Kriged and modeled ambient air levels of benzene in an urban environment: an exposure assessment study

Historically, there has been concern regarding the potential adverse human health effects of ozone, sulfur dioxide, nitrogen oxides, carbon monoxide, particulates, and lead. In 1971, the Clean Air Act was established under which National Ambient Air Quality Standards (NAAQS) were created to regulate ambient air concentrations of these six criteria pollutants [1]. A decline in ambient air concentrations of criteria pollutants has been observed since the induction of this Act [2], and more recently, the focus has shifted to hazardous air pollutants (HAPs), a class of 189 compounds, which are known or suspected to have adverse effects on health [3]. One HAP, benzene, is of particular concern due to its ubiquitous nature and ability to cause cancer in humans [4]. Although the general population is exposed to background levels of benzene, one of the major outdoor sources of personal exposure is vehicular exhaust; additionally, people living near chemical manufacturing facilities or oil refineries may also be exposed to elevated levels of benzene [4].

The potential for human exposure to benzene is well established, but quantifying exposure for population-based epidemiologic studies is problematic and requires immense resources. For this reason, researchers studying the health effects of ambient air levels of benzene often rely on proxy measures of exposure. One potential source of data is routine monitoring data. Tools like geographic information systems (GIS) and spatial interpolation methods such as kriging have helped to utilize these data to estimate levels of ambient air pollutants at unmeasured locations. Previous studies that have used kriging to map air pollution levels include: sulfur dioxide in Instanbul, Turkey [5], ozone in Atlanta [6, 7] and northern Georgia [8], and particulates across the entire U.S. [9] and in Beijing, China [10]. Kriging has also been used in a range of epidemiologic studies of criteria pollutants to examine exposures to NO₂ among pregnant women in Spain [11] as well as associations between: particulates and ozone and mortality in Los Angeles [12], particulates and mortality in Hamilton, Ontario, Canada [13], particulates and ectopy in the U.S. [9], CO, NO₂, CO₂ and preterm birth in Korea [14], ozone and pediatric asthma exacerbation in Atlanta [6], and particulates and low birth weight in Korea [15].

Although kriging appears to have become a useful tool in studies of criteria pollutants, in part because of the availability of monitoring data due to regulatory requirements, a limited number of studies have applied these methods to HAPs. Several studies have used kriging in the exposure assessment of ambient air levels of radon [16–18], but fewer investigations have applied this method to volatile organic compounds (VOCs). Following an extensive ambient air monitoring campaign that was conducted over a two-week period, Miller et al. [19] recently applied ordinary kriging to examine the spatial variability of total VOCs and BTEX (benzene, toluene, ethylbenzene and xylene, combined) in Detroit, Michigan (U.S.) and Windsor, Ontario (Canada). However, the usefulness of kriging to predict annual benzene levels using routine monitoring data, as might be needed in epidemiologic studies of health effects such as cancer, has not been fully explored. In contrast, data from the Assessment System for Population Exposure Nationwide (ASPEN) model, generated from the U.S. EPA has been applied in epidemiologic studies. Previously, for example, we conducted a study of childhood cancer in the Houston metropolitan area that used the ASPEN modeled estimates in the exposure assessment and found increased rates of childhood leukemia among census tracts with the highest levels of benzene [20]. Harris County, in which Houston is located, is home to a large number of petrochemical industries and is the fourth largest metropolitan area of the U.S., with a dense network of roadways. It is also one of the most closely monitored cities in the nation [21]. Given our previous epidemiologic finding, and in light of the fact that the ASPEN modeled estimates are only available for select years, we were interested in whether existing monitoring data could be used to provide additional information regarding exposure assessment of benzene. Hence, we conducted a study to apply kriging to predict ambient air levels of benzene for the years 1998 to 2000 at unmonitored locations in Harris County and to assess the degree of correspondence between the census tract-level kriged benzene levels and estimates of benzene obtained from the ASPEN model.

Figures 2 and 3 provide maps of Harris County and show the kriged values (generated using the 17 monitoring sites in Harris, Galveston, and Brazoria Counties) and ASPEN modeled estimates, categorized by quartiles. Both figures use the ASPEN quartile cut points to aid in the visual comparison of the two methods. Major highways and locations of industrial facilities that, together, contributed to 75% of the reported benzene emissions in 1999 are also represented. The facilities data were obtained from the EPA's Toxic Release Inventory [31]. The darkest areas represent census tracts with the highest estimated (or predicted) ambient levels of benzene, while the lightest areas represent census tracts with the lowest estimated (or predicted) levels. Note in Figure 2 that none of the census tracts' kriged benzene levels were categorized in the lowest ASPEN quartile.

The sensitivity analysis comparing the kriged benzene values using the complete network of all 55 monitors in Texas versus the restricted network of 17 monitors in Harris, Galveston, and Brazoria Counties indicated similar results. The mean ambient benzene level using all 55 monitors was 0.73 (standard deviation (SD) = 0.10) while the mean ambient benzene level using the restricted network was 0.74 (SD = 0.08) and the mean benzene level from the ASPEN model was 0.71 (SD = 0.31). After stratifying by the "near" and "far" groups, the kappa statistic for the "far" census tracts was 0.26 (95% CI = 0.19, 0.33) while the kappa statistic for the "near" census tracts was 0.18 (95% CI = 0.09, 0.27). Although the magnitude of the kappa statistic for the "far" census tracts is greater than that for the "near" census tracts, they were not statistically significantly different (p = 0.3).

