Hazardous air pollutants (HAPs) are a class of 188 compounds, including benzene and 1,3-butadiene, which are known or suspected to have adverse effects on health [U.S. Environmental Protection Agency (EPA) 2008]. Benzene is one of the best understood human carcinogens, and 1,3-butadiene is a suspected human carcinogen; much of the evidence regarding their carcinogenicity, however, comes from animal models and studies of adults exposed occupationally (International Agency for Research on Cancer 1982, 1987, 1998). Moreover, a recent investigation demonstrated that Chinese workers exposed to benzene levels ≤ 1 ppm had significantly decreased white blood cell and platelet counts compared with unexposed workers (Lan et al. 2004), suggesting that benzene exposure occurring at levels comparable with what might be expected of ambient air concentrations has the potential to adversely affect health.

Although the general population is exposed to background levels of benzene and 1,3-butadiene, those living in urban areas with increased traffic density or in close proximity to chemical manufacturing facilities or petroleum refineries are likely exposed to even higher levels of these compounds (Agency for Toxic Substance and Disease Registry 1992, 2007). Despite ubiquitous levels of benzene and 1,3-butadiene in the ambient air environment, there have been few studies of cancer risk among children living in urban communities with potentially high levels of HAPs. Most of these studies have used proxy measures of exposure to traffic-related air pollution, and results have been mixed (Crosignani et al. 2004; Harrison et al. 1999; Knox 2006; Nordlinder and Jarvholm 1997; Pearson et al. 2000; Raaschou-Nielsen et al. 2001; Reynolds et al. 2002; Steffen et al. 2004). Two other studies have reported increased risk of childhood cancers among children whose residence was near a “hot spot” of benzene or 1,3-butadiene emissions (Knox 2005a, 2005b). Reynolds et al. (2003) found 21% [rate ratio (RR) = 1.21; 95% confidence interval (CI), 1.03–1.42] higher rates of childhood leukemia among census tracts with the highest HAP exposure score (for 25 individual pollutants combined) compared with census tracts with the lowest scores. It has been suggested that the inconsistencies in the current air pollution and childhood cancer literature may be attributable to differences in analytic and exposure assessment methodology as well as differences in the sampled populations’ underlying genetic susceptibility (Buffler et al. 2005).

Table 2 presents the distribution of the U.S. EPA ASPEN model estimates for benzene and 1,3-butadiene in the study area. Across the 886 census tracts in our study area, ambient air levels ranged from 0.42 to 9.05 μg/m³ for benzene and from 0.01 to 2.41 μg/m³ for 1,3-butadiene. The median benzene level in the eight-county study area (1.72 μg/m³) is approximately 10 times the median level of 1,3-butadiene (0.16 μg/m³).

We found that census tracts with the highest ambient air levels of benzene had elevated rates of all leukemia (RR = 1.37; 95% CI, 1.05–1.78) and AML (RR = 2.02; 95% CI, 1.03–3.96) compared with census tracts with the lowest estimated levels (Table 3). Additionally, these census tracts had 1.24 (95% CI, 0.92–1.66) times the rate of ALL compared with census tracts with the lowest levels, although this estimate was not statistically significant. We detected a statistically significant trend of increasing incidence rates with increasing estimated levels of benzene for all leukemias combined (p = 0.03) and a borderline significant trend for AML (p = 0.06). We observed no trend for ALL. No associations were observed between benzene levels in ambient air and Hodgkin’s disease or NHL.

Relative to census tracts with the lowest levels, we observed significantly increased rates of all leukemia among census tracts with the highest levels of ambient 1,3-butadiene (RR = 1.40; 95% CI, 1.07–1.81) as well as a significant trend (p = 0.01) (Table 4). We also observed 32% (RR = 1.32; 95% CI, 0.98–1.77) and 68% (RR = 1.68; 95% CI, 0.84–3.35) higher (albeit insignificant) rates of ALL and AML, respectively, in these census tracts compared with census tracts with the lowest levels of 1,3-butadiene. We found a borderline significant trend between rates of ALL (p = 0.06) and 1,3-butadiene levels. We observed no statistically significant trend between rates of AML and 1,3-butadiene levels. We detected no association between estimated 1,3-butadiene levels and incidence of Hodgkin’s disease or NHL.

The few other studies that have explored the potential association between childhood cancer and air pollution have produced equivocal results. Of four studies reporting statistically significant positive results, three used measures of traffic density (Crosignani et al. 2004; Nordlinder and Jarvholm 1997; Pearson et al. 2000), and one study examined residential distance from roadways (Knox 2006) to assess exposure. In contrast, two case–control studies found no association between traffic density and childhood leukemia (Reynolds et al. 2002; Steffen et al. 2004), and one case–control study found no significant risk of leukemia associated with residence within 100 m of either a main road or a gas station (Harrison et al. 1999).

