Modeling of Personal Exposures to Ambient Air Toxics in Camden, New Jersey: An Evaluation Study

Abstract

This study presents the Individual Based Exposure Modeling (IBEM) application of MENTOR (Modeling ENvironment for TOtal Risk studies) in a hot spot area, where there are concentrated local sources on the scale of tens to hundreds of meters, and an urban reference area in Camden, NJ, to characterize the ambient concentrations and personal exposures to benzene and toluene from local ambient sources. The emission-based ambient concentrations in the two neighborhoods were first estimated through atmospheric dispersion modeling. Subsequently, the calculated and measured ambient concentrations of benzene and toluene were separately combined with the time-activity diaries completed by the subjects as inputs to MENTOR/IBEM for estimating personal exposures resulting from ambient sources. The modeling results were then compared with the actual personal measurements collected from over 100 individuals in the field study to identify the gaps in modeling personal exposures in a hot spot. The modeled ambient concentrations of benzene and toluene were generally in agreement with the neighborhood measurements within a factor of 2, but were underestimated at the high-end percentiles. The major local contributors to the benzene ambient levels are from mobile sources, whereas mobile and stationary (point and area) sources contribute to the toluene ambient levels in the study area. This finding can be used as guidance for developing better air toxic emission inventories for characterizing, through modeling, the ambient concentrations of air toxics in the study area. The estimated percentage contributions of personal exposures from ambient sources were generally higher in the hot spot area than the urban reference area in Camden, NJ, for benzene and toluene. This finding demonstrates the hot spot characteristics of stronger local ambient source impacts on personal exposures. Non-ambient sources were also found as significant contributors to personal exposures to benzene and toluene for the population studied.

Introduction

Characterization of potential health risks resulting from exposures to ambient air toxics requires assessing the impact of ambient sources on personal exposures. This is a challenging task because the variables needed to assess individual and population exposures are not always available. Regulatory agencies typically maintain information of ambient levels of air toxics for major urban areas, including source emission data, modeling results, and measured concentrations. However, the concentrations of ambient air toxics in hot spot areas, where there are concentrated local sources on the scale of tens to hundreds of meters, are often much higher than those obtained from regulatory monitors, and individuals living in hot spot areas can experience elevated exposures to multiple agents. In addition, personal exposures to air toxics that people actually breathe are affected by the variety of indoor and outdoor microenvironments they pass through during their daily activities. Therefore, personal exposures to air toxics and the associated health risks in hot spot areas can be underestimated by the information acquired by state and federal agencies. Thus, characterization of ambient and personal exposure levels to air toxics and development of adequate ambient and personal exposure models for air toxics in hot spot areas are critical for conducting accurate risk assessment and the development of effective mitigation strategies by regulatory agencies.

For assessing population exposures to air toxics, the U.S. Environmental Protection Agency (EPA) has jointly utilized modeling approaches from an air quality dispersion model (the Assessment System for Population Exposure Nationwide [ASPEN]¹) and an inhalation exposure model (the Hazardous Air Pollutant Exposure Model [HAPEM]²) in the National Air Toxic Assessment (NATA) studies.^2,3 Rosenbaum et al.¹ compared the ASPEN results with available ambient air toxic monitored concentrations on a national scale. Several local-scale air toxic modeling studies^4–7 evaluated the performances of modeling results for characterizing ambient levels of air toxics using measurements collected from urban areas. However, systematic evaluation of local-scale modeling in characterizing both ambient and exposure concentrations of air toxics has not been conducted previously. This manuscript describes the modeling efforts and model evaluation results for characterizing ambient concentrations and personal exposures to two air toxics—benzene and toluene—using the field measurement data collected from a hot spot for air pollution. Benzene and toluene were selected to represent local impacts of mobile and stationary sources, respectively, on ambient and exposure levels in the study area.

The objective of this study was to assess the impact of local ambient sources of air toxics on personal exposures and air quality of each neighborhood using the anthropocentric or Person Oriented Modeling (POM) approach⁸ for exposure characterization provided by MENTOR (Modeling ENvironment for TOtal Risk studies).⁸ In this study, MENTOR was implemented using Individual-Based Exposure Modeling (IBEM) for characterizing ambient concentrations and personal exposures resulting from ambient sources to benzene and toluene for each of the subjects sampled in the field study. The ambient concentration and personal exposure modeling results were compared with the corresponding actual measured outdoor and personal data for model evaluation. The ambient modeling results were used to identify the limitations of the source emissions data as well as the adequacy of applying standard atmospheric dispersion models such as the Industrial Source Complex Short Term Version 3 (ISCST3) and the AMS/EPA Regulatory Model (AERMOD) in hot spot areas. The personal exposure modeling results were used to quantify contributions of ambient sources to personal exposures as well as identify data needs or gaps (e.g., indoor or occupational source contributions). The impact of neighborhood emissions from specific known ambient sources (i.e., point, area, mobile on-road, and mobile non-road) on the ambient levels and personal exposures of benzene and toluene was also examined.

