Time-Series Analysis of Mortality Effects of Fine Particulate Matter Components in Detroit and Seattle

Background Recent toxicological and epidemiological studies have shown associations between particulate matter (PM) and adverse health effects, but which PM components are most influential is less well known. Objectives In this study, we used time-series analyses to determine the associations between daily fine PM [PM ≤ 2.5 μm in aerodynamic diameter (PM2.5)] concentrations and daily mortality in two U.S. cities—Seattle, Washington, and Detroit, Michigan. Methods We obtained daily PM2.5 filters for the years of 2002–2004 and analyzed trace elements using X-ray fluorescence and black carbon using light reflectance as a surrogate measure of elemental carbon. We used Poisson regression and distributed lag models to estimate excess deaths for all causes and for cardiovascular and respiratory diseases adjusting for time-varying covariates. We computed the excess risks for interquartile range increases of each pollutant at lags of 0 through 3 days for both warm and cold seasons. Results The cardiovascular and respiratory mortality series exhibited different source and seasonal patterns in each city. The PM2.5 components and gaseous pollutants associated with mortality in Detroit were most associated with warm season secondary aerosols and traffic markers. In Seattle, the component species most closely associated with mortality included those for cold season traffic and other combustion sources, such as residual oil and wood burning. Conclusions The effects of PM2.5 on daily mortality vary with source, season, and locale, consistent with the hypothesis that PM composition has an appreciable influence on the health effects attributable to PM.

Over the past few decades, there has been growing interest in the adverse health effects of ambient air particulate matter (PM) on humans. Evidence from epidemiological studies has accumulated to support associa tions between PM and adverse health effects (e.g., Dominici et al. 2006;Peng et al. 2005;Schwartz and Morris 1995;. A number of recent mul ticity timeseries studies have demonstrated associations between daily mortality and the mass concentration of PM 2.5 (PM ≤ 2.5 µm in aerodynamic diameter) (e.g., Laden et al. 2000;Ostro et al. 2006).
There is also growing evidence that PM 2.5 mass alone might not be able to explain the health outcomes (Franklin et al. 2008;Lippmann et al. 2006;Sarnat et al. 2008), because ambient PM 2.5 is chemically nonspe cific and consists of various components and compounds [trace elements, elemental carbon (EC), organic carbon, and sulfate], and the toxicity of each of these chemical components and their mixtures may vary. These different components originate from various sources, such as trafficrelated emissions, biomass com bustion, residual oil burning, and resuspended dust. Metals that are components of ambient PM 2.5 , and especially the transition metals, have been cited as likely toxic components. The focus has often been on iron (Fe), vana dium (V), nickel (Ni), chromium, copper, and zinc (Zn) based on their ability to gener ate reactive oxygen species in biological tis sues. The National Research Council (NRC) has highlighted the importance of investigat ing characteristics and chemical components of PM that contribute to their toxicity (NRC 2004). Focusing regulations on the most toxic PM components could protect public health more effectively and at a lower cost.
The Chemical Speciation Network (CSN) initiated by the U.S. Environmental Protection Agency (EPA) in 2000 has provided new opportunities for studies on PM components' health effects (U.S. EPA 2010b). However, the sampling frequency of only 1in3 or 1in6 days severely limits the statistical power for timeseries analysis. To date, because of this statistical power issue, most studies have used annual or seasonal averages of the PM 2.5 in secondstage regression to explain the cityto city differences in PM health effect estimates obtained in the firststage timeseries analysis in multiple individual cities Franklin et al. 2008;Lippmann et al. 2006;). Direct regressions of chemical species in timeseries models were limited either to a few chemical components that explained a substantial amount of PM 2.5 mass  or to the analysis of multiple cities using earlier data. These analyses produced wide confidence bands, an appar ent reflection of the issue of small sample size (Ostro et al. 2007). In the present study, we analyzed 3 years of daily speciation data in two U.S. cities that allowed an examination of the role of chemical species in PM 2.5 associated health effects with larger sample sizes than those relying on CSN data.

