Fine Particulate Matter Constituents Associated with Cardiovascular Hospitalizations and Mortality in New York City

Background Recent time-series studies have indicated that both cardiovascular disease (CVD)mortality and hospitalizations are associated with particulate matter (PM). However, seasonal patterns of PM associations with these outcomes are not consistent, and PM components responsible for these associations have not been determined. We investigated this issue in New York City (NYC), where PM originates from regional and local combustion sources. Objective In this study, we examined the role of particulate matter with aerodynamic diameter ≤ 2.5 μm (PM2.5) and its key chemical components on both CVD hospitalizations and on mortality in NYC. Methods We analyzed daily deaths and emergency hospitalizations for CVDs among persons ≥ 40 years of age for associations with PM2.5, its chemical components, nitrogen dioxide (NO2), carbon monoxide, and sulfur dioxide for the years 2000–2006 using a Poisson time-series model adjusting for temporal and seasonal trends, temperature effects, and day of the week. We estimated excess risks per interquartile-range increases at lags 0 through 3 days for warm (April through September) and cold (October through March) seasons. Results The CVD mortality series exhibit strong seasonal trends, whereas the CVD hospitalization series show a strong day-of-week pattern. These outcome series were not correlated with each other but were individually associated with a number of PM2.5 chemical components from regional and local sources, each with different seasonal patterns and lags. Coal-combustion–related components (e.g., selenium) were associated with CVD mortality in summer and CVD hospitalizations in winter, whereas elemental carbon and NO2 showed associations with these outcomes in both seasons. Conclusion Local combustion sources, including traffic and residual oil burning, may play a year-round role in the associations between air pollution and CVD outcomes, but transported aerosols may explain the seasonal variation in associations shown by PM2.5 mass.

Recent reviews of the health effects of particulate matter (PM) air pollution suggest evidence of adverse effects on cardiovascular disease (CVD) outcomes [e.g., U.S. Environmental Protection Agency (EPA) 2009]. Although several mecha nisms such as oxidative stress and inflammation have been suggested (Brook et al. 2010;Mills et al. 2009), it is not clear if specific chemical constituents of PM play distinct roles in any of these mechanisms. If PM toxicity could be determined based on specific chemical constitu ents or source types that emit such constituents, the regulation of PM may be implemented more effectively. Also, if a given adverse health outcome is associated with specific PM component(s) but not with others, then a spe cific mechanism may be postulated for further consideration by toxicological studies.
Recent multicity timeseries studies of PM have generated important information regard ing the possible role of sources but have also raised some issues that require further investi gation. For example, the most comprehensive U.S.based multicity study of elderly hospital admissions (Bell et al. 2008) found the largest risk estimates for PM mass with aerodynamic diameter ≤ 2.5 µm (PM 2.5 ; collected with a size selective inlet with 50% cutpoint of 2.5 µm aerodynamic diameter) for both cardiovascular and respiratory hospitalizations in northeast ern cities and during the winter season. The most comprehensive U.S.based multicity study (Peng et al. 2005) of allcause mortality (of which nearly half are from cardiovascular causes) similarly found the largest PM 10 (aerodynamic diameter ≤ 10 µm) risk estimates in the north eastern cities but during the summer season. Thus, although both the hospitalization and mortality multicity studies found the strongest associations in northeastern cities, their seasonal patterns of associations are not consistent.
