Association between Outdoor Air Pollution and Childhood Leukemia: A Systematic Review and Dose–Response Meta-Analysis

Background: A causal link between outdoor air pollution and childhood leukemia has been proposed, but some older studies suffer from methodological drawbacks. To the best of our knowledge, no systematic reviews have summarized the most recently published evidence and no analyses have examined the dose–response relation. Objective: We investigated the extent to which outdoor air pollution, especially as resulting from traffic-related contaminants, affects the risk of childhood leukemia. Methods: We searched all case–control and cohort studies that have investigated the risk of childhood leukemia in relation to exposure either to motorized traffic and related contaminants, based on various traffic-related metrics (number of vehicles in the closest roads, road density, and distance from major roads), or to measured or modeled levels of air contaminants such as benzene, nitrogen dioxide, 1,3-butadiene, and particulate matter. We carried out a meta-analysis of all eligible studies, including nine studies published since the last systematic review and, when possible, we fit a dose–response curve using a restricted cubic spline regression model. Results: We found 29 studies eligible to be included in our review. In the dose–response analysis, we found little association between disease risk and traffic indicators near the child’s residence for most of the exposure range, with an indication of a possible excess risk only at the highest levels. In contrast, benzene exposure was positively and approximately linearly associated with risk of childhood leukemia, particularly for acute myeloid leukemia, among children under 6 y of age, and when exposure assessment at the time of diagnosis was used. Exposure to nitrogen dioxide showed little association with leukemia risk except at the highest levels. Discussion: Overall, the epidemiologic literature appears to support an association between benzene and childhood leukemia risk, with no indication of any threshold effect. A role for other measured and unmeasured pollutants from motorized traffic is also possible. https://doi.org/10.1289/EHP4381


Introduction
Acute leukemia is the most frequent childhood cancer (Kehm et al. 2018a;Siegel et al. 2016). Its age-standardized incidence has been increasing by 0.6% per year from 1975 through most recent years (Isaevska et al. 2017;Noone et al. 2018), up to about five cases per 100,000 population in the United States and other high-income countries (Barrington-Trimis et al. 2017;Isaevska et al. 2017;Noone et al. 2018). The etiology of childhood leukemia is largely unknown, although there is increasing evidence that environmental determinants may play a major role (Metayer et al. 2016a). In addition to established and putative environmental risk factors such as ionizing and nonionizing radiation (Amoon et al. 2018; Bartley et al. 2010), infections (Kreis et al. 2019;Marcotte et al. 2014), pesticides (Malagoli et al. 2016), parental smoking , diet (Henshaw and Suk 2015), and occupation (Spycher et al. 2017), there is concern about outdoor air pollution and, in particular, exposure to contaminants released by motorized traffic, a nearly ubiquitous exposure (De Donno et al. 2018;Landrigan et al. 2018;Montero-Montoya et al. 2018;Suk et al. 2016). In recent years, five systematic reviews have evaluated the association between traffic-related air pollution and acute childhood leukemia (Boothe et al. 2014;Carlos-Wallace et al. 2016;Filippini et al. 2015;IARC 2016;Sun et al. 2014) and one has evaluated the association between benzene exposure and disease (IARC 2018). All but one (Sun et al. 2014) indicated some association. For instance, the 2016 IARC-WHO review of air pollution noted that weak associations with childhood leukemia (especially acute lymphoblastic leukemia) could not be ruled out based on epidemiological studies. Such associations were defined as "suggestive" but "inconsistent" (IARC 2016). However, most of these reviews relied on simple metrics of traffic density, and none addressed the dose-response relation between pollutants and leukemia. Furthermore, several recently published studies were not included in those reviews. In the present review, we include nine additional studies (Houot et al. 2015;Janitz et al. 2016Janitz et al. , 2017Lavigne et al. 2017;Magnani et al. 2016;Raaschou-Nielsen et al. 2018;Spycher et al. 2015;Symanski et al. 2016;Tamayo-Uria et al. 2018). In addition to analyses comparing highest versus lowest categories of exposure, we conducted a dose-response meta-analysis. The dose-response analysis incorporated several proxies to assess exposure to air pollutants besides modeling levels of air contaminants.

