Environ. Sci. Technol. 2007, 41, 8498–8505
Measured and Modeled Personal Exposures to and Risks from Volatile Organic Compounds R O B I N E . D O D S O N , * ,† E . A N D R E S H O U S E M A N , †,‡ JONATHAN I. LEVY,† JOHN D. SPENGLER,† JAMES P. SHINE,† A N D D E B O R A H H . B E N N E T T †,§ Department of Environmental Health, Harvard School of Public Health, Boston, Massachusetts, Department of Work Environment, University of Massachusetts, Lowell, Massachusetts, and Department of Public Health Sciences, University of California, Davis, California
Received May 14, 2007. Revised manuscript received September 12, 2007. Accepted September 17, 2007.
We developed a personal exposure model using volatile organic compound data collected for teachers and office workers as part of the Boston Exposure Assessment in Microenvironments (BEAM) study. We included participantspecific time-activity and concentration measurements of residential outdoor, residential indoor, and workplace microenvironments, along with average concentrations in various dining, retail, and transportation microenvironments. We used a series of time-weighted personal exposure models to compare measured personal concentrations using median regression models, with bias estimates representing the difference between measured and modeled personal exposures. Incorporating only the outdoor microenvironment results in an unbiased estimate of personal exposure only for carbon tetrachloride. Adding the residential indoor microenvironment provides an unbiased estimate for trichloroethene as well. A model incorporating residential outdoor, indoor, and workplace microenvironments provides an unbiased estimate for the above compounds and chloroform, 1,4-dichlorobenzene, benzene, and R-pinene, and adding the transportation microenvironment adds ethylbenzene. A fully saturated model, including outdoor, indoor,workplace,transportation,andallothermicroenvironments, provides an unbiased estimate for the previously listed compounds along with tetrachloroethene and styrene. MTBE, toluene, o-xylene, d-limonene, formaldehyde, and acetaldehyde were not fully characterized even in the saturated model, emphasizing that additional time-activity and concentration information would more fully characterize personal exposure.
Introduction Many volatile organic compounds (VOCs) are classified as hazardous air pollutants (Section 112, U.S. Clean Air Act), and exposure to VOCs can result in a wide range of acute and chronic health effects, such as sensory irritation, nervous system impairment, asthma, and cancer (1–3). The primary emitters of VOCs to air include industrial sources, mobile * Corresponding author e-mail:
[email protected]. † Harvard School of Public Health. ‡ University of Massachusetts. § University of California. 8498
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 41, NO. 24, 2007
sources, various consumer and household products, and building and construction materials. For example, mobile sources are largely responsible for the presence of benzene, toluene, ethylbenzene, xylenes (BTEX), 1,3-butadiene, and methyl tert-butyl ether (MTBE) in the ambient environment. Carpets can release formaldehyde, styrene, and toluene into the indoor environment (4); freshly dry cleaned clothes can contain tetrachloroethene (5); the use of chlorinated drinking water can emit chloroform (5); room deodorizers or mothballs can contain 1,4-dichlorobenzene (5), and personal care products can contain styrene and tetrachloroethene (6). Personal exposure depends largely on locations in which individuals spend time and the type of activities in which they are involved. The National Human Activity Pattern Survey revealed that Americans spend more than 80% of their time indoors (7), in residences and also at work. Over 130 million adults (approximately 55% of the population) are employed full-time, indicating the potential significance of the workplace (8). Since the 1980s, it has been well established that personal exposures to VOCs often exceed outdoor/ambient exposures and indoor residential exposures (9–15). In a recent example, Sexton and colleagues measured personal, indoor, and community outdoor air concentrations for 71 adults in three urban neighborhoods in Minneapolis (9, 16). Personal exposures exceeded indoor air concentrations which exceeded outdoor air concentrations for 13 of the 15 measured VOCs, with the exceptions being carbon tetrachloride and chloroform (9). These researchers used mixed models to estimate the relative difference between outdoor, indoor, and time-weighted outdoor and indoor concentrations and measured personal exposure, concluding that for all compounds studied the outdoor and indoor concentrations underestimated personal exposure, with better performance using time-weighted outdoor and indoor concentrations (16). No other microenvironments were included. In general, most personal exposure studies in the United States have focused only on residential exposures with limited studies on the potential impact of workplace exposures. Concentrations of VOCs in large office buildings