Several studies have examined the annual or seasonal average levels of PM 2.5 chemical components as potential effect modifiers in a secondstage regression in order to try to explain citytocity variation in shortterm PM risk esti mates from the firststage timeseries analysis in individual cities. Lippmann et al. (2006) reported that some of the citytocity varia tion in the PM 10 -mortality risk estimates from the National Morbidity, Mortality, and Air Pollution Study could be explained by the vari ation in nickel (Ni) and vanadium (V) contents in PM 2.5 across cities. However, Dominici et al. (2007) pointed out, in their analysis, that such a finding could be highly influenced by New York City (NYC) data (NYC has a large popu lation and high Ni levels). A recent analysis by Bell et al. (2009) reported that the citytocity variations in PM 2.5 risk for both cardiovascular and respiratory hospitalizations among elderly persons were significantly modified by Ni, V, and elemental carbon (EC), suggesting that PM 2.5 effects were stronger in locations with higher residual oilburning-related and traffic related pollution, highlighting the importance of local combustion sources. Bell et al. (2009) also found that sulfate (SO 4 ) was not a signifi cant predictor of citytocity variation in PM 2.5 risks. In an analysis of 25 U.S. cities, which did not include NYC, Franklin et al. (2008) found that the PM 2.5 -mortality risk estimates were significantly modified by aluminum, arsenic, silicon (Si), SO 4 , and Ni. In their analysis of 27 U.S. cities, Zanobetti et al. (2009) observed that the association between PM 2.5 and CVD hospitalizations was significantly modified when the PM was high in bromine (Br), chro mium (Cr), Ni, and sodium ion (Na + ). Overall, these studies show mixed results, and because PM 2.5 chemical constituents were treated as effect modifiers of PM mass concentration effects on mortality or hospitalizations, they do not necessarily suggest that these individual chemical constituents are associated with the health outcomes on a daytoday basis.
Few studies to date have examined daily PM chemical constituents in relation to health outcomes. In an analysis of Phoenix, Arizona, Mar et al. (2000) examined several key PM 2.5 chemical components in addition to PM 2.5 and gaseous pollutants. They found that EC was significantly associated with CVD mortal ity, but they also found similar associations with nitrogen dioxide (NO 2 ), sulfur dioxide (SO 2 ), carbon monoxide (CO), and PM 2.5 . volume 119 | number 4 | April 2011 • Environmental Health Perspectives Ostro et al.(2007) analyzed data from nine California counties and reported an associa tion between cardiovascular mortality and several chemical components, including EC, organic carbon (OC), nitrate (NO 3 ), iron, and potassium, at various lags. Peng et al. (2009) examined associations between elderly hos pital admissions and seven PM 2.5 chemical constituents that contributed substantive mass fractions to overall mass (EC, OC, SO 4 , NO 3 , Na + , Si, and ammonium ion) in 119 U.S. cit ies and found that EC was significantly associ ated with CVD hospitalizations. These studies either did not examine the PM 2.5 chemical species whose mass contributions were low (e.g., Ni, V) or conduct analyses in the cities where the levels of these species were low.
The mortality and hospitalization studies noted above have examined the same or similar range of categories based on the International Classification of Diseases, 9th Revision [ICD9;World Health Organization (WHO) 1977] and 10th Revision (ICD10; WHO 2007) for CVD, but none of them examined both CVD mortality and hospitalizations in the same anal ysis or compared how these two CVD out comes relate to each other or to other covariates in the regression models. Investigation of the apparent discrepancy in seasonal and regional pattern of associations between PM and these outcomes would benefit from a direct compari son of these CVD outcomes and characteriza tions of their relationships to PM 2.5 chemical constituents and other pollutants.
The objective of our analysis was to exam ine the role of key PM 2.5 chemical constituents on both CVD hospitalizations and mortality in NYC. NYC has one of the nation's highest levels of Ni, presumably from the combus tion of relatively "dirty" residual oils in large buildings for space heating (Peltier et al. 2009), as well as from the ships in the Port of New York ship terminals in Newark and Elizabeth, New Jersey, that burn socalled "bunker fuels," which are heavy residual oils. NYC also has the nation's highest density of traffic. In addi tion, during warm seasons, a major fraction (50-80%) of PM 2.5 in NYC is transported secondary SO 4 and associated chemical con stituents (Lall and Thurston 2006). NYC's very large population also provided sufficient statis tical power to examine the association between the CVD outcomes and PM 2.5 chemical spe ciation data that were collected every third day.

Cardiovascular hospitalization and mortality data.