Search Strategy
We performed a systematic PubMed, Web of Science, and Embase literature database search formatting the research question according to the PECOS statement (Population, Exposure, Comparator(s), Outcomes, and Study design) (Morgan et al. 2018). Full details of the definition of the research question and related database search strings are reported in Table S1. In our latest database search, we looked at all papers up to 20 March 2019, without limiting the literature search to specific languages. We also used extensive search techniques to identify additional references of potential interest, including "snowballing" methods, that is, screening the reference lists of eligible or key articles (backward citations), checking which other articles have cited the eligible or key articles (forward citations), and using the "similar articles" function of online databases, which identifies articles similar to a selected paper [Booth 2008; European network for Health Technology Assessment (EUnetHTA) 2017; Vinceti et al. 2017].

Study Selection and Review
All titles and abstracts were screened by two authors (TF and MV). When there was disagreement or when the abstract was missing, both authors reviewed the full texts to determine eligibility for inclusion. Full-text evaluation of all potentially eligible studies was performed in duplicate, with controversies discussed by the two authors (TF and MV). Inclusion criteria were a) epidemiologic case-control or cohort studies, b) childhood population, c) any type of assessment of exposure to traffic from motorized vehicles, d) reporting of risk estimates or ability to compute them from the reported data. Exclusion criteria included a) ecologic study design, b) studies on only adult populations, c) exposure assessment limited to only occupational activities, and d) a lack of reporting of risk estimates for childhood leukemia.
We then examined all retrieved studies assessing outdoor air pollutant exposure through traffic density, air monitoring data, and dispersion air pollutant models based on motorized traffic. Traffic density included metrics such as traffic count (the estimated number of vehicles per day in the roads within a defined distance from the residence), road density (the sum of the length of roads within a defined area around the residence), and residential distance (the distance between the residence and a major road). We evaluated both maternal exposure during pregnancy and child's exposure at birth, time of diagnosis, and average exposure during the child's lifetime, based on residential addresses during pregnancy, at birth, or at the time of diagnosis. The effect of these different exposure windows was compared in analyses stratified into peri-and postnatal periods. We (TF and AC) also extracted other details of interest such as study size and characteristics, age at diagnosis, and leukemia subtype, that is, acute lymphoblastic leukemia (ALL) and acute myeloid leukemia (AML). We (TF and MV) assessed the characteristics and quality of included studies through the Newcastle-Ottawa Scale (NOS). Briefly, the NOS scale is based on a "star" system in which a study is assessed on three broad perspectives: the "selection" of the study groups, the "comparability" of the groups, and the ascertainment of either the "exposure" or "outcome" of interest for case-control or cohort studies, respectively (Wells et al. 2019). High-quality answers to each NOS scale question are identified with an asterisk/star. In Table S2, we reported details used for the evaluation. Total score is the sum of the score for each answer to the NOS scale. A high score indicates that the study is of high quality.

Statistical Analysis
In our meta-analysis, the ranges of exposure levels for most studies were roughly comparable, although cut points for exposure category identification differed across studies to some extent. We extracted from all studies the risk ratio (RR) in each exposure category, by abstracting the odds ratios in case-control studies and hazard ratios or rate ratios in cohort studies, as well as the number of cases or events and controls or person-years in each exposure category. We used a random-effects model to estimate a summary RR for childhood leukemia, comparing the highest and lowest exposure categories for each metric of interest.
