An investigation of the relationship between air emissions of volatile organic compounds and the incidence of cancer in Indiana counties

ARTICLE IN PRESS

Environmental Research 100 (2006) 242–254 www.elsevier.com/locate/envres

An investigation of the relationship between air emissions of volatile organic compounds and the incidence of cancer in Indiana counties Michael L. Boeglin, Denise Wessels, Diane Henshel Environmental Science Research Center, School of Public and Environmental Affairs, Indiana University, 1315 E. Tenth Street, Bloomington, IN 47405, USA Received 23 November 2004; received in revised form 8 April 2005; accepted 26 April 2005 Available online 29 June 2005

Abstract Cancer is a health endpoint influenced by a multitude of factors, including genetic history, individual behavior, and environmental insults. The ubiquity of toxicants in the environment has raised questions about the extent of their role in causing cancer in humans. More specifically, it is desirable to understand the cancer incidence due to airborne toxicants in anthropogenic pollution. One particular class of such pollutants is volatile organic compounds (VOCs). This paper reports an epidemiological investigation of the incidence of cancer in the 92 counties of Indiana. We evaluated the relationship between the amount of VOCs released in each county, as reported by the Toxic Release Inventory, and the county-by-county incidence of various types of cancer, especially those of less common organ systems not directly associated with the absorption or distribution of toxicants. Our evaluation considered chlorinated versus nonchlorinated emissions as well as stack versus fugitive emissions. We evaluated three models: linear, quadratic, and polynomial. Of these, the quadratic model appeared to be the best predictor (highest r2) for most endpoints for which there was a positive correlation. However, the linear model was the most sensitive (lowest P-value) for skin, melanoma, and endocrine-related cancers, including female genital system cancers. Our results indicate a relationship between emissions of VOCs and the incidence of some types of cancers. Most notable were strong correlations between VOC emissions and cancers of the brain, nervous system, endocrine system, and skin. r 2005 Elsevier Inc. All rights reserved. Keywords: Air emissions; Cancer incidence; Volatile organic compounds; VOC; Indiana

1. Introduction In the United States, nearly 15 million new cancer cases were diagnosed from 1990 to 2000, and roughly 1,268,000 new cases were expected to be diagnosed in 2001. These estimates do not include carcinoma in situ (noninvasive cancer) of any site except the urinary bladder and do not include basal and squamous cell skin cancers. More than 1 million cases of basal and squamous cell skin cancers were expected to be diagnosed in 2001. The National Institutes of Health estimate overall costs for cancer in the year 2000 at $180.2 billion (ACS, 2001). Corresponding author. Fax: +812 855 7802.

E-mail address: [email protected] (D. Henshel). 0013-9351/$ - see front matter r 2005 Elsevier Inc. All rights reserved. doi:10.1016/j.envres.2005.04.004

In response to the prevalence of cancer, most cancer research in recent years has focused on understanding the factors that contribute to the occurrence of this disease. It is common knowledge that cancer is a health endpoint influenced by a multitude of factors, including genetic history, individual behavior, and environmental insults. About 5–10% of cancers are clearly hereditary, in that a faulty gene predisposes a person to a higher risk of certain cancers. The remainder is believed to be a result of the impact of gene damage due either to cellular factors, such as hormones or the digestion of nutrients in a cell, or to environmental factors, such as chemicals or sunlight (ACS, 2001). The ubiquity of toxicants in the environment has raised questions about the extent of their role in causing

ARTICLE IN PRESS M.L. Boeglin et al. / Environmental Research 100 (2006) 242–254

cancer in humans. In particular, it is desirable to understand the cancer incidence due to airborne toxicants in anthropogenic pollution. One particular class of such pollutants is volatile organic compounds (VOCs). Their influence on the incidence of cancer is difficult to assess because toxic pollutants are only one etiologic determinant of cancer, and only about 2% of cancer cases are believed to be attributable to air pollution exposure (Dockery, 1993; National Cancer Institute, 1992). Nevertheless, this challenge merits investigation and has been embraced by others (Goldberg et al., 1995; Moller et al., 1994; Shy, 1993). The Environmental Science Research Center at Indiana University (Bloomington, IN, USA) has focused research efforts into the study of environmental impacts in the Great Lakes and their health implications. One such study completed in Indiana, one of the states bordering the Great Lakes, demonstrated a link between lung cancer and emissions of metals (McCormack and Petro, unpublished data). A second study has shown incidence rates of other types of cancer higher than the national average. These include leukemia, nonHodgkin’s lymphoma, and biliary, kidney, ovarian, rectal, and testicular cancers (Workman, unpublished data). Many studies have investigated the link between lung cancer and VOCs (Chen et al., 2002; Gordon et al., 1985; Khyshiktyev et al., 1994; O’Neill et al., 1988; Phillips et al., 1999; Preti et al., 1988), but a literature review reveals that few have investigated other types of cancers and their relationship to these emissions. While VOCs may be absorbed via the lungs, they are then transported via the bloodstream to all organs of the body and thereby have the potential to act upon many organs and tissues as either the parent compound or the bioactivated metabolite. The study of the incidence of cancer due to general airborne toxicant exposure, including VOCs, requires the analysis of a number of factors. These include exposure assessment, the delineation of dose–response relationships, the transport and chemistry of toxicants in the atmosphere, and various aspects of mobile and stationary sources of airborne toxicants (Moller et al., 1994). Several research groups, as explained below, have explored these factors and the salient issues characterizing them. 1.1. Issues relevant to exposure and endpoints Efforts to explain the incidence of cancer through exposure to toxic air pollutants should consider how to account for both indoor and outdoor exposures when measuring or estimating personal exposure. Many anthropogenic airborne toxicants are released into outdoor air, whereas most Americans spend the majority of their time indoors (Moller et al., 1994).

243

Nevertheless, indoor exposure to toxic chemicals is likely to contribute more to the etiology of malignancies than outdoor exposure (Wallace, 1991). Colome et al. (1992) and Quackenboss et al. (1989) found a wide range of correlation between indoor and outdoor air pollution concentrations. However, measured personal exposure has been found to typically exceed both indoor and outdoor concentrations, and the relationship among these measurements depends on the VOCs of interest and their sources (Payne-Sturges et al., 2004). Personal exposure to chloroform, a byproduct of drinking water chlorination, is driven by indoor exposure, whereas personal exposure to benzene occurs primarily outdoors due to motor vehicle emissions. In light of these considerations, one must determine how to assess personal exposure to airborne toxicants such as VOCs. Tracking the precise exposure of a large number of individuals to specific air toxicants in the different and unique environments to which they subject themselves presents a formidable task. One practical approach is to calculate the average exposure for a population by determining the ambient air concentration of the toxicant(s) of interest (Dockery, 1993). For relative assessments of exposure, a second approach is to calculate how much of a toxicant is released into ambient air. In the absence of specific exposure data, such an approach allows a first approximation of population exposure. Unfortunately, this oversimplified model disguises a small number of extremely high individual exposures that may in fact be responsible for many occurrences of certain cancers (Dockery, 1993). Without sufficient detail in this area, researchers can only conduct ‘‘exploratory studies of carcinogens in ambient air’’ (Shy, 1993). Dose–response relationships tie together exposure information and selected health endpoints; in the studies discussed here, the endpoint is cancer. Epidemiological studies are responsible for establishing dose–response relationships by investigating human toxicology retrospectively. In contemporary toxicology, animal studies are the primary vehicle for obtaining human dose–response relationships, due to the advantages of investigating cancer etiology mechanistically, prospectively, and with control over exposure. In contrast, epidemiological studies—including simplified exploratory studies, such as the one described in this paper—enable researchers to consider the effects of different human environments on cancer etiology without the uncertainties of animal-to-human extrapolation, notwithstanding the limitations outlined in our discussion of exposure assessment. 1.2. Issues relevant to study design Researchers utilize a number of designs for epidemiological studies, the selection of which depends in part on