Hospitalizations and mortality data for five boroughs (Manhattan, Brooklyn, Queens, the Bronx, and Staten Island) of NYC for the years 2000 through 2006 were available at the New York City Department of Health and Mental Hygiene (NYCDOHMH). The hospitalization discharge data were from the Statewide Planning and Research Cooperative System of New York State. The relevant variables available in the electronic discharge abstract for each patient included date of admission, age, sex, and primary ICD9 discharge diagnosis code. The mortality records were obtained from the New York City Office of Vital Statistics. The following variables were used in our study: date of death, age, and the ICD10 codes for underlying cause of death. We extracted hospitalization records whose type of admission was "emergency" or "urgent" only. For both hospitalizations and deaths, we aggre gated daily counts for those ≥ 40 or more years of age for the following categories: hyperten sive diseases (ICD9, code 402; ICD10, code I11); myocardial infarction (ICD9, code 410; ICD10, codes I21-I22); ischemic heart dis ease (ICD9, code 414; ICD10, code I25); dysrhythmias (ICD9, code 427; ICD10, code I48); heart failure (ICD9, code 428; ICD10, code I50); and stroke (ICD9, codes 430-439; ICD10, codes I60-I69). As expected, no mor tality counts were found for dysrhythmias. We combined the daily counts for these specific causes for overall CVD mortality and overall CVD hospitalizations and used them for our main analysis but also examined the specific CVD series separately to check consistency of their results with the overall series.
Air pollution and weather data. Data for PM 2.5 , NO 2 , CO, and SO 2 were retrieved from the U.S. EPA Air Quality System (U.S. EPA 2008). Our previous study in NYC (Ito et al. 2007) found that PM 2.5 and NO 2 show generally high monitortomonitor temporal correlations, whereas CO and SO 2 show gener ally poor monitortomonitor temporal cor relations such that excluding certain monitors was not meaningful. Therefore, all available monitors within the five boroughs of NYC for these pollutants were included, and the aver age of data from multiple monitors was com puted using the 24hr average values. There were only three U.S. EPA PM 2.5 chemical spe ciation monitors; our previous analysis of tem poral correlations of PM 2.5 chemical species at these three monitors (two in the Bronx and one in Queens, several miles apart) found that the extent of monitortomonitor correlation varied across chemical species, with higher cor relations for the species associated with regional pollution and lower correlations for the spe cies associated with local pollution (Ito et al. 2004). However, there was no clear justifica tion for excluding certain monitors given the Table 1. Distribution of daily CVD mortality and hospitalization, weather, and air pollution variables.

All year Warm season Cold season Variable
Mean ± SD IQR a Mean ± SD Mean ± SD Mortality (counts/day) Hypertension 3.5 ± 2.0 3.0 3.3 ± 1.9 3.6 ± 2.0 Myocardial infarction 10.7 ± 3.8 5.0 9.9 ± 3.7 11.5 ± 3.8 Ischemic heart disease 42. closeness of the monitors. Therefore, we also used the average of the three monitors' data for this analysis. The sampling frequency of the chemical speciation data was every third day. This procedure resulted in a reduced sample size for PM components compared with other pollutants. The 24hr average temperature and dew point for LaGuardia Airport were obtained from the National Oceanic and Atmospheric Administration, National Climatic Data Center (2009) Global Summary of the Day database. More than 40 PM 2.5 chemical species were available, but many of them had high percent ages of observations below their detection lim its or poor monitortomonitor correlations. Also, many species were of no interest in terms of source identification or suspected toxicity. Therefore, we selected the following key PM 2.5 chemical components based on past source apportionment studies in NYC (e.g., Ito et al. 2004;Lall and Thurston 2006;Qin et al. 2006) and recent health effects studies discussed in the introductory remarks: EC, OC, Ni, V, zinc (Zn), SO 4 , selenium (Se), Br, NO 3 , and Na + .
Exploratory data analysis. We conducted several exploratory data analyses to describe and characterize the difference in temporal patterns between CVD hospitalizations and mortality and their relationships to weather and air pollution.
To characterize relative variance contribu tions from seasonal trends, dayofweek, and random components to the overall time series, we conducted spectral analysis of CVD hospital izations and mortality. We used modified Daniel smoothers to smooth periodograms, applying several spans of smoothing over frequency, as suggested by Venables and Ripley (2002).
To characterize bivariate temporal rela tionships among weather, air pollution, and health outcome variables, we computed the crosscorrelation function (CCF; correlation with lags) for key variables considered. These crosscorrelations can indicate the sequence of temporal fluctuations (i.e., which variable leads the other in time). The correlation between two time series can be strongly influenced by shared trends, seasonal cycles, and dayofweek patterns that may be confounding. Therefore, in order to remove the influence of these tem poral patterns and to focus on the shortterm relationships between the variables, each time series was first prefiltered in a generalized lin ear model using a natural cubic spline smooth ing function with 8 degrees of freedom per year and a dayofweek variable. A CCF was then computed using pairwise complete obser vations between the lagged residual time series. We chose 8 degrees of freedom per year to be consistent with the extent of adjustment for temporal trends in recent multicity timeseries analyses (e.g., Bell et al. 2009;Samoli et al. 2008). Because relationships among these variables can change across seasons, the CCF was computed in a series of 12 threemonth blocks centered on each month of the year and pooled for the entire 7year study period.
