Source proximity and meteorological effects on residential outdoor VOCs in urban areas: Results from the Houston and Los Angeles RIOPA studies

Abstract

Concentrations of volatile organic compounds (VOCs) measured outside homes in Houston, TX and Los Angeles, CA were characterized by the effects of source proximity and meteorological factors. Benzene, toluene, ethylbenzene, m,p-xylene, o-xylene (BTEX), methyl tert butyl ether (MTBE), tetrachloroethylene (perchloroethylene, PCE), and carbon tetrachloride (CCl₄) were examined. Multiple stepwise regression analysis converged the best-fit models with predictors from meteorological conditions and the proximity to specific point, area, and mobile sources on the residential outdoor VOC concentrations. Negative associations of wind speed with concentrations demonstrated the effect of dilution by high wind speed. Atmospheric stability increase was associated with concentration increase. Petrochemical source proximity was a significant predictor for BTEX and MTBE concentrations in Houston. Ethylbenzene and xylene source proximity was a significant predictor in Los Angeles. Close proximity to area sources such as scrap metal recycling or dry cleaning facilities increased the MTBE, PCE, and CCl₄ concentrations in Houston and Los Angeles. Models for ethylbenzene, m,p-xylene, and MTBE in Houston, and benzene in Los Angeles explained that for the median values of the meteorological factors, homes closest to influential highways would have concentrations that were 1.7–2.2 fold higher than those furthest from these mobile emission sources. If the median distance to sources were used in the models, the VOC concentrations varied 1.7 to 6.6 fold as the meteorological conditions varied over the observed range. These results highlight that each urban area is unique and localized sources need to be carefully evaluated to understand potential contributions to VOC air concentrations near residences, which influence baseline indoor air concentrations and personal exposures. Results of this study could assist in the appropriate design of monitoring networks for community-level sampling. They may also improve the accuracy of exposure models linking emission sources with estimated pollutant concentrations at the residential level.

Graphical Abstract

1. Introduction

The general population in urban settings cannot avoid environmental exposure to complex mixtures of volatile organic compounds (VOCs) generated from industrial and mobile sources because people’s daily activities rely on energy production and chemical use in built environments. Many VOCs that exist ubiquitously in the environment are listed as Hazardous Air Pollutants (HAPs) in the United States Clean Air Act because of their adverse health effects that include irritation of mucous membranes, allergy, aplastic anemia, neurological disorders, leukemia, and other cancers (Kim and Bernstein, 2009; Villeneuve et al., 2014; Weisel, 2010; USEPA, 2015).

Since the Total Exposure Assessment Methodology (TEAM) study conducted in the 1980s, many studies have identified VOC emission sources that impact both indoor and outdoor air thereby contributing to elevated population exposures. Population exposure assessment studies (i.e., TEAM, EXPOLIS, GerES II, MACBETH, NHANES) have consistently reported that indoor VOC concentrations are highly correlated with personal exposures because people spend most of their time indoors, and indoor concentrations are generally higher than outdoor concentrations especially when small sources exist indoors (Edwards and Jantunen, 2001; Gonzalez-Flesca et al., 2000; Hoffmann et al., 2000; Jia et al., 2008; Wallace, 1989, 1995; Wallace et al., 1985). However, the ambient air gasoline-related VOC concentrations immediately adjacent to a residence provides the baseline for the air concentrations typically found indoors.

The significant contribution of outdoor pollution to the indoor concentrations, especially in homes near highways and within close proximity to gas stations, results in elevated exposures to BTEX (Fuller et al., 2013; Jo and Moon, 1999; Lawson et al., 2011). A recent study that analyzed data from the Relationships of Indoor, Outdoor and Personal Air (RIOPA) study concluded that outdoor sources were the major contributors to personal exposures to gasoline-related VOCs (BTEX and MTBE) in the study population even though the participants spent 91% of their time indoors (median of 77% of their time was indoors at home) (Su et al., 2013). The outdoor source contributions to gasoline-related VOCs inside of homes are consistent with findings of previous studies that the majority of indoor fine particulate matter (PM_2.5), polycyclic aromatic hydrocarbons, and elemental carbon concentrations inside homes were found to be of outdoor origin (Hodas et al., 2012; Meng et al., 2007; Naumova et al., 2002; Naumova et al., 2003; Polidori et al., 2007; Polidori et al., 2006).

The characterization of the impact of mobile and point emission sources on residential outdoor concentrations (Kwon et al., 2006; Liu et al., 2006; Polidori et al., 2010) is important for understanding the outdoor source contributions of gasoline-related and other VOCs to indoor concentrations and to personal exposure levels of populations who spend most of their time at home, such as infants, toddlers, children, patients, and senior citizens, particularly for those living near the emission sources. This study aims to identify the major determinants of the residential outdoor concentrations among the local meteorological factors and emission sources near urban residences in different cities. The current analysis evaluated 8 VOCs to examine the influence of the emission sources proximity and meteorological factors on the residential outdoor concentrations (C_out). The approach is to construct models that explain the residential outdoor concentrations of VOCs in Houston and Los Angeles. Results are compared with similar analyses performed previously for the Elizabeth, NJ component of RIOPA (Weisel et al., 2005b).

2. Materials and methods

2.1. Data sources

As part of the RIOPA study, 48-hour integrated VOC samples were collected in three U.S. cities: Elizabeth, NJ, Houston, TX, and Los Angeles, CA in different seasons between the summer of 1999 and the spring of 2001. Homes near emission sources were oversampled based on the distance from those sources in order to estimate contributions to residential outdoor and indoor air and to personal exposures (Weisel et al., 2005b). The RIOPA study design, measurement of air pollutants, and quality assurance/control results are described elsewhere (Weisel et al., 2005b; Weisel et al., 2005a; Turpin et al., 2007). The RIOPA database is publicly available at https://riopa.aer.com/login.php (HEI and NUATRC, 2008).

Corresponding meteorological data were available for 198 samples in Houston and 163 samples in Los Angeles. Eight VOCs were included: benzene, toluene, ethylbenzene, m,p-xylene, o-xylene, methyl tert butyl ether (MTBE), tetrachloroethylene (perchloroethylene, PCE), and carbon tetrachloride (CCl₄). Concentrations below the respective method detection limits were replaced by one-half of the method detection limits before natural log transformation. Distributions of the selected VOCs in the three urban areas are shown in the Supplemental information (Table SI1).

2.2. Study areas

Maps of the three RIOPA study areas are presented in the Supplemental information (Figs. SI1 to SI3). Elizabeth is a city adjacent to the Linden industrial complex to the South. The residential areas are close to the commercial areas and major highways (Weisel et al., 2005b). Houston has the largest density of petrochemical facilities in the world for the production and storage of fuel, chemical precursors, plastics, and solvents. Residential areas around the petrochemical complexes were targeted for recruitment; areas included the Houston Ship Channel, Pasadena, Galena Park, Channelview, Baytown, Deer Park, La Porte, and as a comparison area, the Texas Medical Center (Weisel et al., 2005b). There were four sampling areas in Los Angeles: West Los Angeles, Pico Rivera, Burbank, and Santa Clarita, which were influenced by emissions from at least one major highway (Weisel et al., 2005b).

