Ambient fine particulate matter and breast cancer incidence in a large prospective US cohort

Abstract

Background

Fine particulate matter (PM_2.5) has been inconsistently associated with breast cancer incidence, however, few studies have considered historic exposure when levels were higher.

Methods

Outdoor residential PM_2.5 concentrations were estimated using a nationwide spatiotemporal model for women in the National Institutes of Health–AARP Diet and Health Study, a prospective cohort located in 6 states (California, Florida, Louisiana, New Jersey, North Carolina, and Pennsylvania) and 2 metropolitan areas (Atlanta, GA, and Detroit, MI) and enrolled in 1995-1996 (n = 196 905). Annual average PM_2.5 concentrations were estimated for a 5-year historical period 10 years prior to enrollment (1980-1984). We used Cox regression to estimate adjusted hazard ratios (HRs) and 95% confidence intervals (CIs) for the association between a 10 µg/m³ increase in PM_2.5 and breast cancer incidence overall and by estrogen receptor status and catchment area.

Results

With follow-up of participants through 2017, a total of 15 870 breast cancer cases were identified. A 10 ug/m³ increase in PM_2.5 was statistically significantly associated with overall breast cancer incidence (HR = 1.08, 95% CI = 1.02 to 1.13). The association was evident for estrogen receptor–positive (HR = 1.10, 95% CI = 1.04 to 1.17) but not estrogen receptor–negative tumors (HR = 0.97, 95% CI = 0.84 to 1.13; P_{heterogeneity} = .3). Overall breast cancer hazard ratios were more than 1 across the catchment areas, ranging from a hazard ratio of 1.26 (95% CI = 0.96 to 1.64) for North Carolina to a hazard ratio of 1.04 (95% CI = 0.68 to 1.57) for Louisiana (P_{heterogeneity} = .9).

Conclusions

In this large US cohort with historical air pollutant exposure estimates, PM_2.5 was associated with risk of estrogen receptor–positive breast cancer. State-specific estimates were imprecise but suggest that future work should consider region-specific associations and the potential contribution of PM_2.5 chemical constituency in modifying the observed association.

Breast cancer is the most commonly diagnosed malignancy among women worldwide (1). It is a heterogeneous disease, with established risk factors (2) and survival rates after diagnosis dependent on the hormonal profile of the tumor (3,4). Established risk factors include a woman’s reproductive history, use of exogenous hormones, alcohol intake, and obesity (5). Many of these factors are consistent with the importance of hormones in the etiology of the disease (6) and suggest that environmental chemicals with endocrine-disrupting properties have the potential to influence breast cancer risk (7).

Exposure to fine particulate matter (airborne particles <2.5 µm in aerodynamic diameter; PM_2.5) is widespread and arises from numerous sources, such as motor vehicle exhaust, combustion processes (eg, oil, coal), wood smoke and vegetation burning, and industrial emissions (8). PM_2.5 is classified as a human carcinogen by the International Agency for Research on Cancer based on evidence for lung cancer (9). PM_2.5 is a complex and heterogenous mixture of airborne pollutants, including metals (eg, sodium, nickel), metalloids (eg, silicon), organic compounds (eg, polycyclic aromatic hydrocarbons), ammonium, nitrate, ozone, and sulfate, among others (10). Many of these constituents have endocrine-disrupting properties and thus may be relevant to breast cancer etiology (11,12). Critically, the chemical constituency of PM_2.5 can differ geographically because of varying sources and meteorologic factors that contribute to ambient levels (10).

There has been increasing epidemiologic investigation into the role of air pollution in breast cancer development, with recent meta-analyses supporting a positive association for nitrogen dioxide (NO₂) (13,14). In contrast, the evidence for an association with PM_2.5 is inconsistent. Three recent meta-analyses concluded that there was no overall association (13-15), although a pooled analysis of 6 European cohorts observed a 6% higher breast cancer incidence associated with a 5 µg/m³ increase in PM_2.5 (16).

Most prior studies have assessed breast cancer risk in relation to PM_2.5 exposure at or around the time of study enrollment with few considering more historic exposures. This may be particularly important, as past levels of exposure were higher and may plausibly contribute to breast cancer etiology given the long latency of the disease. Further, evaluation of PM_2.5 associations by tumor histology remains limited despite the well-documented etiologic differences in breast cancer by tumor subtype (2). Finally, few studies have considered geographic variability, which may reflect exposure to differing PM_2.5 component mixtures, in modifying the association with breast cancer risk.

Our objective was to evaluate the association between historic concentrations of PM_2.5 and incident breast cancer, overall and by estrogen receptor status, in a large, geographically spread US cohort.