Regression models. We estimated percent excess risk (%ER) for air pollutants using a Poisson timeseries model, adjusting for temporal trends and seasonal cycles, immedi ate and delayed temperature effects, and day of the week. The extent of lagged days con sidered for air pollution and weather variables were based mainly on recent timeseries stud ies and the exploratory analysis of these data described above. We used natural cubic splines of days to adjust for potentially confounding temporal trends and seasonal cycles using 8 degrees of freedom per year in our base model. To adjust for immediate and delayed nonlin ear temperature effects, our base model also included natural cubic splines of the same day and the average of past 1 through 3day lagged temperature with 3 degrees of freedom over the range for each term. Temperature and dew point were highly correlated (r = 0.93) in this data set, so we did not include them simultaneously. We did, however, compute apparent temperature from temperature and dew point (Steadman 1979) and included it in a separate alternative model for sensitivity analysis. Risks were estimated at lags 0 through 3 days for interquartilerange (IQR) increases of pollutants for allyear data (Table 1)

Results
CVD mortality and hospitalization time series exhibited markedly different characteristics. Figure 1 shows time series and corresponding  power spectra of these time series. CVD mor tality showed a strong seasonal pattern, with broad winter peaks that vary from year to year, apparently reflecting influenza epidem ics. A dayofweek pattern was not apparent in the CVD mortality power spectra, which showed a major variance contribution in the frequency range corresponding to seasonal cycles. In contrast, CVD hospitalizations showed substantial variance contribution from dayofweek cycle frequencies (0.14/day or 7day cycle, and its harmonics), which is stronger than that from seasonal cycles. CVD hospitalizations were lower on Saturday and Sunday, with hospitalizations about 60% higher on Monday than on Sunday. Likewise, the power spectra for specific CVD hospital izations all showed strong dayofweek peaks that were stronger than those for seasonal cycles (data not shown). Table 1 shows distributional characteris tics of CVD outcome, weather, and air pol lution variables. The largest fraction of CVD mortality was from ischemic heart disease, followed by myocardial infarction. CVD hos pitalizations were more evenly divided by the specific categories. Ni, V, Zn, and SO 2 , all likely associated with residual oil burning, showed substantially higher levels in cold seasons than in warm seasons. Correlations among air pollutants by season are shown in Supplemental Material, Table 1 (doi:10.1289/ ehp.1002667).
The following results from CCF and regression analysis necessarily involve multiple testing because of multiple lags, pollutants, and outcomes examined. Therefore, to guard against overinterpretation of chance findings, we focus not on the statistical significance of individual results but instead on identifying consistent patterns and contrasts of results across the lags, pollutants, and outcomes.
In the CCF analysis of key variables, after removing seasonal cycles and dayofweek pat terns, temperature showed different patterns of associations with CVD mortality versus CVD hospitalizations, respectively (Figure 2A,B). With CVD mortality, temperature showed positive associations in summer and lagged negative associations in winter, consistent with fairly wellknown mortality impacts of tem perature reported in the past literature (e.g., Curriero et al. 2002). In contrast, CVD hos pitalizations showed immediate and positive associations with temperature in the cold sea son, clearly not heat events, and biologically implausible. There was also a suggestion of lagged negative associations between tempera ture and CVD hospitalizations in the warm season and the beginning of the cold season. We repeated the analysis using apparent tem perature rather than temperature (data not shown), but the pattern and magnitude of associations were essentially unchanged.