2.3. Meteorological data

To examine the influence of meteorological factors on residential outdoor VOC concentrations (C_out), temperature, dewpoint temperature, relative humidity (RH), wind speed, atmospheric pressure, and precipitation were downloaded from the RIOPA database (HEI and NUATRC, 2008) as candidate variables. The distributions of the meteorological variables by study location are summarized in the Supplemental information (Table SI2). Computed atmospheric Pasquill stability classes, with a time resolution of 3 h, were retrieved from the Air Resource Laboratory, National Oceanic and Atmospheric Administration (NOAA), real-time environmental applications and display (READY) system (NOAA, 2012). Each 3-hour stability class was assigned a code of “1” if the Pasquill category was “stable” or “neutral” (i.e., classes D, E, F, and G), or “0” when it was “unstable” (i.e., classes A, B, and C) to derive the fraction of each 48-hour VOC sampling period with “stable” or “neutral” stability.

2.4. Emission sources

Emission sources of the selected VOCs around the homes were identified by Geographical Information System (GIS) mapping. Emission sources of interest were: 1) point sources listed in the National Emissions Inventory for year 1999 (1999 NEI) (USEPA, 2003); 2) area sources such as gas stations (GS), scrap metal recyclers (SM), and dry cleaning facilities (DCF); and 3) mobile sources, i.e., highways and major arterial roadways. The point sources were located using coordinates in the 1999 NEI. The lists of the area sources were obtained from the Harris County Appraisal District for Houston, and from the South Coast Air Quality Management District for Los Angeles. The mobile sources were identified from the 2000 TIGER (Topologically Integrated Geographic Encoding and Referencing) line files obtained from the U.S. Census Bureau (USCB, 2012).

2.5. Proximity data

The direct distances between sampling locations and emission sources were calculated using the ArcScript extension referred to as “the nearest features” with distances and bearings (version 3.8b, Jenness Enterprises, Flagstaff, AZ) on ArcView 3.2 (ESRI Inc., Redland, CA). To obtain more generalized and consistent model outcomes, the proximity variables under the same source category for each compound were combined as the sum of all distances (km) from the first to the fifth closest facilities (except for CCl₄ in Los Angeles which only had one facility). This combined distance approach was developed to accommodate the wide distribution of every distance between surrounding emission clusters and the sampling homes. For example, the minimum sum of distances to the five closest gas stations in Houston and Los Angeles were 2.1 km and 1.44 km respectively. Those homes were located in the middle of five gas stations with an average distance of 420 m in Houston and 288 m in Los Angeles. This combined distance approach was useful to generate more generalized distributions of distance to sources surrounding homes regardless of the relative direction of the sources or the source clusters. The inverse of the sum of distances (km⁻¹) was calculated as the source proximity variable. The distributions of the proximity variables are shown in the Supplemental information (Tables SI3 and SI4).

2.6. Statistical analysis

All statistical analyses were performed using SAS (version 9.3, SAS Institute Inc. Cary, NC), and SPSS (version 23, SPSS Inc. Chicago, IL). VOC concentrations resembled a log normal distribution; therefore natural-log transformed residential outdoor concentrations (lnC_out) were used as the dependent variables. Bivariate Pearson’s correlations between the lnC_out and proximity and meteorological variables were used to explore the correlations at α = 0.05 (p < 0.05).

Multiple stepwise linear regression analysis was conducted for the selection of predictors from a group of variables (X_i) to explain lnC_out (Y_i). The default entry and inclusion criteria were set at p < 0.15. To avoid over-specified models and potential interaction problems among predictors, an entry criterion of p < 0.10 was used for benzene, ethylbenzene, m,p-xylene, o-xylene, MTBE, and CCl₄ in Houston. Because of the ln-transformation, the additive effects of the independent variables on the regression model predictions become multiplicative. The regression equation can be written as:

$$C_{outi}=ex{p}^{{\beta }_0}\cdot ex{p}^{{\beta }_{1Xi1}}\cdot ex{p}^{{\beta }_{2Xi2}}\cdots ex{p}^{{\beta }_{P-1Xi,P-1}}\cdot ex{p}^{\epsilon }$$ (1)

where C_{out i} is residential outdoor concentration (μg/m³), β₀, β₁ … β_{P − 1} are the regression coefficients, X_i1, X_i2 … X_{i, P − 1} are the values of predictor, and ε is the error term.

To reduce the Type I errors under multiple testing scenarios, preliminary regression analysis was performed for each compound to determine the relative importance of variables within the same type of independent variables, either proximity or meteorological variables separately. The best-fit models selected by the stepwise selection methods and the corresponding statistical results were evaluated for satisfying the major assumptions of linear regression analysis (Polidori et al., 2010).

3. Results

3.1. Statistical summary

Bivariate Pearson’s correlations between the lnC_out and the predictors showed that the correlations between lnC_out and proximity and meteorological variables were in the expected direction. Variables with statistically significant correlations at α = 0.05 (p < 0.05) are sorted in ascending order of p-values in Table 1. The proximity variables, the inverse distances to emission sources, atmospheric stability, and atmospheric pressure were positively correlated with the lnC_out of multiple VOCs. Wind speed, temperature, dewpoint, RH, and precipitation were negatively correlated with lnC_out of multiple VOCs.

Table 1.

Bivariate Pearson’s correlation coefficients (CC) between ln-transformed outdoor concentrations (lnC_out) and predictor variables (P = p-value; N = sample size).