Methods

Study population and cancer ascertainment

Between 1995 and 1996, approximately 567 169 members of the AARP organization, formally named American Association of Retired Persons, living in 6 states (California, Florida, Louisiana, New Jersey, North Carolina, Pennsylvania) and 2 metropolitan areas (Atlanta, GA; Detroit, MI) enrolled in the National Institutes of Health (NIH)–AARP cohort and returned mailed questionnaires to provide information on demographic and anthropometric characteristics, educational attainment, lifestyle factors, and other cancer risk factors (17). Participants consented to the study by returning the questionnaire, and the study was approved by the NIH institutional review board. Participants completed a follow-up questionnaire in 2004-2005, at which time changes in address were recorded, and addresses were also regularly updated (ie, with the US Postal Service National Change of Address database) to send participant newsletters by mail. In 1996 and 1997, a total of 127 167 participants completed a secondary risk factor questionnaire, which included a question regarding history of mammography.

Incident breast cancers were identified through linkage with cancer registries in each of the 8 states, as well as to the 3 states they most commonly relocated: Arizona, Nevada, and Texas (18). Participants were followed from the time of enrollment in the study until the date of first primary breast cancer diagnosis, relocation to a state not included in the registry linkage, or the end of follow-up (December 31, 2017). Tumor characteristics, including estrogen and progesterone receptor status (n = 10 909; 69%) and tumor extent (ie, ductal carcinoma in situ [DCIS] vs invasive; n = 15 464; 97%) were obtained from the cancer registries.

For this analysis, we excluded proxy respondents (n = 15 760); male participants (n = 325 165); those with self-reported breast, colon, prostate, or other cancer on the baseline questionnaire (n = 23 925); or a death certificate diagnosis of cancer (eg, participant not found in a cancer registry; n = 4360). An additional 89 participants were excluded for missing air pollution data and 185 for undetermined follow-up time. The final analytic sample included 196 905 female participants without a prior history of cancer.

Exposure assessment

The validated spatiotemporal PM_2.5 prediction model has been described previously (19). Briefly, data from the US Environmental Protection Agency’s Federal Reference Method network (1999-2010) and the Interagency Monitoring of Protected Visual Environment network (1999-2010) provided measurements of PM_2.5 that were used for model development. Temporal trend estimation for 1980-2010 was determined using available Federal Reference Method and Interagency Monitoring of Protected Visual Environment data and fitted with linear predictors where not available, Clean Air Status and Trends Network annual average PM_2.5 sulfate concentrations (1987-2010), and Weather Bureau Army Navy network visual ranges (1980-2010). Spatially varying long-term mean and trend coefficients were estimated using universal kriging, which incorporated approximately 300 geographic predictors and spatial smoothing. Annual average PM_2.5 levels were estimated at participant residences for each year from 1980 to 2010.

Statistical analysis

We calculated distributions and Spearman correlations (ρ) for PM_2.5 concentrations across three 5-year historical exposure time periods prior to enrollment: 1980-1984, 1985-1989, 1990-1994. We calculated frequencies and proportions for categorical variables and means and standard deviations for continuous variables, overall and by quartiles of PM_2.5 for 1980-1984.

We used Cox proportional hazards regression models to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for the association between levels of PM_2.5 and incident breast cancer, overall and by estrogen receptor status (estrogen receptor–positive vs estrogen receptor–negative) and tumor extent (DCIS vs invasive). We primarily focused on exposures during the period 10-15 years before study start (1980-1984) to allow the longest latent period possible, evaluating associations with PM_2.5 quartiles and as a continuous variable (per 10 µg/m³ increase). We also estimated relationships for PM_2.5 levels in 2 other pre-enrollment time periods: 1985-1989 and 1990-1994. For categorical analyses using quartiles, we evaluated linear trends using the median of each quartile parameterized as a continuous variable. We explored nonlinearity with a restricted cubic spline analysis and tested for linearity using a log-likelihood test. Covariates were selected a priori, and our primary model was adjusted for age (5-year age groups), race and ethnicity (non-Hispanic White; non-Hispanic Black; Hispanic; Asian, Pacific Islander, American Indian, Native Alaskan), smoking (never; former <1 pack/day, former 1-2 packs/day, former 2 or more packs/day; current or quit in past year <1 pack/day; current or quit in past year 1-2 packs/day; current or quit in past year 2 or more packs/day), body mass index (BMI; kg/m²; continuous), catchment area, and educational attainment (high school or less, some college, college or postgraduate degree). Race and ethnicity were included as a confounder, as previous studies have demonstrated that minoritized populations tend to live in areas of higher exposure to PM_2.5 (20-22) and that individuals living in more segregated areas have higher exposure to carcinogenic air toxic pollutants (23).

