We examined personal PB-PAH exposures by major influential factors and developed regression models to estimate PB-PAH exposures based on GPS-tracking data, traffic activity data, and simple questionnaire information (adjusted R2 ranged from 0.58 to 0.75). The strongest predictors of personal PB-PAH exposures were found to be time in-vehicle and the related GPS speed variable, as well as variables describing other exposures to traffic such as traffic density at nearby streets (length-weighted AADT) and work-related exposures to traffic pollutants. The GPS-acquired data made it possible to determine the value of these variables with considerable temporal and spatial accuracy as we reported previously . Our study adds important new findings to the literature on PB-PAH exposure assessment in several ways: (1) it is one of the first studies to model personal PB-PAH exposures at a high-temporal resolution; (2) it demonstrated the usefulness of coupling real-time exposure measures with GPS tracking in personal air pollution exposure assessment; and 3) it confirmed the importance of GPS-based time-activity and GPS speed as a surrogate of on-road exposures in PB-PAH exposure modeling.
Our personal PB-PAH measurements revealed an undeniable contribution from the transport microenvironment. The amount of time in vehicle (on average 4.5% of the total sampling time) explained 48% of the variance in daily personal PB-PAH exposure and 39% of the variance in subject-level exposure. Time in vehicle was the most important determinant of personal PB-PAH exposure, which confirms earlier studies on the relationship between activities and traffic-related air pollution exposure that have suggested an important role for the traffic microenvironment, despite the limited time spent in or near traffic environments [29, 46, 47]. Significant exposure misclassification may occur if only residential exposures are considered since time spent in or near transport may provoke dissimilarity in personal exposure between individuals with similar residential exposures. In addition to in-vehicle time, length-weighted AADT within 500 m explained 8% of the variance in daily PB-PAH exposure. This variable explained approximately 28% of the variance in subject-level indoor PB-PAH exposures (data not shown), which confirms the importance of local traffic emissions to the exposures not only in the commuting environment but also indoors. Finally, the in-vehicle time and GPS speed variables were highly correlated (r > 0.90 for daily and subject-level data). In-vehicle time tended to correlate slightly better with daily and subject-level total exposures, although the square root of GPS speed was a better predictor in the subject-level microenvironmental model (Table 5) and explained 67% of the variance.
The work-related exposure to traffic-related pollutants explained 10% of the variance in PB-PAH exposure at the subject level and 2% of the variance in daily exposure. The work-related exposures were obtained from the questionnaire based on the typical activity patterns in the past three months of the personal sampling, thus this variable was not able to capture substantial day-to-day variation in subjects’ time activity patterns and PB-PAH exposures. Furthermore, approximately 40% of women did not work, thus their activity patterns may vary considerably on a daily basis compared to full-time workers. We also found that subjects tended to have higher PB-PAH exposures during weekdays than weekends and thus the percent of weekday time explained 16% of the variance in PB-PAH exposures at the subject level. Previous studies have reported higher ambient concentrations of traffic-related pollutants on weekdays [48, 49], but little is known about the day of week impact on personal exposures. In addition, we found women who did not speak English at home had marginally higher PB-PAH exposures although this variable was not selected in the final prediction models. Further investigation of our data showed that women who did not speak English at home had a higher percent of in-vehicle time than the others (5.9% vs. 4.1%) although the difference was not significant (p-value = 0.35).
We found that the percent of indoor time was negatively associated with PB-PAH exposure, which is expected due to few PB-PAH sources indoors in this study. Although we did observe and remove high PB-PAH concentrations indoors (e.g. five 1-min exposures exceeding 1000 ng/m3), such events had little influence on the overall average exposures since they occurred occasionally and lasted for a very short period (e.g. one minute). In fact, the models changed little when we included the five outliers with extremely high indoor concentrations (data not shown). This lack of impact on PB-PAH by the indoor environment is likely because we only recruited non-smokers and the prevalence of other indoor sources (e.g. ETS, wood-burning, grilling, and barbecue) was relatively low. California had the 2nd lowest smoking rate (14% in adult population) among all the states in the U.S.  and smoking was prohibited in almost all indoor and outdoor public places in California . Indoor wood-burning or indoor grill/barbecue was not common because of the warm climate of southern California and that our samples were mostly collected in the warm season. However, in other regions or other populations, these non-traffic related indoor sources may contribute more to personal PB-PAH exposures than what was shown in our study.