Houston (N = 198)				Los Angeles (N = 163)
lnCout	Predictor	CC	P	lnCout	Predictor	CC	P
LnBZN	BZN5INVa	0.41	<0.0001	LnBZN	Ua	−0.59	<0.0001
	RHa	−0.27	<0.0001		DewCa	−0.53	<0.0001
	Staba	0.25	0.0004		RH	−0.47	<0.0001
	SM5INV	0.24	0.001		Temp	−0.38	<0.0001
	DewCa	−0.22	0.002		Press	0.38	<0.0001
	Ua	−0.19	0.006		Staba	0.56	<0.0001
	Press	0.19	0.007		GS5inv	−0.17	0.031
	A125INVa	0.18	0.014		A305inva	0.16	0.044
LnTOL	TOL5INVa	0.30	<0.0001	LnTOL	Ua	−0.33	<0.0001
	DewCa	−0.24	0.001		DewCa	−0.29	<0.0001
	SM5INV	0.22	0.002		RH	−0.27	<0.0001
	Temp	−0.21	0.004		Stab	0.25	0.001
	press	0.15	0.030		Press	0.26	0.001
	Staba	0.15	0.034		Temp	−0.20	0.011
	A125INV	0.15	0.035
LnEBZ	Ua	−0.32	<0.0001	LnEBZ	Ua	−0.52	<0.0001
	RHa	−0.32	<0.0001		DewC	−0.47	<0.0001
	EBZ5INVa	0.28	<0.0001		RHa	−0.45	<0.0001
	SM5INV	0.25	0.0004		Temp	−0.30	<0.0001
	Stab	0.23	0.001		Press	0.29	<0.0001
	A125INVa	0.20	0.005		Staba	0.52	<0.0001
					Precip	−0.17	0.03
LnMPX	RHa	−0.31	<0.0001	LnMPX	Ua	−0.51	<0.0001
	Ua	−0.33	<0.0001		RHa	−0.49	<0.0001
	XYL5INVa	0.53	<0.0001		DewC	−0.43	<0.0001
	SM5INV	0.36	<0.0001		Xyl5inva	0.30	<0.0001
	A125INVa	0.27	<0.0001		Staba	0.50	<0.0001
	Stab	0.25	0.0003		GS5inv	−0.25	0.002
	A305INV	0.22	0.002		Press	0.24	0.002
	GS5INV	0.17	0.014		Temp	−0.20	0.01
					A305inv	0.16	0.038
LnOXY	Ua	−0.32	<0.0001	LnOXY	Ua	−0.59	<0.0001
XYL5INVa		0.36	<0.0001		RHa	−0.51	<0.0001
Stab		0.26	0.0003		DewC	−0.47	<0.0001
RHa		−0.25	0.0004		Press	0.36	<0.0001
SM5INV		0.25	0.001		Staba	0.54	<0.0001
A125INV		0.18	0.011		Temp	−0.25	0.002
A305INV		0.17	0.017		Xyl5inva	0.23	0.004
					GS5inv	−0.20	0.011
					Precip	−0.17	0.035
LnMTBE	RHa	−0.33	<0.0001	LnMTBE	Ua	−0.44	<0.0001
	Ua	−0.31	<0.0001		RHa	−0.41	<0.0001
	SM5INVa	0.46	<0.0001		DewC	−0.35	<0.0001
	MTBE5INV	0.38	<0.0001		Stab	0.37	<0.0001
	A305INV	0.26	0.0002		Press	0.18	0.021
	A125INVa	0.24	0.001
	Stab	0.23	0.001
	GS5INV	0.17	0.019
	Stab	0.14	0.048
LnPCE	RHa	−0.19	0.007	LnPCE	Ua	−0.52	<0.0001
	DewCa	−0.21	0.004		RHa	−0.40	<0.0001
	Stab	0.16	0.029		DewC	−0.36	<0.0001
	GS5INV	0.18	0.012		A305inv	0.28	<0.0001
	SM5INVa	0.26	0.0002		Stab	0.46	<0.0001
	PCE5INV	−0.16	0.027		PCE5inva	0.23	0.003
					Press	0.23	0.004
					DCF5inva	0.22	0.005
					Temp	−0.17	0.029
LnCCL4	DewCa	−0.29	<0.0001	LnCCL4	DCF5inva	0.17	0.026
	DCF5INVa	0.26	0.0003
	Temp	−0.24	0.001
	RH	−0.17	0.017
	U	0.16	0.025
	Precipa	−0.14	0.043

Notes: U = wind speed (m/s); Temp = temperature (°C); RH = relative humidity (%); DewC = dewpoint temperature (°C); Press = atmospheric pressure (in Hg); Precip = precipitation (in); Stab = fraction of time atmospheric stability class was stable; GS5inv = inverse distance to the closest 5 gas stations; SM5INV = inverse distance to the closest 5 scrap metal facilities; DCF5Inv = inverse distance to the closest 5 dry cleaning facilities; A125inv = inverse distance to the closest 5 highways (A10, A20); A305inv = inverse distance to the closest 5 arterial roadways (A30); Bzn5inv = inverse distance to the closest 5 benzene point sources; Tol5inv = inverse distance to the closest 5 toluene point sources; Ebz5inv = inverse distance to the closest 5 ethylbenzene point sources; Xyl5inv = inverse distance to the closest 5 xylene point sources; MTBE5inv = inverse distance to the closest 5 MTBE point sources; PCE5inv = inverse distance to the closest 5 PCE point sources; CCL45inv = inverse distance to the closest 5 CCL4 point sources.

^aPredictors included in the model.

Preliminary regression analysis using only the proximity variables (results not shown) indicated that there was model specificity for proximity to emission sources of each VOC. For example, BTEX and MTBE point sources or roadways were included in their respective models, while PCE point source or DCF were not. In contrast, PCE and CCl4 models did not include any other BTEX and MTBE point sources or roadways. Summaries of regression models for each VOC including adjusted R², degree of freedom, and p-value for the overall model, and partial R² and p-value for individual variables are shown in Table 2 for Houston and Table 3 for Los Angeles. The predictors included in the models are sorted by the order of selection (X_ith). A predictor enters the model based on the significance of its association with lnC_out. Most of the predictors included in the models were the variables significantly correlated with lnC_out (Table 1). For comparison, a summary of the Elizabeth model adapted from (Kwon et al., 2006) is shown in Table 4. The F-statistics for all models were statistically significant (p < 0.0001, P_model < F) except for CCl4 in Los Angeles (p = 0.026). The p-values for parameter estimates of the first selected independent variables were statistically significant (p < 0.0001) for BTEX and MTBE. The models explained over 40% of the total variability of outdoor concentrations of m,p-xylene (47%), MTBE (45%) in Houston, and benzene (47%), m,p-xylene (41%), o-xylene (46%), PCE (43%) in Los Angeles. Models explained 39% to 24% of the total variability of outdoor concentrations of benzene (30%), ethylbenzene (30%), o-xylene (25%) in Houston, and ethylbenzene (39%), MTBE (24%) in Los Angeles. < 15% of the total variability of outdoor concentrations was explained for toluene (14%), PCE (10%), CCl4 (14%) in Houston, and toluene (12%), CCl4 (2%) in Los Angeles.

Table 2.

Summary of the best-fit models for VOCs in Houston, TX (Ln-transformed concentrations, μg/m³).