We conducted stratified analyses to evaluate heterogeneity by several different factors: age (50 to younger than 55 years, 55 to younger than 60 years, 60 to younger than 65 years, 65 years and older), BMI (<18.5 kg/m², 18.5 to <25.0 kg/m², 25.0 to <30.0 kg/m², ≥30.0 kg/m²), family history of breast cancer (yes, no), study catchment area, and US census region (Northeast, South, Midwest, West). Age-, BMI-, and family history–stratified analyses were conducted for categorical and continuous PM_2.5. Because of smaller numbers of participants within at least 1 subgroup, analyses stratified by catchment area and census region were considered exploratory and included only continuous exposures to maximize study power. P_{heterogeneity} were estimated using a Wald χ² test comparing models with and without a cross-product term between continuous PM_2.5 and the modifier; heterogeneity by disease characteristics (estrogen receptor and progesterone receptor status and DCIS vs invasive) was evaluated via joint Cox models.

Sensitivity analyses were conducted to explore the robustness of our findings. We evaluated joint estrogen and progesterone receptor status, with women classified as having either 1) estrogen receptor–positive or progesterone receptor–positive tumors or 2) estrogen receptor–negative and progesterone receptor–negative tumors. Most women were postmenopausal at baseline, therefore we restricted to postmenopausal women and re-estimated the main effects (n = 183 261). Other analyses restricted to women 1) with at least 10 years of follow-up or 2) who had received a mammogram in the 3 years prior to baseline (n = 110 231). Three additional adjustment sets were included: 1) removing race and ethnicity from the main adjustment set and 2 models with the inclusion of the following additional covariates in addition to the main model; 2) NO₂ (parts per billion [ppb]), estimated as the annual average at the baseline residence in 1999 (24); and 3) reproductive factors, including age at first live birth (nulliparous, younger than 16, 16-19, 20-24, 25-29, 30-34, 35-39, 40 years and older), parity (0, 1, 2, 3-4, ≥5), oral contraceptive use (ever, never), menopausal hormone use (ever, never), age at menarche (10 years or younger, 11-12, 13-14, 15 years or older), and first-degree relative with a history of breast cancer (yes, no, unknown). The model additionally adjusting for reproductive factors was also used for a sensitivity analysis by estrogen receptor status. Finally, we restricted the analysis to participants with geocodes at the point or street address (n = 179 450).

The threshold for statistical significance in 2-sided tests was an alpha of  .05.

Results

During a median of 20.7 (mean = 16.2) years of follow-up, 15 870 incident cases of breast cancer, of which 14 621 (92%) were among postmenopausal women, were identified. At baseline, the mean age of the study participants was 61.8 years, and most (89%) were non-Hispanic White (Table 1). Approximately one-third (30%) had a college or postgraduate degree, and participants resided primarily in California (32%) or Florida (21%). Mean PM_2.5 concentrations were lower for Florida residents and higher for women living in California or Pennsylvania. Non-Hispanic Black participants comprised only 6% of the cohort overall but made up 12% of women in the highest exposure quartile. Concentrations of NO₂ increased with PM_2.5 exposure (ρ  = 0.67). Mean PM_2.5 concentrations decreased approximately 17% over the 3 time periods (1980-1984 mean = 18.70 µg/m³ vs 1990-1994 mean = 15.60 µg/m³), although levels were highly correlated across the periods (ρ > 0.99; Supplementary Table 1, available online).

Table 1.

Baseline characteristics of female participants (n = 199 750) of the National Institutes of Health–AARP Diet and Health Study, overall and by quartile of PM_2.5 (1980-1984)

