Impact of Covariate Models on the Assessment of the Air Pollution-Mortality Association in a Single- and Multipollutant Context

Abstract

With the advent of multicity studies, uniform statistical approaches have been developed to examine air pollution-mortality associations across cities. To assess the sensitivity of the air pollution-mortality association to different model specifications in a single and multipollutant context, the authors applied various regression models developed in previous multicity time-series studies of air pollution and mortality to data from Philadelphia, Pennsylvania (May 1992–September 1995). Single-pollutant analyses used daily cardiovascular mortality, fine particulate matter (particles with an aerodynamic diameter ≤2.5 µm; PM_2.5), speciated PM_2.5, and gaseous pollutant data, while multipollutant analyses used source factors identified through principal component analysis. In single-pollutant analyses, risk estimates were relatively consistent across models for most PM_2.5 components and gaseous pollutants. However, risk estimates were inconsistent for ozone in all-year and warm-season analyses. Principal component analysis yielded factors with species associated with traffic, crustal material, residual oil, and coal. Risk estimates for these factors exhibited less sensitivity to alternative regression models compared with single-pollutant models. Factors associated with traffic and crustal material showed consistently positive associations in the warm season, while the coal combustion factor showed consistently positive associations in the cold season. Overall, mortality risk estimates examined using a source-oriented approach yielded more stable and precise risk estimates, compared with single-pollutant analyses.

Many early examinations of the association between air pollution and mortality have relied upon single-city time-series studies. Although these studies formed the basis for the linkage between daily changes in air pollution and mortality, it was unclear if data analyzed from multiple locations with different climates and air pollution mixtures with uniform statistical approaches would result in similar associations (1). As a result, multicity studies were used in which site-specific data on air pollution and health were collected under a common framework, and the data were subsequently analyzed using a uniform statistical approach (1). Although these multicity studies consistently reported associations between air pollution and mortality, each used slightly different modeling approaches (2–4).

Within these multicity studies, the examination of the association between air pollution and mortality primarily occurred through the use of single-pollutant regression models. In these studies, it is difficult to determine the extent to which the results of single-pollutant regression models reflect an independent association between the single pollutant in the model and mortality, or if the single pollutant is a marker for a mix of pollutants from a common emission source (e.g., on-road vehicle exhaust emissions, coal combustion). In order to disentangle the health outcomes (e.g., respiratory and cardiovascular morbidity, mortality) attributed to individual pollutants, investigators have used copollutant or multipollutant models. However, the interpretation of results from these models is complicated because regression models become highly unstable when incorporating pollutants that are highly correlated (5). It is because of this complexity in identifying the health outcomes attributed to a single air pollutant, but more importantly the understanding that populations are exposed to a mixture of air pollutants and not one at a time, that associations between air pollution and health outcomes need to be examined in a multipollutant context (6).

One of the difficulties in using multipollutant models in epidemiologic studies is the sheer number of pollutants that could be included in these models. In an attempt to reduce the number of pollutants examined, epidemiologic studies have used dimension reduction techniques, which reduce the data to a set of key predictors and remove those that do not have explanatory power (5, 7). This approach theoretically allows for the data to then be used in regression models to identify sources or groups of components or pollutants that are most toxic (7). Although attractive, dimension reduction approaches do have some limitations, such as not using the response variable (e.g., mortality) to identify the key set of predictors and grouping correlated predictors without incorporating prior scientific knowledge (5). However, to date, limited emphasis has been placed on these types of studies with the majority of research activities driven largely by single-pollutant, single-source regulatory frameworks (8).

As a new multipollutant era in air pollution research begins, it remains unclear if the various regression models used to examine associations between air pollution and health outcomes produce similar risk estimates when examining individual pollutants and, subsequently, when conducting multipollutant analyses. This is especially important considering the lack of systematic comparisons among research models in the air pollution literature (8). To examine whether alternative approaches to controlling for time and weather influence air pollution-mortality associations, we constructed regression models based on the extent of temporal adjustment and weather covariate approaches used in 6 different multicity time-series studies (9–13). We examined the influence of each model on the association between cardiovascular mortality and air pollution in a single-pollutant context using daily air pollution data (i.e., particulate matter with an aerodynamic diameter ≤2.5 µm (PM_2.5), speciated PM_2.5, ozone, carbon monoxide, sulfur dioxide, and nitrogen dioxide) for Philadelphia, Pennsylvania, from May 1992 through September 1995. In contrast to the speciated PM_2.5 data used by most epidemiologic studies that are collected every third day or every sixth day and obtained from the Speciated Trends Network, in this study we used daily speciation data that allow for a clearer interpretation of lagged associations. Additionally, we examined air pollutant-mortality associations in a multipollutant context using source factors identified through a common dimension-reduction method, principal component analysis (PCA).

MATERIALS AND METHODS

Mortality data

Daily deaths for Philadelphia, defined as Philadelphia County, Pennsylvania, for the period May 12, 1992, through September 30, 1995, were obtained from the National Center for Health Statistics, and counts for all ages were constructed for cardiovascular mortality, defined as International Classification of Diseases, Ninth Revision, codes 390–429. We used SAS, version 9.2, software (SAS Institute, Inc., Cary, North Carolina) to construct a daily time series of cardiovascular deaths.