ARTICLE IN PRESS 244

M.L. Boeglin et al. / Environmental Research 100 (2006) 242–254

how groups within the population undergo differential exposure. Exposure status within a group can be a function of geographic, career-related, or time-dependent (i.e., age-related) parameters (Dockery, 1993). In order to draw exposure–response relationships for groups of people living within defined geographic areas at single points in time, the cross-sectional design offers the optimum epidemiological approach. Using this design, health endpoint information in the form of cancer mortality or morbidity rates can be crosscorrelated against exposure information in the form of ambient outdoor air concentrations of certain toxicants, such as VOCs (Dockery, 1993). For less frequently occurring cancers, a large number of cancer cases is needed to confidently identify any correlations between cause and effect. Population studies such as this are beneficial for identifying these cancers (e.g., brain and nervous system, thyroid, skin). A significant challenge in using the cross-sectional design for cancer incidence studies is selection of the point in time for which morbidity rates and toxicant concentrations are gathered (Dockery, 1993). Cancer incidence data, obtained from health registries of some local or regional entity, are often victim to a ‘‘time lag between first diagnosis and entry of cases into the registry’’ (Shy, 1993). Furthermore, cancer as a chronic endpoint presents the challenge of variably long latency between human exposure to the instigating toxicant and first diagnosis (Shy, 1993). Both of these conditions demand consideration of whether the endpoint and exposure information truly represent the incidence of induced cancer at the selected point in time. 1.3. Issues relevant to dispersion modeling Given that ambient air concentrations of VOCs within specific geographic locales can be used to average the exposure of a group of individuals, one must determine how to obtain these concentrations. Dispersion models appropriate for various geographic scales have been devised to account for differential processes guiding the transport, fate, and concentrations of pollutants in ambient air (Moller et al., 1994). The fate and transport of VOCs within an area is dependent on several factors, including geotopography, diurnal cycles, humidity, wind currents, and the distribution of VOC sources, including motor vehicle patterns (Bahrami, 2001). Without a dispersion model that can account for such factors, one would have to assume that toxicants emitted from anthropogenic sources are present in equal concentrations everywhere within a geographic area; however, this assumption is undesired, as a ‘‘cause that is universally present has no influence at all on the distribution of disease’’ (Rose, 1987). Therefore, in the absence of modeling data, a quantitative study of

the distribution of toxicants within a single geographic locale is not feasible and gives way to a qualitative comparison of the uniform toxicant concentrations among many separate locations. Goldberg et al. (1995) addressed the importance of the spatial distribution of toxicants near stationary sources of anthropogenic pollution while investigating how cancer incidence is associated with exposure to biogas from a municipal solid waste landfill in Montreal, Quebec, Canada. Aside from methane and carbon dioxide, typical biogas contaminants might include dimethyl sulfide and carbon disulfide, as well as vinyl chloride, benzene, and other VOCs (Bertacchi et al., 1997; Goldberg et al., 1995). Although Goldberg et al. did not have access to actual exposure data determined by ambient air concentrations of toxicants, they used proximity to the site of the landfill as a surrogate for relative exposure and found an association between the living distance of cancer victims from the site and the incidence of six types of cancer. Cancers of the liver, stomach, bronchus, trachea, and lung in men, as well as cancers of the stomach and cervix in women, were significantly increased and occurred with a relative risk greater than 1.0. Berry and Bove (1997) similarly used residential distance from a hazardous waste landfill in New Jersey (USA), in the absence of actual exposure data, as a surrogate for the inhalation of VOCs and other airborne toxicants measured offsite. During the 5 years of greatest estimated landfill releases (1971–1975), average birth weights in the area nearest and most immediately downwind of the landfill site were found to be statistically significantly lower (192 g) than average birth weights in an adjoining area further from the site. A qualitative question regarding the quantification of airborne toxicants is how one should accurately assess the cancer incidence due to the exposure to complex mixtures of toxicants. Indicator compounds, the presence of which in an environment represents the presence of a larger class of compounds, provide a suitable solution to this problem. Many mutagenic species capable of initiating and progressing cancer are known to be reaction products of VOC degradation, as with the metabolites of trichloroethylene and vinyl chloride (Clewell et al., 1995; Moore and HarringtonBrock, 2000). Current research has shown that semivolatile organic species, which are similar but less volatile than VOCs, may not pose a concern for human health risks (Moller et al., 1994). As such, VOCs serve as a viable group of indicator compounds to represent the presence of potentially carcinogenic compounds in the atmosphere. In the present study, we acknowledge and take into account the many limitations and challenges of cancer incidence assessment identified above. This paper

ARTICLE IN PRESS M.L. Boeglin et al. / Environmental Research 100 (2006) 242–254

reports an epidemiological investigation of the cancer incidence in the 92 counties of the state of Indiana. As a state within the Great Lakes Basin, Indiana serves as an excellent study area in which to investigate the effects of abundant industry and commerce on human health. We identified correlations between the concentration of VOCs in the atmosphere and the incidence of various types of cancer, especially those of less common organ systems not typically directly associated with the absorption or distribution of toxicants. Considering the limited role that airborne toxicants play in the incidence of cancer, we do not conclude any cause–effect relationships, but rather identify potential organs of concern.

2. Methods 2.1. Cancer incidence data The year 1996 cancer rate data were obtained from the Indiana State Department of Health (ISDH) online database ‘‘Incidence Rate by Primary Site and Sex: Cases Diagnosed in Indiana, by County’’ (ISDH, 2001). Indiana cancer incidence numbers and age-adjusted cancer incidence rates are presented for the state as a whole and individually for the 92 counties. By convention, cancer incidence rates do not include carcinoma in

245

situ (with the exception of bladder cancer in situ), nor do they include basal and squamous cell carcinomas of the skin. Table 1 lists the cancer classes for which the ISDH provides cancer rate data. The cancers in bold are those that we evaluated in this study. The cancer incidence rate is the number of new cancers occurring in a specified population during a year, expressed as the number of cancers per 100,000 people. It should be noted that the rate can include multiple primary cancers occurring in one individual. The rates are age-standardized to the US 1970 standard million population to allow for comparisons between groups (geographic or demographic) that have different age distributions. Some counties are omitted from subsets of these analyses. Whenever the number of cases of any type of cancer is less than five at the county level, the actual number is not reported to protect the privacy of these individuals, and for this reason we could not calculate incidence rates for these counties. We calculated the rates in this report using data available in the Indiana State Cancer Registry as of January 24, 2001. These cases represented 88.6% of the estimated number of cases to be diagnosed in Indiana in 1996 (ISDH, 2001). If the number of cases of any type of cancer was less than 20, the rate generated was considered ‘‘unstable’’ and was noted as such in the ISDH cancer incidence tables.