The CCF relationships between PM 2.5 and CVD mortality and hospitalizations also showed contrasting patterns ( Figure 2C,D). PM 2.5 was positively associated with CVD mortality in the warm season with lags 0 and 1 day, whereas it was positively associated with CVD hospitalizations in the cold season, most consistently at lag 0 day. The CCF for chemical components of PM 2.5 versus CVD outcomes showed less distinct patterns (data not shown), likely in part due to the smaller sample size (only every third day). The CCF relationship between temperature and PM 2.5 chemical components showed generally posi tive associations on the same day throughout the year and delayed negative associations in cold seasons, but the relationship between tem perature and Ni showed negative associations with longer lags and no positive associations in the cold season [Supplemental Material, Figure 1d (doi:10.1289/ehp.1002667)].
We found no correlation between CVD hospitalizations and CVD mortality, except for as weak negative correlation in the cold season ( Figure 2E). This is consistent with the contrasting relationships between these CVD outcomes and temperature or PM 2.5 . The CCFs between specific CVD hospitalizations Figure 2. Cross-correlation function (CCF) between X = temperature, Y = CVD mortality (A); X = temperature, Y = CVD hospitalizations (B); X = PM 2.5 , Y = CVD mortality (C); X = PM 2.5 , Y = CVD hospitalizations (D); and X = CVD hospitalizations, Y = CVD mortality (E): a correlation below the centerline (lag 0) indicates that X leads Y, and a correlation above the centerline indicates that Y leads X. Black bars are positive correlations, and white bars are negative correlations (correlations whose absolute values are < 0.1 are not shown). The lengths of reference correlations (0.25 and 0.50) are also shown. CCFs were computed after removing seasonal trends and day-of-week pattern (see "Results"). Abbreviations: X, predictor; Y, outcome. and corresponding CVD mortality (e.g., CCF between heart failure hospitalizations and heart failure mortality) likewise did not show associations (data not shown). Figure 3 shows air pollution-CVD mor tality regression results (see Table 1 for IQRs used to compute %ER). PM 2.5 was associated with CVD mortality in the warm season at lag 0 [%ER = 2.0%; 95% confidence inter val (CI), 0.7-3.3; per 10 µg/m 3 ] and lag 1 day (1.9%; 95% CI, 0.8-3.1) and in the cold season at lag 1 day (1.0%; 95% CI, -0.1 to 2.2). OC and SO 4 , both of which are asso ciated with a major mass fraction of PM 2.5 during the warm season (i.e., secondary aero sols), also showed significant associations with CVD mortality in the warm season at lags 0 and 1 days, with estimated %ERs similar to those for PM 2.5 per IQR increase. Se, primar ily associated with transported coal emissions in the eastern United States, and also associ ated with transported SO 4 , showed a strong association with CVD mortality at lag 1 day. EC was significantly associated with CVD mortality at lag 1 day in the warm season and was nearly significant in the cold season also at lag day 1. Br (perhaps associated with traffic) was strongly associated with CVD mortality at lag 0 in the warm season. Ni, V, and Zn (the elements most associated with residual oil burning in NYC) showed a similar pattern of associations with CVD mortality, with stron ger associations in the cold season at lag 3 days. Na + (considered to be a signature element for sea salt) also showed nearly a significant posi tive association at lag 0 day in both warm and cold seasons. Among the gaseous pollutants, NO 2 showed lag 1day positive associations with CVD mortality in both warm and cold seasons. There are peculiar negative (signifi cant or nearly significant) associations for a number of pollutants at lag 0 day in the cold season (PM 2.5 , OC, Ni, Zn, Si, Se, NO 3 , NO 2 , SO 2 , and CO). These pollutants were not negatively correlated with CVD mortality in the CCF analysis (data not shown), and the negative associations may be induced by the adjustment for temperature. Not surprisingly, among the specific CVD mortality categories, mortality for ischemic heart disease showed a similar pattern of association as that for total CVD mortality (data not shown).