Pollutant	Y			X1st	X2nd	X3rd	X4th	X5th	X6th
Benzene	lnBZN		β0	BZN5INV	RH	STABa	DEWCa	Ua	A125INVa
Adjusted model R2=	0.30	βi	1.021	11.329	−0.010	0.529	−0.021	−0.107	7.769
Enter p<	0.10	Standard error	0.461	2.179	0.003	0.322	0.006	0.050	4.238
Model d.f.=	6	Partial R2		0.168	0.056	0.043	0.026	0.020	0.012
Model p-value	<0.0001	p-value		<0.0001	0.0002	0.001	0.009	0.020	0.068
Toluene	lnTOL		β0	TOL5INV	DEWCa	STAB
Adjusted model R2=	0.14	βi	−0.837	19.497	−0.055	1.982
Enter p<	0.15	Standard error	0.628	5.352	0.018	0.812
Model d.f.=	3	Partial R2		0.089	0.035	0.026
Model p-value	<0.0001	p-value		<0.0001	0.006	0.016
Ethylbenzene	lnEBZ		β0	Ua	RHa	EBZ5INVa	DEWC	A125INV
Adjusted model R2=	0.30	βi	1.425	−0.280	−0.017	7.408	−0.024	12.441
Enter p<	0.10	Standard error	0.423	0.048	0.004	2.693	0.008	5.227
Model d.f.=	5	Partial R2		0.102	0.105	0.052	0.040	0.020
Model p-value	<0.0001	p-value		<0.0001	<0.0001	0.0003	0.001	0.018
m,p-Xylene	lnMPX		β0	XYL5INVa	Ua	RHa	A125INV	DEWC	PRESS
Adjusted model R2=	0.47	βi	36.559	17.104	−0.280	−0.014	10.520	−0.033	−1.153
Enter p<	0.10	Standard error	12.801	2.252	0.043	0.003	4.474	0.009	0.422
Model d.f.=	5	Partial R2		0.282	0.082	0.065	0.020	0.015	0.020
Model p-value	<0.0001	p-value		<0.0001	<0.0001	<0.0001	0.010	0.021	0.007
o-Xylene	lnoXy		β0	XYL5INVa	Ua	RHa
Adjusted model R2=	0.25	βi	1.047	16.901	−0.269	−0.018
Enter p<	0.10	Standard error	0.486	3.334	0.056	0.005
Model d.f.=	3	Partial R2		0.133	0.084	0.047
Model p-value	<0.0001	p-value		<0.0001	<0.0001	0.001
MTBE	lnMTBE		β0	SM5INV	MTBE5INV	Ua	RHa	A125INV
Adjusted model R2=	0.45	βi	1.627	16.001	30.093	−0.231	−0.023	13.556
Enter p<	0.10	Standard error	0.490	3.047	5.422	0.048	0.004	6.143
Model d.f.=	5	Partial R2		0.215	0.097	0.066	0.067	0.014
Model p-value	<0.0001	p-value		<0.0001	<0.0001	<0.0001	<0.0001	0.029
PCE	lnPCE		β0	SM5INV	RHa	DEWC	Ua
Adjusted model R2=	0.10	βi	−0.614	12.906	−0.013	−0.036	−0.188
Enter p<	0.15	Standard error	0.810	5.147	0.008	0.016	0.095
Model d.f.=	4	Partial R2		0.069	0.022	0.012	0.018
Model p-value	<0.0001	p-value		0.0002	0.030	0. 117	0.051
CCl4	lnCCL4		β0	DEWC	DCF5INVa	PRECIP
Adjusted model R2=	0.14	βi	−0.425	−0.007	1.545	−2.129
Enter p<	0.10	Standard error	0.042	0.002	0.454	1.113
Model d.f.=	3	Partial R2		0.086	0.055	0.016
Model p-value	<0.0001	p-value		<0.0001	0.001	0.057

^aPredictor variable also commonly selected in the specific VOC Models in RIOPA CA.

Table 3.

Summary of the best-fit models for VOCs in Los Angeles, CA (Ln-transformed concentrations, μg/m³).

Pollutant	Y			X1st	X2nd	X3rd	X4th	X5th
Benzene	lnBZN		β0	Ua	DewCa	A125Inva	A305Inv	Staba
Adjusted model R2=	0.47	βi	0.161	−0.801	−0.033	15.484	2.429	0.996
Enter p<	0.15	Standard error	0.598	0.157	0.011	5.296	0.863	0.503
Model d.f.=	5	Partial R2		0.347	0.077	0.019	0.026	0.013
Model p-value	<0.0001	p-value		<0.0001	<0.0001	0.022	0.006	0.049
Toluene	lnTOL		β0	U	DewCa
Adjusted model R2=	0.12	βi	2.592	−0.462	−0.020
Enter p<	0.15	Standard error	0.172	0.160	0.010
Model d.f.=	2	Partial R2		0.108	0.022
Model p-value	<0.0001	p-value		<0.0001	0.048
Ethylbenzene	lnEBZ		β0	Stab	Ua	Ebz5inva	RHa
Adjusted model R2=	0.39	βi	0.053	1.592	−0.612	26.621	−0.011
Enter p<	0.15	Standard error	0.592	0.452	0.166	7.758	0.005
Model d.f.=	4	Partial R2		0.273	0.064	0.046	0.020
Model p-value	<0.0001	p-value		<0.0001	0.0001	0.001	0.022
m,p-Xylene	lnMPX		β0	Ua	RHa	Xyl5inva	Stab
Adjusted model R2=	0.41	βi	1.709	−0.674	−0.018	16.201	1.100
Enter p<	0.15	Standard error	0.673	0.191	0.005	3.868	0.520
Model d.f.=	4	Partial R2		0.260	0.080	0.072	0.016
Model p-value	<0.0001	p-value		<0.0001	<0.0001	<0.0001	0.036
o-Xylene	lnoXy		β0	Ua	RHa	Xyl5inva	Stab	A125Inv
Adjusted model R2=	0.46	βi	0.496	−0.800	−0.014	7.050	1.039	10.721
Enter p<	0.15	Standard error	0.616	0.159	0.004	3.599	0.433	5.787
Model d.f.=	5	Partial R2		0.343	0.066	0.039	0.016	0.012
Model p-value	<0.0001	p-value		<0.0001	<0.0001	0.001	0.030	0.066
MTBE	lnMTBE		β0	Ua	RHa
Adjusted model R2=	0.24	βi	4.109	−0.788	−0.018
Enter p<	0.15	Standard error	0.345	0.194	0.006
Model d.f.=	2	Partial R2		0.195	0.051
Model p-value	<0.0001	p-value		<0.0001	0.001
PCE	lnPCE		β0	Ua	PCE5inv	DCF5inv	RHa
Adjusted model R2=	0.43	βi	0.399	−0.916	28.579	2.840	−0.010
Enter p<	0.15	Standard error	0.407	0.153	4.580	0.609	0.004
Model d.f.=	4	Partial R2		0.272	0.067	0.091	0.018
Model p-value	<0.0001	p-value		<0.0001	<0.0001	<0.0001	0.026
CCl4	lnCCL4		β0	DCF5inva
Adjusted model R2=	0.02	βi	−0.599	0.605
Enter p<	0.15	Standard error	0.074	0.270
Model d.f.=	1	Partial R2		0.030
Model p-value	0.026	p-value		0.026

^aPredictor variable also commonly selected in the specific VOC models in RIOPA TX.

Table 4.

Summary of the best-fit models for VOCs in Elizabeth, NJ (Ln-transformed concentrations, μg/m³). Adapted from Kwon et al., 2006.