Table 1 Cancer rate data available from the Indiana State Department of Health Classes of cancer

Types of cancer

All sites Oral cavity and pharynx

Includes all cancer. Lip; tongue; major salivary glands; floor of mouth; gum and other mouth; nasopharynx; tonsil; oropharynx; hypopharynx; pharynx and other oral cavity. Esophagus; stomach; small intestine; colon excluding rectum; rectum and rectosigmoid; anus, anal canal, and anorectum; liver and intrahepatic bile duct; gallbladder; other biliary; pancreas; retroperitoneum; peritoneum; other digestive organs. Nasal cavity, ear and sinuses; larynx; lung and bronchus; trachea, pleura, and other respiratory organs. Bones and joints. Soft tissue (including heart). Melanoma of the skin; other nonepithelial skin cancers. Breast. Cervix uteri; corpus uteri; uterus, NOS; ovary; vagina; vulva; other female genital organs. Prostate gland; testis; penis; other male genital organs. Urinary bladder; kidney and renal pelvis; ureter; other urinary organs. Eye and orbit. Brain; other nervous system. Thyroid gland; other endocrine system (including thymus). Hodgkin’s disease; non-Hodgkin’s lymphoma. Multiple myeloma. Acute lymphocytic; chronic lymphocytic; acute myeloid; chronic myeloid; other leukemias. Other, ill-defined and unknown sites

Digestive system

Respiratory system Bones and joints Soft tissue (including heart) Skin (excluding basal and squamous) Breast Female genital system Male genital system Urinary system Eye and orbit Brain and other nervous system Endocrine system Lymphoma Multiple myeloma Leukemia Other, ill-defined and unknown sites Cancers in bold were evaluated in this study.

ARTICLE IN PRESS 246

M.L. Boeglin et al. / Environmental Research 100 (2006) 242–254

2.2. Data on air quality

2.4. Statistical modeling

The pollutants that we selected for modeling included VOCs reported in the US Environmental Protection Agency’s (US EPA’s) Toxic Release Inventory (TRI), which is available online (TRI Explorer, Version 3.2; US EPA, 2001). TRI Explorer is a database of more than 600 toxic chemicals that are being used, manufactured, treated, transported, or released into the environment. Manufacturing facilities that have 10 or more full-time employees and manufacture or process over 25,000 pounds of the approximately 600 listed chemicals or 28 chemical categories, or use more than 10,000 pounds of any chemical specified in the regulations (40 CFR 372.65), must report their releases annually. Releases for the facilities that report emissions are included in the TRI Explorer database. A baseline year of 1988 was chosen for modeling. This year was the oldest for which TRI data was available and represents the relevant exposure period, which is approximately a decade prior to the initiation of 1996 new cancers. We reviewed TRI emissions data and tallied emissions of VOCs for each county in the state of Indiana. For further evaluation, we divided VOCs into chlorinated and nonchlorinated VOCs and categorized them as either stack or fugitive emissions. Stack emissions are those emissions from point sources and are typically released above human head level. Fugitive emissions are those emissions not captured and emitted from a specific point source. Many fugitive emissions escape closer to human head level.

We performed regression analysis using Statistical Analysis Software (SAS) Release 8.02 (SAS Institute Inc., Cary, NC, USA). All correlation results were obtained using the regression procedure (PROC REG). We evaluated three models in this study to determine whether a correlation exists between emission of VOCs and the incidence rate of cancer: A linear model :

Y ¼ mX þ b;

(1)

where Y is the cancer incidence rate, X the pounds of emissions/square mile, m the slope of the dose–response curve, and b is the y intercept. A quadratic model :

Y ¼ aX 2 þ bX þ C,

(2)

where Y is the cancer incidence rate, X the pounds of emissions/square mile, C is a constant, and a, b are coefficients. A polynomial model :

Y ¼ aX 2 þ C,

(3)

where Y is the cancer incidence rate, X the pounds of emissions/square mile, C is a constant and a is a coefficient. Emission rates entered into the models (for X) included total VOC emissions, total fugitive VOCs, and total stack VOC emissions. We further evaluated emissions for each of these groups by subdividing them into chlorinated versus nonchlorinated VOCs.

3. Results 2.3. Comparison of cancer rates and VOC emissions 3.1. Linear model For each Indiana county, we compared the total TRI VOC emissions rate per square mile to the rate of cancer for the following classes of cancer: all sites, oral cavity and pharynx, digestive system, respiratory system, skin (excluding basal and squamous), breast, female genital system, male genital system, urinary system, brain and other nervous system, endocrine system, lymphoma, multiple myeloma, and leukemia. We evaluated the following specific cancers: lung, thyroid, and melanoma. By subtracting the rate of lung cancer from the rate for all respiratory system cancers, we obtained nonlung respiratory cancer incidence rates. The same applied to nonthyroid cancers of the endocrine system and nonmelanomas of the skin category. Additionally, we compared the rates of cancer for each cancer class to the emission rates for chlorinated and nonchlorinated VOCs for fugitive and stack emission sources separately. The study did not include counties for which there were no reported emissions. Also, counties for which there was no reported cancer rate for a specific cancer class are not included in the evaluation for that class of cancer.

Summary statistics for the three models are shown in Tables 2–7. The most statistically significant correlations from the linear model (P-values less than a ¼ 0:05) are displayed in Table 2. Table 3 ranks the marginally statistically significant correlations (P-values of 0:05pap0:10). The results in Tables 2 and 3, as well as in all subsequent tables for the other models, are listed first alphabetically by cancer type and then in order of decreasing significance within each cancer type. We list and identify coefficients in the tables using the nomenclature provided above. These results indicate that the incidence of some types of cancer can be correlated with emissions of VOCs. These data show that 29 of the 180 correlations from the linear model were statistically significant at the level of a ¼ 0:05, and 8 additional correlations were significant at a ¼ 0:10. The most significant result was the correlation of brain and nervous system cancer with nonchlorinated stack emissions, with a P-value of 0.0006 (n ¼ 24). The R2-value of 0.4232 indicates that approximately 42.32% of the variation in brain and

ARTICLE IN PRESS M.L. Boeglin et al. / Environmental Research 100 (2006) 242–254

247

Table 2 Linear model results with highest statistical significance (a ¼ 0:05) P-value