		PM2.5 (1980-1984)a
Characteristics	Overall	Quartile 1, No. (%) (n = 49 224)	Quartile 2, No. (%) (n = 49 228)	Quartile 3, No. (%) (n = 49 227)	Quartile 4, No. (%) (n = 49 226)
Age, y
50 to younger than 55	27 991 (14)	6416 (13)	7245 (15)	7340 (15)	6990 (14)
55-59	45 444 (23)	10 724 (22)	11 616 (24)	11 562 (23)	11 542 (23)
60-64	55 502 (28)	13 842 (28)	13 813 (28)	13 917 (28)	13 930 (28)
65-69	60 694 (31)	16 294 (33)	14 754 (30)	14 682 (30)	14 964 (30)
70 and older	7274 (4)	1948 (4)	1800 (4)	1726 (4)	1800 (4)
Race and ethnicity
Asian, Pacific Islander, American Indian, Native Alaskan	3168 (2)	469 (1)	787 (2)	776 (2)	1136 (2)
Hispanic	3794 (2)	966 (2)	916 (2)	620 (1)	1292 (3)
Non-Hispanic Black	11 318 (6)	1030 (2)	1805 (4)	2767 (6)	5716 (12)
Non-Hispanic White	175 444 (89)	46 055 (94)	45 021 (91)	44 358 (90)	40 010 (81)
Unknown	3181 (2)	704 (1)	699 (1)	706 (1)	1072 (2)
Educational attainment
High school or less	62 280 (32)	15 395 (31)	15 443 (31)	15 540 (32)	15 902 (32)
Less than college	69 464 (35)	18 704 (38)	17 292 (35)	16 490 (34)	16 978 (34)
College or postgraduate	58 469 (30)	13 384 (27)	14 854 (30)	15 662 (32)	14 569 (30)
Unknown	6692 (3)	1741 (4)	1639 (3)	1535 (3)	1777 (4)
Smoking status
Never smokers	86 878 (44)	20 280 (41)	22 313 (45)	22 437 (46)	21 848 (44)
Former smokers, <1 pack/d	49 035 (25)	12 593 (26)	12 246 (25)	12 266 (25)	11 930 (24)
Former smokers, 1-2 packs/d	17 903 (9)	4962 (10)	4471 (9)	4221 (9)	4249 (9)
Former smokers, ≥2 packs/d	4528 (2)	1296 (3)	1045 (2)	1040 (2)	1147 (2)
Current smokers, quit in the past year, <1 pack/d	22 929 (12)	5875 (12)	5404 (11)	5492 (11)	6158 (13)
Current smokers, quit in the past year, 1-2 packs/d	7846 (4)	2311 (5)	1855 (4)	1815 (4)	1865 (4)
Current smokers, quit in the past year, ≥2 packs/d	606 (0)	174 (0)	137 (0)	162 (0)	133 (0)
Missing	7180 (4)	1733 (4)	1757 (4)	1794 (4)	1896 (4)
BMI (kg/m2), mean (SD)	26.9 (5.6)	26.5 (5.4)	26.8 (5.6)	26.9 (5.7)	27.3 (5.9)
Catchment area
California	63 433 (32)	13 812 (28)	14 289 (29)	12 588 (26)	22 744 (46)
Florida	41 449 (21)	28 582 (58)	11 381 (23)	1485 (3)	<5 (0)
Louisiana	7719 (4)	1609 (3)	3253 (7)	2739 (6)	118 (0)
New Jersey	24 266 (12)	2128 (4)	8305 (17)	9469 (19)	4364 (9)
North Carolina	15 965 (8)	2000 (4)	5619 (11)	6863 (14)	1483 (3)
Pennsylvania	28 659 (15)	874 (2)	4485 (9)	10 264 (21)	13 036 (26)
Atlanta, GA	5501 (3)	<5 (0)	229 (0)	1816 (4)	3455 (7)
Detroit, MI	9913 (5)	218 (0)	1667 (3)	4003 (8)	4025 (8)
NO2, mean (SD), ppb	14.9 (8.3)	8.6 (3.6)	11.8 (4.7)	15.4 (5.4)	23.7 (9.4)
Age at first live birth
Nulliparous	28 057 (14)	6161 (13)	6739 (14)	6892 (14)	8265 (17)
Younger than 16 y	1183 (1)	297 (1)	246 (1)	247 (1)	393 (1)
16-19 y	33 388 (17)	9568 (19)	8323 (17)	7175 (15)	8322 (17)
20-24 y	84 851 (43)	22 379 (45)	21 474 (44)	21 164 (43)	19 834 (40)
25-29 y	34 612 (18)	7638 (16)	8685 (18)	9745 (20)	8544 (17)
30-34 y	8864 (5)	1894 (4)	2303 (5)	2430 (5)	2237 (5)
35-39 y	2186 (1)	427 (1)	595 (1)	580 (1)	584 (1)
40 years and older	400 (0)	75 (0)	94 (0)	126 (0)	105 (0)
Unknown	3364 (2)	785 (2)	769 (2)	868 (2)	942 (2)
Parity
0	29 567 (15)	6496 (13)	7089 (14)	7266 (15)	8716 (18)
1	20 419 (10)	4815 (10)	5019 (10)	5129 (10)	5456 (11)
2	50 760 (26)	12 869 (26)	12 813 (26)	12 829 (26)	12 249 (25)
3-4	71 662 (36)	18 755 (38)	18 244 (37)	17 910 (36)	16 753 (34)
≥5	22 168 (11)	5753 (12)	5524 (11)	5478 (11)	5413 (11)
Unknown	2329 (1)	536 (1)	539 (1)	615 (1)	639 (1)
Oral contraceptive use
Never	116 523 (59)	28 724 (58)	28 670 (58)	29 347 (60)	29 782 (61)
Ever	77 217 (39)	19 760 (40)	19 803 (40)	19 045 (39)	18 609 (38)
Missing	3165 (2)	740 (2)	755 (2)	835 (2)	835 (2)
Menopausal hormone use
Never	92 138 (47)	21 910 (45)	22 540 (46)	23 679 (48)	24 009 (49)
Ever	104 767 (53)	27 314 (55)	26 688 (54)	25 548 (52)	25 217 (51)
Age at menarche, y
10 or younger	13 379 (7)	3452 (7)	3252 (7)	3270 (7)	3405 (7)
11-12	81 956 (42)	20 130 (41)	20 505 (42)	20 602 (42)	20 719 (42)
13-14	809 98 (41)	20 441 (42)	20 337 (41)	20 296 (41)	19 924 (40)
15 and older	18 385 (9)	4693 (10)	4647 (9)	4467 (9)	4578 (9)
Unknown	2187 (1)	508 (1)	487 (1)	592 (1)	600 (1)
First-degree relative with history of breast cancer
No	162 793 (83)	40 923 (83)	40 674 (83)	40 704 (83)	40 492 (82)
Yes	24 054 (12)	5934 (12)	6061 (12)	6031 (12)	6028 (12)
Unknown	10 058 (5)	2367 (5)	2493 (5)	2492 (5)	2706 (6)