Air pollution and meteorologic data

Daily 24-hour average PM_2.5 mass and PM_2.5 component data were collected at one site in downtown Philadelphia from May 1992 to September 1995. From March 1992 through March 1993, the monitoring site was located at 59th Street and Greenway Avenue (Aerometric Information Retrieval System monitor number 397140036 PBY) but was moved to 61st Street and Elmwood Avenue (Aerometric Information Retrieval System monitor number 397140136 ELM) (0.8 km from the original site) from March 1993 to September 1995 because of construction. These sites are ∼6 km west southwest of City Hall. Trace elements from PM_2.5 filters were analyzed by energy-dispersive X-ray fluorescence. Data were collected for the trace elements copper, zinc, bromine, lead, iron, silicon, calcium, manganese, nickel, vanadium, selenium, sulfur, and potassium. The coefficient of haze, an optical measurement of particles deposited on a tape filter, was also monitored. Gaseous pollutants data (i.e., carbon monoxide, nitrogen dioxide, sulfur dioxide, and ozone) were retrieved from the Environmental Protection Agency (EPA) Air Quality System database (http://www.epa.gov/ttn/airs/airsaqs/) for all monitors in Philadelphia. Current National Ambient Air Quality Standard averaging times were used for each pollutant (i.e., daily 1-hour maximum for carbon monoxide, sulfur dioxide, and nitrogen dioxide; daily maximum 8-hour average for carbon monoxide and ozone; and daily 24-hour average for PM_2.5 and sulfur dioxide). Because correlations between 1-hour maximum and maximum 8-hour average carbon monoxide and between 1-hour maximum and 24-hour average sulfur dioxide concentrations were high (r = 0.90 and r = 0.88, respectively), we used 1-hour maximum carbon monoxide and 24-hour average sulfur dioxide concentrations throughout this analysis.

We obtained meteorologic data including daily mean temperature, dew point, and relative humidity from the Philadelphia International Airport (EarthInfo, Boulder, Colorado). Additionally, we calculated the metric of apparent temperature as specified in equation 1. Apparent temperature is used to characterize the physiologic experience of heat and high humidity on the body, using the following formula:

(1)

where Ta is temperature and Td is dew point temperature (14, 15).

Statistical analysis

A time-series analysis was conducted by using a priori regression models from multicity epidemiologic studies that examined the association between short-term particulate matter exposure and mortality. These regression models were selected from studies published by research groups that most commonly examine associations between particulate matter and mortality in a multicity setting. We focused on 6 regression models that adjusted for temporal trends and weather covariates using various approaches referenced from this point forward as (and defined in): Air Pollution and Health: a European Approach 2 (APHEA2) (13), California (12), Canada (2), Harvard (10), a model developed by Harvard University with a covariate for apparent temperature (known as “Harvard AT”) (11), and the National Morbidity, Mortality, and Air Pollution Study (NMMAPS) (16). Table 1 depicts the parameters of each regression model.

Table 1.

Extent of Temporal Adjustment and Weather Covariates Included in Each Regression Model, Philadelphia, Pennsylvania, May 1992–September 1995

Modela	Regression Model	Temporal Adjustment	Weather Covariates
APHEA2b	GAM	Trend (k = 40)c	Same-day temperature (k = 10)c
			Lag 1–3 day temperature (k = 10)c
			Relative humidity (k = 10)c
California	GLM	Trend (4 df/year)	Lag 1 day temperature (3 df)
			Lag 1 day relative humidity (3 df)
Canada	GLM	Trend (8 df/year)	Same-day temperature (4 df)
			Barometric pressure (4 df)
Harvard	GLM	Trend (6 df/year)	Same-day temperature (3 df)
			Lag 1 day temperature (3 df)
Harvard AT	GLM	Trend (8 df/year)	Same-day apparent temperature (4 df)
NMMAPS	GLM	Trend (7 df/year)	Same-day temperature (6 df)
			Lag 1–3 day temperature (6 df)
			Same-day dew point (3 df)
			Lag 1–3 day dew point (3 df)

Abbreviations: APHEA2, Air Pollution and Health: a European Approach 2; GAM, generalized additive model; GLM, generalized linear model; Harvard AT, Harvard model with a covariate for apparent temperature; NMMAPS, National Morbidity, Mortality, and Air Pollution Study.

^a An indicator variable for day of week was also included for each model.

^b The APHEA2 GAM model used thin-plate splines as basis functions for penalized regression splines.

^c k = number of basis functions included for each covariate.

Seasonal analyses were conducted by incorporating an indicator variable for warm (April–September) and cold (October–March) seasons in each regression model similar to the approach discussed in Peng et al. (3).

Single-pollutant regression models examined cardiovascular mortality associations with PM_2.5, ozone, sulfur dioxide, nitrogen dioxide, carbon monoxide, and the PM_2.5 components iron, nickel, sulfur, silicon, selenium, vanadium, copper, zinc, and potassium. We reduced the dimensionality of the full data set (i.e., including all PM_2.5 components measured) by performing a PCA with Procrustes rotation, a targeted oblique rotation, the same method as in Laden et al. (17), to identify source factors. Five groupings of tracer elements consisting of 1) lead, bromine, zinc, and copper; 2) calcium, iron, silicon, and manganese; 3) vanadium and nickel; 4) carbon monoxide, nitrogen dioxide, and sulfur dioxide; and 5) sulfur, selenium, and potassium were selected to identify source factors. All trace elements and the gaseous pollutants carbon monoxide, nitrogen dioxide, and sulfur dioxide were used to identify the source factors.

Associations with individual air pollutants or source factors and cardiovascular mortality were examined using the aforementioned regression models in R, version 2.10.1, statistical software (R Development Core Team) with the R package mgcv (version 1.6-2) developed by Wood (18). Risk estimates for single-pollutant and multipollutant analyses were calculated for an interquartile range increase in the individual pollutant or source factor of interest. Analyses were limited to the examination of associations between cardiovascular mortality and the average of air pollution concentrations on the day of death and 1 day prior to death (lag 0–1) because of previous epidemiologic studies indicating an immediate cardiovascular mortality response after PM_2.5 exposure (4, 10, 19).

Multicollinearity

We examined the potential impact of multicollinearity on the air pollutant-cardiovascular mortality associations observed by calculating the concurvity (i.e., the nonlinear analog of multicollinearity) of each individual pollutant and source factor for all models examined. Concurvity was calculated by regressing each individual air pollutant on the same covariates used in each of the regression models as described by Ito et al. (20). The extent of concurvity represents the correlation between the original series and the predicted series from the regression model.

Sensitivity analysis

To quantitatively examine whether there is evidence of residual confounding or model misspecification in the regression models used, we used the method developed by Flanders et al. (21). As described by these authors (21), we included a term in each regression model to account for future air pollution concentrations (i. e., air pollution concentrations 2 days in the future). We then calculated the I statistic, which provides a statistical test for confounding (equation 2):

(2)

where β_f is the estimated slope of the future air pollution term, and σ_f is the estimated standard error.