Dependent

Independent

R2

n

m (
104)

b

0.0006 0.0014 0.0434 0.0015 0.0056 0.0064 0.0076 0.0098 0.0149 0.0238 0.0281 0.0026 0.0033 0.0036 0.0075 0.0030 0.0064 0.0078 0.0203 0.0224 0.0231 0.0231 0.0039 0.0058 0.0059 0.0129 0.0291 0.0477 0.0489

Brain and nervous system Brain and nervous system Brain and nervous system Melanoma Melanoma Melanoma Melanoma Melanoma Melanoma Melanoma Melanoma Thyroid Thyroid Thyroid Thyroid Skin Skin Skin Skin Skin Skin Skin Endocrine Endocrine Endocrine Endocrine Female genital system Female genital system Female genital system

Nonchlorinated stack Total stack Nonchlorinated total Total emissions Total fugitive Chlorinated total Total stack Nonchlorinated total Chlorinated fugitive Nonchlorinated fugitive Nonchlorinated stack Nonchlorinated stack Total stack Nonchlorinated total Total emissions Total emissions Total fugitive Nonchlorinated total Nonchlorinated fugitive Chlorinated fugitive Nonchlorinated stack Total stack Nonchlorinated stack Total stack Nonchlorinated total Total emissions Nonchlorinated stack Nonchlorinated total Total stack emissions

0.4232 0.3761 0.1727 0.2249 0.1768 0.2044 0.1649 0.1592 0.1716 0.1243 0.1177 0.516 0.4975 0.4917 0.4356 0.1875 0.1603 0.1569 0.1216 0.1401 0.1169 0.1143 0.4147 0.3871 0.3868 0.3284 0.0745 0.0590 0.0602

24 24 24 42 42 35 42 41 34 41 41 15 15 15 15 45 45 44 44 37 44 45 18 18 18 18 64 67 65

7.367 6.652 3.821 5.192 7.770 12.30 8.014 5.971 17.30 9.856 8.419 7.013 6.150 3.810 2.767 4.513 7.319 5.826 9.601 15.5 8.231 5.818 6.151 5.220 3.373 2.410 6.355 4.326 5.250

5.877 5.627 6.408 13.78 13.02 13.29 13.69 13.42 13.00 12.72 13.38 7.533 7.651 7.244 7.230 14.40 13.72 14.14 13.46 13.68 14.12 14.12 7.406 7.501 7.161 7.161 25.67 25.69 25.51

Table 3 Linear model results with less statistical significance (0:05pap0:10) P-value

Dependent

Independent

R2

n

m (
104)

b

0.0609 0.0727 0.0622 0.0730 0.0755 0.0756 0.0979 0.0986

Thyroid Thyroid Skin Endocrine Myeloma Melanoma Oral Urinary system

Nonchlorinated fugitive Total fugitive Chlorinated total Nonchlorinated fugitive Total emissions Chlorinated stack Total stack Nonchlorinated fugitive

0.2446 0.2268 0.0933 0.1871 0.2823 0.1123 0.0887 0.039

15 15 38 18 12 29 32 71

4.410 2.964 6.573 3.960 1.807 19.80 4.617 9.062

6.278 6.349 13.61 6.393 6.163 12.92 12.19 29.62

nervous system cancer among Indiana’s counties in 1996 can be explained by nonchlorinated stack emissions in 1988 when using the linear model. The least significant result at a ¼ 0:10 is the correlation of urinary system cancer with nonchlorinated fugitive emissions, with a Pvalue of 0.0986 (n ¼ 71). The R2-value of 0.0390 indicates that approximately 3.90% of the variation in skin cancer among Indiana’s counties can be explained by nonchlorinated fugitive emissions.

3.2. Polynomial model The most statistically significant correlations from the two-term polynomial model (P-values less than a ¼ 0:05) are summarized and ranked in Table 4. Table 5 ranks the marginally statistically significant correlations (P-values of 0:05pap0:10). These results show that 21 of the 180 correlations from the polynomial model were statistically significant

ARTICLE IN PRESS M.L. Boeglin et al. / Environmental Research 100 (2006) 242–254

248

Table 4 Two-term polynomial model results with highest statistical significance (a ¼ 0:05) P-value

Dependent

Independent

R2

n

a (
108)

C

o0.0001 o0.0001 0.0037 0.0059 0.0106 0.0132 0.0272 0.0386 0.0088 0.0217 0.0293 0.0155 0.0171 0.0184 0.0362 0.0159 0.0163 0.0201 0.0278 0.0462 0.049

Brain and nervous system Brain and nervous system Brain and nervous system Melanoma Melanoma Melanoma Melanoma Melanoma Skin Skin Skin Endocrine system Endocrine system Endocrine system Endocrine system Thyroid Thyroid Thyroid Thyroid Female genital system Female genital system

Nonchlorinated stack Total stack Nonchlorinated total Total emissions Total fugitive Total stack Nonchlorinated total Chlorinated total Total emissions Nonchlorinated total Total fugitive Nonchlorinated stack Total stack Nonchlorinated total Total emissions Nonchlorinated stack Total stack Nonchlorinated total Total emissions Total stack Nonchlorinated stack

0.5696 0.5402 0.3245 0.1748 0.2079 0.199 0.1189 0.1233 0.149 0.1193 0.1057 0.3141 0.3065 0.301 0.2463 0.3714 0.3692 0.3503 0.3206 0.0616 0.0611

24 24 24 42 42 42 41 35 45 44 45 18 18 18 18 15 15 15 15 65 64

4.460 4.293 2.916 2.981 5.893 7.682 4.016 15.73 2.567 3.910 5.488 6.925 5.196 2.098 1.229 7.346 5.612 2.199 1.368 3.356 3.481

6.964 6.746 6.814 12.87 12.45 12.75 12.61 12.58 13.59 13.30 13.18 6.656 6.722 6.509 6.546 6.542 6.592 6.379 6.463 26.25 26.28

Table 5 Two-term polynomial model results with less statistical significance (0:05pap0:10) P

Dependent

Independent

R2

n

a (
108)

C

0.0601 0.0801 0.0665 0.0747 0.0684 0.0709

Skin Skin Melanoma Melanoma Female genital system Brain and nervous system

Nonchlorinated fugitive Nonchlorinated stack Nonchlorinated fugitive Chlorinated fugitive Nonchlorinated total Total emissions

0.0817 0.0711 0.0838 0.0959 0.0502 0.1407

44 44 41 34 67 24

8.646 9.252 8.943 30.06 2.655 1.625

12.98 13.29 12.28 12.45 26.10 6.956

at the level of a ¼ 0:05 and that 6 additional correlations were marginally significant at a ¼ 0:10. These data reveal the smallest P-value of o0.0001 for two correlations: brain and nervous system cancer with nonchlorinated stack emissions (n ¼ 24) and brain and nervous system with total stack emissions (n ¼ 24). The R2-value of 0.5695 for the first of these indicates that approximately 56.95% of the variation in brain and nervous system cancer among Indiana’s counties in 1996 can be explained by nonchlorinated stack emissions in 1988 when using the polynomial model. The least statistically significant result at a ¼ 0:10 is the correlation of skin cancer with nonchlorinated stack emissions, with a P-value of 0.0801 (n ¼ 44). The R2-value of 0.0711 indicates that approximately 7.11% of the variation in skin cancer among Indiana’s counties can be explained by nonchlorinated stack emissions.