a

1980-1984 quartile cut points: quartile 1: <16.0 µg/m³; quartile 2: 16.0 µg/m³ to <18.7 µg/m³; quartile 3: 18.7 µg/m³ to <21.2 µg/m³; quartile 4: >21.2 µg/m³. BMI = body mass index; NO₂ = nitrogen dioxide; PM_2.5 = fine particulate matter; ppb = parts per billion.

A 10 µg/m³ increase in PM_2.5 for 1980-1984 was associated with an 8% increase in overall breast cancer risk (HR = 1.08, 95% CI = 1.02 to 1.13; Table 2). We also observed a statistically significant increased risk across quartiles, with monotonic trend (eg, HR_{Q4 vs Q1} = 1.08, CI = 1.02 to 1.13; P_trend = .004). Overall associations were similar when using the other 2 historical time periods (1985-1989 and 1990-1994). Our spline analysis indicated that the PM_2.5 and breast cancer association did statistically significantly depart from linearity (P = .001) (Supplementary Figure 1, available online), with a plateau observed at about 21 µg/m³, which is approximately the 75th percentile. This is consistent with our findings in our quartile analysis, showing similar hazard ratios for the third and fourth quartiles of exposure (HR_{Q3 vs Q1} = 1.08, 95% CI = 1.03 to 1.14, and HR_{Q4 vs Q1} = 1.08, 95% CI = 1.02 to 1.14).

Table 2.

Associations between PM_2.5 concentrations (1980-1984, 1985-1989, 1990-1994) and incidence of breast cancer in the National Institutes of Health–AARP Diet and Health Study^a

	PM2.5
Historical Exposure Period	Per 10 µg/m3 increase	Quartile 1	Quartile 2	Quartile 3	Quartile 4	Ptrend
1980-1984b
No. of cases	15 870	3779	3945	4066	4080
HR (95% CI)	1.08 (1.02 to 1.13)	1.0 (referent)	1.04 (0.99 to 1.09)	1.08 (1.03 to 1.14)	1.08 (1.02 to 1.14)	.004
1985-1989c
No. of cases	15 870	3771	3966	4034	4099
HR (95% CI)	1.07 (1.02 to 1.13)	1.0 (referent)	1.04 (0.99 to 1.09)	1.08 (1.02 to 1.14)	1.08 (1.03 to 1.15)	.003
1990-1994d
No. of cases	15 870	3744	3996	4043	4087
HR (95% CI)	1.07 (1.01 to 1.14)	1.0 (referent)	1.06 (1.01 to 1.11)	1.09 (1.04 to 1.16)	1.09 (1.03 to 1.15)	.003

a

Models adjusted for age, race and ethnicity, smoking status, body mass index, catchment area, and educational attainment. CI = confidence interval; HR = hazard ratio; PM_2.5 = fine particulate matter.

b

1980-1984 quartile cut points: quartile 1: <16.0 µg/m³; quartile 2: 16.0 µg/m³ to <18.7 µg/m³; quartile 3: 18.7 µg/m³ to <21.2 µg/m³; quartile 4: >21.2 µg/m³.

c

1985-1989 quartile cut points: quartile 1: <14.6 µg/m³; quartile 2: 14.6 µg/m³ to <17.1 µg/m³; quartile 3: 17.1 µg/m³ to <19.5 µg/m³; quartile 4: >19.5 µg/m³.

d

1990-1994 quartile cut points: quartile 1: <13.2 µg/m³; quartile 2: 13.2 µg/m³ to <15.6 µg/m³; quartile 3: 15.6 µg/m³ to <17.7 µg/m³; quartile 4: >17.7 µg/m³.