RESULTS

Study population

In Philadelphia, 17,968 cardiovascular deaths occurred from May 12, 1992, through September 30, 1995, for all ages or ∼14 cardiovascular deaths per day (Table 1). Cardiovascular mortality during this time period accounted for ∼34% of all nonaccidental deaths.

Air pollution data

Air pollution data were available for 94%–100% of the days included in the time series. Table 2 summarizes the air quality statistics for each of the air pollutants included in this analysis. An examination of Pearson correlation coefficients (r) for the entire data set found the strongest correlation between PM_2.5 and sulfur (r = 0.92), with the weakest correlation between PM_2.5 and zinc (Table 3). Seasonal analyses found differential correlations between some pollutant pairs (Web Tables 1 and 2, the first 2 of 6 Web tables and 1 Web figure available at http://aje.oxfordjournals.org/), specifically, PM_2.5 and ozone in the warm season (r = 0.65) and cold season (r = − 0.26) and copper and zinc in the warm season (r = 0.23) and cold season (r = 0.73).

Table 2.

Air Quality, Weather, and Cardiovascular Mortality Summary Statistics for Philadelphia, Pennsylvania, May 1992–September 1995

Pollutanta	Days of Data	Minimum	Median	Mean	IQR	Maximum	Mean (Warm)	Mean (Cold)	SD
PM2.5, µg/m3	1,183	−0.6	14.7	17.3	11.5	72.6	18.7	15.5	9.6
Ozone (8-hour max), ppm	1,234	0.002	0.033	0.036	0.029	0.11	0.049	0.021	0.021
Carbon monoxide (1-hour max), ppm	1,233	0.6	1.8	2.1	1.1	9.5	1.7	2.6	1.1
Sulfur dioxide, ppb	1,237	0.0	6.9	8.4	7.8	45.1	6.0	11.4	6.6
Nitrogen dioxide (1-hour max), ppb	1,233	15.0	46.7	47.4	16.7	146.7	47.9	46.7	14.1
Copper, ng/m3	1,167	−0.8	3.6	4.9	4.1	77.6	4.0	6.1	5.2
Zinc, ng/m3	1,167	0.0	22.8	35.9	28.5	745.3	26.5	47.4	43.7
Iron, ng/m3	1,167	2.5	95.2	2.5	63.0	672.0	107.8	112.1	72.9
Silicon, ng/m3	1,167	−10.0	99.1	126.6	83.9	2,051.5	157.0	89.3	132.8
Nickel, ng/m3	1,167	−0.8	4.8	6.9	6.9	62.0	4.7	9.5	7.9
Vanadium, ng/m3	1,167	−1.6	6.7	8.7	9.1	75.1	7.4	10.3	8.4
Selenium, ng/m3	1,167	−0.5	1.2	1.5	1.5	8.5	1.3	1.7	1.2
Sulfur, µg/m3	1,167	0.1	1.7	2.2	1.7	12.4	2.6	1.6	1.7
Potassium, ng/m3	1,167	2.2	51.2	61.6	36.1	1,570.5	59.1	64.8	62.9
Coefficient of haze	1,237	0.1	0.3	0.4	0.2	1.4	0.3	0.4	0.2
Cardiovascular mortalityb	1,237	3.0	14.0	14.5	5.0	47.0	13.5	15.8	4.5
Temperature, °C	1,237	−17.2	15.6	14.7	16.1	33.3	21.5	6.0	9.7
Apparent temperature, °C	1,237	−11.8	14.4	14.4	19.9	40.9	22.5	4.3	11.5
Dew point, °C	1,237	−22.8	8.9	7.7	16.1	26.1	14.2	−0.57	9.7
Relative humidity, %	1,237	29.0	65.0	66.1	20.0	100.0	66.2	65.9	14.1

Abbreviations: IQR, interquartile range; PM_2.5, particulate matter less than 2.5 µm in aerodynamic diameter; SD, standard deviation.

^a Concentrations for all pollutants represent daily 24-hour average concentrations unless otherwise noted.

^b Deaths per day.

Table 3.

Correlation Coefficients for Air Pollution and Weather Variables for Philadelphia, Pennsylvania, May 1992–September 1995

	PM2.5	Silicon	Sulfur	Potassium	Vanadium	Iron	Nickel	Copper	Zinc	Selenium	Carbon Monoxide	Nitrogen Dioxide	Ozone	Sulfur dioxide	Temperature	Relative Humidity	Dew Point	Coefficient of Haze	Barometric Pressure
PM2.5	1
Silicon	0.44	1
Sulfur	0.92	0.37	1
Potassium	0.35	0.23	0.24	1
Vanadium	0.36	0.14	0.25	0.14	1
Iron	0.48	0.61	0.27	0.33	0.28	1
Nickel	0.21	0.01	0.08	0.12	0.66	0.23	1
Copper	0.23	0.05	0.07	0.29	0.21	0.39	0.40	1
Zinc	0.19	0.06	0.04	0.18	0.29	0.33	0.48	0.51	1
Selenium	0.63	0.21	0.53	0.26	0.43	0.39	0.20	0.17	0.14	1
Carbon monoxide	0.20	0.01	0.02	0.23	0.28	0.42	0.31	0.38	0.34	0.27	1
Nitrogen dioxide	0.48	0.24	0.34	0.27	0.32	0.44	0.22	0.25	0.23	0.39	0.54	1
Ozone	0.43	0.35	0.55	0.06	−0.06	0.00	−0.23	−0.19	−0.22	0.04	−0.35	0.18	1
Sulfur dioxide	0.40	0.12	0.26	0.21	0.52	0.30	0.38	0.19	0.25	0.62	0.41	0.42	−0.19	1
Temperature	0.37	0.36	0.46	0.03	−0.07	0.12	−0.28	−0.15	−0.21	0.02	−0.23	0.17	0.70	−0.28	1
Relative humidity	0.17	−0.03	0.19	−0.02	0.23	0.03	0.25	0.11	0.09	0.13	0.13	−0.01	−0.19	−0.02	0.16	1
Dew point	0.40	0.32	0.49	0.02	0.00	0.11	−0.18	−0.10	−0.16	0.07	−0.17	0.14	0.57	−0.25	0.95	0.47	1
Coefficient of haze	0.50	0.17	0.30	0.26	0.44	0.52	0.35	0.44	0.37	0.49	0.69	0.59	−0.18	0.60	−0.07	0.17	−0.01	1
Barometric pressure	−0.01	−0.09	−0.13	0.07	0.20	0.14	0.26	0.16	0.20	0.03	0.17	0.07	−0.19	0.10	−0.28	−0.23	−0.33	0.15	1

Abbreviation: PM_2.5, particulate matter less than 2.5 µm in aerodynamic diameter.

