Lung Cancer and Exposure to Nitrogen Dioxide and Traffic: A Systematic Review and Meta-Analysis

Background and objective Exposure to traffic-related air pollutants is an important public health issue. Here, we present a systematic review and meta-analysis of research examining the relationship of measures of nitrogen oxides (NOx) and of various measures of traffic-related air pollution exposure with lung cancer. Methods We conducted random-effects meta-analyses of studies examining exposure to nitrogen dioxide (NO2) and NOx and its association with lung cancer. We identified 20 studies that met inclusion criteria and provided information necessary to estimate the change in lung cancer per 10-μg/m3 increase in exposure to measured NO2. Further, we qualitatively assessed the evidence of association between distance to roadways and traffic volume associated with lung cancer. Results The meta-estimate for the change in lung cancer associated with a 10-μg/m3 increase in exposure to NO2 was 4% (95% CI: 1%, 8%). The meta-estimate for change in lung cancer associated with a 10-μg/m3 increase in NOx was similar and slightly more precise, 3% (95% CI: 1%, 5%). The NO2 meta-estimate was robust to different confounding adjustment sets as well as the exposure assessment techniques used. Trim-and-fill analyses suggest that if publication bias exists, the overall meta-estimate is biased away from the null. Forest plots for measures of traffic volume and distance to roadways largely suggest a modest increase in lung cancer risk. Conclusion We found consistent evidence of a relationship between NO2, as a proxy for traffic-sourced air pollution exposure, with lung cancer. Studies of lung cancer related to residential proximity to roadways and NOx also suggest increased risk, which may be attributable partly to air pollution exposure. The International Agency for Research on Cancer recently classified outdoor air pollution and particulate matter as carcinogenic (Group 1). These meta-analyses support this conclusion, drawing particular attention to traffic-sourced air pollution. Citation Hamra GB, Laden F, Cohen AJ, Raaschou-Nielsen O, Brauer M, Loomis D. 2015. Lung cancer and exposure to nitrogen dioxide and traffic: a systematic review and meta-analysis. Environ Health Perspect 123:1107–1112; http://dx.doi.org/10.1289/ehp.1408882


Introduction
Exposure to air pollution from high-density urban traffic is an important public health issue. Traffic-related air pollutants are associated with acute health outcomes as well as chronic respiratory and cardiovascular outcomes [Brunekreef et al. 1997; Health Effects Institute (HEI) 2010; Hoek et al. 2002;Tonne et al. 2007]. Associations of traffic-related air pollution with allergies, asthma, and respiratory infections have also been demonstrated in children (Brauer et al. 2002) and in both developed and developing nations (Brunekreef et al. 2009). Evaluation of carcinogenicity of traffic-based pollutants is complicated by the fact that the diseases are rare and often develop after a long latency period. Thus, multiple large, long-term observational studies are required to demonstrate evidence of carcinogenicity from traffic-related air pollution. A recent systematic review and meta-analysis evaluated the evidence from epidemiology of lung cancer associated with particulate matter exposure (Hamra et al. 2014); importantly, road traffic is an important contributor to urban particulate matter. The evidence supported the recent classification of ambient air pollution and particulate matter as a Group 1 carcinogen by the International Agency for Research on Cancer (IARC).
Here, we present a systematic review and meta-analysis of research examining the relationship of measures of nitrogen oxides [measured as nitrogen dioxide (NO 2 ) or nitrogen oxides (NO x )] and of various measures of traffic exposure and lung cancer. In urban settings, NO 2 is often used as a marker for traffic-based air pollution, which is a complex mixture of many diverse carcinogens, such as volatile organic compounds, metals, and carbonyls (Brook et al. 2007;Valavanidis et al. 2008). We conducted analyses of studies focused on NO x and NO 2 ; the latter is a component of the former, but both have been used as markers of traffic exposure in epidemiological research. We also evaluated studies that used distance from roadways or volume of traffic on nearby roadways as indicators of exposure to traffic-related air pollution (Brook et al. 2007).