Methods

Ambient and Exposure Measurements of Air Toxics

Through the “Personal and Ambient Exposures to Air Toxics in Camden, New Jersey” project, ambient and personal measurements for several air toxics (including fine particulate matter [PM_2.5], volatile organic compounds, carbonyls, and polycyclic aromatic hydrocarbons) were collected from 108 subjects living in the Village of Waterfront South (WFS) and Copewood/Davis Streets (CDS) during the study period from 2004 to 2006 (see Figure 1 for the geographical location of the sampled homes). The WFS neighborhood has been considered as a hot spot of air toxics for over 20 yr.^9,10 CDS is located to the east of WFS and has few local point sources; however, it is affected by major highways and local roads. Each personal exposure measurement had the companion time-location diaries and activity questionnaires, which provided the details of how each subject spent their time in different microenvironments during the personal monitoring period. The local ambient and personal monitoring results comprised a unique dataset that was used in the modeling analyses to improve our understanding of personal exposure levels of air toxics and the impact of ambient sources on these levels in the study area.

Figure 1.

Geographical location of the subjects' homes and the ambient sampling sites in the neighborhoods of WFS and CDS in Camden, NJ.

Source-to-Exposure Individual-Based Modeling

MENTOR can be applied to either populations/subpopulations of interest or to specific individuals.⁸ For the IBEM implementation of MENTOR, a generalized seven-step approach has been developed (the details are described in Georgopoulos and Lioy⁸). It accounts for the processes associated with determining exposures/doses from source to dose and has provided the general modeling framework for the IBEM application of MENTOR to a hot spot area of Camden, NJ. The following describes in detail the implementation of a screening version of MENTOR/IBEM for this study.

Estimation of Local Ambient Pollutant Levels

Estimation of Background Levels of Air Pollutants

Ambient levels of certain toxic air pollutants, including benzene, have been shown through model simulations to have a significant background or regional component (i.e., long-range transport).¹¹ Thus, to accurately estimate total ambient concentrations of benzene, the Camden County specific background level of 0.43 μg/m³ (i.e., an annual average) obtained from the 1999 NATA study³ was used. The NATA county-specific background concentration estimates were derived from available monitored data or extrapolated from measurements in other locations based on a population regression.¹² Alternatively, more spatially and temporally refined background concentrations of air toxics can be derived from grid-based photochemical model simulations.^7,11,13,14 Because toluene is a reactive gas, the background level of toluene was assumed to be zero.

The two standard atmospheric dispersion models (i.e., ISCST3 and AERMOD) were used to estimate the 24-hr average ambient concentrations of benzene and toluene. The estimated concentrations for each compound in each neighborhood were matched both in space and time with the actual ambient measurements collected from the field study. Model receptors were defined as two ambient stationary monitors selected for the overall study (WFS and CDS). The estimated and measured ambient concentrations were used individually to estimate the personal exposures of the study participants resulting from outdoor sources. The following subsections describe how the model inputs of source emissions and meteorological data were prepared for the atmospheric dispersion modeling.

Preprocessing of Emission Inventories

All source categories (including point, area, mobile on-road, and mobile non-road emissions) data were extracted from the 2002 National Emissions Inventory (NEI-2002)¹⁵ for benzene and toluene for an emission modeling domain that included both Philadelphia County, PA, and Camden County, NJ. The emission modeling domain was determined by joining the “buffer zone” of census tracts that are located within a 25-km radius from each sampled home of the field study. Based on the extracted NEI-2002 data, mobile (on-road and non-road) sources account for 76 and 71% of the total emissions for benzene and toluene, respectively, whereas the rest were due to stationary (point and area) sources (24 and 29% for benzene and toluene, respectively).

Because mobile and area source emissions are only reported as county totals in the NEI-2002 database, fine-scale (such as census-tract level) spatial allocation of these emissions are essential for the dispersion modeling to simulate impacts on the neighborhood scale. This is especially important for hot spot analyses. All four source emission categories are also reported as annual totals in the NEI-2002 database. To account for the seasonal and day-type variations, temporal allocation of the annual emissions is required for the adjustments as hourly emission rates. The Emissions Modeling System for Hazardous Air Pollutants Version 3 (EMS-HAP v3.0)¹⁶ was used to process the NEI-2002 emission data for both spatial and temporal allocations. The county-level emissions were spatially apportioned into the census tracts within each county by using spatial allocation factors. These factors were derived from data on the geographic distributions of various spatial surrogates, which have similar geographic variations to the emissions from various source categories. For example, mobile on-road emissions were allocated to census tracts using the geographic distribution of roadway miles. For point sources in the study area, the parameters of source characterization (i.e., emission rate, stack height, diameter, temperature, and exit velocity) were updated with local information provided by New Jersey Department of Environmental Protection (NJDEP).

Several previous studies^4,6,7,17,18 recommended treating mobile sources on major highways as line sources rather than area sources distributed at census tracts for better characterization of the impact of mobile on-road emissions on the nearby receptor location. Therefore, in addition to the spatial allocation of the county-level mobile on-road emissions into census tracts (the standard approach), these emissions were allocated to roadway links; this was determined to improve the mobile source emission attributions on the neighborhood scale for model application. Sensitivity simulations of the ISCST3 modeling were conducted by using the two different mobile on-road emission inputs to examine their impact on the predictions of ambient concentrations.