Associations between cardiovascular mortality and air pollution

Excess cardiovascular mortality risk estimates from all-year analyses for each individual pollutant for the regression models examined are presented in Table 4. Consistent positive associations were observed for all pollutants except carbon monoxide, ozone, and nickel, where, for each pollutant, one model resulted in a negative association (i.e., NMMAPS for carbon monoxide, Harvard for ozone, and Harvard AT for nickel). Additionally for silicon, risk estimates for the APHEA2 and NMMAPS models were less consistent, albeit they remained positive, compared with the other pollutants. In seasonal analysis, ozone, silicon, nickel, and vanadium exhibited at least one negative association across models in the warm season, as did carbon monoxide, copper, and zinc in the cold season (Web Tables 3 and 4). All models produced negative associations for vanadium in the warm season and carbon monoxide and copper in the cold season. These patterns in associations in all-year and seasonal analyses are depicted in Web Figure 1, specifically for the criteria air pollutants ozone, sulfur dioxide, PM_2.5, carbon monoxide, nitrogen dioxide, and the components zinc, silicon, vanadium, and selenium.

Table 4.

Percentage Increase in Cardiovascular Mortality at Lag 0–1 for a Standardized Increase in Daily Pollutant Concentrations in All-Year Analyses, Philadelphia, Pennsylvania, May 1992–September 1995^a

Pollutant	Model
	APHEA2b		Californiac		Canadad		Harvarde		Harvard ATf		NMMAPSg
	%	95% CI	%	95% CI	%	95% CI	%	95% CI	%	95% CI	%	95% CI
PM2.5	1.9	−0.6, 4.4	1.5	−0.9, 3.9	1.8	−0.8, 4.4	1.7	−0.9, 4.3	2.0	−0.6, 4.7	2.0	−0.6, 4.7
Carbon monoxide	0.0	−2.0, 2.0	0.5	−1.4, 2.5	0.3	−1.7, 2.4	0.6	−1.3, 2.7	0.2	−1.8, 2.2	−0.4	−2.4, 1.7
Sulfur dioxide	5.3	1.6, 9.1	4.7	1.1, 8.5	4.7	1.0, 8.6	4.1	0.4, 7.9	4.8	1.1, 8.6	3.8	0.0, 7.9
Nitrogen dioxide	1.0	−0.5, 2.6	1.0	−0.5, 2.5	0.8	−0.8, 2.4	0.8	−0.7, 2.3	0.7	−0.8, 2.2	0.7	−0.9, 2.3
Ozone	1.7	−1.8, 5.3	0.2	−3.4, 3.9	0.5	−3.1, 4.3	−1.6	−5.1, 2.1	1.3	−2.1, 4.9	2.2	−1.8, 6.4
Coefficient of haze	2.1	−0.8, 5.2	1.7	−1.1, 4.5	2.4	−0.6, 5.4	2.2	−0.7, 5.2	2.2	−0.7, 5.2	2.1	−0.9, 5.2
Iron	1.3	−0.7, 3.3	1.9	0.0, 3.8	1.8	−0.2, 3.9	1.9	0.0, 4.0	1.7	−0.2, 3.7	1.0	−1.0, 3.0
Nickel	0.1	−1.9, 2.2	0.2	−1.6, 2.1	0.2	−1.7, 2.1	0.4	−1.5, 2.3	−0.2	−2.0, 3.7	0.2	−1.8, 2.2
Sulfur	2.2	−0.4, 4.8	1.7	−0.8, 4.3	1.8	−0.9, 4.5	1.8	−0.9, 4.6	2.4	−0.3, 5.2	2.8	0.1, 5.6
Silicon	0.5	−1.0, 1.9	1.3	−0.1, 2.8	1.3	−0.2, 2.8	1.3	−0.2, 2.8	1.5	0.0, 3.1	0.1	−1.4, 1.7
Selenium	3.2	0.5, 6.1	3.1	0.4, 5.8	2.9	0.1, 5.7	3.1	0.3, 5.9	3.1	0.3, 5.9	3.0	0.3, 5.9
Vanadium	0.7	−1.9, 3.3	0.6	−1.6, 2.9	0.3	−2.1, 2.9	0.5	−2.0, 3.0	0.0	−2.4, 2.5	0.8	−1.8, 3.4
Copper	0.4	−1.4, 2.3	0.8	−0.9, 2.6	0.6	−1.2, 2.4	0.9	−0.9, 2.8	0.5	−1.3, 2.3	0.4	−1.4, 2.2
Zinc	0.6	−0.8, 2.0	0.9	−0.5, 2.3	0.9	−0.5, 2.4	0.9	−0.6, 2.3	0.7	−0.7, 2.1	0.4	−1.0, 1.9
Potassium	0.9	−0.3, 2.2	0.9	−0.3, 2.1	0.5	−0.7, 1.8	0.9	−0.4, 2.1	0.5	−0.7, 1.8	0.6	−0.6, 1.9

Abbreviations: APHEA2, Air Pollution and Health: a European Approach 2; CI, confidence interval; Harvard AT, Harvard model with a covariate for apparent temperature; NMMAPS, National Morbidity, Mortality, and Air Pollution Study; PM_2.5, particulate matter less than 2.5 µm in aerodynamic diameter.