Lag days
The relationships between air pollutants and CVD hospitalizations showed a pattern of association different than that for CVD mortality (Figure 4). The associations between air pollution and CVD hospitalizations, in contrast to CVD mortality, were generally stronger in the cold season than in the warm season. The exceptions were EC, NO 2 , SO 2 , and CO, which showed similar magnitude and lag structure of associations in both sea sons. PM 2.5 exhibited the strongest association at lag 0 day in the cold season (%ER = 1.1; 95% CI, 0.2-2.0), and most of the PM 2.5 chemical components that showed significant or near significant associations (OC, SO 4 , Ni, Zn, Si, Se, and Br) with CVD hospitaliza tions showed the strongest associations at lag 0 day in the cold season. Of the specific CVD hospitalizations, those for hypertensive dis ease, ischemic heart disease, dysrhythmias, and stroke showed patterns of associations similar to that for total CVD hospitalizations (data not shown). Seasonal patterns of associations, lags, source types, and %ER are summarized for CVD mortality and hospitalizations in Supplemental Material, These results were not sensitive to the alter native model in which we used apparent tem perature, rather than temperature, to adjust for weather effects. Supplemental Material, Figure  2 (doi:10.1289/ehp.1002667), shows com parisons of risk estimates for CVD mortality and hospitalizations using temperature and apparent temperature for the warm and cold Figure 3. %ER for CVD mortality per IQR increase in air pollutant, adjusted for temporal trend, day-of-week, same-day, and delayed temperature effects. seasons. There appears to be no systematic difference in the risk estimates obtained from these models.

Discussion
We found that adult CVD mortality and emergency CVD hospitalizations time series, when constructed with consistent ICD catego ries, were not correlated over time. The CVD mortality series showed a strong influence of seasonal cycles and influenza, whereas CVD hospitalizations were strongly influenced by a dayofweek pattern, with little seasonal pat tern. Their relationships with temperature, as examined in our seasonal CCF analysis, were also different. The strong influence of the dayofweek pattern seen in CVD hospitaliza tions (lower counts on weekends) may imply that they are more strongly influenced by behavioral modification factors than by CVD mortality (e.g., being able to delay seeking treatment over the weekend). Also, it is possi ble that coding practice and conditions under which a "primary cause" is assigned to these outcomes may be different. For example, if a person with a chronic ischemic heart condition dies with an additional complication of influ enzarelated pneumonia, the primary cause of death may still be recorded as a heartdiseaserelated cause, whereas when a person with an acute cardiovascular condition is admitted to a hospital through the emergency department, the primary cause may be less ambiguous. If such differences contribute to the apparent lack of correlation between these two CVD outcomes, then the subgroups at risk for PM effects within these CVD outcomes need not share the same risk characteristics or even the same mechanism(s) of PM effects. However, in this analysis, we did not specifically examine whether some of those who were hospitalized died later, or whether a subgroup's hospitaliza tions were associated with air pollution. If the fraction of each subgroup is relatively small in the overall CVD hospitalizations, then we may not see significant lagged correlation between CVD hospitalizations and mortality, whether or not they are associated with air pollution in the same causal pathway.
CVD mortality's immediate association with high temperature in the warm season and lagged association with cold temperature in the cold season found in our analysis are consistent with expectation and the past lit erature. However, the lack of heat effects on CVD hospitalizations in warm seasons and the presence of positive associations in cold seasons (i.e., the warmer the temperature, the more CVD hospitalizations) found in this study is somewhat surprising, although it is in part consistent with "a protective effect of cold temperature" (i.e., the lack of the left side of the Ushape in the temperature-outcome rela tionship) that was reported by Schwartz et al. (2004) who analyzed temperature and elderly CVD hospitalizations in 12 U.S. cities. Also, in another analysis of CVD hospitalizations in NYC, whereas the overall CVD hospitaliza tions were positively associated with high tem perature, the hospitalizations for CVD, heart failure, and hypertension showed negative risk estimates (Lin et al. 2009). Thus, there may be heterogeneity in temperature effects within subcategories of CVD hospitalizations.
In these analyses, CVD mortality was more strongly associated with PM 2.5 in the warm season than in the cold season, whereas CVD hospitalization was more strongly asso ciated with PM 2.5 in the cold season than in the warm season. These findings are consis tent with the results of the analysis in the U.S. Northeast of (allcause) mortality and PM 10 (Peng et al. 2005) and the analysis of CVD hospitalizations and PM 2.5 (Bell et al. 2008). The lag structures of PM 2.5 associations for CVD mortality (lags 0 and 1 day) and CVD hospitalizations (lag 0 day) are consistent with those found in past multicity studies for mor tality (Peng et al. 2005) and CVD hospitaliza tions (Bell et al. 2008;Dominici et al. 2006;Peng et al. 2009).