Materials and Methods
Detroit's air quality is heavily influenced by a variety of industries, including motor vehicle factories, refinery operations, power plants, and metalworking plants. Accordingly, a number of air pollution-health effect studies have been conducted in Detroit (Harkema et al. 2004;Ito 2003;Morishita et al. 2004;Schwartz 1991;Schwartz and Morris 1995). Seattle, a major seaport for commerce with Asia, has container shipping, as well as ware housing facilities that are located northwest and west of the monitoring site, and the Port of Seattle contributes the largest source of oil combustion effluents in Seattle (Kim et al. 2004;Maykut et al. 2003). Seattle also has some other localized PM sources, such as resi dential wood burning. These two cities also have large populations, different climates, and different pollution mixes. PM 2.5 cumulative filter samples. Two sets of 24hr PM 2.5 samples for the 3year period 2002 through 2004 were collected on 47mm Teflon filters using samplers operat ing at a flow rate of 16.7 L/min. These fil ters were from two monitoring sites: Detroit, Michigan (Allen Park, site ID 261630001), and Seattle, Washington (Beacon Hill, site ID 530330080) and were weighed pre and postsampling to determine daily fine particle mass concentration. We obtained these fil ters from the Washington Department of Ecology (Seattle data) and Department of volume 119 | number 4 | April 2011 • Environmental Health Perspectives Environmental Quality (Detroit data) and subsequently sent them to the Research Triangle Institute (Research Triangle Park, NC) for analysis of trace element concen trations using energydispersive Xray fluorescence (XRF). Black carbon, as a sur rogate index of EC in the filter samples, was determined via reflectometry at New York University Department of Environmental Medicine (Tuxedo, NY).
Choice of PM 2.5 chemical components for data analysis. From the XRF analysis of 48 elements, we chose 10 chemical components for the health effects analysis based on toxi cological consideration: aluminum (Al), Fe, potassium (K), sodium (Na), Ni, sulfur (S), silicon (Si), V, Zn, and EC. Source types were selected a priori, based on previous studies Cooper and Watson 1980;Franklin et al. 2008;Gildemeister et al. 2007;Lippmann et al. 2006;Maykut et al. 2003;Thurston and Spengler 1985), as well as our cursory factor analysis (data not shown). Collectively, these 10 components represent the major emission sources in these two cities: Al and Si for soil, Fe and Zn for smelter efflu ents, Ni and V for residual oil burning, S for coal burning, EC for traffic, Na for sea salt, and K for wood burning. We removed several extreme values of K from the data set (from samples collected around 4 July 20024 July , 20034 July , and 2004 that were influenced by fireworks).
Gaseous pollutants data. We obtained daily concentrations of carbon monoxide (CO), ozone (O 3 ), and nitrogen dioxide (NO 2 ) from the Health Effects Institute (HEI) Air Quality Database (2010) and included these gaseous criteria pollutants into our models as potential confounders of PM components and to esti mate their potential effects.
Mortality data. We obtained National Center for Health Statistics (NCHS) nation wide multiple causeofdeath files with date of death through the U.S. EPA's special arrangement with NCHS to support research (Centers for Disease Control and Prevention 2010). Using codes from the International Classification of Diseases, version 10 (ICD10; World Health Organization 2007), we aggre gated daily death counts to nonaccidental all cause deaths (ICD10, codes A00 through R99), cardiovascular deaths (ICD10, codes I01 through I99), and respiratory deaths (ICD10, codes J00 through J99).
Meteorological data. We retrieved daily average temperature and dew point data for the Detroit Metro Airport and the SeattleTacoma International Airport from the National Oceanic and Atmospheric Administration, National Climatic Data Center (2009) and determined relative humidity with temperature and dew point. All daily mortality, pollutants, and weather data were processed using SAS software (version 9.1; SAS Institute Inc., Cary, NC).
Methods. We used a Poisson regression model to examine shortterm associations between ambient air pollution and mortal ity by cause. The regression model included an indicator for day of week, a smooth func tion of time with 8 degrees of freedom (df) per calendar year (4 df/year for season analy ses) to control for seasonality and longterm trends ), a smooth function of currentday temperature (3 df), a smooth function of delayed temperature (3 df), and a smooth function of humidity (3 df). For all of the smooth functions, we used a natural spline basis. We used consecutive number of study days (from 1 through 1,096 for 3 years) to fit natural cubic splines. We employed 3day average lags (lag 1 to lag 3 days) for delayed temperature effects. The Poisson model also accommodated overdispersion. We exam ined 0, 1, 2, and 3day lag concentrations because previous studies with PM 2.5 total mass showed little evidence of a strong association with health effects beyond 3 days Ostro et al. 2007;Peng et al. 2009). In presenting results, we computed excess risks for an interquartile range (IQR) increase of each pollutant. Because source impacts likely vary across seasons, we stratified the data set by warm (April-September) and cold (October-March) seasons.