Preprocessing of Local Meteorology Information

The meteorological variables needed for atmospheric dispersion modeling include wind speed, wind direction, stability category, and mixing height. The hourly surface meteorological data were obtained from the Philadelphia International Airport, which is located approximately 10 mi from the study area. These data were processed through the meteorological preprocessors of PCRAMMET¹⁹ and AERMET²⁰ to generate the meteorological inputs for ISCST3 and AERMOD, respectively. For the study area of Camden, NJ, southwest or south winds dominated in summer, whereas northwest or west winds were observed in winter during the study period from 2004 to 2006 (see wind rose plots in Figure 2). The mixing heights in winter were generally lower than summer. The meteorological conditions are fairly uniform across the metropolitan area except during frontal passages. Therefore, these meteorological data should be generally representative for the study areas.

Figure 2.

Wind rose plots summarizing the wind speeds and wind directions for the sampling days for (a) summer and (b) winter during the period of April 2004 to July 2006. The data of wind speeds and wind directions were obtained from the meteorological station of Philadelphia International Airport (WBAN no. 13739). The direction of winds shown is the direction from which the wind is blowing.

Microenvironmental and Personal Exposure Modeling

The predicted ambient concentrations of the two air toxics in each neighborhood were used as ambient inputs to estimate the profile of microenvironmental concentrations that each subject experienced during personal monitoring. The personal exposure concentrations were then calculated as the time-weighted averages of microenvironmental concentrations experienced by each subject, where the “time weights” are the durations of time spent by the subject in each microenvironment. The ambient measurements made at the two stationary monitors (WFS and CDS) were used as another option as ambient inputs to calculate the microenvironmental and personal exposure concentrations. For this option, it was assumed that the ambient measurements collected at the two stationary monitors could be used to represent the ambient levels of benzene and toluene for the two corresponding neighborhoods. The results were then compared with the actual exposures measured for each subject on the matched sampling day to examine the percentage contributions of personal exposures resulting from ambient sources.

The subject-specific time-activity diaries collected in the field study provided information about where the subjects spent their time among five microenvironments (home, office/school, other indoors, outdoors (neighborhood and out of neighborhood), and in-vehicle) in a sequence of hourly exposure events during the 24-hr personal monitoring period. These diaries provided the critical information on where and when the exposures occurred and were used in estimating the personal exposures for each subject. A database for time-activity information based on these subject-specific diaries was developed as input variables for MENTOR in a format compatible with the Consolidated Human Activity Database (CHAD),²¹ the default database. For quality assurance of the collected time-activity diaries, 34 subjects were asked to carry the GPS device (i.e., the GeoLogger) to track their movements during the personal monitoring periods. The time-location logs provided by these subjects were generally consistent with the GeoLogger data.

Microenvironmental Modeling Approach

The microenvironmental module of MENTOR was used to derive temporal profiles of five microenvironmental (home, office/school, other indoors, outdoor, and in-vehicle) concentrations of benzene and toluene based on ambient concentration estimates. Different modeling algorithm options for calculating microenvironmental concentrations are available in MENTOR.⁸ In this study, the linear factor approach was used to estimate the microenvironmental concentrations on the basis of the following equation:

$$ME(m,r,t)=\mathit{\text{ADD}}(m)+[\mathit{\text{PROX}}(m)][\mathit{\text{PEN}}(m)][\mathit{\text{AMB}}(r,t)]$$ (1)

where ME(m,r,t) is the concentration of microenvironment m in receptor location r at time t, ADD(m) is the additive factor for microenvironment m, PROX(m) is the proximity factor for microenvironment m, PEN(m) is the penetration factor for microenvironment m, and AMB(r,t) is the ambient concentration in receptor location r at time t.

ADD accounts for the contribution of emission sources from indoors. This term was set to zero, because the focus of this analysis was to estimate the contribution from ambient sources of air toxics to personal exposures. PROX accounts for the relationship between the outdoor concentration in the vicinity of the microenvironment and the ambient concentration at the receptor location represented by AMB(r,t). This factor was set to 1 because the ambient stationary monitors were quite close to the receptor locations of the subjects' residences (see Figure 1). PEN represents the ratio of the microenvironmental concentration to the ambient concentration in the immediate vicinity of the microenvironment, when the microenvironment contains no indoor sources. On the basis of the compilation of extensive literature reviews for the indoor/outdoor (I/O) ratios at different microenvironments, PEN for various air toxics have been developed as part of the 1999 NATA study.³ PEN of benzene and toluene were extracted from the 1999 NATA database to calculate the microenvironmental concentrations. PEN of benzene for the microenvironments of home, office, other indoors, outdoor, and in-vehicle were 0.88, 0.63, 0.78, 1, and 0.96, respectively, whereas PEN of toluene were 0.95, 0.82, 0.81, 1, and 0.88, respectively.