^a Excess risk estimates are for a 10-µg/m³ increase in 24-hour average PM_2.5 concentrations. Gaseous pollutant excess risk estimates are for a 10-ppb increase in sulfur dioxide and nitrogen dioxide concentrations, a 1-ppm increase in carbon monoxide concentrations, and a 0.020-ppm increase in ozone concentrations. Excess risk estimates for particulate matter components are for an interquartile range increase in 24-hour average concentrations.

^b APHEA2: trend (k = 40), temperature (k = 10), lag 1–3 day temperature (k = 10), and relative humidity (k = 10) (k = number of basis functions included for each covariate).

^c California: trend (4 df/year), lag 1-day temperature (3 df), and lag 1-day relative humidity (3 df).

^d Canada: trend (8 df/year), temperature (4 df), and barometric pressure (4 df).

^e Harvard: trend (6 df/year), temperature (3 df), and lag 1–3 day temperature (3 df).

^f Harvard AT: trend (8 df/year) and apparent temperature (4 df).

^g NMMAPS: trend (7 df/year), temperature (6 df), lag 1–3 day temperature (6 df), dew point (3 df), and lag 1–3 day dew point (3 df).

Figure 1.

Factor loadings for principal component analysis for Philadelphia, Pennsylvania, May 1992–September 1995. Br, bromine; Ca, calcium; CO, carbon monoxide; Cu, copper; Fe, iron; Mn, manganese; Ni, nickel; NO₂, nitrogen dioxide; Pb, lead; S, sulfur; Se, selenium; Si, silicon; SO₂, sulfur dioxide; V, vanadium; Zn, zinc. Correlations highlighted with solid bars represent factor loadings ≥0.70.

Principal component analysis

PCA identified 5 source factors. Figure 1 presents the factor loadings for each PM_2.5 component and gas included in the analysis, with only those components or gases with factor loadings ≥0.70 considered representative of each source. Factor 1 had high loadings for copper, zinc, bromine, and lead; factor 2 had high loadings for iron, silicon, and calcium; factor 3 had high loadings for nickel and vanadium; factor 4 had high loadings for nitrogen dioxide and carbon monoxide; and factor 5 had a high loading for selenium. Based on these factor loadings, these groups of components have often been associated with the following sources: traffic (brake/tire and resuspended road dust) (factor 1); crustal materials (factor 2); residual oil combustion (factor 3); traffic exhaust (factor 4); and coal combustion (factor 5) (22, 23).

Source factors estimated from PCA with Procrustes rotation were not highly correlated over the entire data set, but the traffic exhaust factor was moderately correlated (r = 0.55–0.57) with traffic, crustal, and residual oil combustion factors (Table 5). In the warm season, the crustal factor was more strongly correlated with the traffic exhaust factor (r = 0.70) and weakly correlated with the residual oil combustion factor (r = 0.33). Additionally, in the warm season, the correlation between the traffic and residual oil combustion factors was reduced (r = 0.19). In the cold season analysis, the traffic exhaust factor was highly correlated with the crustal factor (r = 0.76). The coal combustion factor was negatively correlated with all of the source factors in both all-year and seasonal analyses.

Table 5.

Correlation Coefficients for Source Factors, Philadelphia, Pennsylvania, May 1992–September 1995^a

	Factor 1	Factor 2	Factor 3	Factor 4	Factor 5
Factor 1	1
Factor 2	0.33	1
Factor 3	0.50	0.18	1
Factor 4	0.55	0.57	0.57	1
Factor 5	−0.18	−0.09	−0.29	−0.37	1

^a Each factor was found to have high loadings (r ≥ 0.70) for the following components: factor 1: copper, zinc, bromine, lead (traffic); factor 2: iron, silicon, calcium (crustal); factor 3: nickel, vanadium (residual oil combustion); factor 4: nitrogen dioxide, carbon monoxide (traffic exhaust); and factor 5: selenium (coal combustion).

Associations between cardiovascular mortality and source factors

In all-year analyses, consistent positive associations were observed across regression models for each source factor (Table 6). For all source factors, except factor 5 (coal combustion), the NMMAPS model produced the smallest risk estimate; however, there was no evidence of one model consistently producing the largest risk estimate. In warm season analyses, consistent positive associations and, in some cases, statistically significant associations were observed for factors 1 (traffic), 2 (crustal), and 4 (traffic exhaust) (Figure 2) (Web Table 5). Negative associations were observed for all models for factor 3 (residual oil combustion) and the Canada and Harvard models for factor 5 (coal combustion). In cold season analyses, consistent positive associations were observed for factors 3 (residual oil combustion) and 5 (coal combustion) with statistically significant associations for factor 5 when using the AHPEA2, California, and NMMAPS models (Figure 2) (Web Table 6). Negative associations were observed across all models for factor 1 (traffic). Unlike the all-year analyses, one model was not found to produce the highest or lowest risk estimate across source factors in either the warm or cold season analyses.

Table 6.

Percentage Increase in Cardiovascular Mortality at Lag 0–1 for an Interquartile Range Increase in Daily Source Factor Scores in All-Year Analyses, Philadelphia, Pennsylvania, May 1992–September 1995^a

Source Factor	Model
	APHEA2b		Californiac		Canadad		Harvarde		Harvard ATf		NMMAPSg
	%	95% CI	%	95% CI	%	95% CI	%	95% CI	%	95% CI	%	95% CI
Factor 1 (copper, zinc, bromine, lead)	0.3	−1.3, 1.9	0.6	−0.9, 2.2	0.5	−1.1, 2.1	0.7	−0.9, 2.3	0.3	−1.2, 1.9	0.2	−1.4, 1.8
Factor 2 (iron, silicon, calcium)	1.3	−0.9, 3.5	2.1	0.0, 4.3	2.0	−0.3, 4.3	1.9	−0.3, 4.2	2.1	−0.1, 4.4	0.9	−1.3, 3.1
Factor 3 (nickel, vanadium)	0.4	−1.9, 2.9	0.8	−1.3, 2.9	0.8	−1.5, 3.2	0.8		0.4	−1.8, 2.7	0.4	−2.0, 2.8
Factor 4 (nitrogen dioxide, carbon monoxide)	1.2	−1.0, 3.4	1.8	−0.3, 3.9	1.6	−0.7, 3.9	1.6	−0.6, 3.8	1.4	−0.8, 3.6	0.6	−1.6, 2.9
Factor 5 (selenium)	2.4	−0.2, 5.1	1.4	−1.2, 4.2	1.4	−1.3, 4.2	1.5	−1.2, 4.3	2.0	−0.8, 4.8	2.8	0.1, 5.6

Abbreviations: APHEA2, Air Pollution and Health: a European Approach 2; CI, confidence interval; Harvard AT, Harvard model with a covariate for apparent temperature; NMMAPS, National Morbidity, Mortality, and Air Pollution Study.