Background and oBjective: Exposure to traffic-related air pollutants is an important public health issue. Here, we present a systematic review and meta-analysis of research examining the relationship of measures of nitrogen oxides (NO x ) and of various measures of traffic-related air pollution exposure with lung cancer. Methods: We conducted random-effects meta-analyses of studies examining exposure to nitrogen dioxide (NO 2 ) and NO x and its association with lung cancer. We identified 20 studies that met inclusion criteria and provided information necessary to estimate the change in lung cancer per 10-μg/m 3 increase in exposure to measured NO 2 . Further, we qualitatively assessed the evidence of association between distance to roadways and traffic volume associated with lung cancer. results: The meta-estimate for the change in lung cancer associated with a 10-μg/m 3 increase in exposure to NO 2 was 4% (95% CI: 1%, 8%). The meta-estimate for change in lung cancer associated with a 10-μg/m 3 increase in NO x was similar and slightly more precise, 3% (95% CI: 1%, 5%). The NO 2 meta-estimate was robust to different confounding adjustment sets as well as the exposure assessment techniques used. Trim-and-fill analyses suggest that if publication bias exists, the overall meta-estimate is biased away from the null. Forest plots for measures of traffic volume and distance to roadways largely suggest a modest increase in lung cancer risk. conclusion: We found consistent evidence of a relationship between NO 2 , as a proxy for trafficsourced air pollution exposure, with lung cancer. Studies of lung cancer related to residential proximity to roadways and NO x also suggest increased risk, which may be attributable partly to air pollution exposure. The International Agency for Research on Cancer recently classified outdoor air pollution and particulate matter as carcinogenic (Group 1). These meta-analyses support this conclusion, drawing particular attention to traffic-sourced air pollution. We abstracted estimates of lung cancer associated with three exposure metrics: NO 2 , NO x (NO + NO 2 ), and volume of traffic and distance to nearest traffic source. Further, we abstracted information on confounders for which the authors adjusted in their analyses. All studies adjusted for age and sex.
Statistical analyses. For consistency, all study estimates were converted to represent the change in lung cancer incidence or mortality per 10-μg/m 3 exposure to NO 2 or NO x . All studies that considered NO x used units of micrograms per cubic meter. Studies of NO 2 measured exposure in units of micrograms per cubic meter or parts per billion. For studies measuring NO 2 in parts per billion, we used a conversion factor of 1 ppb = 1.88 μg/m 3 , which is based on ambient pressure of 1 atmosphere and a temperature of 25°C (Vrijheid et al. 2011). We abstracted the effect estimates that were believed to most effectively adjust for confounding (i.e., the least biased effect estimates), which were largely the "main models" discussed by the authors. There did not appear to be issues of overadjustment or potential adjustment for causal intermediates or colliders in any studies. Most studies used a single-pollutant model; when authors used both a single-and a multi-pollutant model, we chose estimates from the former.
We used random-effects estimation for our meta-analysis. Unlike fixed-effects estimation, random-effects methods include an estimate of the percent of the total variance attributable to between study inconsistency, referred to as the I 2 value (DerSimonian and Laird 1986;Higgins and Green 2008;Higgins et al. 2003). One study (Abbey et al. 1999) presented estimates by sex; these estimates were combined with fixed-effects estimation and then included in analyses of the overall meta-estimates. This assumes that the stratum-specific estimates are collapsible across strata of sex.  Table 2; others are not applicable (NA). b Studies using land use regression, spatiotemporal, inverse distance-weighted, or air dispersion models are more capable of addressing intra-city comparisons of air pollution exposure because they more closely estimate residential exposure to air pollutants. Each technique achieves this goal in a similar but distinct manner: Inverse distance weighting accounts for the distance from monitoring sites to the individual's residence; land use regression includes relevant environmental GIS-derived variables in regression analyses; spatiotemporal models create air pollutant surfaces that vary by space and time using relevant environmental GIS-derived variables; and air dispersion models use relevant environmental GIS-derived variables, but assume that air pollutant transport follows a deterministic process based on Gaussian plume equations. In contrast, fixed-site monitors apply single exposure values obtained at monitoring sites to the populations surrounding the monitors, and are most useful for intercity comparisons. c NR indicates that exposure means (± SDs) were not reported. Nafstad et al. (2003) report the median value of NO x . d Studies that use a case-control design. Unless otherwise noted, all studies are cohort designs. e Studies that examine lung cancer risk associated with distance to roadways or traffic density. These measures are explicitly described in Figure 2. f Studies of NO x . All other studies examine NO 2 exposure. g We report the highest and lowest cohort specific mean (± SD) of measured air pollution from Raaschou-Nielsen et al. (2013) among the cohorts pooled for the ESCAPE study. Katanoda et al. (2011) report the range of observed NO 2 values.