Personal Exposure Modeling

In this study, MENTOR/IBEM used the dispersion model predictions and the ambient measurements as the options of ambient inputs for calculating microenvironmental concentrations and the subsequent personal exposures. Furthermore, the option of using ambient measurements would be reflective of the “actual” ambient contributions to exposures and would yield information on ambient contributions to personal exposures not accounted for in the emission inventories. Personal exposure concentrations were estimated by combining the information from time-activity diaries with the estimated microenvironmental concentrations. Specifically, the time and location recorded for each exposure event in the time-activity diary were first used to extract the corresponding microenvironmental concentrations each subject experienced during the personal monitoring period. The time series of hourly personal exposure concentrations was then generated for each subject when the atmospheric dispersion modeling results were used as ambient inputs. Finally, the average of the hourly personal exposure concentrations during the personal monitoring period was calculated and compared with the corresponding 24-hr integrated personal air measurement obtained for each subject, which included contributions from both ambient and non-ambient sources, to estimate the contribution from ambient sources to personal exposures. When the ambient measurement obtained on the same day as the personal exposure monitoring was used as the ambient input, the time-weighted average of personal exposure concentration was calculated instead of the hourly time series, because the measured ambient concentration was determined by the 24-hr integrated sample.

Model Performance Evaluation

To evaluate the performance of model estimates for ambient and personal concentrations of the two air toxics, linear regression analysis was conducted to assess the agreement between the model estimates and actual measurements. If the model estimates consistently agree with the actual measurements, the slope of the regression line should be close to 1 if the sources were ambient and could be fully taken into account by the emissions input to the dispersion model. In addition to linear regression, three model performance statistics (i.e., the mean difference between the modeled concentrations and the measurements [mean error], the root mean square error [RMSE], and the fractional bias [FB]) were calculated. The mean error provides the sign of the bias. RMSE provides a measure of the deviations from the 1:1 relationship between the modeled concentrations and the measurements and also preserves the scale of the measurements. FB is the statistic recommended by EPA as a screening metric for model performance evaluation.²² FB is calculated as follows:

$$\mathit{\text{FB}}=2\left[\frac{\overline{\mathit{\text{OB}}}-\overline{\mathit{\text{PR}}}}{\overline{\mathit{\text{OB}}}+\overline{\mathit{\text{PR}}}}\right]$$ (2)

where $\overline{\mathit{\text{OB}}}$ and $\overline{\mathit{\text{PR}}}$ refer to the averages of measurements and modeled concentrations, respectively. EPA guidance suggests using the highest 25 concentrations that are unmatched in space and time for calculating FB. However, a more stringent test of using all of the matched pairs of measurements and modeled concentrations was used in this study for assessing the model performance. If the FB values are in the range of +0.67 (underprediction) to −0.67 (overprediction), the model predictions are equivalently in agreement with the measurements within the factor of 2 acceptance criterion recommended by EPA.²²

Results

The source-to-exposure modeling results are presented in three parts: (1) dispersion model predictions of ambient concentrations, (2) personal exposure predictions using MENTOR/IBEM, and (3) ambient source contributions for ambient levels and personal exposures to benzene and toluene.

Ambient Concentrations of Air Toxics

The use of the two different sets of spatially allocated mobile on-road emissions (census-tract-based vs. road-link-based) for dispersion modeling was examined first by comparing the corresponding ISCST3 predictions of benzene and toluene ambient concentrations with ambient measurements. As shown in the time-series profiles of Figure 3, the finer spatial allocation of mobile on-road sources to roadway links increased the predicted hourly ambient concentrations of benzene and toluene due to mobile on-road sources within WFS. The increase was a factor of 2 during the evening rush hours within WFS when compared with the census-tract-based modeling results. Furthermore, the 24-hr averages of the road-link-based modeling results are closer to the ambient measurement results collected at the WFS stationary site than the census-tract-based estimates. This is reasonable because the WFS stationary site was immediately adjacent to busy roads with truck traffic (Broadway and Ferry Street). Because of its proximity to major roadways, the use of road-link-based spatial allocation for mobile on-road emissions improves the model performance of ISCST3 predictions at the WFS location. However, both of the census-tract- and road-link-based approaches have similar ISCST3 model performance at CDS (see Figure 4). This result is probably due to the CDS monitoring site not being immediately adjacent to major roadways. The road-link-based approach was selected for the spatial allocation of the county-level mobile on-road emissions for all of the ISCST3 and AERMOD model estimations on the basis of the model performance. For the area and mobile non-road sources, the census-tract-based approach was still used for the fine-scale allocation of county-level emissions.

Figure 3.

Comparison of the time-series profiles of hourly ISCST3 predictions of (a and b) benzene and (c and d) toluene ambient concentrations from mobile on-road sources at WFS using the census-tract-based spatial allocation and the road-link-based spatial allocation for mobile on-road emission sources, as well as the corresponding comparison between the 24-hr averages of ISCST3 predictions from all source and ambient measurements for the sampling date of July 15, 2006.

Figure 4.

Comparison of the time-series profiles of hourly ISCST3 predictions of (a and b) benzene and (c and d) toluene ambient concentrations from mobile on-road sources at CDS using the census-tract-based spatial allocation and the road-link-based spatial allocation for mobile on-road emission sources, as well as the corresponding comparison between the 24-hr averages of ISCST3 predictions from all source and the ambient measurements for the sampling date of July 15, 2006.

