Particulate Matter and Traffic-Related Exposures in Relation to Breast Cancer Survival

Among women, breast cancer is the most commonly diagnosed cancer, excluding non-melanoma skin cancer, and a leading cause of cancer-related death (1, 2). In the United States, several million women at any given time are living with breast cancer (2). Although the global burden of breast cancer–specific mortality has decreased over time in most high-income countries, there is a growing burden of breast cancer–specific mortality in developing countries (3–5). There is also wide international variation in air pollution, with high and increasing levels of particulate matter (PM) less than 2.5 microns in diameter (PM_2.5) observed in parts of Latin America, Asia, and Africa (6). The International Agency for Research on Cancer (IARC) recently declared that outdoor air pollution and PM are human carcinogens, mainly based on evidence with lung cancer incidence (7) and several cohort studies have since reported generally null associations between PM and breast cancer incidence (8–11); however, research on PM after diagnosis in relation to breast cancer survival is limited.

There is convincing evidence that PM induces systemic inflammation, oxidative stress, and epigenetic changes (12–15), which are hypothesized to be underlying mechanisms that may drive breast tumor progression (16–19). To date, one study using California Surveillance Epidemiology and End Results (SEER) registry data reported that county-based PM_2.5 levels at diagnosis were associated with higher breast cancer–specific mortality rates, particularly among patients with breast cancer with localized disease (16). In a smaller study within a northern Italian province, a similar association between PM_2.5 and breast cancer–specific mortality was reported (20); however, these studies did not account for important clinical and lifestyle predictors of breast cancer–specific mortality, particularly post-diagnostic lifestyle factors such as body mass index (BMI), weight gain, physical activity, and aspirin use that are relevant to breast cancer survival (21–23), or have updated PM exposure. These findings have yet to be validated in broader geographic settings and potential confounding factors remain to be considered, particularly related to lifestyle and clinical factors.

Given the nationwide geographic scope of the Nurses' Health Study (NHS) and Nurses' Health Study II (NHSII) and the ability to account for individual-level clinical and lifestyle factors, we investigated associations between PM exposures and distance to major roadways in relation to breast cancer survival among women with Stage I–III disease residing across the conterminous United States. We hypothesized that exposure to PM_2.5, which has a host of biologic effects and a longer atmospheric half-life (12, 24), following a diagnosis would be associated with higher breast cancer–specific death and overall mortality, particularly among women with localized, Stage I disease as observed in the California SEER study.

The NHS cohort was established in 1976, enrolling 121,700 married, female nurses between the ages of 30–55 years who resided in 11 states (California, Connecticut, Florida, Massachusetts, Maryland, Michigan, New Jersey, New York, Ohio, Pennsylvania, and Texas; ref. 25). The NHSII cohort was established in 1989, enrolling 116,430 female nurses between the ages of 25–42 years residing in 14 states (California, Connecticut, Indiana, Iowa, Kentucky, Massachusetts, Michigan, Missouri, New York, North Carolina, Ohio, Pennsylvania, South Carolina and Texas), without a prior history of cancer, excluding non-melanoma skin cancer. Briefly, participants in both cohorts completed baseline and subsequent biennial questionnaires that were mailed to their residential addresses to collect information on medical history, anthropometrics, reproductive history, lifestyle factors, and medications (26).

Participants with self-reported breast cancer gave consent for study staff to review medical records to confirm the diagnosis. In this study, we included participants without a previous report of cancer and with confirmed Stage I–III primary breast cancer diagnosed between 1988 and 2008 who had PM exposure data (NHS n = 6,499; NHSII n = 2,437); the large majority of participants in these cohorts have information on stage, 92% in NHS and 87% in NHSII. Return of the questionnaires implied informed consent, and all participants or next-of-kin provided written approval to obtain medical records. The study protocol was approved by the institutional review boards of the Brigham and Women's Hospital, Harvard T.H. Chan School of Public Health and those of participating registries as required.