Forest and funnel plots were created to provide a visual representation of the distribution of study-specific effect estimates. In addition, we conducted trim-and-fill analyses, which, in the presence of funnel plot asymmetry, recalculate the meta-estimate based on hypothetical, unobserved studies that would have been necessary to create a symmetrical funnel plot (Duval and Tweedie 2000;Higgins and Green 2008). Finally, we conducted influence analyses, where the meta-estimate is recalculated excluding one study at a time to test if the overall metaestimate is robust to exclusion of any single study. We did not include formal assessments of study quality. Analyses were conducted using STATA software (version 12.1; StataCorp).

Studies included.
We identified 20 studies that met our initial inclusion criteria through PubMed (http://www.ncbi.nlm.nih.gov/ pubmed) and three additional studies through discussion among co-authors or by searching article references (Cao et al. 2011;Hart et al. 2011;Puett et al. 2014). Six of the 23 studies had at least partially overlapping populations; of these, the three judged to be less informative were excluded (Beeson et al. 1998;Jerrett et al. 2013;Vineis et al. 2006). More informative studies are those that have either better information on estimated individual exposure or a greater number of observed lung cancers. Of particular note, although the exposure assessment for an updated American Cancer Society Cancer Prevention II Study was more accurate (Jerrett et al. 2013), we considered a slightly older study that benefits from a much larger sample size (Krewski et al. 2009). The final data set included 20 studies that met all required criteria for inclusion in our meta-analyses (Abbey et al. 1999;Beelen et al. 2008;Cao et al. 2011;Carey et al. 2013;Cesaroni et al. 2013;Filleul et al. 2005;Hart et al. 2011;Heinrich et al. 2013;Hystad et al. 2013;Katanoda et al. 2011;Krewski et al. 2009;Lipsett et al. 2011;Nafstad et al. 2003;Nyberg et al. 2000;Puett et al. 2014;Raaschou-Nielsen et al. 2010Villeneuve et al. 2014;Yorifuji et al. 2013). Table 1 summarizes the 20 studies included in our analyses, and reports mean (± SD) of annual, individual exposure to NO 2 or NO x in units of micrograms per cubic meter or parts per billion, depending on what the authors originally reported. Of these studies, 15 and 4 provide estimates of the change in lung cancer incidence or mortality associated with NO 2 and NO x , respectively; Puett et al. (2014) provide information only for distance from roadways. Because the ratio of NO 2 to NO x is complex and location-specific, and depends on the time-varying levels of atmospheric oxides, there is no simple conversion factor to estimate the percent of NO x represented by NO 2 . We therefore summarize meta-estimates from studies of NO x separately.