were measured in the U.S. Environmental Protection Agency’s (EPA) Building Assessment Survey and Evaluation (BASE) study. Of the 48 VOCs found at detectable levels, all had median indoor/outdoor ratios exceeding 1, indicating the presence of indoor sources (17), in agreement with other studies of workplace concentrations (13, 18). In a smaller San Francisco Bay area workplace study, trichloroethene and d-limonene were characterized as being predominantly from indoor sources while benzene, xylenes, and tetrachloroethene were characterized as having a mix of both indoor and outdoor sources (19). In a Los Angeles classroom study (representing a workplace environment for some individuals), the six compounds detected with the highest prevalence were formaldehyde, acetaldehyde, toluene, m,p-xylene, R-pinene, and d-limonene, likely due to cleaning and personal care products and furnishing or teaching materials (20). Given both time-activity patterns and the presence of indoor sources, inclusion of the workplace may be necessary to accurately characterize personal exposures using microenvironmental models. However, additional microenvironments may be necessary to include as well. A recent European study found that for participants in Helsinki, Finland, the geometric mean of the measured personal air concentrations could be reasonably estimated from the time-weighted concentrations within residential indoor and workplace microenvironments (13). 10.1021/es071127s CCC: $37.00
2007 American Chemical Society
Published on Web 11/15/2007
However, when data were combined across Europe and into chemical classes, approximately 30% of the measured personal exposure could not be fully explained by the timeweighted average indoor, outdoor and workplace concentrations (21), indicating the potential for other sources of exposure. The researchers further noted that the concentrations measured in Europe are not comparable to those in the U.S., suggesting a difference in product composition and use (13), and the need for similar studies in the U.S. To further understand the impact of workplace and various other microenvironments on personal exposure, we used data collected as part of the Boston Exposure Assessment in Microenvironments (BEAM) study, which was designed to examine concentration distributions in various microenvironments in an effort to reduce uncertainties in predicting exposure levels. Measured personal concentrations are compared to time-weighted concentration models using median regression models to determine if time-weighted microenvironmental concentrations can be used as a proxy for personal exposure and to determine which microenvironments are needed to produce unbiased estimates of personal exposure for a suite of VOCs.
Materials and Methods Study Design and Sampling Methods. In the BEAM study, personal sampling was completed on a total of 55 nonsmoking participants living in and around Boston, Massachusetts, across two seasons (summer and winter), with 34 participants included in both seasons on the basis of availability, for a total of 89 sampling visits. All sampling occurred midweek, generally Tuesday evening to Thursday evening. All enrolled participants worked in nonmanufacturing office or school (primary and secondary public schools) environments and were recruited as a convenience sample, including through the Harvard School of Public Health alumni network, local boards of health, and environmental organizations. Participants wore a personal sampling backpack and had samples taken inside their home, just outside their home, and in their workplace. All VOCs were collected actively as integrated samples, with personal and home samples collected over the 48 h sampling period, and workplace samples collected over the participant’s planned working hours (generally a total of 16 h). Another phase of the BEAM study included collecting samples in a variety of retail stores, smoking and nonsmoking dining establishments, and multiple modes of transportation in the Boston area, as reported elsewhere (22). These concentrations were used to assess likely exposures in microenvironments not measured for each individual. We measured 18 compounds (2 carbonyls, 14 other VOCs, and 2 terpenes), selected because they were found at elevated levels in previous studies or pose a health risk at observable levels. Carbonyls were collected actively using 2,4-dinitrophenylhydrazine-coated silica cartridges (Waters Corp, Milford, MA), while the remaining VOCs were collected actively using custom-made thermal desorption tubes (200 mg of Carbopack B, 230 mg of Carbopack X, and 170 mg of Carboxen 1001; Supelco/Perkin-Elmer, Bellefonte, PA). Ozone scrubbers were used on all outdoor aldehyde samplers during summer sampling. Samples were analyzed following U.S. EPA methods. Personal sampling backpacks contained a personal BGI pump (BGI4004: BGI, Inc., Waltham, MA), flow restriction valves, a temperature and humidity HOBO data logger (Onset Computer Corp., Bourne, MA), and a motion logger to confirm compliance. Sampler inlets were secured to the shoulder strap of the backpack. Participants were asked to wear, or keep near them at breathing height, the personal backpacks as they underwent their normal activities. All other samples were collected using stationary samplers housing MEDO air pumps (Medo, Hanover Park, IL) with sampler