3.3. Quadratic model The most statistically significant correlations from the three-term quadratic model (P-values less than a ¼ 0:05) are summarized in Table 6. Table 7 shows the marginally statistically significant correlations (P-values of 0:05pap0:10). These results show that 28 of the 180 correlations from the quadratic model were statistically significant at the level of a ¼ 0:05 and that 13 additional correlations were significant at a ¼ 0:10. These data reveal the smallest P-value of o0.0001 for the same two correlations that were most significant using the polynomial model: brain and nervous system cancer with nonchlorinated stack emissions (n ¼ 24) and brain and nervous system cancer with total stack (n ¼ 24). The R2value of 0.6019 for the first of these indicates that approximately 60.19% of the variation in brain and

ARTICLE IN PRESS M.L. Boeglin et al. / Environmental Research 100 (2006) 242–254

249

Table 6 Three-term quadratic model results with highest statistical significance (a ¼ 0:05) P

Dependent

Independent

R2

n

a (
108)

b (
103)

C

o0.0001 o0.0001 0.0011 0.0055 0.0068 0.0106 0.0132 0.0219 0.0339 0.0076 0.0084 0.0124 0.0132 0.0081 0.011 0.0112 0.0153 0.0273 0.0281 0.0478 0.0152 0.0214 0.02395 0.0315 0.0195 0.0397 0.0308 0.0395

Brain and nervous system Brain and nervous system Brain and nervous system Melanoma Melanoma Melanoma Melanoma Melanoma Melanoma Thyroid Thyroid Thyroid Thyroid Skin Skin Skin Skin Skin Skin Skin Endocrine system Endocrine system Endocrine system Endocrine system Leukemia Leukemia Lymphoma Urinary system

Nonchlorinated stack Total stack Nonchlorinated total Total emissions Chlorinated total Total fugitive Total stack Chlorinated fugitive Nonchlorinated total Nonchlorinated stack Nonchlorinated total Total stack Total emissions Chlorinated total Total emissions Total fugitive Total stack Chlorinated fugitive Nonchlorinated total Nonchlorinated fugitive Nonchlorinated stack Nonchlorinated total Total stack Total emissions Chlorinated fugitive Chlorinated total Total emissions Total emissions

0.6019 0.5748 0.4783 0.2344 0.2681 0.2079 0.199 0.2186 0.1632 0.5566 0.5492 0.5189 0.5137 0.2406 0.1934 0.1927 0.1804 0.1909 0.1599 0.1378 0.4279 0.4012 0.3921 0.3695 0.2244 0.1879 0.103 0.087

24 24 24 42 35 42 42 34 41 15 15 15 15 38 45 45 45 37 44 44 18 18 18 18 34 34 67 74

7.440 7.004 8.909 2.170 30.20 9.729 12.66 55.94 2.171 7.249 2.863 3.807 2.630 31.31 1.653 9.821 14.33 57.44 1.643 11.49 3.980 1.404 1.814 1.844 40.29 13.14 11.71 6.752

0.6061 0.5424 1.14 0.8349 2.98 1.78 1.75 3.97 0.8522 1.254 0.779 0.951 0.718 3.03 0.6887 1.74 1.78 3.87 0.7678 1.92 0.903 0.526 0.673 0.542 3.39 1.90 1.89 1.39

8.029 7.901 9.169 14.24 14.01 13.59 14.46 13.46 13.69 8.124 7.942 8.096 8.164 15.10 14.75 14.30 15.19 14.25 14.30 13.73 7.671 7.453 7.674 7.759 12.935 13.101 22.31 31.40

Table 7 Three-term quadratic model results with less statistical significance (0:05pap0:10) P

Dependent

Independent

R2

n

a (
108)

b (
103)

C

0.0502 0.0886 0.0984 0.0525 0.0555 0.0561 0.0659 0.0777 0.0804 0.0697 0.0726 0.0943 0.0979

Myeloma Myeloma Myeloma Skin Melanoma Melanoma Urinary system Urinary system Urinary system Brain and nervous system Thyroid Female genital system Lymphoma

Total emissions Total fugitive Nonchlorinated Nonchlorinated Nonchlorinated Nonchlorinated Total fugitive Nonchlorinated Nonchlorinated Total emissions Total fugitive Nonchlorinated Total fugitive

0.4856 0.4164 0.4027 0.1339 0.1411 0.1407 0.0758 0.0724 0.0695 0.2241 0.3541 0.0745 0.0711

12 12 12 44 41 41 72 71 73 24 15 64 66

3.299 8.905 12.46 11.92 14.90 12.42 19.95 30.38 8.345 5.954 8.386 0.01063 28.69

0.7678 1.26 1.42 1.56 1.77 2.04 2.61 3.21 1.47 0.8335 1.249 0.6632 2.95

7.761 6.716 7.672 14.55 13.96 13.09 30.73 30.45 30.90 9.044 7.155 25.51 21.08

stack stack stack fugitive fugitive total

stack

nervous system cancer among Indiana’s counties in 1996 can be explained by nonchlorinated stack emissions in 1988 when using the quadratic model. The least statistically significant result at a ¼ 0:10 is the correlation of myeloma with nonchlorinated stack emissions, with a P-value of 0.0984 (n ¼ 12). The R2-value of 0.4027 indicates that approximately 40.27% of the variation in myeloma among Indiana’s counties can be explained by nonchlorinated stack emissions.

4. Discussion 4.1. Trends in the results We have found correlations between VOCs and the incidence of some types of cancer. A review of data in Tables 2, 4 and 6 reveals that cancers of the brain and nervous system produced the most significant correlations in all three models. Aside from these, the