The geometric mean of PB-PAH exposures tended to be higher in rush hours than non-rush hours at the minute level (Additional file 1: Table S1), but the pattern was opposite at the subject level (not statistically significant) (Table 1). We found that the geometric mean of one-minute PB-PAH concentrations started to increase remarkably at 6 AM, peaked at 7 AM (geometric mean = 5.1 ng/m3), dropped gradually from 9 AM to 2 PM, peaked again at 4 PM (geometric mean = 3.4 ng/m3), and started to drop gradually from 5 PM till 1 AM in the morning (Additional file 1: Table S6). We observed a late afternoon peak (3–5 PM) in PB-PAH exposure, but not the evening rush hour peak that occurs at 4–7 PM. Since about 40% of the study participants did not work, they may have picked up their children or done errands in the late afternoon or other times of the day. We found that on average the subjects spent approximately 9.8% and 7.1% of the time traveling in vehicles during 3–5 PM and 4–7 PM, respectively. Thus, the rush hour variable may not appropriately capture their time in traffic.
No meteorological variables were entered into the predictive models. This is likely because we modeled personal exposure rather than ambient outdoor pollutant concentrations. Personal exposures are strongly influenced by near-source activities of human subjects. Although subject-level PB-PAH exposure was significantly higher (r = 0.02) in the group with lower relative humidity (Table 1), the continuous measure of relative humidity was only marginally correlated with subject-level exposure (r = 0.10) and was not selected in the final models. In addition to wind, temperature, and relative humidity, we also examined the usefulness of the atmospheric stability class data modeled every three hours at 40 km by 40 km resolution from the nearest EDAS modeling grid of the National Oceanic and Atmospheric Administration (http://www.arl.noaa.gov/ready.html). However, the stability variables were not significantly associated with PB-PAH exposures, likely due to substantial uncertainties associated with the modeled stability estimate.
A major limitation of the study is the semi-quantitative feature of the PAS sampler. The PAS may respond differently to individual PAH species thus the PAS signal may not be directly proportional to the concentration of individual species . In addition, the components of the PAH mixtures may differ by emission sources (e.g. traffic, tobacco smoke, wood combustion, food grill), thus the PAS measurements reflect not only the total concentration but also the nature of the PAH mixtures in different microenvironments. This creates uncertainty in the exposure measures among different microenvironments. Despite the limitations, the PAS sampler is the only available instrument that is capable of continuous personal PB-PAH monitoring. The highly informative nature of the predictive models (adjusted R2: 0.58-0.75) is a testament to the approach, which could be adapted to methods using more accurate instruments in the future if they become available.
Other time-activity patterns such as biking and travel by bus and subways may also be associated with high levels of exposure to traffic-related air pollutants including PB-PAH [52–54]. However, no subjects reported traveling in an underground train, by bus, or biking based on our questionnaire on the means of transportation. Therefore, we did not examine the other travel modes in this study although they may be important in other studies and regions where subjects may engage in these activities frequently.
We did not use diaries to track subjects’ activities that may significantly influence their exposure levels (e.g. near a smoker, cooking) but are not easily obtained from the GPS data alone, mainly because the collection of such detailed information may significantly increase the burden to the subjects. Combining the GPS data with simple questionnaire or diary data may further improve the model performance. Additionally, our subjects were only pregnant women or women who had delivered babies within one year of the sampling dates. Other population groups (e.g. children, men, other women) and subjects in other regions may have different time-activity patterns than our study participants. However, we believe that the method of coupling real-time exposure sampling with GPS time-activity tracking and the application of GPS data in exposure modeling can be easily adapted to other populations in different studies.
We found higher PB-PAH exposures in the winter with a geometric mean (based on one-minute data) of 4.8 ng/m3 and 2.3 ng/m3 in the cool and warm season, respectively (Additional file 1: Table S2). Unfortunately, less than 5% of our data were collected during the cool season. Thus we could not examine the seasonal difference. Future research may improve model prediction by sampling in different seasons and measuring a more diverse and larger number of subjects.