inlets placed approximately at breathing height. Additional information regarding sampling and analytical methods as well as quality assurance and control information is provided in the Supporting Information. Time-activity questionnaires were administered after each 24 h period. Seasonal differences in time-activity patterns were assessed using a nonparametric, two-sided Wilcoxon rank sum test. Model Development. Personal exposure is a function of the concentrations within the various microenvironments visited as well as the time spent in those microenvironments. M )
∑Ct
(1)
j j
j
where M is the time-weighted modeled proxy for personal exposure to a pollutant, Cj is the concentration of a pollutant in microenvironment j, and tj is the fraction of modeled time spent in microenvironment j. Our objective is to estimate the systematic difference between M and the actual exposure to a pollutant, P. We therefore model the difference P – M as follows: P - M ) βo + ε
(2)
where intercept βo is the systematic bias term and ε is the random error term. If βo is significantly different from zero, the difference between measured and modeled personal exposure is significant and the time-weighted modeled surrogate is biased. Modeled personal exposure was first set equal to the participant-specific outdoor concentration as a baseline estimate of personal exposure. This is because personal exposures depend, in part, on outdoor concentrations, since, for VOCs, we can assume that the penetration efficiency to the indoor environment is near 1 (23). If the penetration efficiency is not near 1 due to the reactivity of the compound, the result could be a slight overprediction of exposure at the baseline. The residual, or difference between measured and modeled personal exposure, was then assumed to be a result of all other microenvironment-specific or activity-specific exposures. This modeling structure assumes that omitted microenvironmental concentrations at each model stage have higher concentrations than those observed outdoors. Additional microenvironments were subsequently added to the model, and results were compared to measured personal concentrations. First, each individual-specific concentration was included, specifically the residential indoors followed by workplace microenvironments, as outlined in Table 1. Nonindividual-specific measured microenvironments were then added, resulting in the fully saturated model including residential outdoor, residential indoor, workplace, and transportation microenvironments and a category of microenvironmental exposures called “all other” that includes dining and retail exposure. For various types of retail stores, restaurants, and transportation modes, concentration geometric means were multiplied by the participant-specific time spent in that microenvironment. See Supporting Information for additional details. We used median regression models to model P – M, the difference between the measured personal concentration and personal exposure calculated by each of the timeweighted microenvironmental exposure models, as an intercept representing a bias term plus a linear combination of regression parameters and covariates (e.g., season). The random error terms (ε) were assumed to have a zero median, unlike in ordinary least-squares regression where the error terms are assumed to have a zero mean. Median regression models were used due to the skewed nature of the personal exposure data since they are insensitive to distributional assumptions (24, 25), unlike ordinary least-squares regresVOL. 41, NO. 24, 2007 / ENVIRONMENTAL SCIENCE & TECHNOLOGY
9
8499
TABLE 1. Time-Weighted Models Used in Analysis of Personal Exposurea model
N
microenvironments included
O
66
O+I
62
O+W
66
O+I+W
62
O+I+W+T
62
O+I+W+T+A
62
residential outdoor residential outdoor and residential indoor residential outdoor and workplace residential outdoor, residential indoor, and workplace residential outdoor, residential indoor, workplace, and transportationb residential outdoor, residential indoor, workplace, transportationb, and all otherb,c
model formulation M ) Cout M ) (Cintin) + [Cout(1 – tin)] M ) (Cworktwork) + [Cout (1 – tin)] M ) (Cintin) + (Cworktwork) + [Cout(1 – tin – twork)] M ) (Cintin) + (Cworktwork) + (Ctransttrans) + [Cout(1 – tin – twork – ttrans)] M ) (Cintin) + (Cworktwork) + (Ctransttrans) + (Cothertallother) + [Cout(1 – tin – twork – ttrans – tallother)]
a N indicates the sample size available for regression models; M ) time-weighted modeled proxy for personal exposure (P); Cj ) concentration measured in that microenvironment; tj ) fraction of time spent within that microenvironment. b Concentration measurement is not a participant-specific measure, although time fractions are participant-specific. c “All other” includes retail store and dining microenvironments.