The linear regression results for assessing the agreement between model predictions and ambient measurements are shown in Table 1. For benzene, the slopes of regressing model predictions (ISCST3 and AERMOD) versus ambient measurements were 0.92 and 0.93, respectively. Furthermore, both of the 95% confidence intervals of the two slopes covered positive values, indicating statistical significance for the slopes. For toluene, the slopes of ISCST3 and AERMOD regressed against ambient measurements were 0.94 and 1, respectively. The width of the corresponding 95% confidence intervals for toluene were narrower than those for benzene, suggesting that the performances of the two dispersion models (ISCST3 and AERMOD) were better for toluene than benzene. However, the R² values were relatively small for the ISCST3 and AERMOD predictions (e.g., 0.08 and 0.22 for ISCST3 predictions of benzene and toluene, respectively), indicating that missing emission sources or missing source contributions (local traffic patterns) can contribute to daily variability.

Table 1.

Linear regression coefficients (slope and intercept), the associated 95% confidence intervals, and R² obtained from regressing the ambient benzene and toluene measurements against the corresponding atmospheric dispersion modeling results (ISCST3 and AERMOD) at the two stationary sites.

Chemical	Model	Regression Coefficients	95% Confidence Interval	R2
Benzene	ISCST3	Slope: 0.92	(0.40, 1.45)	0.08
		Intercept: 0.08	(0.02, 0.15)
	AERMOD	Slope: 0.93	(0.40, 1.49)	0.07
		Intercept: 0.10	(0.04, 0.17)
Toluene	ISCST3	Slope: 0.94	(0.68, 1.21)	0.22
		Intercept: −0.01	(−0.11, 0.09)
	AERMOD	Slope: 1.00	(0.72, 1.28)	0.23
		Intercept: −0.01	(−0.1, 0.1)

Notes: The total numbers of matched pairs between ambient measurements and model predictions are 139 and 171 for benzene and toluene, respectively.

Table 2 shows the comparison of the three model performance statistics (mean error, RMSE, and FB) of the ISCST3 and AERMOD modeling results for benzene and toluene in the two neighborhoods (WFS and CDS). In general, the dispersion modeling results underestimated the ambient measurements of benzene and toluene, because the mean errors were all negative for the ISCST3 and AERMOD results. For benzene, the performances of the dispersion modeling results were better in WFS than CDS, on the basis of smaller values of RMSE and FB. For toluene, the opposite trend was observed, in which the performance of the dispersion modeling results was better in CDS than WFS; however, the performances of the ISCST3 and AERMOD predictions for benzene and toluene were in agreement with the ambient measurements within the acceptable criterion of the factor of 2 (i.e., the FB values were in the range of −0.67 to +0.67), except the ISCST3 predictions of benzene in CDS (FB = 0.74).

Table 2.

Mean concentrations (μg/m³) of ambient benzene and toluene measurements at the two stationary sites (CDS and WFS) in Camden, NJ, and the corresponding model performance statistics (mean error, RMSE, and FB^a) of the ISCST3 and AERMOD predictions.

Chemical		Metric	Location of Stationary Site
			CDS	WFS
Benzene	Measurement	Mean	2.03	1.81
	ISCST3	Mean error	−1.10	−0.70
		RMSE	3.62	2.27
		FB	0.74	0.47
	AERMOD	Mean error	−1.14	−0.76
		RMSE	3.64	2.30
		FB	0.67	0.53
Toluene	Measurement	Mean	2.14	3.71
	ISCST3	Mean error	−0.22	−1.2
		RMSE	1.42	4.42
		FB	0.11	0.39
	AERMOD	Mean error	−0.32	−1.37
		RMSE	1.47	4.42
		FB	−0.10	0.45

Notes:

^aDefinition is given in the text.

On the basis of the model evaluation results presented in Tables 1 and 2, the ISCST3 and AERMOD predictions have comparable model performances in characterizing ambient levels of benzene and toluene in the study area. Thereinafter, only the AERMOD results are presented in the following modeling analysis. To examine the agreement between the ambient measurements and the modeled concentrations for the whole distribution, the scatterplot of the ambient measurements versus the matched AERMOD predictions is presented in Figure 5. In general, the AERMOD predictions were in agreement with the ambient measurements of benzene and toluene within the factor of 2, except for the high-end concentrations (i.e., concentrations >5 μg/m³) for which the modeled results significantly underestimated the pollutant levels. Furthermore, the magnitudes of the high-end toluene measurements in WFS were higher than those in CDS. This is mainly due to the presence of local stationary sources close to the WFS site, which can contribute significantly to the daily variation in neighborhood-scale levels. In contrast, the CDS monitoring site is relatively far away from these sources (see Figure 1).

Figure 5.

Scatterplots of the ambient measurements vs. the AERMOD predictions at the two stationary sites (CDS and WFS) for (a and b) benzene and (c and d) toluene for all of the sampling days from 2004 to 2006.