The patterns of lagged associations between individual PM 2.5 chemical components and CVD mortality were not always consistent with those for PM 2.5 mass concentrations. It is not surprising that SO 4 and OC were asso ciated with CVD mortality on the same lag days (0 and 1) as PM 2.5 because both SO 4 and OC contribute major mass fractions to PM 2.5 . Si, a marker of soil whose associated species together contribute significant amounts of par ticle mass, also showed the same lag structure of associations. However, other PM chemical components with smaller mass contributions showed different lagged associations with CVD mortality compared with PM 2.5 . Ni, V, and Zn (i.e., the species associated with residual oil burning in NYC) all showed delayed associa tions (largest estimates at lag 3 days) with CVD mortality. However, this may reflect the rela tionship between temperature and these species in the cold season [see Supplemental Material, Figure 1 (doi:10.1289/ehp.1002667)], which was negative with long lags and had no posi tive associations on the same day. Therefore, caution is required in interpreting these lagged associations between PM components and health outcomes, because they may have varying relationships with other covariates in the regression model. Also, the 3day lagged association between Ni and CVD mortality is not necessarily consistent with Ni's effect modification of 1day lagged PM 10 -mortality associations suggested in Lippmann et al.'s (2006) result.
We observed the associations between CVD hospitalizations with air pollutants pri marily in the cold season, except for EC, Na + , NO 2 , SO 2 , and CO, which were associated in both seasons. Because most of the PM 2.5 chemical components and gaseous pollut ants examined similarly yielded significant or nearly significant associations at lag 0 day, it is difficult to draw inferences on which chemical components or sources may be more impor tant for the PM 2.5 association with CVD hos pitalizations in this data set. These results are, in part, consistent with the significant effect modifiers (EC and Ni) of PM 2.5 effects on elderly CVD hospitalizations reported in the Bell et al. (2009) analysis or those (Br and Ni) reported in the Zanobetti et al. (2009) analy sis but show a broader spectrum of PM 2.5 source components contributing to effects.
Given the lack of correlation between CVD hospitalizations and CVD mortality over time, we did not expect "coherence" in specific PM 2.5 chemical components' association with these two CVD outcomes. Of the pollutants examined, EC and NO 2 showed the most con sistent associations with these CVD outcomes, in that these pollutants showed the same extent of associations in both warm and cold seasons with the same lags (lag 1 day for CVD mortal ity and lag 0 day for CVD hospitalizations). Both EC and NO 2 are often considered to be "signature" tracers for traffic sources. However, EC may be emitted from other fuel combus tions, including oil boiler emissions (Schauer 2003). Likewise, NO 2 is also emitted from a variety of combustion sources. In fact, in the New York City Community Air Survey con ducted during the winter of 2008 and 2009, which measured PM 2.5 , its chemical compo nents, and gaseous pollutants at 150 locations in NYC, landuse regression analysis found that both traffic and building density (i.e., proxy for building space heating) were signifi cant predictors of EC and NO 2 in this city (NYCDOHMH 2009). This does not directly imply that these pollutants per se are respon sible for the health effects, but they may be important indicators of sources relevant to the PM 2.5 effects on CVD outcomes.
The relative significance of the associations of the PM components with CVD outcomes could have been influenced by their corre sponding exposure misclassification errors that include the error associated with the locations of monitors relative to the population. An investigation related to this issue involving the same three speciation monitors (Ito et al. 2004) found that the PM 2.5 chemical spe cies associated with regional secondary aero sols (e.g., SO 4 and NO 3 ) tend to have higher monitortomonitor temporal correlations than do the species associated with more local combustion sources (e.g., EC and Ni). Thus, the interpretation of our study findings is lim ited by the uncertainty regarding the influence of relative exposure error across the species on the observed associations.

Conclusion
We found that CVD hospitalizations and CVD mortality time series are not correlated, but they are each independently associated with a number of PM 2.5 chemical compo nents, including both regional secondary aero sols and local combustion sources. EC and NO 2 showed the most consistent associa tions with these CVD outcomes throughout the year, but the PM components associated with the regional transported aerosols from coal combustion (Se and SO 4 ) exhibited the seasonspecific association pattern most simi lar to that of PM 2.5 mass. We conclude that local combustion sources, including traffic and residual oil burning, may be important sources affecting CVD adverse effects in NYC.