We also employed a constrained (second degree polynomial) distributed lag model using   Gasparrini et al. (2010), to estimate cumu lative effect over multiple days. Because most of the significant associations in individual lag models occurred at lags of 0 through lag 2 days, we summed up the cumulative effects from lag 0 to lag 2 days in the distributed lag model. Because the risk estimates can change considerably depending on the model speci fications (Schwartz 1994a(Schwartz , 1994bTouloumi et al. 2004) in timeseries models, we per formed additional sensitivity analyses: use of alternative degrees of freedom (5 df/year and 12 df/year) for temporal adjustment and use of alternative degrees of freedom (4, 5, and 6 df over the temperature range) for smooth tem perature terms. In addition, we also examined "other mortality" [allcause minus (cardio vascular plus respiratory)] to check consistency of results with causal inference. All analyses were conducted using R (version 2.11.1; R Development Core Team 2010).

Results
Tables 1 and 2 show the descriptive statistics of the ambient pollutants, daily death, and weather for Detroit and Seattle, respectively. In Detroit, PM 2.5 and most components (Al, Fe, K, Ni, S, Si, V, Zn, and EC) had some what higher concentrations in the warm sea son. Na, CO, and NO 2 levels were higher in the cold season. In Seattle, PM 2.5 , Fe, K, Zn, EC, CO, and NO 2 showed higher levels in the cold season, whereas the levels of Al, Na, Ni, S, Si, and V were higher in the warm season.
The risk estimates at individual lag days often showed positive associations at multi ple lag days, especially at lag 0 through lag 2 days. Thus, the results from the distributed lag models exhibited patterns of associations generally similar to those from the individual lag analysis. Therefore, for clarity, we present the results from the distributed lag models here but also describe some aspects of indi vidual lag results [see Supplemental Material (doi:10.1289/ehp.1002613)] that were lost in the distributed lag results. Figures 1-4 summarize the results of dis tributed lag model that produced the cumu lative effects from lag 0 to 2 days. The most prominent contrast between the two cities was the seasonal pattern of the associations-gen erally, in Detroit, some pollutants were more positively associated with mortality outcomes in the warm season, whereas in Seattle, some pollutants showed stronger associations in the cold season.
In Detroit, we found significant positive associations for PM 2.5 , S, and O 3 for allcause and cardiovascular mortality in the warm sea son (Figure 1), suggesting a role of secondary pollutants. Si was significantly negatively asso ciated with allcause mortality. Only O 3 was significantly associated with respiratory mor tality in the warm season. We observed no sig nificant association in the cold season, except for a significant negative association between NO 2 and cardiovascular mortality (Figure 2 In Seattle, in contrast to Detroit, we observed no significant association in the warm season for any mortality category (Figure 3), and multiple pollutants (PM 2.5 , Al, K, S, Si, Zn, EC, CO, and NO 2 ) were sig nificantly positively associated with allcause or cardiovascular mortality in the cold season ( Figure 4). We observed no significant associa tion for respiratory mortality. In the individual lag model results [see Supplemental Material, Figures 3, 4 (doi:10.1289/ehp.1002613)], Al, Fe, Ni, (all at lag 1 day), and V (at lag 2 days) also showed positively significant associations in the cold season. In the warm season, O 3 was significantly positively associated with all cause mortality at lag 2 days. We observed inexplicable significantly negative associations at lag 0 day for PM 2.5 , Al, Fe, K, Na, Si, and CO for allcause mortality in the warm season. Such negative associations might have been caused by overadjustment of the sameday temperature, and most of them disappeared when we removed the sameday temperature adjustment from the model as part of sensitiv ity analyses (data not shown).
The estimated risks were not sensitive to alternative degrees of freedom per year (5 df/year and 12 df/year vs. 8 df/year for the main analysis). We observed no indica tion of systematic difference in risk estimates [see Supplemental Material, Figures 5, 6 (doi:10.1289/ehp.1002613)]. Likewise, we examined the sensitivity of risk estimates to adjustment of degrees of freedom for tempera ture and delayed temperature, using 4, 5, and 6 df, as opposed to 3 df used in the original model, and found that the risk estimates were robust to the change in degrees of freedom to fit temperature effects (data not shown). The fitted temperature terms showed generally posi tive slopes for the sameday temperature and negative slopes for the average of 1 to 3daylag temperature for both cities for allcause and car diovascular deaths (data not shown).
In the analysis of "other cause" of deaths, although we observed a few significant asso ciations (e.g., NO 2 at lag 3 days in the cold season in Seattle, and at lag 1 day in the warm season in Detroit), the air pollutants' associa tions with "other cause" deaths were, in gen eral, much lower than those for cardiovascular and respiratory deaths (data not shown).

