Air Pollution, Clustering of Particulate Matter Components, and Breast Cancer in the Sister Study: A U.S.-Wide Cohort

Air pollution is classified by the International Agency for Research on Cancer (IARC) as a Group 1 carcinogen (Loomis et al. 2013), consistent with the epidemiologic evidence for the role of air pollution in lung cancer incidence (Hamra et al. 2015). However, less is known about the association between air pollution and breast cancer. Air pollution exposure is widespread and thus has the potential to have a substantial impact on the incidence of breast cancer, which is the most common cancer diagnosed among women in the United States (Siegel et al. 2019).

Air pollution contains many carcinogens and other compounds that may act as endocrine disruptors—including polycyclic aromatic hydrocarbons (PAHs), metals, and benzene—which may influence breast cancer risk. Ecologic studies suggest that breast cancer risk is elevated in urban areas with higher air pollution in comparison with rural areas (Chen and Bina 2012; Wei et al. 2012). Some population studies have reported associations between air pollution and breast cancer, as reviewed by White et al. 2018, especially in studies that consider markers of traffic-related pollution such as nitrogen dioxide (${\mathrm{NO}}_2$), nitrogen oxides (${\mathrm{NO}}_{\mathrm{x}}$), and PAH exposure (Bonner et al. 2005; Hystad et al. 2015; Mordukhovich et al. 2016; Nie et al. 2007; Reding et al. 2015). In the Sister Study cohort, Reding et al. (2015) reported a modest association between residential ${\mathrm{NO}}_2$ levels and risk of estrogen and progesterone receptor-positive (${\text{ER}}^{+}{\text{PR}}^{+}$) breast cancer. However, associations with measures of particulate matter (PM) $<2.5\mathrm{\mu }\mathrm{m}$ and $<10\mathrm{\mu }\mathrm{m}$ in aerodynamic diameter (${\mathrm{PM}}_{2.5}$ and ${\mathrm{PM}}_{10}$, respectively) have not been consistently observed (Andersen et al. 2017a, 2017b; Hart et al. 2016; Reding et al. 2015; Villeneuve et al. 2018).

Fine particulate matter (${\mathrm{PM}}_{2.5}$) is a complex mixture that varies in composition geographically due to varying sources, differences in meteorology, and other factors (Bell et al. 2007). Regional differences in particulate matter have been shown to modify the association with breast density, an important predictor of breast cancer risk (DuPre et al. 2017). Associations between ${\mathrm{PM}}_{2.5}$ and health effects such as blood pressure (Keller et al. 2017), cardiovascular disease (Brook et al. 2010), and mortality (Franklin et al. 2008) have been shown to vary significantly by ${\mathrm{PM}}_{2.5}$ component profiles. In this report, we have extended our prior research on the relationship between air pollutants and breast cancer risk (Reding et al. 2015) with additional years of follow-up and case accrual and expanded this work to include consideration of effect measure modification by ${\mathrm{PM}}_{2.5}$ components and breast cancer risk using predictive k-means clusters (Keller et al. 2017). We hypothesized that air pollution would be related to breast cancer risk and that associations for ${\mathrm{PM}}_{2.5}$ would vary by ${\mathrm{PM}}_{2.5}$ component cluster. Breast cancer is a heterogenous disease (Polyak 2011). Associations with established breast cancer risk factors have been shown to vary by hormone receptor status [often defined by the presence or absence of the estrogen receptor (ER) and progesterone receptor (PR)] (Anderson et al. 2014) as well as by menopausal status at diagnosis (White et al. 2015). In addition, risk factors may vary by whether the tumor is invasive or ductal carcinoma in situ (DCIS) (Barclay et al. 1997). Previous research on the association between air pollution and breast cancer has been inconclusive on whether associations vary by these different outcome classifications; therefore, we also evaluated the risk associated with air pollutant exposure considering these different outcome definitions.