Pollutant	Y			X1st	X2nd	X3rd	X4th	X5th	X6th
Benzene	lnBZN		β0	K	U	GS−1	Stab	FC14−1
Adjusted model R2=	0.38	βi	11.4	−0.04	−0.11	17.9	0.54	5.4
Enter p<	0.15	Standard error	1.4	0.005	0.04	5.6	0.26	3.4
Model d.f.=	5	Partial R2		0.25	0.07	0.05	0.02	0.01
Model p-value	<0.0001			<0.0001	<0.0001	0.0004	0.0465	0.1109
Toluene	lnTOL		β0	Stab	FC14−1	RH	K	U
Adjusted model R2=	0.25	βi	7.3	0.76	18.6	0.02	−0.02	−0.10
Enter p<	0.15	Standard error	2.0	0.37	4.6	0.005	0.007	0.05
Model d.f.=	3	Partial R2		0.12	0.06	0.03	0.05	0.02
Model p-value	<0.0001			<0.0001	0.0007	0.0129	0.0011	0.0679
Ethylbenzene	lnEBZ		β0	Stab	K	U	FC14−1
Adjusted model R2=	0.14	βi	6.9	1.18	−0.03	−0.11	9.6
Enter p<	0.15	Standard error	2.3	0.43	0.01	0.07	5.7
Model d.f.=	5	Partial R2		0.08	0.05	0.015	0.015
Model p-value	0.0001			0.0002	0.0033	0.0924	0.0915
m,p-Xylene	lnMPX		β0	Stab	K	GS−1	U	Precip	FC14−1
Adjusted model R2=	0.29	βi	9.2	0.92	−0.03	18.5	−0.14	0.006	8.4
Enter p<	0.15	Standard error	1.6	0.30	0.005	6.4	0.05	0.003	4.5
Model d.f.=	5	Partial R2		0.11	0.11	0.05	0.03	0.02	0.01
Model p-value	<0.0001			<0.0001	<0.0001	0.0014	0.0120	0.0499	0.0669
o-Xylene	lnoXy		β0	Stab	K	U	GS−1	FC14−1	RH
Adjusted model R2=	0.43	βi	7.1	1.04	−0.03	−0.14	9.9	9.7	0.007
Enter p<	0.15	Standard error	1.3	0.25	0.004	0.04	5.3	4.4	0.003
Model d.f.=	3	Partial R2		0.24	0.09	0.06	0.03	0.01	0.02
Model p-value	<0.0001			<0.0001	<0.0001	0.0001	0.0066	0.0427	0.0331
MTBE	lnMTBE		β0	U	GS−1	K	FC11−1
Adjusted model R2=	0.24	βi	−0.9	−0.24	35.5	0.01	22.8
Enter p<	0.15	Standard error	2.0	0.05	8.2	0.007	14.4
Model d.f.=	5	Partial R2		0.12	0.09	0.01	0.01
Model p-value	<0.0001			<0.0001	<0.0001	0.0885	0.1142
PCE	lnPCE		β0	U	DCF−1	RH	K	Stab
Adjusted model R2=	0.31	βi	3.0	−0.15	32.5	0.01	−0.01	0.38
Enter p<	0.15	Standard error	1.2	0.04	12.4	0.003	0.004	0.25
Model d.f.=	4	Partial R2		0.18	0.04	0.03	0.05	0.01
Model p-value	<0.0001			<0.0001	0.0056	0.0100	0.0014	0.1367
CCl4	lnCCL4	No model

U = wind speed (m/s); K = temperature °K; RH = relative humidity (%); Precip = precipitation (mm); Stab = fraction of time atmospheric stability class was stable; GS⁻¹ = inverse distance to the closest gas station (km⁻¹); DCF⁻¹ = inverse distance to the closest dry cleaning facility (km⁻¹); FC11⁻¹ = inverse distance to the closest highways (km⁻¹); FC14⁻¹ = inverse distance to the closest arterial roadways (km⁻¹).

3.2. Meteorological variables

Wind speed was selected as the first predictor (X_1st, p < 0.0001) in models for ethylbenzene in Houston, benzene, toluene, m,p-xylene, o-xylene, MTBE, and PCE in Los Angeles, and MTBE and PCE in Elizabeth (Kwon et al., 2006). Wind speed entered as the second predictor in 4 models: m,p-xylene and o-xylene in Houston, ethylbenzene in Los Angeles, and benzene in Elizabeth. The partial R² of the wind speed represented N 50% of the model R² in 8 models: benzene, toluene, m,p-xylene, o-xylene, MTBE, and PCE in Los Angeles, and MTBE and PCE in Elizabeth (Kwon et al., 2006). Among the Los Angeles VOC models that had adjusted model R² higher than 30%, the wind speed explained from 26% to 35% of the overall variability of residential outdoor VOC concentrations. The effect of the wind speed was less significant in Houston, where it explained 10% of the overall variability of residential outdoor ethylbenzene concentrations. As expected, wind speed was inversely associated with the lnC_out having a negative coefficient in all included models.

Atmospheric stability was selected as a predictor for 12 models: benzene and toluene in Houston, benzene, ethylbenzene, m,p-xylene, and o-xylene in Los Angeles, and BTEX and PCE in Elizabeth. Stability was selected as the first predictor in 5 models: ethylbenzene (p < 0.0001) in Los Angeles, and toluene, ethylbenzene, m,p-xylene, and o-xylene in Elizabeth. The partial R² of stability in these VOC models accounted for 68, 44, 50, 34, and 53% of each model R², respectively. The partial R² of the stability in 7 other models was relatively small, ranging from 2 to 17% of model R². Stability was positively associated with the concentrations.

Temperature was a strong predictor included in all 7 Elizabeth models except for CCl_4, which did not converge for any of the models. However, temperature was not selected in any of the models in Houston and Los Angeles. Instead, dewpoint temperature was selected for six models in Houston and two models in Los Angeles. Dewpoint was included in benzene, toluene, ethylbenzene, m,p-xylene, PCE, and CCl₄ models in Houston, and in benzene and toluene models in Los Angeles. RH was selected in 6 models in Houston: benzene, ethylbenzene, m,p-xylene, o-xylene, MTBE, and PCE, and 5 models in Los Angeles: ethylbenzene, m,p-xylene, o-xylene, MTBE, and PCE. Dewpoint and RH were both negatively associated with C_out in Houston and Los Angeles.

3.3. Source proximity

Point source proximity was selected as the first predictor in 5 models of BTEX in Houston. The MTBE model included point source proximity as the second predictor entered into the model. The partial R² of the point source proximity variable for benzene, toluene, ethylbenzene, m,p-xylene, o-xylene, and MTBE accounted for 52, 59, 16, 58, 50, and 21% of each model’s R², respectively. Among the Houston VOC models that had adjusted model R² > 30% and included proximity as the first predictor, the point source proximity explains 17% to 28% of the overall variability of residential outdoor VOC concentrations of benzene, m,p-xylene, and MTBE. In Los Angeles, point source proximity was included as the second predictor (X_2nd, p < 0.0001) entered into the PCE model, and the third predictor (X_3rd) in the ethylbenzene (p = 0.001), m,p-xylene (p < 0.0001), and o-xylene (p = 0.001) models. The point source proximity in Los Angeles models of ethylbenzene, m,p-xylene, o-xylene, and PCE explained 5, 7, 4, and 7% of overall variability of residential outdoor concentrations, respectively.

