoor particles indoors varies seasonally.
We first applied season and cityspecific Poisson regression models, controlling for longterm trend and seasonality with a natural cubic regression spline with 1.5 degrees of freedom for each season for each year; day of the week using indicator variables; and for weather using a natural cubic spline with three degrees of freedom for the sameday tempera ture and for the previousday temperature.
We first applied singlepollutant models and then fit a model including both PM 2.5 and PM coarse. To test the hypothesis that coarse particles may affect mortality with a  longer lag, making the choice of lags 0 and 1 less appropriate, we also fit a distributed lag model for 4 days, from the same day and up to 3 days earlier for coarse particles. The cityspecific results were then com bined with a random effects metaanalysis (Berkey et al. 1998). To be conservative, we report the results incorporating a random effect, whether or not there was a significant heterogeneity.
The pooled analysis was done for each outcome separately, with a total of 448 city and seasonspecific coefficients for PM 2.5 and 188 for PM coarse. In the twopollutant model, the metaanalysis was done for each pollutant separately.
We used the I 2 statistic to assess the pro portion of total variation in effect estimates that was attributed to betweencity heteroge neity (Higgins and Thompson 2002). The I 2 statistic is a generalization of the C 2 or Q test for heterogeneity and expresses the proportion of variance explained. We used the following formula: where Q is the Qtest for heterogeneity and k is the number of community. If Q/(k-1) is < 1, then the I 2 is null and indicates that no vari ability is attributable to heterogeneity (zero in the tables).
We analyzed the data using R version 2.7.2 (R Development Core Team 2008). The effect estimates were expressed as a percent increase in mortality for a 10µg/m 3 increase in PM 2.5 mass or PM coarse concentration.

Results
In the 112 cities during the study period 1999-2005, there were 5,609,349 total deaths, 1,787,078 for CVD, 397,894 for MI, 330,613 for stroke, and 547,660 for respira tory disease. The biggest cities are Los Angeles, California; New York City, New York; and Chicago, Illinois.
The cities with higher levels of PM 2.5 were in California, with maximum concentrations of PM 2.5 > 100 µg/m 3 , whereas the cities with the lower maximum levels were in Oklahoma. The median PM 2.5 ranged from 5.6 µg/m 3 in Albuquerque, New Mexico, followed by Eugene and Bend, Oregon, with 6 µg/m 3 , whereas the highest median concentrations were 21.5 µg/m 3 in Rubidoux, California, and 17.4 µg/m 3 in Los Angeles. Table 1 in Supplemental Material (avail able online at http://www.ehponline.org/ members/2009/0800108/suppl.pdf) presents the following for each of the 112 cities: the years of study, the daily mean concentration levels for PM 2.5 and PM coarse, and the daily mean number of death by cause. Figure 1 shows the location of the 112 U.S. cities included in the study; the symbol size represents the population in each city, and the color represents the PM 2.5 concen trations. High levels of PM 2.5 (red) are in California and in the industrial Midwest. Table 1 shows the percent increase in mortality for a 10µg/m 3 increase in PM 2.5 for the mean of lags 0 and 1 (henceforth mean01), for the mean01 by season.
We found significant associations with all the analyzed causes of death and PM 2.5 , with the highest effect for stroke with a 1.78% increase [95% confidence interval (CI), 0.96-2.62], and respiratory mortality with a 1.68% increase (95% CI, 1.04-2.33) for a 10µg/m 3 increase in the mean01 of PM 2.5 . When looking at the results by season, the high est effects are in spring, with > 2% increases. Table 2 shows the percent increase in mortality for a 10µg/m 3 increase in PM coarse across the 47 cities for the sum of the 4 days distributed lag model and by sea son. We found significant associations with total mortality, stroke, CVD, and respira tory mortality, for which we had the largest effect: a 1.2% increase (95% CI, 0.4-1.9) for a 10µg/m 3 increase in PM coarse. The effect sizes per unit of mass for PM coarse are about half those for PM 2.5 .
When we examined the distributed lag for PM coarse (Figure 2), we found little evidence that the effects were at longer lags. Figure 2 reports the percent increase in causespecific mortality at each lag from the distributed lag model, combined across the 47 cities. There were indications of some effect at lag 2, but the sum of the distributed lag was not higher than the results for mean01.