When evaluating estrogen receptor–positive and estrogen receptor–negative tumors separately, PM_2.5 was associated with a higher incidence of estrogen receptor–positive (per 10 µg/m³ increase, HR = 1.10, 95% CI = 1.04 to 1.17) but not estrogen receptor–negative tumors (HR = 0.97, 95% CI = 0.84 to 1.13; P_{heterogeneity} = .3; Table 3). Associations for estrogen receptor–positive or progesterone receptor–positive tumors compared with estrogen receptor–negative and progesterone receptor–negative tumors were similar with associations evident for the hormone receptor–positive subtype but not hormone receptor–negative (Supplementary Table 2, available online). We saw no statistically significant difference in the estimates by tumor extent, although the association for DCIS (HR = 1.02, 95% CI = 0.90 to 1.16) was attenuated and less precise compared with invasive tumors (HR = 1.08, 95% CI = 1.02 to 1.14; Table 3).

Table 3.

Associations between historic PM_2.5 concentrations and incidence of breast cancer in the National Institutes of Health–AARP Diet and Health Study, by tumor characteristics and study catchment area

		PM2.5 (1980-1984)a
Tumor characteristics and study area		Per 10 µg/m3 increase	Quartile 1	Quartile 2	Quartile 3	Quartile 4	Ptrend
By estrogen receptor statusb
Estrogen receptor–positive	No. of cases	9185	1835	2338	2519	2493
	HR (95% CI)	1.10 (1.04 to 1.17)	1.0 (referent)	1.10 (1.03 to 1.17)	1.15 (1.07 to 1.23)	1.11 (1.03 to 1.19)	.007
Estrogen receptor–negative	No. of cases	1731	379	423	463	466
	HR (95% CI)	0.97 (0.84 to 1.13)	1.0 (referent)	0.94 (0.81 to 1.09)	0.98 (0.84 to 1.15)	0.97 (0.83 to 1.14)	.9
By tumor extentb
DCIS	No. of cases	2510	591	618	655	646
	HR (95% CI)	1.02 (0.90 to 1.16)	1.0 (referent)	1.02 (0.90 to 1.15)	1.06 (0.93 to 1.21)	1.01 (0.88 to 1.16)	.8
Invasive	No. of cases	12 955	3113	3232	3277	3333
	HR (95% CI)	1.08 (1.02 to 1.14)	1.0 (referent)	1.04 (0.98 to 1.09)	1.07 (1.01 to 1.13)	1.08 (1.02 to 1.15)	.008
By study catchment areab,c
California	No. of cases	5522
	HR (95% CI)	1.06 (1.00 to 1.13)	—	—	—	—
Florida	No. of cases	3150	—	—	—	—
	HR (95% CI)	1.13 (0.93 to 1.37)	—	—	—	—
Los Angeles	No. of cases	588	—	—	—	—
	HR (95% CI)	1.04 (0.68 to 1.57)	—	—	—	—
New Jersey	No. of cases	1999	—	—	—	—
	HR (95% CI)	1.11 (0.90 to 1.36)	—	—	—	—
North Carolina	No. of cases	1132	—	—	—	—
	HR (95% CI)	1.26 (0.96 to 1.64)	—	—	—	—
Pennsylvania	No. of cases	2315	—	—	—	—
	HR (95% CI)	1.08 (0.88 to 1.32)	—	—	—	—
Atlanta, GA	No. of cases	432	—	—	—	—
	HR (95% CI)	1.22 (0.68 to 2.19)	—	—	—	—
Detroit, MI	No. of cases	732	—	—	—	—
	HR (95% CI)	1.14 (0.77 to 1.67)	—	—	—	—

a

All models adjusted for age, race and ethnicity, smoking status, body mass index, catchment area, and educational attainment. 1980-1984 quartile cut points: quartile 1: <16.0 µg/m³; quartile 2: 16.0 µg/m³ to <18.7 µg/m³; quartile 3: 18.7 µg/m³ to <21.2 µg/m³; quartile 4: >21.2 µg/m³. CI = confidence interval; DCIS = ductal carcinoma in situ; HR = hazard ratio; PM_2.5 = fine particulate matter.

b

Heterogeneity P values for continuous PM_2.5: estrogen receptor status, P = .3; tumor extent, P = .6; study catchment area, P = .9.

c

No estimates provided for quartile-specific PM_2.5 and breast cancer by study catchment area because of small sample sizes.