We also performed a dose-response meta-analysis to understand the shape of the curve relating air pollution and disease risk. To do that, we used methodology developed by Greenland and Longnecker (1992) and Orsini et al. (2012) that has been applied in other contexts (Crippa et al. 2018b;Vinceti et al. 2016) to estimate the trend from the RRs across categories of pollutant exposure levels and their approximate pointwise 95% confidence intervals (CIs) based on asymptotic normality. For each of the exposure strata, we abstracted the mean or, if means were unavailable, the median. If neither mean nor median were stated, we used the midpoint. When the highest and lowest exposure categories were "open," that is, mean or median or extreme values were not presented, we entered a value that was 20% higher or lower than the closest cut point. We based this decision on the studies reporting both cut point values and median/mean values for the extreme categories and which found a difference ranging between 15% and 20%. We excluded from this dose-response meta-analysis the studies not reporting any exposure category cut points for the investigated pollutant and the studies providing only RR estimates based on 1-unit increment in exposure based on a linear model because these studies could not contribute to the assessment of departure from linearity. However, all those studies were included in the analysis of overall studies comparing highest versus lowest exposure. We then investigated the shape of the relation between traffic or pollutant exposures and childhood leukemia risk with either one-stage and two-stage doseresponse meta-analysis, using restricted cubic splines with 3 knots at fixed percentiles (10, 50, and 90%) of the exposure distribution (Crippa et al. 2018a;Orsini et al. 2012). The restricted cubic spline model was fit with a generalized least-squares regression taking into account the correlation within each set of published RRs, and combining the study-specific estimates using the restricted maximum likelihood method in a multivariable random-effects meta-analysis (Greenland and Longnecker 1992;Jackson et al. 2010;Orsini et al. 2006Orsini et al. , 2012. This dose-response meta-analysis was applied to three traffic-related metrics (traffic count, road density, and residential distance from a major road) and to the only two pollutants for which enough data for measured or monitored air levels were available [benzene and nitrogen dioxide (NO 2 )]. For one study , some data needed for the meta-analysis were not reported in the publication but were provided by two coauthors (JEH and ASP). In these analyses, estimates for benzene were adjusted for 1,3-butadiene (and vice versa).
Heterogeneity was taken into consideration in the estimation of the random-effects model in the highest versus lowest analysis. We reported the I 2 and s 2 measures for each analysis. In order to explore the source of heterogeneity, we conducted stratified analyses, by leukemia subtype (ALL and AML), age at diagnosis (<6 and ≥6 y of age), exposure time window (residence during pregnancy, at birth, and/or diagnosis), and continent. Because the fixed effect (average regression coefficients across studies) and variance/covariance structure of the random effects in spline models are not easily interpreted, we provided a graphical overlay of the marginal and conditional (best linear unbiased prediction of study-specific) dose-response trends we computed for each study (Crippa et al. 2018a). We also re-ran the analyses repeatedly, each time without one of the studies, to assess the missing study's influence and to characterize the source and magnitude of any heterogeneity of results.
Environmental Health Perspectives 046002-2 127(4) April 2019 Finally, we checked for the possible presence of publication bias using funnel plots for studies reporting highest versus lowest exposure. We used Stata software (release 15.1; Stata Corp.) for all data analyses and specifically the "metan," "metaninf," "metafunnel," "mkspline," and "drmeta" routines. Figure 1 reports the flowchart of the literature search. We retrieved 29 papers eligible for this review, including 26 case-control and three cohort studies. Detailed characteristics of included studies are summarized in Table 1. Overall, they included over 13,000 cases and (for the case-control studies) 145,000 controls worldwide. The year of diagnosis of the cases ranged from 1960 to 2012. Seven studies limited their study population to children <6 y of age (Ghosh et al. 2013;Heck et al. 2013Heck et al. , 2014Lavigne et al. 2017;Reynolds et al. 2001Reynolds et al. , 2004Symanski et al. 2016), and seven presented age-stratified analyses (Badaloni et al. 2013;Houot et al. 2015;Janitz et al. 2016Janitz et al. , 2017Savitz and Feingold 1989;Spycher et al. 2015;Vinceti et al. 2012).