Area sources of the target VOCs during the study periods were GS for gasoline-related compounds, DCF for PCE, and SM for gasoline-related compounds and solvents. The GS proximity variable was selected as a predictor for Elizabeth models of benzene, m,p-xylene, o-xylene, and MTBE (p < 0.01) (Kwon et al., 2006). The SM proximity was selected as the first predictor entered in Houston models of MTBE (p < 0.0001) and PCE (p = 0.0002). The partial R² accounted for 47% and 57% of the model variability respectively, which explained 22% and 7% of overall variability of residential outdoor concentrations. The DCF proximity was selected as a predictor for 4 models: PCE in Elizabeth (p = 0.0056), CCl₄ in Houston (p = 0.001), and PCE (p < 0.0001) and CCl₄ (p = 0.026) in Los Angeles.

In the prior Elizabeth study, arterial roadway proximity was a significant predictor in lnC_out models of BTEX, and highway proximity for MTBE (Kwon et al., 2006). The highway proximity was a significant predictor in Houston models of benzene, ethylbenzene, m,p-xylene, MTBE, and in Los Angeles models of benzene and o-xylene. The Los Angeles benzene model included both highways and arterial roadways as predictors. Among the 24 models, 7 VOC models included the highway proximity, and 6 models included the arterial roadway proximity. < 3% of overall variability of residential outdoor concentrations was explained by the roadway proximity variables. The roadway proximity was generally selected as the 4th or 5th predictor. The highway proximity and arterial roadway proximity were positively associated with the lnC_out in all models. Therefore models confirm that close proximity to traffic increases the lnC_out, and concentrations decrease as the distance from the highways or arterial roadways increases for mobile source VOCs.

4. Discussion

This study demonstrates the significant influence of source proximity on outdoor residential concentrations of BTEX, MTBE, PCE, and CCl₄ measured in Houston and Los Angeles, and the influence of site-specific meteorological conditions. The partial R² for the meteorological variables was typically larger than that for the proximity variables for the Elizabeth and Los Angeles models, implying that a greater percentage of the explanatory power was due to changes in the meteorological conditions rather than to the distance from emission sources. However, for Houston, the partial R² values for the source proximity variables were larger than or similar to those for the meteorological variables for benzene, toluene, m,p-xylene, o-xylene, MTBE, and PCE. This suggests that the impact of emissions from petrochemical facilities on the lnC_out near the Houston Ship Channel and surrounding areas, where the Houston homes were mostly located, can be significant.

Although toluene levels were higher compared to the other VOCs in this study, the percentage of concentrations above the limit of detection was only 41%. In addition, toluene has various sources that are not listed in inventories. Therefore, those factors may have affected the lower explanatory power of toluene models than other gasoline-derived compounds.

4.1. Wind speed and stability

The most influential meteorological variable explaining the variability of C_out was wind speed. Among the 24 models determined for the three urban areas, wind speed was selected for 20 models. More than half of the models (13/24) included the wind speed as either first or second predictor. For 8 models in Los Angeles and Elizabeth, wind speed (p < 0.0001) explained 11% to 35% of the overall variability of the residential outdoor VOC concentrations. The VOC models that did not include wind speed were toluene in Houston, and CCl₄ in all three locations. The consistent inclusion of wind speed confirmed the inverse association between wind and lnC_out. The results are consistent with the conclusion from the RIOPA data analysis by Su et al. (2013) showing that wind speed was negatively associated with personal exposure concentrations of benzene, toluene, ethylbenzene, m,p-xylene, o-xylene, MTBE, styrene, TCE, PCE, and α-pinene. Stability was included as the first influential predictor in four Elizabeth models, while only ethylbenzene among Los Angeles models included stability as the first predictor. The twelve models including stability explained that the longer the period of time with a stable or neutral atmospheric stability class, the higher the estimated residential outdoor concentration.

4.2. Model-explained effect of wind

Relationships between explained C_out and the effect of wind speed in Houston and Los Angeles estimated from the multiple regression models are illustrated in Figs. 1 and 2, when all other variables in each model are held constant using observed median values. Although the models and individual variables were statistically significant, still a larger proportion of the C_out variability remains unexplained. To avoid misleading interpretations of model explanatory power, explanatory effects were limited to the VOCs with adjusted model R² larger than 0.3 and individual variable with p < 0.05. The explained C_out decreased with increasing wind speed. The explained C_out in Los Angeles decreased more rapidly than that of Houston per unit increase of wind speed. Across the VOCs evaluated the C_out estimated at the minimum wind speed were 3.5- to 6.6-fold higher in Los Angeles and 1.7- to 4.2-fold higher in Houston than the C_out estimated at the maximum wind speed for each site for homes at the median distance from sources. For example, the C_out of benzene in Los Angeles estimated at the minimum wind speed of 0.26 m/s is 2.94 μg/m³, which is 5.2-fold higher than the estimated benzene concentration of 0.57 μg/m³ at the maximum wind speed of 2.26 m/s. The overall higher wind speeds in Houston likely contributed to lower C_out than were measured in Los Angeles. The declines of concentrations with wind speed in Houston was comparable to that observed from the Elizabeth results.

Fig. 1.

Simulated effect of wind speed on residential outdoor air concentrations of benzene, ethylbenzene, m,p-xylene, and MTBE in Houston areas estimated from the best fit models using the median values for other variables in each model.

Fig. 2.

Simulated effect of wind speed on residential outdoor air concentrations of benzene, ethylbenzene, m,p-xylene, and o-xylene in Los Angeles areas estimated from the best fit models using the median values for other variables in each model.

4.3. Temperature, dewpoint, and RH

RH was negatively associated with lnC_out for all Houston and Los Angeles models. In contrast, temperature was associated negatively with C_out of BTEX and PCE in Elizabeth, implying that higher C_out on cold days may be caused by incomplete combustion and exhaust byproducts from mobile sources or that there are seasonal differences in fuel content (Kwon et al., 2006). The only exception, i.e., a positive temperature association, in Elizabeth was observed in the MTBE model, which did not include RH. When the RIOPA study was conducted, MTBE was added at higher percentages in winter-time gasoline fuel blends which could balance the greater evaporative emissions at warmer temperatures, confounding any expected relationship between temperature, RH and MTBE concentration. When temperature and RH were concurrently included in the same model, RH was associated positively with toluene, o-xylene, and PCE models in Elizabeth.

Because RH depends on temperature, the two parameters may have adjusted interactively within the model producing a positive coefficient for RH in Elizabeth. Inclusion of a variable into a model depends on the correlation with previously included variables in multiple linear regression. The bivariate correlations between meteorological variables by study areas are shown in the Supplemental information (Tables SI6 and SI7). The directions of correlation coefficients between meteorological variables could be different by location because correlations was derived only for the weather conditions during sampling days. For example, wind speed was negatively correlated with temperature in Elizabeth and Houston indicating wind speed increase was correlated with decreased temperature. In contrast, wind speed was positively correlated with temperature in Los Angeles, indicating increased wind speed with increased temperature. This indicates that regional climate factors interacted with meteorological variables differently by urban area. As a result, inconsistent direction of association in the model parameters was observed for temperature and RH in Elizabeth. However, the direction of association with C_out in models was consistent for wind speed, stability, and dewpoint.