Our investigation made comparisons between a complex dispersion model developed by the U.S. EPA (ASPEN model) and a method of spatial interpolation (kriging) using routine monitoring data for benzene, a HAP. Overall, we found a lack of correspondence between kriged benzene predictions and ASPEN modeled estimates. The kriging model that we applied did not use data that could potentially affect variability in ambient benzene levels, including meteorology and source emissions from roadway traffic or industrial sites. The ASPEN model, in contrast, relies on more complex data sources, incorporating not only meteorological and pollutant source data, but also fate and transport data (including information about reactive decay of pollutants, deposition, and secondary formation) and monitoring data to estimate background concentrations of the pollutant of interest [25]. The discrepancies in the input data used by each method may contribute to the discordant results we observed. Although kriging has been used to interpolate ambient air levels of criteria air pollutants in exposure assessment and epidemiologic studies [6, 9, 11–15], few other studies have applied kriging to benzene [19, 32].

The range of predicted benzene values observed in our study is smaller than that estimated from the ASPEN model. This may be due to the fact that kriging is a spatial smoothing tool that provides an optimal average over the region based on sampled observations. We also found poor agreement between the two metrics irrespective of whether the continuous or discretized values were evaluated. While significant, there was only modest correlation between the modeled and predicted values. Based on the quartile analysis, approximately 44% of census tracts classified by the ASPEN model as having the lowest ambient air levels of benzene were classified into a higher category using the predicted kriged results. Similarly, about 72% of census tracts in the highest quartile based on the ASPEN model were classified into lower categories using the kriging metric.

Measurements of ambient air pollutants using standard reference methods represent the optimal data source. Several studies have compared estimates from the ASPEN model to routine monitoring data and found reasonably good agreement between the two for ambient benzene levels [33–35]. Therefore, under the assumption that the ASPEN model yields reasonable estimates of ambient air pollutants at the census tract level, our results suggest that relying on kriged predictions instead would result in extensive misclassification had this exposure metric been used in an epidemiologic study. Although Houston, Texas is considered a closely monitored area [21], in our study, use of kriging that relies upon routine monitoring data for epidemiologic purposes or exposure assessment is still limited by a low ratio of monitoring sites to total land area. Given that we found similar agreement between the ASPEN modeled estimates and the kriged predictions for census tracts that were within 5 miles of a monitoring site as compared with census tracts further (>5 miles) from one or more monitoring sites, it appears that greater resolution in terms of the number of benzene monitors per area would be required for kriging, as has been suggested recently by the work of Cocheo and colleagues [36].

There has been success in applying kriging to predict ambient air levels of the criteria pollutants [9, 11–15] which may be due, in part, to the regional nature of some of these pollutants as well as the relatively large monitoring networks utilized by many of these investigations. For example, the networks used in these studies ranged from 23 monitoring sites measuring PM_2.5 and 42 ozone sites in the Los Angeles area [12] to 93 monitors measuring NO₂ in Valencia, Spain to an average of 456 sites measuring PM_2.5 across the U.S. [9]. Spatial variability over such expansive geographic regions are likely influenced more by meteorological and topographic factors than by specific emission sources and dispersion characteristics [37].

Another issue that may have affected the lack of agreement observed in the present study is the placement of monitors. Because the monitors are used primarily for regulatory purposes, they are not uniformly spatially located. The majority of the monitors used in our study are located near a large petrochemical complex in Houston (known as the Houston Ship Channel) and largely situated away from major transportation corridors in the county. As a result, the kriging predictions generally do not account for HAP levels arising from mobile sources, which make a significant contribution to ambient air levels of HAPs such as benzene [38, 39]. In a recent study that applied kriging methodology to BTEX and other VOCs [19], after placement of 100 monitors across two cities in North America, the authors concluded that there was considerable intraurban variability in VOC levels in Detroit, Michigan (U.S.), where emissions of these chemicals are large and from similar sources to those in the greater Houston metropolitan area. Thus, we expect that interpolating ambient air levels of a localized (rather than regional) air pollutant will remain problematic in the absence of additional and more equitably distributed monitors across a geographic locale.

To obtain stable estimates of the theoretical semivariogram function, we used data from all ambient monitors in Texas. We assumed that the underlying spatial correlation between monitoring sites in our study area was similar to that of all monitoring sites in Texas. This is similar to others who have generated semivariograms for ozone and particulates [9, 40] over five regions in the U.S., and for NO₂, PM₁₀, and O₃ across the European Union [37]. Further, when we conducted a sensitivity analysis comparing the kriged predictions using the complete monitoring network of all Texas monitors versus only the 17 monitoring sites in Harris, Galveston, and Brazoria counties, the results were similar. At the time the present study was conducted, the ASPEN estimates were only available for 1999; thus, we restricted our use of the monitoring data to 1998-2000. Although we considered using only monitoring data from 1999, there were not enough data in this one year to accurately estimate the semivariogram. This is not likely to have introduced much error in our evaluation given that the annual mean benzene levels (mean ± 2SD) using all 55 monitoring sites in Texas for 1998 (N = 46), 1999 (N = 48), and 2000 (N = 46) were: 0.8 ppbV (0.1, 1.4), 0.9 ppbV (0.2, 1.5), and 0.8 (0, 1.6), respectively.

This study compared predicted annual levels of benzene in ambient air to those generated by the ASPEN model to assess the value of kriging in epidemiologic investigations of HAPs. Due to the discrepancies found in these two methods, we feel that until improvements are made regarding the placement and number of monitors collecting HAP data, researchers interested in studying the health effects of these air pollutants must rely on data sources such as the ASPEN modeled estimates or other exposure models and innovative exposure assessment strategies such as generalized additive models or land use regression.