The primary outcome of interest was death due to breast cancer (n = 1,211: NHS n = 899; NHSII n = 312). Deaths were reported by next-of-kin, the post office, or by searches of the National Death Index for nonresponders; it has been estimated that ascertainment of vital status using the National Death Index in this cohort was 98% (27, 28). The date and cause of death were confirmed by study physicians using information from medical records, the National Death Index, tumor registries and death certificates. As secondary analyses, we examined associations with all-cause mortality (n = 2,614 deaths from to all-causes: NHS n = 2,241; NHSII n = 373) and breast cancer recurrence. A participant was considered to have distant recurrence if she reported (i) a distant metastasis to the lung, liver, brain or bone on a supplemental questionnaire sent to women with breast cancer, or (ii) another cancer in the lung, liver, or brain on the main questionnaires that was confirmed by medical record review to be a metastasis and not a primary cancer. For women who died from breast cancer without reporting recurrence, we included them as having recurred two years before death for those with at least 2 years of follow-up or as having recurred at the time of death for those with less than two years of follow-up.

As part of the questionnaire mailing process, residential addresses were updated biennially in both cohorts. By the mid-1990s, participants resided in all 50 states (29). Study participants' residential addresses were geocoded and linked to predicted estimates of PM and to proximity to various types of roadways over the course of the study.

PM is classified into the following three size fractions: particles less than 2.5 μm in aerodynamic diameter (PM_2.5), particles less than 10 μm (PM₁₀), and particles between 2.5 and 10 μm (PM_2.5-10; ref. 24). PM comes from various sources and the varying contribution of these sources is different across regions of the United States. We defined the regions based on the Census Bureau designated regions: Northeast, Midwest, South, and the West (30).

Predictions of monthly average PM_2.5 and PM₁₀ were generated using spatio-temporal models that accounted for spatial and meteorological variation over time (31). The monthly estimates were linked with participants' residential addresses between 1988 through 2007 (31). As PM_2.5 data were not directly measured before 1999, we derived PM_2.5 levels before 1999 from the PM₁₀ levels before 1999 using the 1999 ratio of PM_2.5:PM₁₀ (31). PM_2.5-10 was calculated as the difference between PM₁₀ and PM_2.5 estimates. We evaluated the models for predictive accuracy using a 10-set cross-validation approach; cross-validation correlation coefficients were high for PM_2.5 (R² = 0.77) and moderate for PM₁₀ (R² = 0.58) and PM_2.5-10 (R² = 0.46; ref. 31). Because PM predictions were not available after 2007, we carried forward the 2007 average PM levels for subsequent years of follow-up. In the current study, the primary exposure of interest was 2-year average PM that was updated every questionnaire cycle from the report of diagnosis until the year of an endpoint or the end of follow-up.

A secondary exposure of interest was proximity of each residential address to various types of major roadways as a proxy for traffic-related exposures. Distance (in meters) from each residential address to nearest roadways was calculated using GIS software and the ESRI StreetMap Pro 2007 road network. Distances were calculated from three different types of roadways U.S. Census Feature Class Code roadways: A1 (primary roads, typically interstate highways, with limited access, division between opposing directions of traffic, and defined exits); A2 (primary major, non-interstate highways and major roads without access restrictions); and A3 (smaller, secondary roads, usually with more than two lanes).

We considered the following demographic and clinical predictors of breast cancer–specific mortality in multivariable models stratified on time since diagnosis: age at diagnosis, calendar year of diagnosis, disease stage (I, II, III), grade (1, 2, 3, missing), treatment (radiation only, chemotherapy only, both radiation and chemotherapy, neither radiation or chemotherapy—of whom, 99% had surgery—or missing), hormonal treatment (yes, no, missing), estrogen receptor (ER) status (ER positive, ER negative, missing), race/ethnicity (White, Black, Hispanic, Other), region of residence at diagnosis (Northeast, Midwest, South, West), and Census tract-level median income based on values from the 2000 Census. We additionally considered individual level markers of socioeconomic status gathered from the main questionnaires in NHS (work status, marital status, and husband's education) and NHSII (living arrangements, marital status and personal income). Stage and grade data were abstracted from review of medical records. Tumors ≤2.0 cm without lymph node involvement were classified as Stage I; Stage II tumors were ≤2.0 cm with 1–3 lymph nodes involved, 2.1–4.0 cm with <3 lymph nodes involved, or >4.0 cm without lymph node involvement; tumors of any size with ≥4 lymph nodes involved or tumors >4.0 cm with 1–3 lymph nodes involved were classified as Stage III tumors. Information on breast cancer treatment was self-reported at the time of giving permission to obtain medical records. Information on ER status was obtained from pathology reports and tissue microarrays (TMA; refs. 32, 33).