In our exploratory analysis by study catchment area, we observed differences in the magnitude of associations, although we had limited statistical power to detect effect measure modification (P_{heterogeneity} = .9). Estimates were lowest for Louisiana (HR = 1.04, 95% CI = 0.68 to 1.57) and California (HR = 1.06, 95% CI = 1.00 to 1.13; Table 3). In contrast, estimates were highest for North Carolina (HR = 1.26, 95% CI = 0.96 to 1.64) and Atlanta (HR = 1.22, 95% CI = 0.68 to 2.19).

There was some evidence of heterogeneity by BMI (P_{heterogeneity} = .2), with elevated associations between continuous PM_2.5 and incident breast cancer observed for women with a BMI of 18.5 kg/m² to less than 25.0 kg/m² (HR = 1.10, 95% CI = 1.02 to 1.19) or a BMI of 25.0 kg/m² to less than 30.0 kg/m² (HR = 1.10, 95% CI = 1.01 to 1.21) but a null association for women with a BMI of at least 30.0 kg/m² (Supplementary Table 3, available online).

Analyses restricting to postmenopausal women (HR = 1.07, 95% CI = 1.02 to 1.13) were similar to those of the full analytic sample (Supplementary Table 4, available online). Although overall breast cancer estimates for continuous PM_2.5 were slightly attenuated when reproductive factors were added to the adjustment set (HR = 1.06, 95% CI = 1.00 to 1.11), the associations for estrogen receptor–positive tumors remained statistically significant (HR = 1.08, 95% CI = 1.01 to 1.15). Associations remained similar when restricting to women with at least 10 years of follow-up (HR = 1.08, 95% CI = 1.01 to 1.16), who had a mammogram in the past 3 years (HR = 1.11, 95% CI = 1.04 to 1.18), and when NO₂ was added to the adjustment set (HR = 1.09, 95% CI = 1.01 to 1.18). Associations also slightly attenuated when removing race and ethnicity from the adjustment set. Restricting to better quality geocodes resulted in similar findings (Supplementary Table 5, available online).

Discussion

In this large, prospective cohort of women across the United States, we observed an 8% increase in breast cancer risk for a 10 µg/m³ increase in estimated historic PM_2.5 concentrations during a period of 10-15 years before enrollment. This association was evident for estrogen receptor–positive but not estrogen receptor–negative tumors.

PM_2.5 has established genotoxic effects (25), and several components of PM_2.5 or chemicals that attach to PM_2.5, such as polybrominated diphenyl ethers, phthalates, and metals (26,27), have been associated with endocrine-disrupting behaviors. Estrogen plays an important role in breast cancer initiation, promotion, and progression (28), specifically for estrogen receptor–positive breast cancer (29). There is also emerging evidence for a role of PM_2.5 on breast tissue characteristics that may predispose women to being at a higher risk. Higher PM_2.5 concentrations were cross-sectionally associated with higher counts of terminal duct lobular units, which are the endothelial tissue sites where breast cancer is most likely to arise (30). In an analysis that used machine learning methods to evaluate the different tissue types within nonmalignant breast tissue samples, an increase of 5 µg/m³ in PM_2.5 was associated with a greater percent increase in stroma and in the proportion of epithelial-to-stromal tissue, which may create a tissue environment that is conducive to tumor development by allowing epithelial proliferation (31).

Two recent meta-analyses on PM_2.5 and breast cancer included analyses by hormone receptor status, and neither observed an association for estrogen receptor–positive and progesterone receptor–positive tumors (13,14). Our study was better powered to detect an association by estrogen receptor and/or progesterone receptor status [eg, Gabet et al. (13); n = 5917 cases vs 9321 in our study], although that is unlikely to be the sole explanation for the difference in findings, as other differences in study population characteristics or exposure assessment may be playing a role. The ELAPSE pooled cohort analysis, which was not included in these meta-analyses, also observed an overall association with PM_2.5 and breast cancer risk but was unable to evaluate associations by hormone receptor status (16).

Although these analyses were underpowered, potential geographic variation in the association between PM_2.5 and breast cancer is worth further study, given the known heterogeneity in PM_2.5 chemical composition across the United States (32). In our exploratory analyses, we found a suggested stronger association within certain catchment areas, including North Carolina and Atlanta. In the Sister Study, participants diagnosed with invasive breast cancer residing in the western United States were found to have a hazard ratio of 1.21 (95% CI = 1.03 to 1.43) per interquartile range increase in PM_2.5 (3.6 µg/m³), whereas negligible associations were observed in other regions (33). The Black Women’s Health Study observed an elevated risk for women living in the Midwest (HR = 1.18, 95% CI = 1.00 to 1.39) per interquartile range increase in PM_2.5 (2.87 µg/m³) with no associations for the other regions (34). In our analysis, we observed a slightly higher association between PM_2.5 and breast cancer in Detroit, the one catchment area in the Midwest (HR = 1.14, 95% CI = 0.77 to 1.67). One possible explanation for the inconsistent regional findings between studies is the difference in the timing of PM_2.5 assessment; all used concentrations estimated for different years: the Sister Study (2006), the Black Women’s Health Study (1999-2008), and our analysis in the NIH-AARP Diet and Health Study (1980-1994). It could be that temporal and regional changes in PM_2.5 component concentrations have not occurred at the same pace across the United States.