^a Excess risk estimates are for an interquartile range increase in source factors. Factor 1 = 0.58; factor 2 = 3.5; factor 3 = 1.3; factor 4 = 2.8; and factor 5 = 7.9.

^b APHEA2: trend (k = 40), temperature (k = 10), lag 1–3 day temperature (k = 10), and relative humidity (k = 10) (k = number of basis functions included for each covariate).

^c California: trend (4 df/year), lag 1-day temperature (3 df), and lag 1-day relative humidity (3 df).

^d Canada: trend (8 df/year), temperature (4 df), and barometric pressure (4 df).

^e Harvard: trend (6 df/year), temperature (3 df), and lag 1 day temperature (3 df).

^f Harvard AT: trend (8 df/year) and apparent temperature (4 df).

^g NMMAPS: trend (7 df/year), temperature (6 df), lag 1–3 day temperature (6 df), dew point (3 df), and lag 1–3 day dew point (3 df).

Figure 2.

Percentage increase in cardiovascular mortality at lag 0–1 for an interquartile range increase in daily source factor scores, all-year and seasonal analyses, Philadelphia, Pennsylvania, May 1992–September 1995. APHEA2, Air Pollution and Health: a European Approach 2; Harvard AT, Harvard model with a covariate for apparent temperature; NMMAPS, National Morbidity, Mortality, and Air Pollution Study. Excess risk estimates are for an interquartile range increase in source factors. Factor 1 = 0.58; factor 2 = 3.5; factor 3 = 1.3; factor 4 = 2.8; and factor 5 = 7.9. Each factor was found to have high loadings ≥0.70 for the following components: factor 1: copper, zinc, bromine, lead (traffic); factor 2: iron, silicon, calcium (crustal); factor 3: nickel, vanadium (residual oil combustion); factor 4: nitrogen dioxide, carbon monoxide (traffic exhaust); and factor 5: selenium (coal combustion).

Extent of multicollinearity

Tables 7 and 8 present the concurvity calculated for each individual pollutant and source factor across the regression models examined. In analyses using all-year data, the extent of concurvity was relatively low (R² ≤ 0.57) for all pollutants except ozone, where it ranged from 0.77 to 0.84. Although the difference in concurvity across models was small for ozone, the Harvard AT model was the only model with an R² < 0.80 (i.e., R² = 0.77). An examination of the concurvity for each portion of the regression model (i.e., time only, time + weather covariates, and time + weather covariates + day of week) found additional evidence that ozone is most sensitive to adjusting for time (R² = ∼ 0.70 for all models) and time plus weather covariates (R² = 0.76–0.83) (results not reported). In the warm season, the extent of concurvity across pollutants and models was relatively low (R² ≤ 0.60) for all models and pollutants except ozone. For ozone, all models had an R² ≥ 0.60 except the Harvard AT model (R² = 0.52). During the warm season, the California model primarily exhibited the lowest concurvity for all pollutants except the gaseous pollutants, where the Harvard AT model had the lowest concurvity. In the cold season, the extent of concurvity was similar to that of the warm season with an R² < 0.60 for all pollutants and models except ozone. The California and Harvard AT models exhibited the lowest concurvity across pollutants. However, the low extent of concurvity observed for the Harvard AT model was not limited to the gaseous pollutants as was found in the warm season.

Table 7.

Concurvity (R²) of Individual Pollutants for Each Regression Model in All-Year and Seasonal Analyses, Philadelphia, Pennsylvania, May 1992–September 1995

Pollutant	All-Year						Warm (April–September)						Cold (October–March)
	APHEA2	California	Canada	Harvard	Harvard AT	NMMAPS	APHEA2	California	Canada	Harvard	Harvard AT	NMMAPS	APHEA2	California	Canada	Harvard	Harvard AT	NMMAPS
PM2.5	0.50	0.37	0.47	0.45	0.46	0.49	0.58	0.45	0.53	0.55	0.54	0.56	0.33	0.17	0.22	0.19	0.16	0.27
Carbon monoxide	0.41	0.30	0.36	0.33	0.35	0.41	0.37	0.33	0.31	0.31	0.31	0.37	0.37	0.19	0.28	0.26	0.23	0.33
Sulfur dioxide	0.57	0.45	0.48	0.46	0.46	0.55	0.44	0.40	0.41	0.44	0.35	0.45	0.49	0.33	0.31	0.33	0.29	0.45
Nitrogen dioxide	0.32	0.18	0.29	0.24	0.26	0.36	0.32	0.27	0.28	0.28	0.25	0.36	0.41	0.20	0.32	0.29	0.27	0.39
Ozone	0.81	0.80	0.80	0.80	0.77	0.84	0.64	0.62	0.60	0.62	0.52	0.68	0.71	0.63	0.60	0.60	0.58	0.67
Coefficient of haze	0.52	0.35	0.44	0.42	0.42	0.49	0.59	0.50	0.54	0.54	0.54	0.58	0.49	0.25	0.36	0.34	0.31	0.42
Iron	0.31	0.20	0.30	0.27	0.27	0.32	0.31	0.18	0.26	0.24	0.24	0.29	0.39	0.29	0.38	0.34	0.31	0.40
Nickel	0.45	0.27	0.33	0.32	0.29	0.45	0.31	0.14	0.15	0.14	0.17	0.25	0.44	0.21	0.26	0.26	0.20	0.43
Sulfur	0.54	0.46	0.51	0.51	0.53	0.54	0.55	0.44	0.50	0.51	0.52	0.53	0.29	0.21	0.17	0.16	0.15	0.25
Silicon	0.44	0.33	0.40	0.40	0.39	0.44	0.41	0.30	0.36	0.36	0.34	0.40	0.28	0.28	0.26	0.25	0.24	0.34
Selenium	0.33	0.20	0.27	0.26	0.26	0.31	0.47	0.27	0.42	0.41	0.41	0.43	0.24	0.16	0.15	0.16	0.13	0.23
Vanadium	0.47	0.24	0.37	0.35	0.36	0.45	0.41	0.20	0.32	0.31	0.36	0.41	0.51	0.26	0.32	0.35	0.28	0.44
Copper	0.25	0.18	0.22	0.22	0.20	0.27	0.17	0.14	0.14	0.14	0.14	0.18	0.28	0.15	0.21	0.22	0.16	0.31
Zinc	0.24	0.15	0.21	0.20	0.18	0.26	0.13	0.08	0.08	0.08	0.08	0.16	0.26	0.13	0.16	0.18	0.12	0.27
Potassium	0.16	0.06	0.16	0.12	0.14	0.18	0.23	0.08	0.11	0.11	0.10	0.13	0.29	0.15	0.16	0.17	0.09	0.28