During an average of 8.4 y of follow-up, there were 2,852 incident breast cancer cases (2,225 invasive and 623 DCIS). Study participant baseline characteristics have been previously published (Sandler et al. 2017). Briefly, the median age at enrollment was 55.6 y. Women in the study are predominately non-Hispanic white (83.7%), reported being married or living as married (74.7%), and over half have a bachelor’s degree or higher. The Sister Study includes participants from each of the contiguous states, with representation ranging from 0.2% participants from Wyoming to 8.5% from California. Participant characteristics by geographic region are displayed in Table 1.

An IQR increase in ${\mathrm{NO}}_2$ ($5.8\text{ppb}$) was associated with breast cancer risk overall [$\text{HR}=1.08$ (95% CI: 1.03, 1.13)] (Table 2). We observed substantial heterogeneity when stratifying by invasive disease versus DCIS and therefore show these results separately. This association was stronger for DCIS [$\text{HR}=1.23$ (95% CI: 1.12, 1.35)] than for invasive breast cancer [$\text{HR}=1.02$ (95% CI: 0.96, 1.07)]. Similarly, ${\mathrm{PM}}_{2.5}$ ($\text{IQR}=3.6{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$) was positively associated with DCIS incidence [$\text{HR}=1.16$ (95% CI: 1.02, 1.31)] but not invasive breast cancer [$\text{HR}=1.03$ (95% CI: 0.96, 1.09)]. No elevated HRs were observed in relation to ${\mathrm{PM}}_{10}$ ($\text{IQR}=5.8{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$). Further adjustment for other known and established breast cancer risk factors and other markers of socioeconomic status, including household income, census-tract income, marital status, parity, and BMI, did not materially change the point estimates.

An IQR increase in ${\mathrm{NO}}_2$ was inversely associated with ${\text{ER}}^{–}{\text{PR}}^{–}$ breast cancer [$\text{HR}=0.87$ (95% CI: 0.73, 1.04)] but not with ${\text{ER}}^{+}{\text{PR}}^{+}$ breast cancer [$\text{HR}=1.03$ (95% CI: 0.95, 1.10)] (Table 3). Associations for ${\mathrm{PM}}_{2.5}$ and ${\mathrm{PM}}_{10}$ did not vary by ER/PR status of the tumor. We did not observe notable heterogeneity in the observed associations by menopausal status at diagnosis (see Table S1).

Associations for invasive breast cancer and exposure to ${\mathrm{PM}}_{2.5}$ ($p_{\text{heterogeneity}}=0.04$), ${\mathrm{PM}}_{10}$ ($p_{\text{heterogeneity}}=0.04$), and ${\mathrm{NO}}_2$ ($p_{\text{heterogeneity}}=0.05$) all varied notably by geographic region (Table 4). An IQR increase in ${\mathrm{PM}}_{2.5}$ [$\text{HR}=1.14$ (95% CI: 1.02, 1.27)] was associated with invasive breast cancer in women residing in the West but not other geographic regions [Northeast $\text{HR}=0.89$ (95% CI: 0.73, 1.07); Midwest $\text{HR}=0.93$ (95% CI:0.81, 1.08), South $\text{HR}=1.03$ (95% CI: 0.90, 1.17)]. A similar trend, with a slightly higher HR among women in the Western United States was observed for ${\mathrm{PM}}_{10}$ exposure. An IQR increase in ${\mathrm{NO}}_2$ was similarly associated with breast cancer among women living in the West [$\text{HR}=1.09$ (95% CI: 0.99, 1.21)] as well as for women residing in the South [$\text{HR}=1.16$ (95% CI: 1.01, 1.33)]. For DCIS, in general we observed associations to be more pronounced in women living in the Northeast or the Midwest. For example, for an IQR increase in ${\mathrm{PM}}_{2.5}$, we observed an $\text{HR}=1.35$ (95% CI: 0.97, 1.88) for women in the Northeast and $\text{HR}=1.68$ (95% CI: 1.21, 2.34) for women in the Midwest. The pattern was similar for ${\mathrm{PM}}_{10}$ ($p_{\text{heterogeneity}}=0.01$). For ${\mathrm{NO}}_2$, risk of DCIS also varied by region ($p_{\text{heterogeneity}}=0.01$), with the highest HRs observed in the Midwest [$\text{HR}=1.73$ (95% CI: 1.39, 2.14)]. These associations persisted with further covariate adjustment and when including ${\mathrm{PM}}_{2.5}$ component clusters in the model. These associations were also robust to the inclusion of additional interaction terms with cluster, BMI, and education in the model (see Table S2).