sion, which can behave poorly for skewed or heavy-tailed distributions. Another typical approach for analyzing skewed data is to model the logarithm of the response, entailing an underlying lognormality assumption; however, in our setting, the response is a difference between two log-normal variables, which can take negative values, thereby invalidating this simpler approach. Since many previous personal exposure studies have used ordinary least-squares (OLS) linear regression models either with transformed or untransformed response variables (14–16), we conducted additional analyses to understand the potential implications of using a median regression model over an OLS linear model. Generalized estimating equations (GEE) models were used to estimate the population means for the OLS models. A total of 89 sampling events occurred, with 34 participants sampled across two seasons; however, the actual number of observations used in the regression models was less due to some analytical and equipment issues. All available sampling data were used as individual observations in the regression models. Season was initially included as a covariate in the regression models; however, it was removed due to a lack of significance in all of the models, which could potentially be a result of an insufficient sample size. Since an estimation of the regression parameter βo involved repeated measures, standard errors were adjusted using a modified version of the pivotal estimating equations method of Parzen et al., where we applied random perturbations at the subject rather than the observation level (26); this approach is analogous to the standard error adjustment implemented in GEEs. The presence of repeated measures necessitated the use of median regression models over more simple median difference models. A stepwise approach was used to evaluate the influence of each microenvironment. All models were fit using the quantreg package in R. The influence of each microenvironment was evaluated by comparing the change in the intercept or bias estimates. When intercepts were significantly different than zero, significant, unexplained exposure remained, referred to as a biased estimate of personal exposure. Risk Calculations. Along with determining the microenvironments needed to best characterize personal exposures to individual VOCs, we used identical methods to determine the microenvironments needed to provide unbiased estimates of total cancer risk associated with these VOCs. Risk estimates were calculated by summing across compounds using inhalation unit risk factors (IURs) obtained from the State of California’s Office of Environmental Health Hazard 8500
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 41, NO. 24, 2007
Assessment (OEHHA) for those compounds believed to be at least potentially carcinogenic. The IURs represent highto low-dose, nonthreshold, linear extrapolations from animal studies or epidemiological studies using upper 95% confidence limits or maximum likelihood estimates, respectively. We used California OEHHA values as opposed to the U.S. EPA’s values from their Integrated Risk Information System database because the U.S. EPA’s value for formaldehyde is undergoing a re-evaluation (27, 28). California has a lower IUR, which research indicates may be appropriate (28, 29). Both 1,3-butadiene and methylene chloride, despite not being evaluated as individual compounds in the personal exposure models due to a substantial number of measurements below the limit of detection, were included in the estimate of total risk given their potentially significant contributions.
Results and Discussion Study Population. The study population included slightly more men (55%) than women (45%), was mostly Caucasians (91%), was highly educated (69% with some postcollege education), and was fairly evenly distributed across the working age distribution (ages 21–30, 25%; 31–40, 31%; 41–50, 20%; and >50, 24%). All of the participants were employed in nonmanufacturing workplaces, including schools and offices, and lived in suburban or urban locations in the greater Boston area. Thus, this analysis is likely only directly applicable to populations of office workers or teachers living in or around a metropolitan area. Concentrations and Risks. Concentrations measured with the personal sampling backpacks, inside the home, outside the home, and at the participant’s workplace, are presented in the Supporting Information. Outdoor concentrations were lower than indoor concentrations for all compounds, except carbon tetrachloride, and indoor residential concentrations were fairly comparable to workplace concentrations. Note the heavy right-tailed skewed nature of many of the data, in particular 1,4-dichlorobenzene. At the median, the concentrations measured in the office environments tended to be higher than those measured in the school environments. Measured median indoor and outdoor residential concentrations were slightly lower than previous studies, suggesting a general decreasing trend in ambient VOC concentrations or differences across sampling locations. The measured median personal concentrations were generally lower (with some exceptions) than those reported for teenagers in New York City and Los Angeles (12), for children in the Minnesota Children’s Pesticide study (15), and adults in the EXPOLIS-Helsinki, Finland study (13). The concentra-
FIGURE 1. Relative percent difference between the measured personal exposures and the time-weighted modeled personal exposures. Note: Boxplots represent the 5th, 25th, 50th, 75th, and 95th percentiles. tions, however, were fairly comparable to those measured in two other Minnesota-based studies (9, 14). Workplace concentrations measured in other studies were compared to those in our study. The measured office concentrations, with the exception of d-limonene, were almost a factor of 2 lower than those measured in the U.S. EPA’s BASE study (17). One possible explanation is that office concentrations are influenced, in part, by outdoor concentrations, which were likely lower in our study. However, our results were comparable to the concentrations reported in a California office study, except that our concentration distributions tended to be slightly more skewed (19). The concentrations measured in school classrooms were similar to those measured in main school buildings in Los Angeles County with the exception of acetaldehyde and MTBE, which were lower in our study, and d-limonene, which was higher (20). Cancer risk estimate distributions across study participants calculated on the basis of the measured personal concentrations are presented in the Supporting Information. The median total personal cancer risk was 3.2 × 10-4 (5th and 95th percentiles of 1.7 × 10-4 and 8.9 × 10-4, respectively), which is lower than the median estimates calculated for innercity teenagers in New York City and Los Angeles (6.66 × 10-4 and 4.86 × 10-4, respectively) (12) and the mean total personal risk estimated in a U.S.-wide study (6 × 10-4) using similar IURs (30). The most influential compound on the total cancer estimate was formaldehyde (33% of total risk at the median) followed by benzene (15%), 1,3-butadiene (12%), carbon tetrachloride (9%), acetaldehyde (9%), tetrachloroethene (2%), 1,4-dichlorobenzene (2%), chloroform (1%), MTBE (