ARTICLE IN PRESS 250

M.L. Boeglin et al. / Environmental Research 100 (2006) 242–254

remaining significant correlations involve cancers almost exclusively related to the skin and endocrine systems (skin, melanoma, endocrine system, and thyroid cancers). The other cancers for which there were significant correlations in at least one model include urinary system, female genital system, lymphoma, leukemia, and oral cavity. Cancers not significantly predicted by any TRI emission rates include pharynx, digestive, lung, respiratory, male genital, and breast cancers. An unexpected finding is that respiratory cancer (respiratory/nonlung and lung) was not significantly correlated with any TRI emissions (a ¼ 0:10). This may be due to an insufficient time lag between the exposure data (1988) and the health data (1996), as respiratory cancers are believed to have a latency period of one to three decades (Bilello et al., 2002; Hauptmann et al., 2001; Steenland and Palu, 1999). Following absorption in the lungs, VOCs are distributed systemically throughout the body and bioactivated by ubiquitous cytochrome P450 enzymes, including CYP1A1, CYP1A2, and CYP2E1 (Eaton and Klaassen, 2001). Byproducts of these P450 reactions can include reactive oxygen species, which can also induce cellular damage. It is reasonable to suppose that the metabolites or byproducts of VOC biotransformation can inflict carcinogenic insults in tissues of the brain, nervous system, endocrine system, and skin at least as well as in respiratory tissue; however, a review of current literature reveals that knowledge in this area is limited. All three models produced a number of significant correlations between TRI emissions and cancers of the endocrine system and thyroid. In contrast, no significant correlations were produced with any model for nonthyroid endocrine system cancers. This trend suggests that cancers of the endocrine system produced significant correlations largely, if not exclusively, due to their inclusion of thyroid cancers. Other trends in the results pertain to emissions. Only nine of the most significant correlations from the three models involved purely chlorinated VOCs, and these were almost exclusively in the linear and quadratic models. Furthermore, all but one of those nine results were correlated with cancers of the skin (melanoma and skin cancer). The one exception is leukemia, which was correlated with chlorinated fugitive emissions by the three-term quadratic model. This suggests that the only cancers that may be significantly predicted by exposure to chlorinated VOC emissions are those of the skin. A second implication of this pattern is that all of the remaining types of cancers significantly correlated with TRI emissions are predicted by exposure to nonchlorinated rather than chlorinated emissions. Even though the three total emissions categories (total fugitive, total stack, and total) include chlorinated VOCs, it appears that their presence among significant correlations may

be due only to their nonchlorinated components. In other words, it is possible that total stack, total fugitive, and total emissions can be used to significantly predict cancer only because they include nonchlorinated VOCs. The disparity between significant correlations involving chlorinated and nonchlorinated VOCs suggests merit in distinguishing between them when investigating their relationship to cancer incidence. While toxic compounds from both classes share many of the same target tissues and biotransformation pathways, it is reasonable to suppose that unique reactive intermediates from some chlorinated VOCs may mediate carcinogenic insults differently from nonchlorinated VOCs and associated metabolites. Trichloroethylene and perchloroethylene, two of the most prevalent chlorinated hydrocarbons still used extensively in industry during the time frame of our analysis, are thought to mediate toxicity primarily via their metabolites (Clewell et al., 1995; Eaton and Klaassen, 2001). In the case of our results, more focused research is needed to investigate whether toxicology can explain our significant correlations between chlorinated VOCs and cancers of the skin. 4.2. Choice of significance levels To choose a significance level for the interpretation of the results of the regression analyses, we considered the difficulty inherent in conducting a meaningful study of the relationship between VOC emissions and cancer rates, as previously discussed. Due to this difficulty, we cannot expect most of the variation in cancer rates among counties to be explained by the three models. Thus, in the statistical computations performed with SAS, the variation reflected by the sum of squares from error is expected to grossly surpass the sum of squares from regression, giving a small f-value for the F-test. Because the expectation for strong correlations is minimal, a very small F statistic is desired in order to identify the most important correlations, relatively speaking, even if those correlations do not reveal a great deal about the relationship of cancer rates to emission of VOCs. Thus, we selected a relatively highlevel a test for the study, with a ¼ 0:10. However, in the interest of singling out unexpectedly strong correlations, a tabulation of correlations with an F statistic corresponding to a ¼ 0:05 was also completed. By the above rationale, the results of this study were surprising. The number of correlations deemed significant by the a ¼ 0:05 test, for all three models, was not much smaller than the number of less significant correlations by the a ¼ 0:10 test. That is to say, a larger number of correlations fell within the category of a Pvalue less than a ¼ 0:05 than correlations with a P-value between a ¼ 0:05 and 0.10. Overall, most of the results were either very strongly or very weakly correlated, with few intermediate correlations. It is likely that factors

ARTICLE IN PRESS M.L. Boeglin et al. / Environmental Research 100 (2006) 242–254

other than VOC emissions, but aligned closely to emissions in their effect on cancer incidence, play a strong role as confounding factors in the relationship between emissions and cancer incidence. Such factors may include exposure to airborne VOCs emitted by nonTRI sources (e.g., mobile and stationary sources) and exposure to industrial pollutants through drinking water. Without evaluating confounding factors in this study, we acknowledge the challenge of determining why such factors have a strong presence in some correlations and not others.

4.3. Comparison of regression models Of the three models tested, the linear and quadratic models produced similar results, in terms of both the number (linear, 29; quadratic, 28) and breadth of health endpoints for which there was a significant correlation. The polynomial model did not generate results for any endpoints, of any statistical significance, that were not also significant at some level in the linear and quadratic models. Urinary system cancer was significantly correlated to total, total fugitive, nonchlorinated fugitive, and nonchlorinated total emissions in the linear and quadratic models but was not at all significantly correlated to any emissions in the polynomial model. Similarly, myeloma was significantly correlated only marginally to total, total fugitive, and nonchlorinated stack emissions in the linear and quadratic models, while it was not at all significant in the output of the polynomial model. The linear and quadratic models produced other unique correlations on an individual basis. Lymphoma (correlated to total emissions) and leukemia (correlated to chlorinated fugitive and chlorinated total emissions) were only statistically significant in the quadratic model. Similarly, oral cancers (correlated to total stack emissions) were only even marginally significant in the linear model. Comparisons between the models in terms of correlation coefficients for groups of endpoints reveal that the quadratic model was clearly superior for cancers of the brain and nervous system. For all significant correlations of thyroid and endocrine system cancers to TRI emissions, the quadratic model produced the highest correlation coefficients, followed consistently by the linear model and then the polynomial model. Precisely the same pattern was obtained for skin cancer and melanoma, with exceptions being the correlations of melanoma to total fugitive and total stack emissions. Comparisons between the models with regard to Pvalue for groups of endpoints also reveal a consistent pattern. For all significant correlations of thyroid, endocrine system, melanoma, and skin cancers, the linear model produced lower P-values than the quad-

251

ratic or polynomial models, making it the most sensitive model for these groups of endpoints. When TRI emissions are used as a surrogate for exposure to VOCs, the three models represent dose– response relationships with different curvature, each intended to capture a different sensitivity of cancer morbidity rates to variable VOC emissions. The quadratic and polynomial models, with a term of squared emissions, potentially embody a greater sensitivity of cancer rate toward emissions than does the linear model. Our findings, however, overshadow this distinction, as evidenced by the similarity of cancer endpoints for which we found significant correlations in the quadratic and linear models. As discussed above, only a subset of endpoints commonly significant in the linear and quadratic models was also significant in the polynomial model. Furthermore, among significant correlations in the quadratic model, the coefficients for the linear term consistently dominated coefficients for the squared term by approximately five orders of magnitude. The linear term, common to the linear and quadratic models but not the polynomial model, appears to drive the shape of our dose–response relationships, making the relationships more or less linear. Overall, the quadratic model was the most predictive model for the relationship between cancer and TRI emissions for all classes of endpoints except cancers of the oral cavity and female genital system. The quadratic model produced only one result for the female genital system that was even marginally significant (ap0:10), which was the correlation to nonchlorinated stack emissions, whereas the linear and polynomial models each produced two significant results (ap0:05) for this endpoint (nonchlorinated total and total stack emissions, and total stack and nonchlorinated stack emissions, respectively.) 4.4. Limitations of the research A limitation of the research is the incompleteness of the cancer incidence rate data obtained from the ISDH. The ISDH does not report rates for counties in which the incidence of a particular cancer is less than 5. Additionally, the ISDH considers the incidence rate data unstable when the incidence is less than 20. Thus, our power to detect significant correlations decreased for cancers with a low incidence rate in many of the counties in Indiana. Other limitations pertain to the use of TRI data. TRI only applies to facilities that emit large volumes of specific contaminants, such that the data include only emissions from facilities with at least 10 employees that release in excess of 25,000 pounds of a listed chemical annually. Smaller facilities, and those that emit less than 25,000 pounds of a chemical annually, do not report, but they are nonetheless likely to contribute substantial