Description of Included Studies
The various methods used in these studies for the assessment of exposure to traffic exhaust contaminants are summarized in Figure 2. Traffic density was assessed in 20 studies: as traffic count (n = 6), road density (n = 6), residential distance from a major road (n = 7), or a combination of them (n = 6). Another 16 studies utilized measured or modeled levels of traffic-related air contaminants, mainly benzene (n = 8), 1,3-butadiene (n = 2), NO 2 (n = 10), and particulate matter (PM) (n = 4). Other pollutants considered in these studies were ozone, other nitrogen oxides, carbon monoxide, and polycyclic organic matter. These pollutants were generally included in a multivariate model with other pollutants. Sixteen studies utilized only one method of exposure assessment, whereas the remaining 13 studies used two or more methods. NOS scale scores of included studies are reported in Table S3. The median value of the total score was 9 for both case-control and cohort studies, the highest level of this scale, indicating a generally very good quality and a substantially low risk of bias. Most studies had a population-based design, with 2 being hospital-based. Among the 27 population-based studies overall, 19 (including all those assessing benzene exposure) did not depend on voluntary participation of children and their families. These studies performed an exposure assessment, without contacting the participants, based on modeling applied to location information for the residence of study participants.
To assess exposure, most studies used either the residential address at the time of diagnosis or the address retrieved from the  death certificate. Some studies used residential address during pregnancy or at birth; only two studies evaluated exposures both at and after birth (Figure 2). Badaloni 2013 Crosignani 2004 Feychting 1998 Ghosh 2013 Harrison 1999 Heck 2013   Characteristics of included studies according to traffic pollution assessment, and other characteristics. Composite index, combination of various measures of traffic density; D, residential address retrieved from death certificate; distance, distance between the residence and a major road; NO 2 , nitrogen dioxide; PM, particulate matter; road density, sum of the length of roads within a defined area around the residence; traffic count, estimated number of vehicles per day in the roads within a defined distance from the residence. For population-based studies, we indicated whether exposure assessment depended upon acceptance to participate in the study (voluntary) or it was based solely on residential address, without contacting the participants (involuntary). variables, and all but four included a measure of socioeconomic status in the regression model. The databases with extracted data used for both highest versus lowest (see Excel Table S1) and dose-response meta-analysis (see Excel Table S2) are included and described in the Supplemental Excel File.

Highest versus Lowest Meta-Analysis
Tables 2 and 3 present the summary RR and 95% CIs comparing the highest with lowest exposure categories for each exposure assessment method and the corresponding forest plots are shown in Figures S1-S15. As shown in Table 2, the summary RR comparing the highest to lowest exposure categories from studies using traffic density to assess air pollution exposure and leukemia risk was slightly elevated, being RR = 1.09 (95% CI: 1.00, 1.20). Of the eight studies that investigated benzene exposure, two involved the same population. For the remaining seven, the summary RR = 1.27 (95% CI: 1.03, 1.56). Of the eligible studies assessing NO 2 exposure, two were updated versions of previous reports. The eight independent studies of NO 2 exposure yielded a summary RR = 1.04 (95% CI: 0.90, 1.19). We observed a moderate association for PM 10 and childhood leukemia [two studies, summary RR = 1.20 (95% CI: 0.70, 2.04)], but the association was much weaker for PM 2:5 [three studies, summary RR = 1.05 (95% CI: 0.94, 1.16)]. 1,3-butadiene was associated with increased disease risk, based on two studies [summary RR = 1.45 (95% CI: 1.08, 1.95)] (Table 3).