We considered the following pre- and post-diagnosis lifestyle factors: pre-diagnosis body mass index (BMI), weight change between the cycle of diagnosis compared to the cycle after the diagnosis (stable weight change within 5%, moderate weight gain of 5-<10%, large weight gain ≥10%, moderate weight loss 5-<10%, large weight loss of ≥10%), pre- and post-diagnosis smoking status (never, past, current, missing), pre- and post-diagnosis physical activity categories (<3, 3 to <9, 9 to <15, 15 to <24, ≥24 MET-hours/week, missing), and post-diagnosis aspirin use (nonuser, current user, and missing). The post-diagnosis variables were updated every 2 years or as they became available.

We conducted Cox proportional hazards models for time-varying covariates (34) to estimate Hazard Ratios (HR) and 95% confidence intervals (CI) for each 10 μg/m³ increase in updated PM exposures and by categories of proximity to A1, A1-A2, and A1-A3 roadways (≥250, 150–249, 50–149, and <50 m). Person-time started at the questionnaire date when the participant reported her diagnosis and ended at the date of the endpoint (breast cancer–specific death, death, or distant recurrence), or end of follow-up (June 2014). Violation of the proportional hazards assumption was tested using the likelihood ratio test (LRT). We considered adjustment for the aforementioned covariates and allowed the baseline hazard to vary by time since diagnosis to finely control for confounding. Four models are presented showing adjustment for age and diagnosis date in Model 1, additional adjustment for demographic factors (i.e., race/ethnicity, region, Census-tract median income) and pre-diagnosis lifestyle factors (i.e., physical activity and BMI) in Model 2, additional adjustment for clinical information in Model 3 (i.e., stage, ER status, treatment, hormone treatment, and grade), and adjustment for post-diagnosis factors in Model 4 (weight change, physical activity, and aspirin use) given previous findings with their associations with breast cancer survival (21–23, 35). Estimates from the cohorts were meta-analyzed to present pooled estimates. Statistically significant heterogeneity between the estimates across cohorts was assessed using the Q statistic. We assessed whether associations differed by stage, ER status, grade, region of residence, aspirin use, and BMI categories using likelihood ratio tests. In sensitivity analyses, we restricted to women who did not move to another state after diagnosis.

All analyses were conducted in SAS version 9.4.

Among the 8,936 participants with Stage I–III breast cancer diagnosed between 1988 and 2008 (NHS n = 6,499; NHSII n = 2,437), there were 1,211 confirmed breast cancer–specific deaths over follow-up (NHS n = 899; NHSII n = 312) and 2,614 deaths from to all-causes (NHS n = 2,241; NHSII n = 373). The median date of diagnosis for NHS was 1998 and for NHSII was 2001. The proportion of cases alive after five-years since diagnosis in NHS was 89% and in NHSII was 93%. The median follow-up time in NHS was 13.25 years [standard deviation (SD) = 6.4] and in NHSII was 12.0 years (SD = 5.2). Among those who died from breast cancer, the average time from diagnosis to breast cancer–specific death was 6.6 years in NHS (IQR 3.0–9.4) and 5.9 years in NHSII (IQR 2.6–8.3). In NHS, the annual average levels of PM_2.5, PM_2.5-10, and PM₁₀ at diagnosis were 13.3 μg/m³ (SD = 3.5), 8.9 μg/m³ (SD = 4.8), and 22.2 μg/m³ (SD = 6.9), respectively. In NHSII, the annual average levels of PM_2.5, PM_2.5-10, and PM₁₀ at diagnosis were 12.9 μg/m³ (SD = 3.1), 8.4 μg/m³ (SD = 4.7), and 21.3 μg/m³ (SD = 6.2), respectively. Age-standardized characteristics are presented in Table 1. Overall, no consistent patterns were observed across cohorts between breast cancer mortality predictors and calendar-year adjusted PM_2.5 quintiles at diagnosis.