Personal Concentrations of Air Toxics

Personal exposures to benzene and toluene can be affected by ambient and non-ambient (indoor, occupational) sources, as well as environmental tobacco smoke exposure. The focus of this paper is to assess the impact of ambient sources on personal exposures. To achieve this goal, the percentage contributions of ambient sources to personal concentrations were calculated by dividing the MENTOR predictions with the personal measurements for each subject at WFS and CDS (i.e., the MENTOR/personal measurement ratios). Figure 6 shows the box plots of the calculated MENTOR/personal measurement ratios for benzene and toluene, in which two different options of ambient inputs were used for the MENTOR predictions—ambient measurements and AERMOD predictions. In general, the MENTOR/personal measurement ratios were higher in WFS than in CDS for benzene and toluene, indicating higher percentage contributions of ambient sources to personal exposures for the WFS subjects than for the CDS subjects. Furthermore, the MENTOR predictions based on the ambient measurements generally showed higher MENTOR/personal measurement ratios than the MENTOR predictions based on dispersion modeling results. The exception was the personal exposures to toluene for the CDS subjects, in which the two sets of MENTOR predictions had similar results. This was mainly because the best model performance of the dispersion model predictions was observed for the ambient toluene concentrations predicted for CDS.

Figure 6.

Box plots of the MENTOR/personal measurement ratios for (a) benzene and (b) toluene estimated for the subjects at the two neighborhoods (CDS and WFS) of Camden, NJ, where two different options of ambient inputs were used for the MENTOR predictions: ambient measurements (noted as field in the box plot) and AERMOD predictions.

Based on the MENTOR predictions with the inputs of ambient measurements, the median level of percentage contributions from local ambient sources to the personal benzene exposures in the WFS neighborhood was estimated around 61%, whereas 43% was estimated as the median level for the CDS neighborhood (see Figure 6). For toluene, a lower contribution from ambient to personal exposure was observed, with 55 and 30% for the WFS and CDS neighborhoods, respectively. This may be due to the stronger influences of indoor sources of toluene in both neighborhoods; however, the toluene personal exposures did appear to have occupational contributions for CDS residents.²³

Linear regression analysis was conducted to examine the extent of association between the MENTOR predictions and the personal measurements. Because the MENTOR predictions only accounted for the contributions from ambient sources, the regression slopes would be significantly less than 1. However, cross-comparison of linear regression results would provide further insights on the adequacy of using different ambient input options for personal exposure modeling. For benzene (see Table 3), the MENTOR predictions based on the inputs of ambient measurements (slope = 0.51, R² = 0.27) have a much larger slope and R² than those predictions based on dispersion modeling results (slopes = 0.20, R² = 0.0040 for the AERMOD predictions). These results suggest that the dispersion modeling results for benzene might not be adequate for exposure modeling and further improvement is needed.

Table 3.

Linear regression coefficients (slope and intercept), the associated 95% confidence intervals, and R² obtained from regressing the personal benzene and toluene measurements against the corresponding MENTOR exposure modeling results with two different options of ambient inputs: ambient measurements (noted as field in the table) and AERMOD predictions.

Chemical	Model	Regression Coefficients	95% Confidence Interval	R2
Benzene	MENTOR (AERMOD)	Slope: 0.20	(−0.17, 0.58)	0.0040
		Intercept: 0.41	(0.36, 0.46)
	MENTOR (field)	Slope: 0.51	(0.41, 0.61)	0.27
		Intercept: 0.37	(0.33, 0.40)
	Ambient measurements	Slope: 0.51	(0.41, 0.61)	0.27
		Intercept: 0.34	(0.30, 0.38)
Toluene	MENTOR (AERMOD)	Slope: 0.46	(0.24, 0.68)	0.05
		Intercept: 0.62	(0.55, 0.69)
	MENTOR (field)	Slope: 0.35	(0.24, 0.47)	0.10
		Intercept: 0.64	(0.59, 0.70)
	Ambient measurements	Slope: 0.36	(0.24, 0.47)	0.10
		Intercept: 0.64	(0.58, 0.69)

Notes: Regression analyses were also performed for personal measurements vs. ambient measurements of benzene and toluene, respectively.

For toluene (see Table 3), the MENTOR predictions based on the inputs of actual ambient measurements (slope = 0.35, R² = 0.10) have closer regression slopes and R² values to the predictions based on dispersion modeling results (slopes = 0.46, R² = 0.05 for the AERMOD predictions). This was due to the following two factors: (1) weaker association between ambient and personal measurements (R² = 0.10) of toluene than benzene (R² = 0.27), in which non-ambient sources have larger impacts on toluene personal levels than benzene; and (2) better model performance of dispersion model predictions of ambient concentrations for toluene.

Analysis of Ambient Source Contributions to the Predictions of Ambient Concentrations and Personal Exposures

To assess the impact of local ambient sources of air toxics on ambient levels at the two neighborhoods (WFS and CDS), the contributions from ambient sources were apportioned and categorized by five emission groups (i.e., background, point, area, mobile on-road, and mobile non-road) using 24-hr average AERMOD predictions for benzene and toluene. Table 4 presents the source contribution analysis results for benzene and toluene for the sampling dates during the 2005 winter and 2006 summer seasons. The results were typical of data and simulations obtained on other sampling dates.