When looking at the heterogeneity by sea son, we found no significant heterogeneity in general in summer (Table 1), whereas sig nificant heterogeneity was seen in spring and autumn. For allcause mortality and PM 2.5 , 32% in spring and 15% in autumn of the total variability in cityspecific coefficients was attributable to betweencommunity differ ences (as opposed to stochastic variation). For PM coarse, we only found significant city specific heterogeneity for allcause and respira tory mortality. Again, the highest percentage of explained total variability was in spring. Table 3 shows the results from the mul tipollutant model, which included both PM 2.5 and PM coarse. There were only minor changes in the effect size estimates for either pollutant, and both remained significant for allcause mortality, although some of the causespecific results had increased CIs that included no effect. Table 4 shows the effects for PM 2.5 and PM coarse by region. The number of cities  Values are percent increase (95% CI) for 10-µg/m 3 increase in PM 2.5 . I 2 statistics and significance level (*p < 0.05) for heterogeneity. Mean01 is the mean of lags 0 and 1 (overall and by season).
varied in each region and by pollutant. For PM 2.5 in the six regions, we have 47, 28, 17, 2, 3, and 15 cities, respectively. For PM coarse, we have 15, 11, 11, 2, 3, and 5 cities. The regions with dry climates and together with continental climate are the regions with the lower number of cities because they are not very populated, but they have the same number of cities for both PM 2.5 and PM coarse. The effects of PM 2.5 on allcause mortality are similar for all regions except for the last (Mediterranean), which include California, Oregon, and Washington. The results were more varied for the specific causes of death, but the precision of the estimates was also less. There was a consistent trend for lower effects in the Mediterranean region for each cause as well.
In contrast, the pattern was different for coarse particles. First, there was consider ably more variation in general in the allcause mortality effects by region. Not only was the Mediterranean region different (as for PM 2.5 ); there was no effect in the dry region as well. In addition, the effect size in the dry conti nental region was double that in the humid subtropical region for allcause mortality and triple for CVD deaths.
The percentage of explained total variabil ity differed between regions and between the two pollutants, with significant heterogene ity in the "warm summer, continental," "hot summer, continental," and "Mediterranean" regions when looking at allcause mortality and PM 2.5; for PM coarse significant hetero geneity was found in "warm summer, conti nental" and "dry" regions.

Discussion
In this national, multicity study we found a significant association between fine particulate air pollution and the risk of mortality for all causes, MI, CVD, stroke, and respiratory dis ease. We also found a significant association of coarse PM with daily deaths. Both effects were little changed after controlling for the other pollutant. There were several other fea tures of our results worth noting.
These associations were higher during spring, which is consistent with the findings of the effects of PM 10 on mortality by Zeka et al. (2006) but at variance with the findings of the National Morbidity, Mortality, and Air Pollution Study for PM 10 (Peng et al. 2005), which found stronger effect during summer. It is also consistent with the report of Franklin et al. (2008), who showed that mean tempera ture in a given city had an inverted Ushaped association with the seasonspecific coefficient for PM 2.5 : Mild temperatures, which were associated with greater indoor penetration, were associated with higher PM 2.5 effects, whereas both hot and cold temperatures were associated with lower effects. In our study, the higher effect sizes in the spring may reflect the same pattern.
In contrast to the PM 10 association, where substantial regional differences have been reported, our analysis showed most climatic regions had very similar effect size estimates for PM 2.5 , except for the Mediterranean climatic region. However, the coarse particle effects varied much more, and the pattern of which regions were higher and lower differed between fine and coarse mode particles. This suggests that there are regional variations in the toxicity of coarse particles that require further study. It may be possible that coarse particles are coated with different substances in different regions, for example. In some previous studies, PM 10 showed substantially lower effects and more regional heterogeneity than we see for PM 2.5 and PM coarse. However, if the relative toxic ity of PM fine and PM coarse varies differently by region, models fit with PM 10 as the expo sure index, which is essentially the sum of the two, may be effectively inducing measurement error in the exposure variable, which likely con tributes to a downward bias in effect size, com pared with treating them separately. Similarly, the different regional variability in effect size for fine and coarse mass may result in greater regional variability of PM 10 coefficients.
One possible explanation for the lower effect in the Mediterranean region, which includes California, is more measurement error due to the extremely large counties in California, where people may live far away from the monitors. Moreover, in California there is substantial withincounties gradient in particle concentrations, as shown by Jerrett et al. (2005).
The magnitude of the PM 2.5 risk estimates reported in our study are similar to the esti mates from the two studies of Franklin and coauthors. In the first paper, Franklin et al. (2007) analyzed 27 cities that had PM 2.5 . The results from the case-crossover analysis for the previous day were a 1.21% (95% CI, 0.29 to 2.14%) increase in allcause mortality, a 1.78% (95% CI, 0.20 to 3.36%) increase in respiratoryrelated mortality, and a 1.03% (95% CI, 0.02 to 2.04%) increase in stroke related mortality with a 10µg/m 3 increase in previous day PM 2.5 . These results are gener ally comparable but slightly higher than ours, although we report results for a 2day average.