Overall, the associations for ${\mathrm{PM}}_{2.5}$ using 2010 air pollution estimates ($2010\text{IQR}=2.9{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$) were similar to those from our main results using data from 2006 [e.g., 2010 invasive $\text{HR}=1.01$ (95% CI: 0.95, 1.07) vs. 2006 invasive $\text{HR}=1.03$ (95% CI: 0.96, 1.09)] (Table 5). Consistent with the results stratified by geographic region, invasive breast cancer risk also varied by ${\mathrm{PM}}_{2.5}$ component cluster ($p_{\text{heterogeneity}}=0.3$) (Table 5). Specifically, we observed an elevated risk of invasive breast cancer associated with ${\mathrm{PM}}_{2.5}$ exposure for both Cluster 4 (California; Figure 1) and Cluster 7 (West; Figure 1) but no increase in risk for women in any of the other clusters. The California monitors were captured in Cluster 4 (Figure 1), which was characterized by having low S fractions and large fractions of Na and ${\mathrm{NO}}_3^{-}$ (Figure 2), indicating exposure to marine aerosols and agricultural emissions (Keller et al. 2017). For an IQR increase in ${\mathrm{PM}}_{2.5}$ for women who were assigned to Cluster 4, we observed a 25% higher risk of invasive breast cancer [$\text{HR}=1.25$ (95% CI: 0.97, 1.60)]. Cluster 7 was also centered in the Western United States (Figure 1), and was defined by high fractions of Si, Ca, K, and Al (Figure 2), consistent with the surface soil in this geographic region (Shacklette and Boerngen 1984). For women in Cluster 7, we also observed an elevated risk associated with an IQR increase in ${\mathrm{PM}}_{2.5}$ [$\text{HR}=1.60$ (95% CI: 0.90, 2.85)], but the estimate for this cluster was imprecise due to the small number of cases ($n=59$). These associations remained similar with further adjustment for additional covariates and inclusion of geographic region in the adjustment set.

For DCIS, although sample sizes were small, there was less evidence of risk heterogeneity by cluster ($p_{\text{heterogeneity}}=0.9$) (Table 5). Across the clusters, ${\mathrm{PM}}_{2.5}$ was positively associated with DCIS in all but Cluster 7. For example, a higher risk of DCIS in relation to an IQR increase in ${\mathrm{PM}}_{2.5}$ was observed for women in Cluster 1 [$\text{HR}=1.38$ (95% CI: 1.02, 1.86)] and Cluster 2 [$\text{HR}=1.37$ (95% CI: 1.03, 1.83)]. Cluster 1 is in the Midwest and Mid-Atlantic region (Figure 1) with above-average ${\mathrm{NO}}_3^{-}$ and ${\mathrm{SO}}_4^{2-}$ (Figure 2), which is consistent with high ambient ammonia levels from agriculture. Cluster 2 is in the Northeast (Figure 1) and is characterized by higher fractions of Cd, V, and Ni (Figure 2). Elevated HRs, but with wide CIs, were also observed for women in Cluster 3 [$\text{HR}=1.22$ (95% CI: 0.75, 1.96)], Cluster 4 [$\text{HR}=1.33$ (95% CI: 0.80, 2.22)], Cluster 5 [$\text{HR}=1.18$ (95% CI: 0.67, 2.07)], and Cluster 6 [$\text{HR}=1.22$ (95% CI: 0.35, 4.26)] in relation to DCIS.