Table 4.

Estimated average percentage contributions of ambient sources categorized by five groups (background, point, area, mobile non-road, and mobile on-road) to the AERMOD predictions of benzene and toluene ambient concentrations at the location of two stationary sites for the sampling dates during the 2005 winter and 2006 summer seasons.

Chemical	Location	Background (%)	Point (%)	Area (%)	Mobile Non-Road (%)	Mobile On-Road (%)
Benzene	CDS	46.4	1.4	10.8	15.2	26.2
	WFS	43.8	1.3	13.4	15.8	25.7
Toluene	CDS	0.0	4.0	47.5	14.8	33.7
	WFS	0.0	23.5	28.8	13.5	34.3

For benzene, the average background contributions were approximately 44% and 46% for WFS and CDS on the basis of the AERMOD predictions. These background contribution estimates were relatively high because of the assumption of using the annual average county-specific NATA background concentration. However, more spatially and temporally refined background concentration estimates can be obtained through the grid-based photochemical model applications as described in the methods section. For the local source impacts, the major contributors were from local mobile sources, where the mobile on-road sources appeared dominant in both CDS and WFS. The contributions from stationary (point and area) sources were relatively small for benzene in both locations. For toluene, there was no contribution from background to local ambient levels because it is a more reactive chemical compared with benzene. The local stationary and mobile sources had approximately equal contributions to the local ambient levels of toluene in both CDS and WFS. However, the estimated contribution from point sources was much higher at the WFS site than the CDS site, which was clearly due to the WFS site being closer to the industrial facilities than the CDS site (see Figure 1), a reason for selecting WFS as a hot spot for this study.

Table 5 presents the source contribution analysis results for personal exposures to benzene and toluene for the subjects sampled during the 2005 winter and 2006 summer seasons. The estimated non-ambient source contributions to personal exposures were significantly higher in CDS (benzene: 51.5%, toluene: 60.4%) than in WFS (benzene: 36.1%, toluene: 25%). On the other hand, the ambient source contributions to personal exposures were higher in WFS than in CDS. The estimated local mobile source contributions to personal exposures were higher in WFS than in CDS for both benzene and toluene, which was not clearly observed for the ambient levels of these two pollutants. The estimated local stationary source contributions to toluene exposures were much higher in WFS than in CDS, a point that appears to be related to the hot spot characteristics of the study design.

Table 5.

Estimated average percentage contributions of ambient sources (background, point, area, mobile non-road, and mobile on-road) and non-ambient sources to the personal measurements of benzene and toluene for the subjects sampled during the 2005 winter and 2006 summer seasons.

Chemical	Neighborhood	Ambient Sources					Non-Ambient Sources (%)
		Background (%)	Point (%)	Area (%)	Mobile Non-Road (%)	Mobile On-Road (%)
Benzene	CDS	22.5	0.7	5.2	7.4	12.7	51.5
	WFS	28.0	0.8	8.6	10.1	16.4	36.1
Toluene	CDS	0	1.6	18.8	5.9	13.3	60.4
	WFS	0	17.6	21.6	10.1	25.7	25.0

Discussion and Conclusions

Ambient Concentrations of Air Toxics

The general factor of 2 agreement between modeled ambient concentrations and the measurements, which is similar to the results reported by the studies of Pratt et al.⁴ and Wheeler et al.,⁵ indicated that the two standard dispersion models (ISCST3 and AERMOD) can be used to characterize the baseline ambient benzene and toluene levels in a hot spot area of Camden, NJ. However, the modeled ambient concentrations of benzene and toluene were significantly underestimated at the high-end concentrations (i.e., concentrations >5 μg/m³).

The performance of dispersion modeling has rarely been evaluated by comparing the predictions with air toxics measurements, especially over an averaging time period shorter than 1 yr.⁴ The dispersion modeling analyses presented in this study with augmented emissions factors for mobile on-road sources (i.e., line estimates) provided new insights into the model performance through comparison with ambient measurements over a 24-hr time period, which is shorter than in the previous studies,^4,5 in which 48-hr and monthly/annual averaging time periods were used, respectively. In addition to the factor of 2 agreement between the model results and the measured values, the median levels of benzene and toluene ambient concentrations predicted by ISCST3 and AERMOD were also comparable to the medians of the measurements collected at the two corresponding stationary sites. Furthermore, both of the ISCST3 and AERMOD dispersion modeling results have similar model performances.

The limitations of local-scale dispersion models may be their inputs (i.e., emissions, source characteristics, and meteorology) as well as model formulation.^11,13,24,25 Sax and Isakov²⁴ have shown that uncertainties associated with the estimates of emission rates were identified as the largest source of uncertainty in model inputs for simulating local-scale air toxic impacts. The meteorology, spatial, and temporal allocations of sources and source release parameters also contributed to the uncertainty in air toxics concentrations.^13,26 The underestimated modeled concentrations of benzene and toluene at the high-end percentiles may be due to the inadequate inputs (i.e., NEI-2002) for local small emission sources and lack of local traffic emission data (e.g., heavy truck traffic and local congestion). The effect of photochemistry was not taken into account by the two dispersion models, such that the overestimated toluene concentrations were also observed at the low-end (i.e. ∼0 to 20th) percentiles. Furthermore, the simple assumption of steady-state meteorology (i.e., model formulation) in dispersion models might also affect their model performances in complex urban topography and can be improved through parameterized or computational fluid dynamics-based street canyon models,^27,28 which can be considered for future studies.