ARTICLE IN PRESS 252

M.L. Boeglin et al. / Environmental Research 100 (2006) 242–254

levels of air pollutants when considered collectively. Facilities required by regulation to report may in fact not report emissions. Emission rates are self-reported and therefore subject to individual facility estimations and assumptions in calculating emissions. In addition, the addresses provided in TRI reports sometimes do not match the physical locations of polluting facilities, although we do not know the extent of this limitation in the case of Indiana facilities. The use of TRI data is also limited by the exclusion of mobile sources of VOC emissions. Jo and Park (1999) and Mohamed et al. (2002) discuss the importance of mobile source emissions of VOCs relative to stationary and natural sources. Lee et al. (2002) found in Hong Kong that the concentration of ambient VOCs varies throughout the region depending on the local mix of stationary and mobile sources. Using ambient air emissions as a surrogate for exposure fails to account for occupational exposures to VOCs. A small percentage of the population is employed as parking attendants, service station attendants, and roadside storekeepers, all of whom experience much higher exposures to nonchlorinated VOCs, such as benzene, toluene, ethyl benzene, and xylenes (Jo and Song, 2001). Workers in other industries are likewise exposed to various subsets of VOCs and account for a fraction of cancer morbidity, the extent of which was not considered in this study. The cross-sectional model used in this study inherently assumes that the cancers reported in 1996 were initiated during a year when emissions were at the 1988 level. Due to the latency of many cancers, such as lung cancer, a wider separation between the two sets of data may have been more appropriate; however, neither ISDH data more recent than 1996 nor an inventory of statewide VOC emissions in Indiana prior to 1988 was available at the time of this study. Further, the model does not correct for the population that may have moved into or out of a county or the state. We also assumed that individuals were diagnosed and treated in the same county in which they were exposed to VOC emissions. In fact, the latter two assumptions are complicated by a large inconsistency among individuals in the time between diagnosis and the entry of a cancer in case registries (Shy, 1993). The nature of emissions data also presented limitations. To compare emissions to cancer rates, the crosssectional model applies a single quantity of VOCs to all people living within a given county. Thus, the model assumes that all people diagnosed with cancer in a county were exposed to the same quantity of VOCs, regardless of the area of the county over which the VOCs migrated. As such, no distinction is made among counties having very different exposure situations. For example, people spread throughout a large county in which most of the facilities reside near one another will

experience a wide range of exposures, whereas those living in a small county with diffuse facilities will likely experience a more uniform exposure. Another limitation inherent in the emissions data is the assumption that emissions from one county make a negligible contribution to the exposure conditions in neighboring counties. In reality, a facility located near the border between two counties may well contribute equally to the exposure of people in both counties. Prevailing wind speed and direction determine the proportion of emissions from a border facility that will reach adjacent counties. The above two limitations of TRI emissions data can be addressed in more focused future studies by incorporating a more formal spatial analysis. Hopenhayn-Rich et al. (2002), who also conducted an epidemiological assessment using county-level data, emphasize the need to use more ‘‘individual-level data’’ to estimate exposures with more accuracy. Reynolds et al. (2003) modeled census tract-level concentrations of hazardous air pollutants as a more geographically specific surrogate for exposure in their study of childhood cancer. As Lee et al. (2002) demonstrated, spatial analyses with high resolution are needed within urban areas to differentiate exposures driven primarily by mobile sources from those controlled more by stationary sources of VOCs. The US Environmental Protection Agency’s National Air Toxics Assessment (NATA) provides an approach for modeling ambient concentrations of air pollutants, as well as population exposures, at the census-tract level (US EPA, 2002). The NATA is conducted for only 32 pollutants from the US Clean Air Act’s list of air toxics, including chlorinated and nonchlorinated VOCs, making it a good candidate for obtaining exposure estimates of high spatial resolution in future studies focusing on the toxic effects of selected pollutants.

5. Conclusions Our results indicate a correlation between TRI emissions of VOCs and the county-level incidence of some types of cancers in Indiana. Most striking were the strong correlations between VOC emissions and cancers of the brain and nervous system, thyroid and other endocrine system, and skin. The paucity of studies on the carcinogenic potential of VOCs and other airborne toxicants in these organ systems suggests that more research is needed to better understand the mechanisms of chronic insults in these areas. A second implication of our results follows from the discrepancy between chlorinated and nonchlorinated emissions in the most significant correlations. The fact that nonchlorinated emissions dominate the results points to the need for greater attention to the control

ARTICLE IN PRESS M.L. Boeglin et al. / Environmental Research 100 (2006) 242–254

of emissions from industrial sources that emit nonchlorinated VOCs as a byproduct of their production processes. These results may be due to chance associations between emissions and some types of cancer due to the limitations inherent in our study design. However, we believe this is unlikely and that these results suggest a course for future epidemiological research. Future epidemiological studies should attempt to account for confounding factors that affect the relationship between VOCs and cancer and evaluate trends over time for these two parameters. A time-series analysis of the cancer incidence rate and VOC emission rate would show the appropriate lag time between rising and falling trends. An analysis of specific VOCs, rather than of groupings of VOCs, could indicate whether a particular chemical is responsible for the cancer incidence and would allow an investigation of the unique roles played by individual chlorinated and nonchlorinated VOCs in cancer incidence. More formal spatial modeling that includes emissions from neighboring counties would account for the effects of emissions entering a county from nearby sources in other counties. A location analysis of counties with a significant correlation of cancer incidence and emission rate would indicate whether there are any geographical relationships that could provide insight into other potential factors in the cancer incidence for such counties.