We observed no significant effect measure modification of the associations between any of the air pollutants and breast cancer risk by time spent living at the baseline residence (see Table S3). However, we did note an elevated HR for invasive breast cancer was observed for ${\mathrm{PM}}_{2.5}$ in women who lived in their residences for $\ge 10\mathrm{y}$ [$\text{HR}=1.07$ (95% CI: 0.98, 1.17)]. We observed modification by obesity; women who had a BMI $\ge 30{\mathrm{kg}/\mathrm{m}}^2$ had a higher risk of invasive breast cancer associated with ${\mathrm{PM}}_{2.5}$ [$\text{HR}=1.19$ (95% CI:1.06, 1.34), $p_{\text{heterogeneity}}=0.02$], and ${\mathrm{NO}}_2$ [$\text{HR}=1.11$ (95% CI: 1.01, 1.21), $p_{\text{heterogeneity}}=0.1$] (see Table S4). We observed no significant effect measure modification of the associations for air pollutants and breast cancer risk by extent of breast cancer family history or hormone therapy use (see Tables S5 and S6). As expected, there was substantial overlap between clusters and geographic region (see Table S7).

In this large, U.S.-wide prospective cohort study, we evaluated the association between air pollutants and breast cancer risk and demonstrated that air pollution levels were related to both invasive breast cancer and DCIS in certain geographic regions. For example, exposure to ${\mathrm{PM}}_{2.5}$ tended to be related to invasive breast cancer risk in the Western United States, whereas for DCIS, the associations were most evident among women in the Northeast and Midwest. These results were consistent with our analysis utilizing predictive k-means clustering to evaluate ${\mathrm{PM}}_{2.5}$ component mixtures in relation to breast cancer risk. ${\mathrm{PM}}_{2.5}$ levels in two Western-based clusters were related to the risk of invasive breast cancer, whereas ${\mathrm{PM}}_{2.5}$ exposure in other clusters were more strongly related to the risk of DCIS. Together, these results suggest that consideration of geographic variability in air pollution is crucial when evaluating associations with breast cancer. This is the first U.S.-based study to evaluate the relationship between PM components and breast cancer risk.

Air pollution is plausibly related to breast cancer given that it is a complex mixture containing numerous carcinogens and endocrine disruptors (Loomis et al. 2013). In breast cancer cell lines, PM has been shown to have estrogenic properties and oxidative stress–related DNA-damaging activity (Chen et al. 2013). Inhaled toxicants can reach the breast tissue (Hill and Wynder 1979) and traffic-related air pollution has been associated with aberrant DNA methylation in breast cancer–related genes measured in tumor tissue (White et al. 2016). Air pollution has also been related to higher breast density (DuPre et al. 2017; White et al. 2019c; Yaghjyan et al. 2017), a marker of breast cancer risk.

Markers of traffic pollution such as ${\mathrm{NO}}_2$, ${\mathrm{NO}}_{\mathrm{x}}$, and PAH exposure have been found to be associated with breast cancer risk (Bonner et al. 2005; Hystad et al. 2015; Mordukhovich et al. 2016; Nie et al. 2007; Reding et al. 2015), whereas results for measures of PM have been mostly null (Andersen et al. 2017a, 2017b; Hart et al. 2016; Reding et al. 2015; Villeneuve et al. 2018). However, these studies have largely not considered the impact of geographic variability or PM heterogeneity. For example, although we too saw little consistent evidence of an association with ${\mathrm{PM}}_{2.5}$ or ${\mathrm{PM}}_{10}$ and invasive breast cancer in our nationwide study population, stratifying by region elucidated significant variability in the associations.