The analysis of ambient source contributions provided insights about the major contributors to the ambient levels of benzene and toluene in the study area. The outcomes of these analyses can be used as guidance for improving the air toxic emission inventory so that the ambient concentrations of air toxics on the neighborhood scale can be better characterized. For instance, the source contribution analysis results showed that the major contributors to the benzene ambient levels were from mobile sources, whereas mobile and stationary (point and area) sources contributed equally to the toluene ambient levels in the study area. This finding is also supported by the correlation analysis results of the spatial variability study,²⁹ in which high correlation was found between ambient benzene and methyl tertiary butyl ether (MTBE; a surrogate of mobile sources) measurements and relatively lower correlation between toluene and MTBE. Therefore, future efforts to improve the air toxic emission inventory should focus on mobile sources for benzene and mobile and stationary sources for toluene in the study area of Camden, NJ.

Personal Concentrations of Air Toxics

Several previous studies^30–33 have shown that personal and indoor concentrations of benzene and toluene were higher than outdoor concentrations, suggesting that indoor sources dominated personal and indoor concentrations of benzene and toluene. However, the Toxic Exposure Assessment Columbia/Harvard (TEACH)–New York City study³⁴ reported more consistent indoor, outdoor, and personal concentrations for benzene and toluene, suggesting ambient concentrations as the driving force for personal exposures to benzene and toluene for the community of high school students studied in New York City. On the basis of MENTOR predictions of personal exposures and the comparison with personal measurements in the current study, it was suggested that ambient and indoor sources were important contributors to personal exposures of benzene and toluene, whereas non-ambient sources had stronger impacts on personal exposures of toluene than benzene, especially in CDS. These observations are consistent with source information at the two locations. The results from this study and previous studies suggest that relationships between emission sources and personal exposures vary by location and personal characteristics. Therefore, characterization of ambient levels and personal exposures, particularly in hot spots, are necessary to define the contribution of ambient air pollution on personal exposures.

The effect of location factor (WFS vs. CDS) was revealed in percentage contributions to personal exposures from the local ambient concentrations (stationary site data), in which the percentages in WFS were generally higher than those in CDS for benzene and toluene. This finding demonstrates the hot spot characteristics of the WFS neighborhood, where the impacts of local ambient sources on personal exposures of benzene and toluene were larger for WFS subjects than for CDS subjects. The measured toluene personal concentrations were higher on weekdays than those on weekends, which might be caused by the weekly operating cycle of the facilities associated with stationary sources of toluene. The MENTOR predictions of toluene personal exposures also showed the same trend as in the personal measurements. There was no day-of-the-week effect observed for the personal measurements and the MENTOR predictions of benzene.

In calculating the microenvironmental concentrations, an assumption was made that the ambient stationary measurements, made on the same day as personal monitoring, were representative estimates of the ambient concentrations of benzene and toluene in the locations each subject visited during personal monitoring. Based on the large spatial variability of ambient benzene and toluene concentrations in these two neighborhoods, as described in detail by Zhu et al.,²⁹ this assumption is not valid. Therefore, the actual ambient contributions to the personal exposures of benzene and toluene could be underestimated for residents living close to stationary sites. On the other hand, if the ambient concentrations outside of the nonresidential microenvironments (e.g., office, school, in-vehicles, and so on) visited by the subjects were much lower than the stationary site measurements, the calculated ambient source contributions to personal exposures would be significantly overestimated. This was shown for some MENTOR predictions, in which the calculated MENTOR/personal measurement ratios were greater than 1. Therefore, further research is needed to characterize spatial variability of ambient air toxic concentrations in hot spot areas to better quantify ambient contributions to personal exposures of air toxics and to better deal with the need for such information before completing risk assessments.

In conclusion, person-oriented exposure modeling when coupled with the types of data acquired in the hot spot study can provide valuable information on the ambient source types contributing to personal exposure within a population living in a hot spot. The data gaps identified in this analysis showed a need for: (1) better characterization of spatial variability of ambient air toxic concentrations in hot spot areas through spatial saturation sampling²⁹ or refinements in estimating local source (mobile and stationary) emissions and micrometeorology effects, and (2) better estimates of indoor or other non-ambient source emission rates that contribute to personal exposure. However, the ambient contributions to personal exposure were reasonably characterized, which was the focus of the study. Finally, if the performance of the emission-based dispersion modeling for a hot spot is improved, a more precise characterization of contributions of different source emission categories (such as point, area, mobile on-road, and mobile non-road) on personal exposure levels to air toxics can be achieved. Then an anthropocentric or person-oriented source-to-exposure application of MENTOR/IBEM can lead to a much better selection of control strategies that can be formulated for reducing personal exposure to ambient sources of air toxics on the neighborhood scale.