After adjustment for demographic, clinical, and pre- and post-diagnostic lifestyle factors, none of the PM exposures were statistically significantly associated with breast cancer–specific death overall (Table 2). The pooled HR of breast cancer–specific death for a 10 μg/m³ increase in PM_2.5 was 1.09 (95% CI, 0.87–1.36), PM_2.5-10 was 1.03 (95% CI, 0.85–1.24) and PM₁₀ was 1.05 (95% CI, 0.89–1.24). The proportional hazards assumption was not violated in the models with LRT P > 0.05. The change in estimates from Model 1 to Model 2 in NHSII was driven by pre-diagnosis physical activity and BMI. The strongest contributors to the change in estimates from Model 2 to Model 3 for NHS were stage, ER status and hormonal therapy. No significant heterogeneity was found between cohorts, and adjusting for individual-level markers of socioeconomic status did not materially change the estimates or conclusions. Proximity to A1-A3 roadways was not associated with breast cancer mortality (see Supplementary Table S1). However, the associations between PM_2.5 and breast cancer survival differed by disease stage at diagnosis (Table 3). Among women with Stage I breast cancer, the HR of breast cancer–specific death for a 10 μg/m³ increase in PM_2.5 was 1.69 (95% CI, 1.16–2.47) in the pooled multivariable model that did not adjust for post-diagnostic lifestyle factors. The associations were null among women with Stage II or III breast cancer (Table 3). After multivariable adjustment that accounted for post-diagnostic lifestyle factors, the pooled HR among Stage I participants was similar (HR, 1.64; 95% CI, 1.11–2.43) and remained null among women with Stage II and III disease (Table 3). Similar patterns were observed for PM exposures and breast cancer recurrence (see Supplementary Table S2) and when we restricted to women who did not move (see Supplementary Table S3). The associations between PM exposures and breast cancer mortality did not differ by grade, ER status, region of residence at diagnosis, aspirin use, or BMI (see Supplementary Table S4).

Among women diagnosed with breast cancer in the NHS who died over follow-up (n = 2,241), the five most common causes of death were breast cancer (40.1% of deaths), dementia (5.7%), unknown causes (4.8%), lung cancer (4.4%), and myocardial infarction (3.3%). The remaining deaths were due to other cancers (9.6%), cardiovascular diseases (6.3%), respiratory conditions (4.4%), Parkinson's disease (0.7%), and other causes (20.8%). Among the NHSII breast cancer participants who died over follow-up (n = 373), the five most common causes of death were breast cancer (83.9%), lung cancer (1.6%), unknown causes (1.1%), ovarian cancer (0.8%), and lymphomas (0.8%). The remaining deaths were due to other cancers (2.4%), cardiovascular diseases (1.9%), respiratory conditions (0.5%), and other causes (7.0%). There was no statistically significant heterogeneity across cohorts for PM exposures and all-cause mortality. The pooled HRs of all-cause mortality for a 10 μg/m³ increase in PM_2.5 was 1.12 (95% CI, 0.96–1.30), PM_2.5-10 was 1.12 (95% CI, 1.00–1.24) and PM₁₀ was 1.09 (95% CI, 1.01–1.18; Table 4).

We observed moderate positive associations between PM_2.5 and breast cancer–specific mortality only among women with Stage I breast cancer, but not among those with Stage II or III disease. The hazard of breast cancer–specific death for women with Stage I breast cancer was 1.64 times higher (95% CI, 1.11–2.43) per 10 μg/m³ increase in PM_2.5. The associations were specific to PM_2.5 and were not found for larger PM size fractions. Similar to a previous study of all-cause mortality in the NHS (36), we observed modest associations between PM exposures and all-cause mortality among women diagnosed with breast cancer.