Dewpoint was selected for six Houston models and two Los Angeles models, while temperature was a strong predictor for C_out in Elizabeth. Houston and Los Angeles model results indicated that an increase in dewpoint and RH was associated with decreased C_out. Dewpoint, and RH are indicators of the moisture in the atmosphere. It seems likely that the effect of water vapor content in the atmosphere influenced the C_out in Houston and Los Angeles. In Houston, 4 models of benzene, ethylbenzene, m,p-xylene, and PCE included both RH and dewpoint as predictors, while no model included RH and dewpoint concurrently in Los Angeles (2 dewpoint vs. 5 RH). The concurrent inclusion of RH and dewpoint in Houston may imply a potentially greater impact of the humid climate in Houston on C_out compared to Elizabeth or Los Angeles. The results are also consistent with the RIOPA data analysis conclusion that RH was negatively associated with concentrations of benzene, ethylbenzene, m,p-xylene, o-xylene, MTBE, styrene, and β-pinene (Su et al., 2013).

4.4. Petrochemical emission sources

The influences of point source proximity are much higher for BTEX and MTBE in Houston and ethylbenzene, xylenes, and PCE in Los Angeles, compared to the Elizabeth models. Petrochemical facilities are sources for multiple VOCs in Houston. The total annual tons of target VOCS generated from point sources within the municipalities of sampled homes in Houston, Los Angeles and Elizabeth were 1858 tons, 111 tons and 0.04 tons, respectively (Table 5). Additional generation in adjacent municipalities in Elizabeth was 37 tons (Table 5). In Houston, point source proximity was included as the most influential predictor (X_1st, p < 0.0001) of benzene, m,p-xylene, o-xylene with 13% 28% of overall variability of C_out explained by models. A source apportionment study concluded that emissions from the Houston Ship Channel accounted for two-thirds of VOC mass contributions to the Houston urban area along with mobile and evaporative emissions (Leuchner and Rappenglück, 2010). In contrast, Elizabeth models included proximity to area sources such as GS or DCF for benzene, m,p-xylene, o-xylene, MTBE, and PCE. In Los Angeles, the PCE model included PCE point source proximity as the 2nd important predictor (X_2nd, p < 0.0001) followed by the DCF proximity as the 3rd predictor (X_3rd, p < 0.0001). It is notable that the annual PCE emissions were 69.8 tons in Los Angeles, and 316.4 tons in adjacent municipalities based on the NEI 1999 (Table 5). In Elizabeth and Houston, PCE emissions were insignificant at 0.16 and 0.42 tons respectively including adjacent cities. It is therefore reasonable that PCE point source proximity was not a predictor included in Elizabeth and Houston. The different variables in Houston and Los Angeles models compared to the previous Elizabeth models (Polidori et al., 2010; Liu et al., 2006; Kwon et al., 2006) suggests that the multiple linear regression models reflect the site-specific characteristics such as emission profiles. Thus, the models weigh the relative importance and influence of sources differently, in accordance with the corresponding meteorological conditions in different areas during the sampling period.

Table 5.

Annual emissions (Tons) reported in the 1999 NEI from the point sources within the cities that included the RIOPA homes, and within adjacent cities. A more detailed version is available in the Supplemental information (Table SI5).

	Benzene	Toluene	Ethylbenzene	Xylenes	MTBE	PCE	CCl4	Total
NJ, RIOPA	0.01	0.02	0	0.01	0	0.003	0.003	0.04
NJ, adjacent	2.19	6.53	0.77	2.75	24.29	0.15	0	36.67
NJ, sum	2.19	6.54	0.77	2.76	24.29	0.16	0.003	36.71
TX, RIOPA	350.35	419.27	159.65	519.31	380.72	0.42	27.95	1857.66
TX, adjacent	0.01	0.08	0.0002	0.03	0	0	0	0.12
TX, sum	350.36	419.35	159.65	519.34	380.72	0.42	27.95	1857.78
CA, RIOPA	1.87	23.54	0.01	15.69	0	69.79	0.02	110.92
CA, adjacent	29.46	444.44	60.08	200.37	15.41	316.40	0.02	1066.18
CA, sum	31.33	467.98	60.09	216.06	15.41	386.19	0.05	1177.10

4.5. Model-explained effect of xylene point source proximity

The model-explained effects of the xylene point source proximity variable in Houston and Los Angeles on the C_out of m,p-xylene and o-xylene are illustrated in Fig. 3. Explanatory interpretations were limited to the VOCs with adjusted model R² larger than 0.3 and individual variable with p < 0.05. When the median values of all other variables were used in each model, while the xylene point source proximity was varied from the closest to farthest distances observed at each site, the explained C_out of m,p-xylene and o-xylene decreased with increasing distance to emission sources. The C_out of m,p-xylene estimated at the closest distance were 2.5- and 2.6-fold higher in Los Angeles and Houston, respectively, compared to the C_out estimated at the farthest distance, while the C_out for o-xylene were 1.5 fold higher in Los Angeles.

Fig. 3.

Simulated effect of sum of distance to 5 closest xylene emissions on residential outdoor air concentrations of m,p-xylene, and o-xylene in Los Angeles and m,p-xylene in Houston areas estimated from the best fit models using the median values for other variables in each model.

4.6. Influence of proximity to SM facilities on MTBE and PCE

During the Houston Exposure to Air Toxics Study (HEATS), which investigated VOC exposure in Houston, the coordinates of many personal properties used as junkyards or auto parts shops in local neighborhoods were identified in order to investigate the association of the SM proximity and elevated C_out at homes (Morandi et al., 2009). The MTBE and PCE models in Houston included the proximity to SM as a predictor. MTBE, one of the largest components of reformulated gasoline, was used as an oxygenate in gasoline during the RIOPA study years from 1999 to 2001. Its source at SM facilities could be evaporation of residual gasoline left in salvage automobiles. PCE is a common degreasing solvent for metal parts, which could also be used at these facilities.

SM proximity in the Houston MTBE model explained 22% of the overall variability, while point source proximity and highway proximity explained 10 and 1.4%, respectively. This is a good example of uncontrolled area sources near homes potentially contributing more to exposure variability than larger and more distant point sources and mobile sources. Overall, proximity to three different types of sources explained 33% of the overall variability of C_out of MTBE in Houston. The MTBE model in Los Angeles did not include source proximity; it only included the meteorological variables, wind speed, and RH, which indicates that the two models reflect variations of the two different study areas. This is supported by the 1999 NEI which did not include any MTBE point source near Los Angeles homes. MTBE emissions were 15.4 tons per year in the municipalities adjacent to Los Angeles homes, which is very small compared to the 380 tons per year of MTBE emissions in Houston (Table 5).