Our study has the benefit of being population based and not subject to selection bias. With more than a decade of cancer incidence data for the greater Houston area, we also had sufficient power to evaluate risk associated with HAP exposure. Additionally, we chose to study childhood cancers, which have a shorter latency period than adult cancers and therefore present fewer methodologic challenges when studying environmental exposures. We used an innovative method of assigning cases with “difficult” or “nongeocodable” addresses, most of which were post office boxes, to census tracts by “matching” them to another case with similar demographic characteristics and reporting the same residential ZIP code. This method of assignment is an improvement over discarding data for subjects for whom an exact match cannot be made or by assigning these cases to the ZIP code centroid, which may create an artificial cluster of cases in one location and result in nondifferential exposure misclassification (Hurley et al. 2003).

A limitation of our study is the potential for ecologic fallacy, which may result when using grouped versus individual-level data. Nonetheless, an ecologic study design is efficient when little is known about the association under study (Friis and Sellers 2004), as is the case for childhood cancer and HAPs. Moreover, the smaller the geographic unit of analysis, the more likely bias resulting from aggregation will be minimized (Diggle and Elliott 1995). We chose to analyze data at the census tract level, which we felt to be the smallest possible spatial resolution at which to aggregate without sacrificing so much power that we could no longer detect an association. Because of the aggregate nature of our composite measure of socioeconomic status, there was likely misclassification of this variable. Also, the U.S. Census does not provide intercensus estimates at the sub-county level. Moreover, in our study, it was necessary to have population totals by race/ethnicity, age, and sex for each census tract (32 groups in total), and the accuracy of population projections for small areas has been called into question (Smith and Shahidullah 1995). Given these difficulties, we chose to develop our 10-year subpopulation estimates using the 2000 U.S. Census (U.S. Census Bureau 2006) even though we recognize that there was likely inter-census variability in population growth or decline during this period.

We based our exposure assignment on address at diagnosis, which may not represent the address at which the case or case mother resided during the prenatal or postnatal periods, which may be etiologically more relevant. Few studies have examined the degree of residential stability in epidemiologic investigations (Canfield et al. 2006; Hurley et al. 2003), and we are aware of none that have examined this issue in studies of childhood cancer. To examine possible effects of exposure misclassification due to residential mobility, we ran sensitivity analyses restricting cases to those diagnosed at ages < 5 years. For NHL, which had a majority of cases diagnosed at later ages, we found increased rates compared with the analyses including children of all ages (benzene: RR = 1.49; 95% CI, 2.04–4.54; 1,3-butadiene: RR = 1.61; 95% CI, 0.53–4.90). We also repeated the sensitivity analysis for all leukemias and each subtype (there were too few cases of Hodgkin’s disease diagnosed before 5 years of age to analyze) and found results similar to those presented among all children. These results indicate that mobility is likely not as large a problem for childhood cancers that are diagnosed earlier, such as leukemia, compared with those that are diagnosed at later ages (e.g., NHL).

The ASPEN data provide the advantage of a complex dispersion modeling approach using myriad information about sources and fate and transport of pollutants. The ASPEN model accounts for multiple factors that influence HAP levels in the environment (e.g., meteorologic conditions, emission height, deposition, and secondary formation) and represents a more robust measure of potential exposure than has previously been used in studies of childhood cancer and HAPs. Despite this advantage, our exposure assessment is subject to error. Limitations of the modeling approach include uncertainties in the emissions inventory and the inherent inability to capture variability within a census tract (Ozkaynak et al. 2008). Additionally, we relied upon levels of benzene and 1,3-butadiene in ambient air in 1999. Because cases were diagnosed in 1995–2004, the etiologic window of exposure varied over calendar time, so some misclassification of exposure is likely. On the one hand, data for a single year in the midpoint of this interval may be suitable surrogates for levels over longer periods because the relative ranking of geographic areas of high versus low ambient air pollutant levels (because of their proximity to roadways and point sources) likely remains the same. On the other hand, significant changes in air pollutant emissions due, for example, to the opening or closing of a large industrial facility during the study period could potentially affect the relative ranking of census tracts and result in some misclassification.

Because the ASPEN data represent levels of HAPs in the outdoor environment, data regarding levels in the home or other indoor environments were not available. Because children tend to spend a great deal of time indoors (Adgate et al. 2004), indoor sources of benzene and 1,3-butadiene, such as environmental tobacco smoke, may be a significant contributor to their personal exposure. However, parental smoking has been weakly and inconsistently associated with childhood leukemias in the epidemiologic literature (Belson et al. 2007). Nonetheless, we attempted to control for smoking prevalence by creating a census-tract–level variable using information from county-level rates on smoking by ethnicity, available from the Texas Department of State Health Service’s Behavioral Risk Factor Surveillance System (Texas Department of State Health Services 2005). However, effects associated with air pollutant levels were essentially unchanged with the inclusion of this variable in the model (data not shown), and we chose to report on the results from the more parsimonious model.