In the second paper, Franklin et al. (2008) reported the association from the time series analysis by season for the 2day averaged PM 2.5 concentrations. They found a 0.74% (95% CI, 0.41 to 1.07%) increase in total mortality; a 0.47% (95% CI, 0.02 to 0.92%) increase in CVD; a 0.67% (95% CI, -0.21 to 1.07%) in stroke; and a 1.01% (95% CI -0.03 to 1.57%) increase in respiratory   Dominici et al. (2007) examined the years 1999-2000 and reported a PM 2.5 effect at lag 1 day of 0.29% (posterior interval, 0.01, 0.57) per 10µg/m 3 increase for all causes. This is lower than our estimate but was done using data for 2 years only. One other possible explanation for the difference is that Dominici et al. (2007) esti mated an effect for PM 2.5 in a single model for all seasons. In contrast, our analysis and the Franklin timeseries analysis fit separate regressions in each city for each season. This allows for seasonal differences in the effect of day of the week terms and of the splines for temperature. Hence, it allows, for example, for different effects of a hot day in May ver sus August. In contrast, Dominici et al. used many more degrees of freedom to control for weather variables, but fit those curves for the entire year. Thus, control for confound ing was different. In addition, we fit different effect size estimates for PM 2.5 for each sea son. Although we averaged those estimates to obtain a yearly average effect, because that is relevant to a pollutant with an annual average standard, we began with different effects by season, which were significant. In contrast, the Dominici analysis effectively assumes the same size effect in each season. If that assumption is false, it could result in a type of measurement error such as attenuation in the effect size estimate, even for the annual aver age effect, because seasonal changes in toxicity of 1 µg/m 3 of particles can be considered as a seasonal variation in measurement error [in the use of the same exposure metric, when the true exposure (toxicity weighted) is different]. Moreover, the differences in the effect sizes could be explained by differences in exposure metrics, where we also used the mean of lags 0 and 1 instead of lag 1 alone as exposure metric.
There is considerable toxicologic sup port for these findings. Animal experiments indicate that reactive oxygen species, which have established relevance in the pathogenesis of CVD and aging (Dhalla et al. 2000), are affected by particles, which represents one pathway for their cardiovascular and lung effects (Brook et al. 2004;Gurgueira et al. 2002;Nel 2005;Rhoden et al. 2004). Diesel particles increase oxidative stress in endothe lial tissue, inducing the production of heme oxygenase1, a rapid response part of the body's defense system against oxidative stress (Furuyama et al. 2006). The viability of cell cultures of microvascular endothelial cells was impaired by diesel particles, with an accom panying large increase in induction of heme oxygenase1 (Hirano et al. 2003).
The implication of oxidative stress mecha nisms has epidemiologic support as well. A recent report showed that subjects who were GSTM1 null or had the long variant of HMOX-1 had enhanced effects of particles on heart rate variability, including a threeway interaction (Chahine et al. 2007). Rossner et al. (2008aRossner et al. ( , 2008b examined bus drivers in Prague and reported increased levels of indica tors of oxidative stress such as F2 isoprostane and 8hydroxydeoxyguanosine (8OHdG) in drivers compared with controls. Particles, and particularly the metals on particles, have also been associated with an increased production of 8OHdG (Chuang et al. 2007;Kim et al. 2004;Knaapen et al. 2002;Prahalad et al. 2000Prahalad et al. , 2001, including specifically exposure to traffic pollution (Chuang et al. 2003;Lai et al. 2005;Ma and Ma 2002;Tokiwa et al. 1999). 8OHdG was also elevated in urban children compared with rural children (Tondel et al. 2005).
Other mechanisms have also been impli cated. Particles have been shown to increase sICAM1 (soluble intercellular adhesion molecule1) and sVCAM1 (soluble vascular adhesion molecule 1) in diabetics (O'Neill et al. 2007), a finding confirmed in a con trolled human exposure chamber study (Salvi et al. 1999). Both animal and controlled human exposure studies have demonstrated that ambient particles can increase prothrom botic (clotforming) activity and even induce thrombosis in acute exposures (Mills et al. 2007;Mutlu et al. 2007;Nemmar et al. 2003). In a relevant epidemiologic study, Baccarelli et al. (2007Baccarelli et al. ( , 2008 reported an asso ciation of airborne particles with decreased clotting time as well as the risk of deep vein thrombosis. This is consistent with the results of a controlled exposure study to diesel par ticles, which reported increased ST depres sion (a sign of ischemic heart disease) and alterations in fibrinolytic capacity (the ability to break up clots that have formed in blood vessels) (Mills et al. 2007).

Conclusions
In summary, there is a strong association of both fine and coarse particles with daily deaths. These associations are biologically plausible and, at the mean concentrations in the United States, suggest tens of thou sands of early deaths per year, which could be avoided by reducing particle concentrations. Because coarse particles are not currently reg ulated by U.S. EPA and many power plants and pre2007 diesel engines are grandfathered from having to retrofit controls to reduce fine particles, considerable public health improve ment may be possible.