4.7. CCl₄ and DCF

CCl₄ has very limited production and use. This is a reasonable explanation for CCl₄ not producing a significant model in Elizabeth. However, with very small variance explained by models, CCl₄ models for Houston (adjusted R² = 0.14) and Los Angeles (adjusted R² = 0.02) showed increased concentration with decreased distance to DCF. A series of studies have reported that CCl₄ and other halogenated VOCs are formed by mixing surfactants with sodium hypochlorite (NaOCl) the major ingredient of bleach (Odabasi et al., 2014; Odabasi, 2008). In those studies, the headspace concentration of CCl₄ observed was as high as 101 mg/m³, while the highest indoor air concentration of CCl₄ observed was 1124 μg/m³. If DCF used significant amount of detergents and bleach products then CCl₄ may have been produced at the DCF and emitted to nearby neighborhoods.

4.8. Model-explained effect of highway proximity

The model-explained effects of highway proximity in Houston and Los Angeles on C_out are illustrated in Fig. 4. All other predictors in each model are held constant using observed median values while the distance to highways was changed from the closest to farthest distance observed at each site. Explanatory interpretations were limited to the VOCs with adjusted model R² larger than 0.3 and individual variable with p < 0.05. The explained C_out of benzene in Los Angeles and ethylbenzene, m,p-xylene, and MTBE in Houston increased with decreased distances to highways. The sum of the distances to the 5 closest highways ranged between 12.7 and 34 km in Los Angeles, while it ranged between 13.1 and 42 km in Houston. The C_out of benzene estimated at the closest distances were 2.2-fold (2.2 μg/m³ vs. 1.0 μg/m³) higher in Los Angeles compared to the C_out estimated at the farthest distances. The results are consistent with the findings from Valach et al. (2014) who reported that benzene concentrations measured in congested urban areas were more than double the concentrations measured in urban locations away from major sources. The C_out of ethylbenzene, m,p-xylene, and MTBE in Houston estimated at the closest distances were 1.9-, 1.7-, and 2.0-fold higher, respectively, compared to the C_out estimated at the farthest distances.

Fig. 4.

Simulated effect of sum of distance to 5 closest highways on residential outdoor air concentrations of benzene in Los Angeles areas, and ethylbenzene, m,p-xylene, and MTBE in Houston areas estimated from the best fit models using the median values for other variables in each model.

4.9. Advantages and limitations

Multiple regression analyses on each VOC in residential ambient air converged to statistically significant models that included proximity and meteorological variables. The overall models explained from 25% to 40% of the variability of the VOC residential outdoor concentrations for these urban areas. The unexplained variance may be the result-of the RIOPA study design which sampled at each location only twice on different 48-hour periods, thus reducing the models’ ability to predict the temporal variation The samples were 48-hour integrated samples with changing wind direction and speed over that time period. This increases the uncertainty in the models’ capability to account for the meteorological effects and impacts of area sources on the measured concentrations. The model for each VOC was unique for each urban area sampled, and was based on each region’s emission profiles and meteorological conditions that were specified with data identified for the region and time period of sampling. There are many traffic-related air pollutants studies focused on the criteria air pollutants such as PM_2.5, NO₂, and CO. However few studies have been done of VOCs levels of concentrations near roadways or have performed inter-urban comparisons (Cook et al., 2008; Batterman et al., 2010; Zhang and Batterman, 2013; Fujita et al., 2011; Hoek et al., 2008).

The RIOPA study was designed to investigate the relationship of outdoor-indoor-personal air concentrations. The air pollutants were measured in different 48-hour periods throughout different seasons between the summer of 1999 and spring of 2001 to characterize personal exposures. The current analysis is an effort to connect the background ambient pollution from surrounding emission sources to the residential outdoor levels. Several RIOPA papers consistently indicated the residential outdoor air contaminants contribute to indoor air pollutant concentrations through penetration and ventilation (Hodas et al., 2012; Meng et al., 2007; Naumova et al., 2002; Polidori et al., 2007).

Although the model analyses show that only small fraction of the variability (1–3%) in residential outdoor VOC concentration for these urban settings is significantly associated with roadway proximity, people who live/work in the close proximity to such sources will be exposed to higher levels than found in background urban air (Kwon et al., 2006). Approximately 11.3 million persons (or 3.7% of the 308.7 - million U.S. population) live within 150 m of a major highway (Boehmer et al., 2013). The current results indicate that large petrochemical emission sources impact outdoor levels of VOCs in nearby residences, contributing to an individual’s life-time VOC exposures. This study confirms that the spatial variation of the air pollutants is difficult to estimate without considering urban local meteorology and without well-defined emission source data. Traffic count on roadways, or area of roadways surrounding homes may have been useful to better characterize the mobile source contribution. We observed strong associations of outdoor residential benzene levels with surrounding area covered with roadways in Elizabeth, NJ. For the RIOPA study, traffic counts were not measured with sufficient spatial and temporal frequency in each roadway in the study areas.

5. Conclusions

Multiple regression models selected the relatively more influential factors from the meteorological variables and proximity to point, area, and mobile sources to explain residential C_out in Houston and Los Angeles. The multiple regression models explained 12–47% of the total variability of C_out of BTEX. Difference in sources and meteorological conditions among the different sites led to some differences in variables included in the explanatory models. Thus, mobile sources were not dominant for all models for gasoline constituents. Petrochemical point and area source proximity were significant predictors for BTEX and MTBE concentrations in Houston. Ethylbenzene and xylene source proximity were significant predictors in Los Angeles. Close proximity to area sources such as SM or DCF increased the predicted concentrations of MTBE, PCE, and CCl₄ in Houston and Los Angeles. Highway proximity was selected for 6 VOC models with the concentration increased 1.5–2.2 fold with decreasing distance to roadways. This study characterized residential outdoor concentrations of VOCs in urban residential areas in different cities with distinctively different climates and different local geographic profiles of emission sources in each study area successfully. This approach can be used for improving multi-pollutant exposure estimates for epidemiological research in complex urban environments when site-specific meteorological and emission source information are available. Results of this study could assist in the appropriate design of monitoring networks for community-level sampling. They may also improve the accuracy of exposure models linking emission sources with estimated pollutant concentrations at the residential level.

HIGHLIGHTS.

• Source proximity explained more variance in TX while meteorology explained more in CA.

• The range in wind speed caused greater variation in VOC concentrations than proximity.

• Uncontrolled area sources can dominate the ambient VOC levels near residences.

• MTBE and PCE concentrations were elevated in close proximity to scrap metal recyclers.

• Urban site-specific regression VOC models can improve ambient air exposure estimates.

Abbreviations:

1999 NEI

National Emission Inventory of 1999

BTEX

benzene, toluene, ethylbenzene, m,p-xylene, o-xylene

C_out

residential outdoor VOC concentrations

DCF

dry cleaning facilities

GS

gas stations

MTBE

methyl tert butyl ether

PCE

tetrachloroethylene (perchloroethylene)

RH

relative humidity

RIOPA

Relationship among Indoor, Outdoor, and Personal Air

SM

scrap metal recyclers

VOCs

volatile organic compounds