Abbreviations: APHEA2, Air Pollution and Health: a European Approach 2; Harvard AT, Harvard model with a covariate for apparent temperature; NMMAPS, National Morbidity, Mortality, and Air Pollution Study; PM_2.5, particulate matter less than 2.5 µm in aerodynamic diameter.

Table 8.

Concurvity (R²) of Source Factors for Each Regression Model in All-Year and Seasonal Analyses, Philadelphia, Pennsylvania, May 1992–September 1995

Factora	All-Year						Warm (April–September)						Cold (October–March)
	APHEA2	California	Canada	Harvard	Harvard AT	NMMAPS	APHEA2	California	Canada	Harvard	Harvard AT	NMMAPS	APHEA2	California	Canada	Harvard	Harvard AT	NMMAPS
1	0.31	0.18	0.23	0.23	0.20	0.31	0.20	0.12	0.13	0.13	0.14	0.20	0.30	0.12	0.16	0.19	0.11	0.30
2	0.38	0.29	0.37	0.35	0.34	0.39	0.38	0.27	0.34	0.33	0.32	0.38	0.33	0.28	0.33	0.29	0.27	0.33
3	0.51	0.33	0.44	0.42	0.41	0.51	0.26	0.15	0.20	0.19	0.20	0.26	0.51	0.25	0.34	0.36	0.28	0.51
4	0.35	0.23	0.32	0.28	0.29	0.37	0.28	0.23	0.25	0.25	0.23	0.28	0.36	0.19	0.30	0.27	0.23	0.36
5	0.42	0.37	0.42	0.39	0.41	0.42	0.48	0.41	0.44	0.44	0.45	0.48	0.27	0.26	0.25	0.22	0.20	0.27

Abbreviations: APHEA2, Air Pollution and Health: a European Approach 2; Harvard AT, Harvard model with a covariate for apparent temperature; NMMAPS, National Morbidity, Mortality, and Air Pollution Study.

^a Each factor was found to have high loadings (≥0.70) for the following components: factor 1: copper, zinc, bromine, lead (traffic); factor 2: iron, silicon, calcium (crustal); factor 3: nickel, vanadium (residual oil combustion); factor 4: nitrogen dioxide, carbon monoxide (traffic exhaust); factor 5: selenium (coal combustion).

When examining source factors in all-year analyses, we found that the California model exhibited the lowest concurvity (R² = 0.18–0.37), but overall across source factors and models the extent of concurvity was small, not exceeding 0.51. In seasonal analyses, the extent of concurvity was similar to that of the all-year analyses with R² ≤ 0.48 in the warm season and R² ≤ 0.51 in the cold season.

Sensitivity analysis

Examination of the results from the single-pollutant and source factor analyses indicated that ozone risk estimates were most sensitive to the various approaches used to adjust for temporal trends and weather covariates. Therefore, an examination of potential residual confounding was conducted for ozone. Table 9 presents the results of including ozone concentrations 2 days in the future in regression models for all-year and seasonal analyses. These analyses found similar risk estimates and standard errors for each regression model compared with those obtained in the main analysis.

Table 9.

Comparison of Percentage Increase and Standard Error of Ozone Excess Risk Estimates in Regression Models With and Without 2-Day Future Ozone Concentrations and I Statistic for All-Year and Seasonal Analyses, Philadelphia, Pennsylvania, May 1992–September 1995

Model	All-Year					Warm					Cold
	Lag 0–1 Day		Lag 0–1 Day With 2-Day Future Ozone Concentrations			Lag 0–1 Day		Lag 0–1 Day With 2-Day Future Ozone Concentrations			Lag 0–1 Day		Lag 0–1 Day With 2-Day Future Ozone Concentrations
	%	SE	%	SE	I Statistic	%	SE	%	SE	I Statistic	%	SE	%	SE	I Statistic
APHEA2	1.7	0.892	1.7	0.894	−0.666	1.1	0.958	1.2	0.963	−0.588	3.5	1.801	3.4	1.835	−0.352
California	0.2	0.926	0.4	0.935	−0.492	−1.1	0.992	−1.0	1.004	−0.490	5.2	1.721	5.4	1.801	−0.135
Canada	0.5	0.939	0.4	0.945	−0.287	0.2	1.042	0.1	1.048	−0.183	1.3	1.938	1.1	1.971	−0.281
Harvard	−1.6	0.944	−1.4	0.956	−0.396	−2.7	1.035	−2.5	1.047	−0.390	2.8	1.935	2.9	1.974	−0.012
Harvard AT	1.3	0.881	1.3	0.889	−0.353	1.1	0.968	1.1	0.978	−0.308	1.9	1.909	1.7	1.937	−0.149
NMMAPS	2.2	1.021	2.2	1.027	−0.593	1.9	1.140	2.0	1.149	−0.491	2.8	2.006	2.7	2.030	−0.346

Abbreviations: APHEA2, Air Pollution and Health: a European Approach 2; Harvard AT, Harvard model with a covariate for apparent temperature; NMMAPS, National Morbidity, Mortality, and Air Pollution Study; SE, standard error.