In analyses stratified by disease subtype (Tables 2 and 3; see also Figures S1-S6), RR estimates were found to differ markedly only for benzene exposure, for which the summary RR became 1.09 (95% CI: 0.88, 1.36) for ALL, based on seven studies, and 1.84 (95% CI: 1.31, 2.59) for AML, based on five studies. No such change in RR according to leukemia subtype was found for traffic indicators, for NO 2 or for PM. 1,3-Butadiene exposure was associated with a higher RR for AML compared with ALL, but the summary estimates were statistically unstable, being based on two studies only. For the aforementioned pollutants, RRs according to disease subtype were roughly the same after restricting the analysis to children <6 y of age, with the exception of the summary RRs associated with benzene exposure, showing a larger difference between AML and ALL in this younger subgroup (Tables 2 and 3; see also Figures S7-S11). Data for older children, however, were based on one study only, and the RR was imprecise (Tables 2 and 3; see also Figures S12-S15). In general, for other air pollution exposure metrics, results were somewhat stronger among younger children, but there were few studies with data on children ≥6 y of age, and the latter results were statistically unstable.
In analyses stratified by study region, the North American studies showed a summary RR = 1.02 (95% CI: 0.89, 1.16) and 1.21 (95% CI: 1.04, 1.41) for traffic density and benzene exposure, respectively (Tables 2 and 3; see also Figures S1-S6). For European studies, the corresponding estimates were higher, that is, RR = 1.25 (95% CI: 1.05, 1.49) for traffic indicators and 1.36 Table 2. Summary risk ratios (RRs) of childhood leukemia in the highest exposure category versus the lowest one for traffic density, benzene, and nitrogen dioxide (NO 2 ) exposure, for all studies and stratified by age at diagnosis, leukemia subtype, exposure timing, and region. Note: -, data not available; ALL, acute lymphoblastic leukemia; AML, acute myeloid leukemia; CI, confidence interval; I 2 , I-squared statistic; RR, risk ratio; s 2 , tau-squared statistic.
(95% CI: 0.92, 2.00) for benzene exposure. When exposure timing in relation to leukemia occurrence was taken into account, studies assessing exposure on the basis of either child's residence at diagnosis or the longest place of residence yielded a higher summary RR for traffic density of 1.32 (95% CI: 1.12, 1.55) and for benzene 1.36 (95% CI: 0.92, 2.00). When exposure assessment was based on maternal residence at birth or during pregnancy, the RR = 0.98 (95% CI: 0.90, 1.06) for traffic indicators and 1.21 (95% CI: 1.04, 1.41) for benzene.

Dose-Response Meta-Analysis
In the dose-response meta-analysis, five studies (Abdul Rahman et al. 2008;Harrison et al. 1999;Houot et al. 2015;Janitz et al. 2016;Langholz et al. 2002) had no value assigned to open-ended categories at the end of the exposure scale. For these we entered a value 20% higher or lower than the closest cut points. In the analysis of traffic density indicators, the restricted cubic spline analysis ( Figure 3) showed little association between leukemia RR and the number of vehicles per day (Langholz et al. 2002;Pearson et al. 2000;Raaschou-Nielsen et al. 2001;Reynolds et al. 2004) in the street closest to the child's residence, except at the highest exposure levels where a small and statistically imprecise excess risk emerged ( Figure 3A). Three studies used a similar measure of road density around the child's residence (Houot et al. 2015;Magnani et al. 2016;Reynolds et al. 2004), yielding results comparable with those obtained for the vehicles per day metric ( Figure 3B). Distance from a major nearby road (Abdul Rahman et al. 2008;Badaloni et al. 2013;Crosignani et al. 2004;Harrison et al. 1999;Houot et al. 2015;Spycher et al. 2015) showed little association with RR until 150 m from the road's edge, below which an indication of a higher RR emerged and increased steeply with decreasing distance ( Figure 3C). For benzene (Crosignani et al. 2004;Heck et al. 2014;Houot et al. 2015;Janitz et al. 2017;Raaschou-Nielsen et al. 2018;Vinceti et al. 2012), there was an approximately linear increase in estimated risk starting from the lowest levels of benzene exposure (1-15 lg=m 3 ), as shown in Figure 4. After stratifying by leukemia subtype, we found a considerably stronger association between benzene exposure and risk of AML Houot et al. 2015;Janitz et al. 2017;Raaschou-Nielsen et al. 2018;Vinceti et al. 2012) than we found for ALL (Crosignani et al. 2004;Heck et al. 2014;Houot et al. 2015;Janitz et al. 2017;Raaschou-Nielsen et al. 2018;Vinceti et al. 2012). Concerning NO 2 (Badaloni et al. 2013;Feychting et al. 1998;Heck et al. 2014;Houot et al. 2015;Janitz et al. 2016;Weng et al. 2008), the doseresponse meta-analysis showed some evidence of an excess risk starting from 40 lg=m 3 to the maximum amount investigated, 60 lg=m 3 , although the increase was statistically unstable ( Figure 5). Subgroup analysis according to disease type showed that this excess risk at the higher exposure levels was limited to ALL ( Figure 5). The dose-response analysis could not be carried out for 1,3-butadiene or for PM, due to the limited number of studies reporting the information needed for such an analysis.