Air pollution is a complex mixture and it is important to address the heterogeneity of this exposure and to evaluate how that may impact breast cancer risk. Only one prior study has evaluated PM components with breast cancer. In a pooled analysis of European cohorts, Andersen et al. (2017b) considered PM components individually in relation to postmenopausal breast cancer risk. They observed a higher breast cancer risk for exposure to both ${\mathrm{PM}}_{2.5}$ and ${\mathrm{PM}}_{10}$ V and ${\mathrm{PM}}_{10}$ Ni levels. Importantly, considering a single PM component at a time does not address the correlated nature of the PM components. To better capture this heterogeneity, we utilized predictive k-means clustering, which is a data reduction technique that identifies subgroups of individuals who are exposed to similar combinations of PM components. This permits the identification of PM component mixtures and consideration of how these complex mixtures influence the association between ${\mathrm{PM}}_{2.5}$ and breast cancer risk.

We observed heterogeneity by geographic region and ${\mathrm{PM}}_{2.5}$ component cluster, individually and after simultaneous adjustment, in the associations between air pollutants and breast cancer risk. Although this geographic variability has not been explicitly considered previously in relation to breast cancer, DuPre et al. (2017) observed geographic variation in that ${\mathrm{PM}}_{2.5}$ in the Nurses’ Health Study was related to breast density only among participants living in the Northeast. In our study, ${\mathrm{PM}}_{2.5}$ was related to DCIS across most of the clusters despite lower power to detect associations. In contrast, ${\mathrm{PM}}_{2.5}$ was associated with invasive breast cancer only in women assigned to two Western-based clusters (Clusters 4 and 7), consistent with our regional results finding a higher risk among women living in the Western United States. Cluster 4, which encompassed the California monitors, was characterized by having low fractions of S and large fractions of Na and ${\mathrm{NO}}_3^{-}$, indicative of marine aerosols and agricultural emissions. Airborne exposure to pesticides from agricultural practices may contribute to cancer risk (Engel et al. 2005; Lee et al. 2002; Lerro et al. 2015). Cluster 7, which was more widely spread across the Western United States, had high fractions of Si, Ca, K, and Al, consistent with the surface soil in this region (Shacklette and Boerngen 1984). In a subset of our study population with DNA methylation data, among women in Clusters 4 and 7, ${\mathrm{PM}}_{2.5}$ was also associated with DNA methylation-based biologic age acceleration (White et al. 2019a), a marker of future breast cancer risk (Kresovich et al. 2019). These consistent findings support a role for these clusters of ${\mathrm{PM}}_{2.5}$ components in breast carcinogenesis.

Differences between overall results for invasive breast cancer and DCIS were unexpected. DCIS is generally thought to be a precursor to invasive breast cancer, and risk factor profiles for DCIS and invasive disease are similar although there are some differences (Reeves et al. 2012). However, it is possible that variation in socioeconomic status by region may have contributed to differences in access to health care that could have influenced the associations observed with DCIS, which is primarily detected by screening (Virnig et al. 2010). To address this, we further adjusted our models for risk of DCIS for individual and census tract–level socioeconomic variables, but we did not observe a change in results. It is unlikely that screening practices explain these results because over 92% of women in our study population were screened within the last 2 y. This high rate of screening may not be too surprising given that our study population consists of women with a family history of breast cancer among whom regular screening is very common. In addition, mammographic screening did not vary by geographic region, so geographic differences in screening behaviors or access cannot explain observed differences in associations by region or cluster. Despite extensive efforts to address potential residual confounding, it remains possible that there is some unaddressed confounding from other factors such as noise or other pollutants that may be driving the differences in DCIS/invasive disease risk by region. Another potential explanation is that these mixtures of pollutants simply contribute differently to breast cancer risk by stage of disease, perhaps by influencing tumor growth rate. Our results of a higher risk of DCIS in relation to air pollutants in the Northeast are consistent with results from a study of women on Long Island, New York, for whom higher vehicular traffic air pollution was similarly associated with DCIS (Mordukhovich et al. 2016).