DISCUSSION

Multicity time-series studies have consistently provided evidence of associations between PM_2.5 mass and daily mortality (10, 24). Within these studies, a uniform statistical approach was used to examine the relation between short-term PM_2.5 exposure and mortality. As evidence has arisen suggesting that some components may be better predictors of particle toxicity than mass, and as air pollution epidemiology begins to shift to examining multipollutant exposures, the question remains whether a uniform statistical model is appropriate for examining single-pollutant and multipollutant exposures (4, 12).

Of the gaseous pollutants, PM_2.5 components, and sources examined, variability in risk estimates across models was observed only in single-pollutant analyses for ozone. These findings may be attributed to ozone's being more temperature dependent than the other air pollutants and sources examined. Welty and Zeger (25) found that models extensively controlled for the potential confounding effect of weather did not result in qualitatively different PM₁₀ mortality risk estimates. Additionally, Welty and Zeger (25) suggest that other air pollutants that have strong temperature dependence, such as warm-season ozone, may require more extensive control for the potential confounding effect of weather to estimate an association between an air pollutant and health outcome. However, additional control of the potential confounding effects of weather may be problematic as it remains unclear whether temperature terms actually “control” for weather effects or instead act as surrogates for pollutants in the middle range of temperature (20). The sensitivity of ozone-mortality risk estimates across the models examined suggests that the uniform statistical approaches used in multicity ozone-mortality studies to adjust for weather covariates may not be applicable and could potentially explain the regional heterogeneity observed in these studies (26, 27). As a result, a more thorough understanding of the temperature and weather factors that may confound the ozone-mortality relation on a city-to-city basis may be needed.

When examining the magnitude and direction of risk estimates for the other gaseous pollutants, PM_2.5 components, and sources, we observed relative consistency across regression models. In the few instances where the direction of the risk estimate changed (i.e., carbon monoxide and nickel in all-year analyses, silicon and nickel in the warm season, and zinc in the cold season), only in the case of silicon and nickel was there evidence of risk estimates, specifically for the APHEA2 and NNMAPS models, deviating away from the results of the other models.

In an attempt to explain the variability in the magnitude of associations observed across regression models, we examined the concurvity of individual air pollutants and source factors. When individual pollutant mortality risk estimates were examined, those observed to be outliers were often found to have a higher concurvity with a specific model, but overall there was limited evidence of concurvity with R² < 0.60 in all cases except for ozone. The source factors identified were not found to exhibit evidence of concurvity with the various temporal and weather covariate approaches used across models (R² < 0.51) in all-year and seasonal analyses.

Although the R² for ozone exceeded 0.60 for all models in all analyses, except Harvard AT in seasonal analyses, the extent of concurvity across models did not explain the variability in ozone risk estimates. Further quantitative analysis of the ozone results found no evidence of residual confounding in any of the regression models. These results indicate that there is no unmeasured confounder leading to the variability in results across models, and that the covariate structure is more than likely producing the variability. For example, the Harvard and Harvard AT models account for temporal trends similarly, but the Harvard model adjusts only for temperature (i.e., same-day and 1 day lag). The lack of a covariate to account for dew point in the Harvard model may explain the difference in results observed between the Harvard and Harvard AT models.

Previous studies that examined the relation between short-term exposure to PM_2.5 components and mortality have been limited because of the rather sparse component data available as a result of the mostly every-third or every-sixth-day PM_2.5 sampling schedule. Although the air quality data for this analysis are from 1992 to 1995, they are unique in that they consist of daily PM_2.5 component data, which at this time are still not monitored on a daily basis. Because this analysis focuses on a detailed examination of the influence of alternative approaches used to control for the potential confounding effects of time and weather on air pollution-cardiovascular mortality associations, the relative concentrations of the pollutants examined are not as important as having daily air quality data to identify potential air pollution sources and to appropriately examine whether air pollution-mortality associations are influenced by different model parameters.

Of note are several limitations to this study. The first limitation is the use of a centrally located monitor. To try and minimize the exposure misclassification due to the use of one monitor, we restricted the study area to Philadelphia County instead of the entire metropolitan statistical area. We recognize that the temporal and spatial distribution of some PM_2.5 components and gaseous pollutants will not be adequately captured through the monitoring approach used in this study. Additionally, this study encompasses only one city and, as such, we did not expect any of the results to reach statistical significance. However, the limitation of using only one city was minimized by the robust air quality data set available. Although the study focuses on the utility of the various regression models for examining single-pollutant and multipollutant exposures, it is possible that a city with different spatial and temporal pollutant profiles or weather characteristics could lead to different results.

An additional limitation of this study is the identification of potential sources of air pollution using PCA. PCA, like the other current dimension reduction techniques, identifies only a few sources and may omit other important, but more difficult to measure or less understood sources (5). Additionally, it is difficult to generalize results from source-based analyses to other cities because of the location-specific nature of most source signals (5). However, because this study does not focus on the source(s) most strongly associated with cardiovascular mortality but, instead, on whether potential associations between sources and cardiovascular mortality are sensitive to alternative model specifications, the conclusions of this study could be applicable to other cities with different source profiles.

Overall, the results of this study indicate that the various approaches used in previous multicity mortality studies to account for temporal trends and the potential confounding effects of weather result in relatively consistent risk estimates for most air pollutants and sources. In single-pollutant models, gaseous and PM_2.5 component risk estimates were consistent across models, but the various modeling approaches were found to influence the ozone-cardiovascular mortality relation, which could not be explained through an examination of multicollinearity or residual confounding. The difference in risk estimates observed for ozone is more than likely due to the variability in the covariate structure across models. When examining associations between air pollutants and cardiovascular mortality in a multipollutant context, using a source-oriented approach, we found the risk estimates to be more stable and concise and, therefore, less sensitive to model choice than in single-pollutant analyses.