Sensitivity Analyses and Publication Bias
We repeated all analyses after systematically excluding each study in turn from the meta-analysis (see Table S4-S5 and Figures S16-S18). None of these analyses appreciably changed the various summary estimates. Use of alternative estimates for Table 3. Summary risk ratios (RRs) for association of childhood leukemia with particulate matter (PM 2:5 =PM 10 ) and 1,3-butadiene comparing the highest versus the lowest exposure categories for all studies and stratified by age at diagnosis, leukemia subtype, exposure timing, and region. the highest and lowest exposure categories with unknown mean/ median values, that is, entering a ± 15% value instead of ± 20% with relation to the closest (lower and upper) available boundary, also had little effect on the results (see Figures S19-S21). Funnel plots based on the different exposure assessment methods showed a slightly asymmetric distribution, especially for traffic and benzene, and therefore we could not entirely rule out the occurrence of publication bias ( Figure S22). Finally, we report study-specific dose-response trends in addition to the overall dose-response meta-analyses in Figures S23-S25.

Discussion
The reviewed literature of epidemiologic studies and the combined dose-response meta-analysis appear to support a relation of air pollution, and particularly benzene emissions from motorized traffic, with the risk of childhood leukemia (Table 2 and Figure 4); the strongest associations were found with AML and there was little evidence of any threshold of exposure. For exposure to NO 2 , little evidence for an association was found except at the highest levels of exposure (Table 3 and Figure 5). Based on a limited number of studies, there was some association with 1,3butadiene (Table 3) and a weak association with PM 10 (Table 3), although the dose-response curve could not be assessed for these contaminants. Air pollution has not been unanimously considered to be associated with childhood leukemia (Sun et al. 2014) due to some inconsistencies across epidemiologic studies (Schüz and Erdmann  (Langholz et al. 2002;Pearson et al. 2000;Raaschou-Nielsen et al. 2001, Reynolds et al. 2004), (B) road density in km=km 2 (Houot et al. 2015;Magnani et al. 2016;Reynolds et al. 2004), and (C) residential distance from a major road in meters (Abdul Rahman et al. 2008;Badaloni et al. 2013;Crosignani et al. 2004;Harrison et al. 1999;Houot et al. 2015;Spycher et al. 2015). Spline curve (black solid line) with 95% confidence limits (gray dashed lines). RR, risk ratio.  (Crosignani et al. 2004;Heck et al. 2014;Houot et al. 2015;Janitz et al. 2017;Raaschou-Nielsen et al. 2018;Vinceti et al. 2012), (B) acute lymphoblastic leukemia only (Crosignani et al. 2004;Heck et al. 2014;Houot et al. 2015;Janitz et al. 2017;Raaschou-Nielsen et al. 2018;Vinceti et al. 2012), and (C) acute myeloid leukemia only Houot et al. 2015;Janitz et al. 2017;Raaschou-Nielsen et al. 2018;Vinceti et al. 2012). Spline curve (black solid line) with 95% confidence limits (gray dashed lines). RR, risk ratio. 2016), although most of the recent reviews tend to support such a relation (Boothe et al. 2014;Carlos-Wallace et al. 2016;Filippini et al. 2015;IARC 2016IARC , 2018Spycher et al. 2017;Steinmaus and Smith 2017). In some instances, the reevaluation of previous databases by conducting stratified analyses for leukemia subtypes has identified associations between benzene exposure from motorized traffic and disease risk, particularly AML, that had previously gone undetected (Raaschou-Nielsen et al. 2001. In the present review, we took advantage of the availability of nine recently published studies, a reanalysis of a former study , and a new method for dose-response spline regression analysis (Crippa et al. 2018a) to assess for the first time, to the best of our knowledge, the shape of the relation of exposure to benzene, NO 2 , and traffic density with childhood leukemia risk. We also attempted to identify thresholds of exposure, timing of exposure, and the roles of some potential effect modifiers.