NO 2 /NO x and lung cancer. Figure 1 and Table 2 summarize the studies examining the change in lung cancer incidence and mortality associated with traffic exposure as measured by NO 2 . The overall metaestimate of the change in lung cancer incidence or mortality per 10-μg/m 3 increase in exposure is 4% [95% confidence interval (CI): 1%, 8%], with an I 2 estimate of 72.8%. Region-specific estimates vary; for Europe, North America, and Japan, the metaestimates are 2% (95% CI: -1%, 6%), 7% (95% CI: 0%, 14%), and 11% (95% CI: 3%, 20%), respectively. The funnel plot is visually asymmetrical (Figure 2). Trim-andfill analyses indicate that five hypothetical studies would need to have been observed to create a symmetrical funnel plot; the metaestimate for the change in lung cancer per 10-μg/m 3 increase in NO 2 recalculated based on inclusion of these unobserved studies was attenuated toward the null: 1% (95% CI: -2%, 5%). A cohort study by Naess et al. (2007) of all residents in Oslo, Norway, who were 51-90 years of age met all inclusion criteria, but could not be included in our analyses because risk estimates from this study were based on quartiles of exposure. However, the results of that work support an Figure 1. Forest plot of study-specific estimates of relative risk (RR) of lung cancer associated with a 10-μg/m 3 increase in exposure to NO 2 . The meta-estimate and weights in the forest plot are estimated from random effects meta-analyses. increase in lung cancer risk associated with NO 2 ; percent change in lung cancer mortality associated with a 1-quartile increase in NO 2 for men and women age 51-70 years were 7% (95% CI: -3%, 18%) and 23% (95% CI: 10%, 38%), and for men and women age 71-90 years were 9% (95% CI: -2%, 20%) and 12% (95% CI: -2%, 27%), respectively (Naess et al. 2007). The meta-estimate for the change in lung cancer associated with a 10-μg/m 3 increase in measured NO x is 3% (95% CI: 1%, 5%), based on five relative risk estimates for lung cancer associated with a 10-μg/m 3 increase in NO x derived from four studies: 1.03 (0.99, 1.07) (Cao et al. 2011), 1.08 (1.02, 1.15) , 1.00 (0.98, 1.03) (Raaschou-Nielsen et al. 2013), and 1.04 (1.01, 1.08) and 1.00 (0.93, 1.08) (Raaschou-Nielsen et al. 2010). Figure 3 provides a visual summary of these studyspecific relative risk estimates. The smaller number of risk estimates available does not allow for detailed analyses of sensitivity to confounder adjustment or other subgroupspecific analyses. However, it is notable that the meta-estimate for lung cancer associated with NO x is similar to that of NO 2 , but slightly more precise. Table 2 summarizes sensitivity and subgroup analyses for lung cancer associated with measured NO 2 . When restricted to studies that adjusted for confounding due to smoking status, socioeconomic status/income, education, or occupation, meta-estimates are largely unchanged; point estimates are similar, and confidence intervals largely overlap with that of the overall meta-estimate. Further, when we divide studies by the exposure assessment technique used, the meta-estimates are identical, or nearly identical, to the overall meta-estimate, but with slightly different confidence intervals. Influence analyses suggested that no single study influenced the overall meta-estimate (see Supplemental Material, Table S1).
Evidence from studies that consider distance from traffic sources or volume of traffic at roads nearest to residence is summarized in Figure 4. Each study uses a unique exposure contrast; thus, we do not provide a meta-estimate of these study specific measures of lung cancer associated with traffic exposure. We only note that many of the estimates reported are either close to or include a null association within the bounds of 95% confidence intervals reported.

Discussion
We provide a systematic review and metaanalyses of the evidence regarding the association of NO 2 , NO x , and distance to roadways or volume of vehicles as markers of exposure to traffic-based air pollution. The studies we considered represent many regions of the world and vary by the level of exposure measured, exposure assessment technique, and potential confounders considered. The results of the overall meta-analyses support a relationship between lung cancer and exposure to air pollution from traffic sources.
Analyses stratified by continent suggest the possibility of geographic variability in the magnitude of the relative risk. However, confidence intervals largely overlap, which limits the ability to draw conclusions about heterogeneity. In addition, exposure lags varied in studies from 0 years to 20 years, but were not reported in some studies; Raaschou-Nielsen et al. (2010) found no difference in risk estimates when using a 0-year versus a 10-year exposure lag. Trim-and-fill analyses for NO 2 meta-estimates indicated the potential for publication bias, which, if present, would shift results away from the null. To obtain the null result suggested by the   Figure 2. Funnel plot of study-specific estimates of the relative risk of lung cancer associated with a 10-μg/m 3 increase in exposure to NO 2 . The metaestimate represented by the vertical solid line of the funnel plot is based on a fixed-effects metaanalysis. Fixed-effects meta-analyses are required to assess the potential for publication bias, and assume that there is no between-study variation. trim-and-fill analyses, we would need to assess five hypothetical (i.e., unobserved) studies that obtained results that are the inverse of five studies with the largest and least precise risk estimates. Otherwise, the meta-estimate for NO 2 was robust to confounders considered and methods of exposure assessment used by the authors of each study. No formal assessment of study quality was conducted as part of this meta-analysis. Epidemiologic evidence alone does not allow us to assess nitrogen oxides as a causative agent in development of lung cancer. Nitrogen oxides are not thought to be carcinogenic agents (Valavanidis et al. 2008). Rather, they serve as a marker for other pollutants formed in high-temperature combustion of fossil fuels, a mixture that also includes fine particles. The role of nitrogen oxides as a marker for traffic-related air pollution is most plausible in urban settings, where traffic is often the primary source of NO x in the atmosphere and the main source of variability in NO x levels. NO x is the sum of NO (nitrogen oxide) and NO 2 ; because NO 2 is a secondary oxidation product of atmospheric reactions involving NO, the latter of which is directly emitted in combustion, NO x is considered a better indicator of traffic pollution than is NO 2 alone. Because NO 2 is a component of NO x and will, therefore, always be present at a lower concentration, one would expect a lower effect estimate for NO x (per 1-μg/m 3 ) compared with that for NO 2 , assuming that either was the true carcinogenic agent in the traffic pollution mixture. Recent studies have shown that NO x is highly correlated with combustion by-products, such as volatile organic compounds and carbonyls (Brook et al. 2007;Curren et al. 2006), many of which are known carcinogens. Combustion of fuel from motor vehicles can contribute a significant portion of outdoor air pollution, particularly in urban areas.