Our observed HRs of breast cancer–specific mortality were attenuated compared to those reported in the California SEER study (16) and the Varese, Italy cancer-registry study (20), and we did not observe associations among Stage II or Stage III patients. The California SEER study reported an 86% higher rate of overall breast cancer death (95% CI, 1.12–3.10) for each 5 μg/m³ increase in PM_2.5 and a weaker association for a 10 μg/m³ increase in PM₁₀ (HR, 1.13; 95% CI, 1.02–1.25; 16), although we did not find an association with PM exposures and overall breast cancer–specific mortality (PM_2.5 HR, 1.04; 95% CI, 0.84–1.28; PM₁₀ HR, 1.02; 95% CI, 0.89–1.18). The California SEER study observed positive associations for a 5 μg/m³ increase in PM_2.5 among breast cancer patients with localized disease (HR, 2.13; 95% CI, 1.15–3.95), regional disease (HR, 2.07; 95% CI, 1.11–3.84) and distant stage disease (HR, 1.62; 95% CI, 1.05–2.51); however, we only observed positive associations for PM_2.5 among Stage I patients with breast cancer (HR, 1.64; 95% CI, 1.11–2.43) and not among Stage II (HR, 0.88; 95% CI, 0.60–1.30) or Stage III (HR, 1.08; 95% CI, 0.69–1.70) participants. Of note, the California SEER registry data could not consider important clinical (e.g., treatment, grade, and ER status) or pre- or post-diagnosis lifestyle factors (e.g., body mass index, physical activity, and aspirin use) as we have done in the current study that may contribute to the observed differences across studies. Similarly, in a population-based cancer registry cohort in Varese, Italy, an elevated HR of breast cancer–specific death was also observed (HR, 1.72; 95% CI, 1.08–2.75) comparing women residing in the highest quartile of PM_2.5 (≥26.5 μg/m³) to the lowest quartile of PM_2.5 (<21.1 μg/m³) after adjusting for age, stage, grade, diagnosis date, and screening participation (20).

The associations between PM and breast cancer mortality were specific to PM_2.5. PM_2.5 has an atmospheric half-life ranging from days to weeks (24) and comes primarily from combustion sources, organic compounds, and metals, which can penetrate the small airways and alveoli deep in the lung (15). In the United States, approximately 80% of PM_2.5 composition consists of sulfates, nitrates, ammonium, elemental carbon, organic carbon, Na⁺, and silicon and the remaining approximately 20% is a catch-all category consisting largely of many minerals and metals (e.g., lead, cadmium, vanadium, nickel, copper, zinc, manganese, and iron; refs. 24, 37). On the other hand, PM_2.5-10 does not penetrate as deep into the lung as PM_2.5, has a shorter atmospheric half-life ranging from minutes to days, and comes from the breaking of large crustal material as well as ocean spray and organic materials (24). The biological effects of exposure to PM_2.5 go beyond the lungs, inducing systemic inflammation, oxidative stress, and epigenetic changes, though no studies have examined the influence of PM_2.5 on breast tissue (12–15, 38–43).

Inflammation may be a key driver of the actions of PM_2.5 on breast tumor progression. Epidemiologic studies of aspirin and other NSAID use after a breast cancer diagnosis reported improved breast cancer survival, suggesting the importance of inflammatory processes after diagnosis (44, 45). Similarly, protein expression of cyclooxygenase-2 (COX-2), a marker of inflammation and target for aspirin, in breast tissue specimens was associated with worse breast cancer prognosis (46). In human lung cells, benzo[a]pyrene, a major component of air pollution, induced COX-2 expression at the RNA and protein levels (47). Although we did not observe differences in the associations of PM and breast cancer–specific mortality among aspirin users and among non-users, the role of air pollution and inflammation in the breast remains to be elucidated and other biological mechanisms may be at play. Early-life exposures to ambient total suspended particles—particles of larger size—and traffic emission estimates were associated with differential epigenetic methylation patterns for a few genes within breast tissue in the Western New York Exposures and Breast Cancer study and the Long Island Breast Cancer Study Project case–control study (48, 49). To date, no studies have explored whether ambient PM_2.5 influences molecular changes in breast tissue.