A role of outdoor air pollution in childhood leukemia etiology is supported by several laboratory studies, which have provided biological plausibility for this association (Andreoli et al. 2012;Arayasiri et al. 2010;Bollati et al. 2007;Buthbumrung et al. 2008;Carlos-Wallace et al. 2016;D'Andrea and Reddy 2018;Elliott et al. 2017;Jiang et al. 2016;Loomis et al. 2013;McKenzie et al. 2017;Suk et al. 2016;Vattanasit et al. 2014). A review published by the IARC in 2016 defined air pollution as carcinogenic due to its ability to induce lung cancer in humans. This review reported evidence from animal and mechanistic studies supporting the carcinogenicity of air pollution for lung cancer and identified the several inorganic and organic components of air pollution classified as established, probable, or possible carcinogens (IARC 2016). However, with reference to the above evidence, it should be noted that many of the laboratory studies were carried out using mixtures, precluding identification of the specific pollutants responsible for the increased risk. Evidence from animal carcinogenicity data and mechanistic data has also recently been made available, specifically for benzene (IARC 2018), showing the capacity of this compound to generate reactive electrophilic metabolites, induce oxidative stress, and DNA damage as well as immunotoxic and hematotoxic effects.
Although exposure to benzene, an established leukemogen in adults for the acute myeloid form (IARC 2018), was associated with disease risk in our analysis, exposure assessed through traffic indicators such as road density or distance showed limited association with childhood leukemia (Figure 3). This discrepancy may indicate that traffic-related metrics are less sensitive as a measure of outdoor air pollutants than modeled levels of carcinogenic pollutants such as benzene and 1,3-butadiene, as has been reported elsewhere (Wu et al. 2011). This would explain the smaller estimates of effect seen when using surrogate measures of exposure ( Table 2). The higher risk observed for families living within 150 m of a major road is of considerable interest, however, because it suggests a distance potentially usable by policymakers considering placement of schools or other facilities for children. This 150-m corridor is characterized by high levels of carcinogenic air pollutants, including heavy metals and benzene, with a sudden drop in their concentration outside that range (Karner et al. 2010).
A few of the studies we reviewed have suggested that other air pollutants, such as 1,3-butadiene and selenium Knox 2006;Symanski et al. 2016;Vinceti et al. 2012), may be associated with childhood leukemia risk. However, few studies have controlled for other pollutants simultaneously Symanski et al. 2016). Therefore, we acknowledge that an etiologic relation between air pollution and childhood leukemia risk, which is supported by our findings and appears to be mainly attributable to benzene, might be at least partly due to other pollutants in outdoor air that covary with benzene emissions (Ghosh et al. 2012;Heck et al. 2014;Vinceti et al. 2012;Wilhelm et al. 2011), acting either alone or as mixtures. In the two published studies on 1,3-butadiene, correlations with benzene were high Symanski et al. 2016). Exposure assessment methods based on denser air pollution monitoring networks or more sophisticated air pollution models may be better able to differentiate between the effects of benzene and 1,3-butadiene and, more generally, to tease apart the influence of each toxic agent.