Discussion
Our timeseries analysis in two U.S. cities indi cates considerable risk heterogeneity from pol lutant to pollutant. This is consistent with the results of previous studies (Burnett et al. 2000;Mar et al. 2000;Metzger et al. 2004). Laden et al. (2000) examined sourceoriented com binations of PM 2.5 species from the Harvard Six Cities study and found that motor vehicle exhaust and coal combustion were associated with mortality, whereas the soil factor was not. In our allyear analysis, we found no associa tion with Al; however, we observed significant associations with mortality during the cold sea son in Seattle. This is consistent with previous study results of Ostro et al. (2007) for PM 2.5 and mortality in six California counties. We observed excess risks of 1.5-7.4% for cardio vascular and respiratory death in the Poisson regression model based on IQR increase of each pollutant, and fewer, yet larger, excess risks in the distributed lag model. In addition, the cardiovascular and respiratory mortality series exhibited different source and seasonal patterns in each city. The PM components and gaseous pollutants associated with mor tality in Detroit appear most associated with warmseason secondary aerosols and traffic markers. In Seattle, the species associated with mortality include those for coldseason traffic and other combustion sources, such as residual oil and wood burning.
PM components have been examined as independent predictors in several epidemio logical studies on local and regional scales, and the epidemiological evidence linking specific PM components to health risks is mixed (Anderson et al. 2001;Burnett et al. 1997;Chuang et al. 2007;Sarnat et al. 2006). Disparities in findings may result from the diversity of the study locations and their differ ent pollutant concentrations and ratios, health outcomes, or the analytic methods. However, some studies (Laden et al. 2000;Metzger et al. 2004;Ostro et al. 2007Ostro et al. , 2008Tolbert et al. 2000;Zanobetti and Schwartz 2006) found associations between cardiovascular outcomes and EC, and the evidence of increased risk of cardiovascular mortality and EC reported in our studies is consistent with these previous findings. Potassium (K) is generally considered a reasonable marker for biomass combustion, including residential wood burning (Maykut et al. 2003;Watson et al. 2001), which is an important contributor to air pollution in Seattle. Our results showed a significant asso ciation between K and cardiovascular mortal ity, but only in Seattle, despite the fact that the concentration of K was higher in Detroit during the study period.
It is also noteworthy that the annual average PM 2.5 concentration in Detroit (15.1 g/m 3 ) is at the current National Ambient Air Quality Standard (NAAQS) (U.S. EPA 2010a), whereas the annual average PM 2.5 concentration in Seattle (9.7 g/m 3 ) would be in compliance with even the lowest concentra tion limit being considered as a revised annual average for PM 2.5 (11-13 g/m 3 ) (Samet 2010). Because cardiovascular mortality is consider ably more closely associated with PM 2.5 in Seattle than in Detroit, it suggests that PM 2.5 pollution is associated with increases in daily mortality even at currently observed low lev els (a mean of 9.7 g/m 3 ) in Seattle, although further analyses are needed to shed light on the association between a low level of PM 2.5 and health effects. The PM 2.5 components and gaseous pollutants most closely associated with cardiovascular mortality in Seattle were K, S, Si, Zn, EC, NO 2 , and CO. Thus, reduction of cardiovascular mortality risk in Seattle may need to focus on residual oil combustion, the major source of S in this city; biomass burn ing, the major source of K; metal processing industries and incinerators, important sources of Zn; and motor vehicles, the major source of EC, NO 2 , and CO. The PM components and gaseous pollutants associated with daily mortality in Detroit (in the warm season only) were S and O 3 . Thus, reduction of cardiovas cular and respiratory mortality risk in Detroit may need to focus most on traffic emissions and power plants, which are sources of sulfate and O 3 precursors.
The mortality effects exhibited differing seasonal patterns in these two cities. In Detroit, significant associations between mortality and PM 2.5 components and gaseous pollutants appeared only in the warm season, when there is more infiltration of outdoor air into occu pied spaces. Detroit and Seattle showed similar temperature effects: Warm weather had posi tive associations with sameday health effects and negative associations with delayed health effects. However, the mortality showed oppo site seasonal patterns in these cities, although the pollutant concentration variations did not indicate obvious seasonal patterns in either of the two cities. In Seattle, where the annual range of temperature was smaller, we observed significant associations between mortality and air pollutants in the cold season, when some pollutants had higher concentrations, for exam ple K, EC, and CO, and others had lower con centrations, such as Al, S, and Si. Becker et al. (2005) similarly reported seasonal variation in the toxicity of PM, which they justified on the basis that PM in different seasons contains dif ferent elemental compositions that are affected by changes of local environment and weather conditions.