A number of studies have observed racial and ethnic disparities in exposure to PM_2.5 (20,22), with historic redlining potentially a contributing factor (35). We observed a notable difference in the distribution of exposure across racial and ethnic groups in our population, with non-Hispanic Black participants more likely to be exposed to the highest levels than non-Hispanic White participants. However, we were underpowered to investigate effect measure modification of this association by race and ethnicity in this cohort.

This study included a large, geographically diverse population with more than 15 000 breast cancer cases identified, providing sufficient power to conduct stratified analyses, particularly by tumor characteristics. Only 1 prior registry-based study of residents of Ontario, Canada, had more cases of breast cancer (n = 91 146) (36); however, it was lacking in individual-level data apart from age at baseline. Although our population did not cover the entire United States, we had participants in all 4 US census regions, enabling us to evaluate geographic variability in the PM_2.5–breast cancer association. An additional strength of our study is the use of historical PM_2.5, which provides a more temporally relevant exposure window for the development of breast cancer with its estimated minimum latency of 15 years (37). Application of air pollution exposure models that predict historic PM_2.5 concentrations in epidemiologic studies is limited, and ours is the first US-based cohort investigation of breast cancer to utilize this approach.

Despite numerous strengths, our study had limitations. Because the NIH-AARP study recruited adults aged older than 50 years, our analysis consisted primarily of postmenopausal breast cancer cases, precluding analyses of premenopausal breast cancer. Although the historical PM_2.5 estimates allowed us to consider exposure 10-15 years prior to enrollment in the cohort, we were unable to capture early life exposures or exposures during developmental windows that might be especially important (eg, puberty, pregnancy) (38,39). Finally, historic PM_2.5 concentrations were estimated at the baseline addresses, which may result in some nondifferential misclassification of exposure if participants moved prior to enrolling in the study. Although some participants likely moved between 1980 and 1995, we expect the proportion to be relatively small based on US census data by age and time period (40). In residence history estimation for a subset of the California participants, the median duration at the enrollment address prior to study start was 13 years (41). Thus, although we can expect some exposure misclassification occurred because of residential mobility, its influence on our results should be minimal. We observed similar associations across the different time periods of air pollution exposure assessment explored in this study (1980-1984, 1985-1989, 1990-1994), as there is a high correlation in PM_2.5 levels over time at a single location. With presumed residential stability, this provides some support for studies evaluating this association with exposures estimated closer to study enrollment. Moreover, we expect that on average this misclassification would be nondifferential with respect to breast cancer incidence, likely resulting in attenuation of estimates of risk. Future studies should explore how residential mobility may influence air pollutant exposure levels and resulting associations with breast cancer incidence.

In this large prospective study, long-term historical exposure to ambient PM_2.5 at the residence was associated with a higher risk of estrogen receptor–positive breast cancer. Future work should emphasize evaluation of historic exposures and consider region-specific associations and the potential contribution of PM_2.5 chemical constituency in modifying the observed association with breast cancer.

Data availability

Data described in the manuscript, code book, and analytic code will be made available upon request, pending approval from the NIH-AARP Diet and Health Study Steering Committee. Further details are provided at https://www.nihaarpstars.com/.

Author contributions

Alexandra J White, PhD (Methodology; Supervision; Writing—original draft; Writing—review & editing), Jared A Fisher, PhD (Data curation; Formal analysis; Writing—review & editing), Marina Sweeney, PhD (Writing—original draft; Writing—review & editing), Neal Freedman, PhD (Writing—review & editing), Joel Kaufman, MD (Methodology; Writing—review & editing) Debra Silverman, ScD (Supervision; Writing—review & editing), and Rena R Jones, PhD (Conceptualization; Methodology; Supervision; Writing—review & editing).

Funding

This research was funded by the National Institutes of Environmental Health Sciences and the National Cancer Institute Intramural Program, ZIA CP010125–28, Z1AES103332.

Conflicts of interest

The authors report no conflict of interests.

Acknowledgements

The funders had no role in the study, including collection, analysis, and interpretation of the data, writing of the manuscript, and the decision to submit the manuscript for publication.