We did not observe substantial evidence of variability in the associations of overall air pollutant exposures and breast cancer risk by menopausal status or by tumor subtype. However, a limitation of this study was that, despite our large sample size, we were unable to explore effect measure modification by cluster with consideration of tumor subtype.

We observed that invasive breast cancer risk associated with exposure to ${\mathrm{PM}}_{2.5}$ and ${\mathrm{NO}}_2$ was higher among women with a BMI $\ge 30{\mathrm{kg}/\mathrm{m}}^2$, suggesting a possible synergistic relationship between obesity and air pollution. Components of air pollution, such as PAHs, are lipophilic (IARC 2010), whereas other components, such as metals, have been detected in visceral fat (Qin et al. 2010). Thus, fat tissue may serve as a possible reservoir for which the constituents of air pollution may accumulate. This finding is consistent with prior research on PAHs (Niehoff et al. 2017) and airborne metals (White et al. 2019b).

A strength of this study was the use of predictive k-means clustering to determine subgroups of women who were exposed to different ${\mathrm{PM}}_{2.5}$ component mixtures. Consideration of the mixture is important because PM is not a homogenous exposure and our approach permitted a more refined and nuanced exposure assessment. The predictive k-means approach used to identify and assign PM component clusters in the Sister Study was an unsupervised method, meaning that the clusters identified are useful for a public health-focused approach to identify existing air pollution mixtures and determine how they are related to health outcomes. However, given that breast cancer case status was not included in the identification of these clusters, it is possible that there are some groups of pollutants that may be more strongly related to breast cancer risk that were not identified. Although these clusters incorporate 22 different ${\mathrm{PM}}_{2.5}$ components, it is possible that these clusters may be influenced by other correlated unmeasured air pollutants. In addition, the accuracy of the concentration measurements may vary for some of the ${\mathrm{PM}}_{2.5}$ components and thus may result in differential measurement error. Furthermore, we classified individuals into the cluster for which each person had the highest probability of membership, and there is uncertainty in the cluster predictions that could also lead to exposure measurement error. Finally, we cannot rule out the possibility of residual spatial confounding.

The Sister Study is a prospective cohort with extensive covariate information. A strength of this study is the use of land-use regression models with spatial smoothing to assess exposure to air pollution at the level of cohort enrollment residence. However, a limitation of this approach is that we used air pollution measures estimated around the time of enrollment in the study (on average, 8 y prior to breast cancer diagnosis). This measurement may not represent the most relevant time period of exposure with respect to breast cancer etiology. We did, however, consider duration of residence at the current residence. It is noteworthy that most results did not differ for women with $<10$ or $\ge 10\mathrm{y}$ at their enrollment address. It is possible that more long-term exposure, or exposure occurring during hypothesized susceptible windows of exposure including childhood (Bonner et al. 2005; Nie et al. 2007; Shmuel et al. 2017), or exposure during the reproductive time period may be more relevant.

In conclusion, in this large, prospective U.S.-wide cohort, we observed that measures of air pollution, including ${\mathrm{NO}}_2$, ${\mathrm{PM}}_{2.5}$, and ${\mathrm{PM}}_{10}$, were related to both invasive and DCIS breast cancer when stratifying by geographic region. Using predictive k-means clusters to consider the potential modifying role of ${\mathrm{PM}}_{2.5}$ components, we observed that the risk of breast cancer varied based on ${\mathrm{PM}}_{2.5}$ component clusters, which were also correlated with geographic region. This study supports a relationship between air pollution and both invasive breast cancer and DCIS risk within certain geographic subgroups and emphasizes the need to consider variability in air pollution measures by geographic region and composition of the mixture, as well as by tumor staging, when assessing associated risks with breast cancer.

This research was supported by the Intramural Research Program of the National Institute of Environmental Health Sciences, National Institutes of Health (Z01-ES044005 and ES103332-01).

The authors declare they have no actual or potential competing financial interests.