Airconditioned buildings may keep peo ple indoors from being exposed to high lev els of outdoor air pollutants. According to the American Housing Survey (U.S. Census Bureau , 2005, the percentages of housing with air conditioning were 84.4% in Detroit in 2003, and14.7% in Seattle in 2004. Homes in Seattle had a much lower percentage of air conditioning than did those in Detroit, which may have been due to the milder weather in Seattle. In our case, the average temperatures in the warm and cold seasons were 64.4°F and 35.7°F in Detroit and 59.6°F and 45°F in Seattle, respectively. Many air pollutants are highly correlated spatially across a metropolitan area, but they may not permeate equally in all buildings (Wechsler et al. 1989). High percentages of airconditioned homes and buildings may reduce the average ability of air pollutants to reach residents for much of the day, which may diminish the overall population adverse health implications of outdoor air pollution in those cities versus cities without extensive air conditioning. As a result, it is likely that, in cities that have extensive use of air condition ing, the estimated health effects on exposed individuals may be underestimated compared with effects estimated in communities having milder climates and more limited use of air conditioning. The influence of such potential effect modifiers needs further investigation.
It is important to note the limitations of our study. First, the use of a single location in each city to represent the concentrations of PM 2.5 components is likely to lead to ran dom measurement error and the potential for downwardly biased effect estimates that may vary with PM source. Therefore, one of our concerns is a monitor's representation of regional, subregional, and local air pollution exposures for the population in these two met ropolitan areas. Some investigators have also questioned whether the observed associations are plausible given these findings. However, Schwartz et al. (1996) noted that daily mortal ity is calculated over the population, and the relevant exposure measure is the mean of per sonal exposures on that day, which is probably more tightly correlated with central station monitoring of outdoor air pollution than with individual exposures. A previous study (Mage et al. 1999) also indicated that such data from central monitoring sites are usually, although not always, highly correlated with the average of individual exposures in a population. We should also note that an application of such timeseries epidemiological analyses is often to help set ambient standards that will ultimately be monitored at a central site, and from a reg ulatory perspective, the central site ambient pollution levels are most relevant.
Another limitation on the interpretation of our results is that our definitions of car diovascular and respiratory mortalities reduce the specificity of the estimated associations, compared with more narrowly defined cate gories. However, the countervailing benefit of this approach is the increased power to detect associations between PM components and health outcomes. As in many other studies, we had to strike a balance between statistical power and specificity of the outcome. A third limitation is that we conducted this research for only two cities with daily speciation data. More states should follow these two states' lead in retaining daily sample filters to allow distributed lag modeling and to improve the power of future analyses.
Our results add to those previously reported in two important ways. First, in our study, with 3 years of daily PM 2.5 spe ciation data for Detroit and Seattle, we were able to examine the distribution of mortality effects over time, characterize the shortterm health effects of PM 2.5 more thoroughly, and potentially provide insights for better tar geted regulation to reduce source emissions. Second, this study indicates sources of PM air pollution that could be targeted as part of a comprehensive air quality control strat egy. Ambient PM 2.5 is produced by numerous emission sources, and individual PM 2.5 com ponents identified as having associations with health outcomes may be acting as markers for other components or a set of components with similar sources. For example, EC results from combustion of fossil fuels, but also from the combustion of biomass and from industry (HEI 1995). Our results showed that PM 2.5 components and gaseous pollutants associated with cardiovascular mortality in Detroit appear to be associated with secondary aerosols and traffic markers in the warm season, whereas in Seattle, the PM components and gaseous pollutants associated with cardiovascular mor tality include traffic and other combustion sources in the cold season, such as residual oil burning, incinerators, and wood smoke.

Conclusion
In summary, we analyzed data from Detroit and Seattle to assess the associations between daily mortality and PM 2.5 and its elemental components, as well as gaseous pollutants. We found some inconsistent results: although we observed slightly more daily deaths in the cold season than in the warm season in both cities, the daily mortality associations in these cities showed opposite seasonal patterns, whereas the pollutants did not indicate obvious sea sonal concentration patterns in either city. It is generally accepted that weather conditions affect mortality rates, and it is difficult to rule out the possibility that there are other causes of mortality. Our results were quite different depending on the location, so we could not draw a definitive conclusion as to whether this is due to differences in the composition of PM 2.5 and other air pollutants, differences in weather, or some other variables.
Most important, our study found a major contrast in seasonal pattern of association between air pollutants and daily mortality in two major U.S. cities. Along with the previous works of others, our results provide substan tial evidence that a more refined ambient air quality control strategy, based on the effects of PM 2.5 components, can and should be developed. The range of chemical components and sources linked to various health responses supports the hypothesis that the toxicity of various components may vary. Although some sources, such as vehicular traffic and power plants, have been indicated as important con tributors in this and other published analyses, understanding the health effect contributions of various pollutants and their sources still needs further investigation in more cities with daily speciation data.