References ACS, 2001. Cancer Facts & Figures. Am. Cancer Soc. (Atlanta). Bahrami, A.R., 2001. Distribution of volatile organic compounds in ambient air in Tehran. Arch. Environ. Health 56, 380–383. Berry, M., Bove, F., 1997. Birth weight reduction associated with residence near a hazardous waste landfill. Environ. Health Perspect. 105, 856–861. Bertacchi, M., Capuano, F., Fornaciari, S., Franzoni, C., Poluzzi, V., Renna, E., Vivi, B., Bertolini, E., Meglioli, E., 1997. Control of the air quality and research of smelling substances in landfills. In: Environmental Impact, Aftercare and Remediation of Landfills, pp. 155–169. Bilello, K.S., Murin, S., Matthay, R.A., 2002. Epidemiology, etiology, and prevention of lung cancer. Clin. Chest Med. 23, 1–25. Chen, S.J., Wang, J.L., Chen, J.H., Huang, R.N., 2002. Possible involvement of glutathione and p53 in trichloroethylene- and perchloroethylene-induced lipid peroxidation and apoptosis in human lung cancer cells. Free Radical Biol. Med. 33, 464–472. Clewell, H.J., Gentry, P.R., Gearhart, J.M., Allen, B.C., Andersen, M.E., 1995. Considering pharmacokinetic and mechanistic information in cancer risk assessments for environmental contaminants: examples with vinyl chloride and trichloroethylene. Chemosphere 31, 2561–2578. Colome, S.D., Kade, N.Y., Jaques, P., Kleinman, M., 1992. Indoor–outdoor air pollution relations: particulate matter less than 10 mm in aerodynamic diameter (PM10) in homes of asthmatics. Atmos. Environ. A 26, 2173–2178. Dockery, D.W., 1993. Epidemiologic study design for investigating respiratory health effects of complex air pollution mixtures. Environ. Health Perspect. 101 (Suppl. 4), 187–191.

253

Eaton, D.L., Klaassen, C.D., 2001. Chapter 2: principles of toxicology. In: Classen, C.D. (Ed.), Casarett & Doull’s Toxicology: The Basic Science of Poisons, sixth ed. McGraw-Hill, New York. Goldberg, M.S., Al-Homsi, N., Goulet, L., Riberdy, H., 1995. Incidence of cancer among persons living near a municipal solid waste landfill site in Montreal, Quebec. Arch. Environ. Health 50, 416–424. Gordon, S.M., Szidon, J.P., Krotoszynski, B.K., Gibbons, R.D., O’Neill, H.J., 1985. Volatile organic compounds in exhaled air from patients with lung cancer. Clin. Chem. 31, 1278–1282. Hauptmann, M., Berhane, K., Langholz, B., Lubin, J., 2001. Using splines to analyse latency in the Colorado Plateau uranium miners cohort. J. Epidemiol. Biostat. 6, 417–424. Hopenhayn-Rich, C., Stump, M.L., Browning, S.R., 2002. Regional assessment of atrazine exposure and incidence of breast and ovarian cancers in Kentucky. Arch. Environ. Contam. Toxicol. 42, 127–136. ISDH, 2001. Incidence rate by primary site and sex: cases diagnosed in Indiana, by county. Available at URL http://www.IN.gov/isdh/ dataandstats/cancerinc/cancer_inc_index.htm (accessed 15 November 2001). Indiana Department of Health, IN, USA. Jo, W.K., Park, K.H., 1999. Commuter exposure to volatile organic compounds under different driving conditions. Atmos. Environ. 33, 409–417. Jo, W.K., Song, K.B., 2001. Exposure to volatile organic compounds for individuals with occupations associated with potential exposure to motor vehicle exhaust and/or gasoline vapor emissions. Sci. Total Environ. 269, 25–37. Khyshiktyev, B.S., Khyshiktueva, N.A., Ivanov, V.N., Darenskaia, S.D., Novikov, S.V., 1994. Diagnostic value of investigating exhaled air condensate in lung cancer (in Russian). Vopr. Onkol. 40, 161–164. Lee, S.C., Chiu, M.Y., Ho, K.F., Zou, S.C., Wang, X., 2002. Volatile organic compounds in urban atmosphere of Hong Kong. Chemosphere 48, 375–382. Mohamed, M.F., Kang, D., Aneja, V.P., 2002. Volatile organic compounds in some urban locations in United States. Chemosphere 47, 863–882. Moller, L., Schuetzle, D., Autrup, H., 1994. Future research needs associated with the assessment of potential human health risks from exposure to toxic ambient air pollutants. Environ. Health Perspect. 102 (Suppl. 4), 193–210. Moore, M.M., Harrington-Brock, K., 2000. Mutagenicity of trichloroethylene and its metabolites: implications for the risk assessment of trichloroethylene. Environ. Health Perspect. 108, 215–224. National Cancer Institute (NCI), 1992. Cancer Statistics Review: 1973–1989. Publication No. 92-2789. Division of Cancer Prevention and Control, NCI, Bethesda, MD, USA. O’Neill, H.J., Gordon, S.M., O’Neill, M.H., Gibbons, R.D., Szidon, J.P., 1988. A computerized classification technique for screening for the presence of breath biomarkers in lung cancer. Clin. Chem. 34, 1613–1618. Payne-Sturges, D.C., Burke, T.A., Breysse, P., Diener-West, M., Buckley, T.J., 2004. Personal exposure meets risk assessment: a comparison of measured and modeled exposures and risks in an urban community. Environ. Health Perspect. 112, 589–598. Phillips, M., Gleeson, K., Hughes, J.M.B., Greenberg, J., Cataneo, R.N., Baker, L., et al., 1999. Volatile organic compounds in breath as markers of lung cancer: a cross-sectional study. Lancet 353, 1930–1933. Preti, G., Labows, J.N., Kostelc, J.G., Aldinger, S., Daniele, R., 1988. Analysis of lung air from patients with bronchogenic carcinoma and controls using gas chromatography mass spectrometry. J. Chromatogr. 432, 1–11.

ARTICLE IN PRESS 254

M.L. Boeglin et al. / Environmental Research 100 (2006) 242–254

Quackenboss, J.J., Lebowitz, M.D., Crutchfield, C.D., 1989. Indoor– outdoor relationships for particulate matter: exposure classifications and health effects. Environ. Int. 15, 353–360. Reynolds, P., Von Behren, J., Gunier, R.B., Goldberg, D.E., Hertz, A., Smith, D.F., 2003. Childhood cancer incidence rates and hazardous air pollutants in California: an exploratory analysis. Environ. Health Perspect. 111, 663–668. Rose, G., 1987. Environmental factors and disease: the man-made environment. Br. Med. J. 294, 963–965. Shy, C.M., 1993. Epidemiologic studies of neurotoxic, reproductive, and carcinogenic effects of complex mixtures. Environ. Health Perspect. 101 (Suppl. 4), 183–186.

Steenland, K., Palu, S., 1999. Cohort mortality study of 57,000 painters and other union members: a 15 year update. Occup. Environ. Med. 56, 315–321. US EPA, 2001. TRI Explorer, Version 3.2. Available at URL http:// www.epa.gov/triexplorer/reports.htm [accessed 15 November 2001]. Office of Environmental Information, US Environmental Protection Agency, Washington, DC. US EPA, 2002. National Air Toxics Assessment. Available at URL http://www.epa.gov/ttn/atw/nata/ [accessed 16 February 2005]. US Environmental Protection Agency, Washington, DC. Wallace, L., 1991. Comparison of risks from outdoor and indoor exposure to toxic chemicals. Environ. Health Perspect. 95, 7–13.