This study has several limitations. Exposure measurement error is often a challenge. Instead of collecting personal exposure data, which is not feasible for a decades-long large epidemiologic scale, we used predictions from spatio-temporal modeling. The PM data are subject to Berkson error as the PM monitors measure only part of the true exposure that results in imprecision, as well as classical error, which usually results in attenuated estimates (50, 51). Despite the limitations inherent in the exposure assessment, the PM predictions were more strongly correlated with personal PM exposure than PM values from a nearest monitor (52) and have been shown to be associated with various outcomes such as mortality (36), lung cancer (29), and coronary heart disease (53, 54). Although missing data are often a limitation in epidemiologic studies, the missing data of the covariates in this study were not related to ambient PM exposure. Residual confounding cannot be fully ruled out, though we adjusted in stages to demonstrate the influence of individual-level demographic factors, pre-diagnosis lifestyle factors, clinical factors, and post-diagnosis lifestyle factors as well as markers of socioeconomic status. Additionally, the participants of this study were predominately white and results of this study may not be generalizable to other populations if there is reason to believe that the associations would be different by race/ethnicity. Although the overall 5-year survival proportion in NHS and NHSII were similar to overall 5-year survival in SEER from corresponding years (55).

There are also several strengths of this study. We used a high-resolution spatio-temporal model of PM estimates across a large geographic area of the contiguous United States that were updated from 1988 to 2007. The PM estimates that have been associated with mortality, respiratory, and cardiovascular diseases in these cohorts (29, 36, 53, 54, 56–58). The NHS and NHSII participants reside across the United States, making this the most geographically expansive study of PM and breast cancer survival to date, as far as the authors are aware. The large size of the study allowed for stratified analyses to explore the associations separately by stage and other potential effect modifiers. Furthermore, the wealth of longitudinal data on lifestyle, clinical factors, and demographics allowed for finer control of potential confounding that was not feasible in previous registry-based studies.

In conclusion, PM_2.5 was associated with worse breast cancer prognosis only among women with locally spread breast cancer, but not advanced tumors—even after consideration of important clinical and pre- and post-diagnostic lifestyle factors. This study confirms in a large geographic context that even after adjustment for lifestyle and clinical factors that PM_2.5 was associated with worse breast cancer–specific mortality among women with localized disease, demonstrating that air pollution may have broader health effects than previously recognized.

No potential conflicts of interest were disclosed.

Conception and design: N.C. DuPré, J.E. Hart, F. Laden, R.M. Tamimi

Development of methodology: P. Kraft, F. Laden, R.M. Tamimi

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J.E. Hart, M.D. Holmes, F. Laden, R.M. Tamimi

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): N.C. DuPré, J.E. Hart, M.D. Holmes, E.M. Poole, P. Kraft, F. Laden

Writing, review, and/or revision of the manuscript: N.C. DuPré, J.E. Hart, M.D. Holmes, E.M. Poole, P. James, P. Kraft, F. Laden, R.M. Tamimi

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): R.M. Tamimi

Study supervision: F. Laden, R.M. Tamimi

We would like to thank the participants and staff of the Nurses' Health Study and Nurses' Health Study II for their valuable contributions as well as the following state cancer registries for their help: AL, AZ, AR, CA, CO, CT, DE, FL, GA, ID, IL, IN, IA, KY, LA, ME, MD, MA, MI, NE, NH, NJ, NY, NC, ND, OH, OK, OR, PA, RI, SC, TN, TX, VA, WA, and WY. This research was supported by the UM1 CA186107, UM1 CA176726, P01 CA87969, K99 CA201542, National Institute of Health Epidemiology Education Training Grant (NIH T32 CA09001; to N.C. DuPré), and the Susan G. Komen for the Cure (IIR13264020).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.