Our meta-analysis suggested a differential effect of trafficrelated benzene exposure according to the clinical subtype of leukemia, that is, a much higher risk associated with AML compared with ALL (Table 2 and Figure 3). Conversely, little evidence of such a differential relation with disease type emerged when assessing exposure through traffic indicators and PM levels. For . Dose-response meta-analysis of childhood leukemia risk from NO 2 exposure for (A) all leukemias (Badaloni et al. 2013;Feychting et al. 1998;Heck et al. 2014;Houot et al. 2015;Janitz et al. 2016;Weng et al. 2008), (B) acute lymphoblastic leukemia only Houot et al. 2015;Janitz et al. 2016), and (C) acute myeloid leukemia only Houot et al. 2015;Janitz et al. 2016). Spline curve (black solid line) with 95% confidence limits (gray dashed lines). RR, risk ratio. NO 2 , a pollutant so far not recognized as a carcinogen, the association appeared to be limited to ALL and to the highest exposure levels only (Table 3 and Figure 5). We observed higher RRs associated with benzene exposure in the postnatal period compared with the perinatal window of exposure (Table 2). We found higher estimates for benzene exposure in younger (<6 y of age) compared with older children, although for the latter, little data were available. There was also an indication of an interaction between exposure window and age given that the excess disease risk associated with benzene exposure at diagnosis compared with exposure at birth was limited to the youngest children, although based on only one study. A previous meta-analysis investigating the association between traffic exposure and childhood leukemia risk found similar results, that is, a higher summary RR for exposure at diagnosis than for exposure at birth or gestation; effect-measure modification by age was not investigated in that study (Carlos-Wallace et al. 2016).
We also detected slightly stronger RRs for benzene exposure in Europe compared with North America (Table 2). This difference might result from misclassification of exposure stemming from greater residential mobility in North America (Lupo et al. 2010;Tee Lewis et al. 2019;Urayama et al. 2009) or by the larger influence of postnatal exposure assessment in European studies. We also note that there is a much larger percentage of diesel cars in Europe than in North America (Neumaier 2014), possibly reflecting differences in pollutant exposure characterizing the two regions. Another potential source of discrepancy might be the use in some of the U.S. studies of methods based on a "nearest-neighbor" exposure assessment that has been carried out at the census tracts level or using a limited number of air monitors Symanski et al. 2016). These methods likely yield a lower spatial resolution compared with personlevel exposure data based on air pollution dispersion modeling at the residential address, thus adding to exposure misclassification and likely diluting the RRs.
Some degree of unmeasured confounding may have occurred in the studies, due to sources of outdoor air pollution such as oil and gas development (Elliott et al. 2017;McKenzie et al. 2017) and industrial sources (García-Pérez et al. 2015;Park et al. 2017), indoor air pollution from heating sources and dust (Whitehead et al. 2011(Whitehead et al. , 2015, passive smoking (Metayer et al. 2016b;Pyatt and Hays 2010), magnetic field exposure (Amoon et al. 2018), and socioeconomic factors (Kehm et al. 2018b). However, most of these factors are unlikely to play a major role in disease etiology, and some studies we included in the metaanalysis, particularly the most recent ones, took into account several potential confounding factors, yielding results in line with the overall meta-analysis (Table 1 shows the confounding factors considered in each study). In addition, we systematically used the most adjusted RRs to carry out our meta-analysis. The occurrence of selection bias could also be ruled out particularly for the most recent studies, including all those assessing benzene exposure in relation to disease risk given that they were population-based and not dependent on voluntary participation.
In conclusion, in this systematic review and dose-response meta-analysis, we found that traffic-related air pollution, particularly exposure to benzene, was associated with excess risk of childhood leukemia. No apparent minimal threshold of exposure emerged for benzene, whereas analyses for traffic density and NO 2 gave evidence of such a threshold. Disease subtype, windows of exposure and child's age appeared to modify these associations.