Because cancer is an outcome that is believed to develop over many years, exposure modeling that better approximates longterm, individual-level exposures may provide more valid estimates of the carcinogenic potential of traffic, or other, air pollution. This is distinct from studies of outcomes with sudden onset; the latter can rely on aggregate measures, such as fixed-site monitors and short-term exposure information, because these measures adequately describe relevant, short-term exposure episodes such as high traffic or temperature days . Even when modeling techniques are used, the residential-level exposures assigned to an individual are, at best, an approximation of individual-level exposure. Further, each modeling technique will vary in its ability to provide individual-level estimates of exposure to pollution. Land use regression, used in four studies considered here, can provide better prediction around major highways where traffic patterns are better known, but not in areas without dense monitoring systems (Jerrett et al. 2005). Alternately, air dispersion models can account for spatiotemporal differences in pollution concentrations, but rely on assumptions that dispersion patterns are Gaussian, which may not be valid (Bellander et al. 2001). Both of these techniques, as well as other spatiotemporal models, attempt to assess within-city contrasts that cannot be accounted for with fixed-site monitors. In short, all exposuremodeling techniques provide, at best, an approximation of individual exposure. A detailed comparison of the most commonly used exposure assessment techniques was provided by Jerrett et al. (2005). Results of our meta-analyses for NO 2 were robust to stratification by exposure assessment technique; that is, results were nearly identical regardless of whether or not an exposure modeling technique was used in place of fixed site monitors. It is noteworthy that the meta-estimate restricted to exposure modeling techniques was slightly more precise than that obtained from studies using fixed-site monitors.
Similarly, risk estimates for distance to traffic or roadways, summarized in Figure 2, are crude in that they do not postulate what an individual's exposure at a residence would be from traffic sources. Distance to roadways can be a marker of socioeconomic status, the influence of which can be difficult to fully take into account. Also, the exposure measures for which the relative risks are estimated are inconsistent across studies; thus, it is impossible to summarize them in any quantitatively meaningful way.
A recent meta-analysis (Hamra et al. 2014) showed a strong and consistent association between exposure to particulate matter and lung cancer; the results, like those presented here, were robust to different confounder sets considered, as well as variation in exposure assessment techniques and study locations (Hamra et al. 2014). In the present review we considered indicators of exposure to motor vehicle traffic (nitrogen oxides and distance). Vehicle traffic also generates particulate matter, which can also arise from numerous anthropogenic and natural processes outside of traffic. By considering markers of traffic exposure, we hope to provide a better understanding of the association of lung cancer with pollutant mixtures that are characteristic of urban and highly populated environments. The results for NO x and NO 2 are compatible with the hypothesis that traffic-related air pollution increases the risk of lung cancer. Further, the IARC has classified exposure to diesel exhaust as a Group I carcinogen (Benbrahim-Tallaa et al. 2012). The results of our meta-analyses are consistent with this previous determination, and provide further support for the carcinogenicity of air pollution. Table 2 and an incorrect table title. The unit of increase in NO 2 was corrected from 10 meta 3 to 10-μg/m 3 . The correct version is included in this article.