Ambient Air Pollution and Cancer Mortality in the Cancer Prevention Study II

Abstract

Background:

The International Agency for Research on Cancer classified both outdoor air pollution and airborne particulate matter as carcinogenic to humans (Group 1) for lung cancer. There may be associations with cancer at other sites; however, the epidemiological evidence is limited.

Objective:

The aim of this study was to clarify whether ambient air pollution is associated with specific types of cancer other than lung cancer by examining associations of ambient air pollution with nonlung cancer death in the Cancer Prevention Study II (CPS-II).

Methods:

Analysis included 623,048 CPS-II participants who were followed for 22 y (1982–2004). Modeled estimates of particulate matter with aerodynamic diameter $<2.5\mathrm{\mu }\mathrm{m}$ (${\text{PM}}_{2.5}$) (1999–2004), nitrogen dioxide (${\text{NO}}_2$) (2006), and ozone (${\mathrm{O}}_3$) (2002–2004) concentrations were linked to the participant residence at enrollment. Cox proportional hazards models were used to estimate associations per each fifth percentile–mean increment with cancer mortality at 29 anatomic sites, adjusted for individual and ecological covariates.

Results:

We observed 43,320 nonlung cancer deaths. ${\text{PM}}_{2.5}$ was significantly positively associated with death from cancers of the kidney {adjusted hazard ratio (HR) per $4.4{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3=1.14$ [95% confidence interval (CI): 1.03, 1.27]} and bladder [$\text{HR}=1.13$ (95% CI: 1.03, 1.23)]. ${\text{NO}}_2$ was positively associated with colorectal cancer mortality [HR per $6.5\text{ppb}=1.06$ (95% CI: 1.02, 1.10). The results were similar in two-pollutant models including ${\text{PM}}_{2.5}$ and ${\text{NO}}_2$ and in three-pollutant models with ${\mathrm{O}}_3$. We observed no statistically significant positive associations with death from other types of cancer based on results from adjusted models.

Conclusions:

The results from this large prospective study suggest that ambient air pollution was not associated with death from most nonlung cancers, but associations with kidney, bladder, and colorectal cancer death warrant further investigation. https://doi.org/10.1289/EHP1249

Introduction

The International Agency for Research on Cancer (IARC) classified outdoor air pollution and airborne particulate matter as carcinogenic to humans (Group 1) for lung cancer based on findings from observational and experimental studies as well as from strong mechanistic evidence (IARC 2013). Recent meta-analyses reported positive associations of particulate matter with aerodynamic diameter $<2.5\mathrm{\mu }\mathrm{m}$ (${\text{PM}}_{2.5}$ {relative risk (RR) per $10{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3=1.09$ [95% confidence interval (CI): 1.04, 1.14]} and nitrogen dioxide (${\text{NO}}_2$) [RR per $10{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3=1.04$ (95% CI: 1.01, 1.08) with lung cancer risk (Hamra et al. 2014, 2015). Similar findings were reported in other recent reviews (Cui et al. 2015; Yang et al. 2016).

Ambient air pollution represents a complex mixture of a broad range of carcinogenic and mutagenic substances that may play a role in chronic systemic inflammation, oxidative stress, and DNA damage in tissues other than the lung (Brook et al. 2010; Crouse et al. 2010; IARC 2012, 2013; Newby et al. 2015). As such, there may also be associations between ambient air pollution and other types of cancer; however, the epidemiological evidence for these associations is limited (IARC 2013).

Some previous studies reported positive associations of ambient air pollution with fatal bladder cancer (Liu et al. 2009), pancreatic cancer (Ancona et al. 2015), upper digestive tract, digestive accessory organs, and breast cancer (Wong et al. 2016), and with incident brain and cervical cancer (Raaschou-Nielsen et al. 2011), breast cancer (Crouse et al. 2010; Reding et al. 2015), hepatocellular carcinoma (Pan et al. 2015), and prostate cancer (Parent et al. 2013). Results from other studies of fatal brain cancer (McKean-Cowdin et al. 2009), incident brain tumors (Poulsen et al. 2016), and incident leukemia (Winters et al. 2015) were mixed or null. Previous studies are difficult to compare given variation in the specific cancers studied; whether outcomes were incident versus fatal cancers; and differences in study design, sample size, exposure assessment, and the availability of data on potential confounders.

We previously reported positive associations between long-term ambient air pollution exposure and lung cancer mortality in the American Cancer Society (ACS) Cancer Prevention Study II (CPS-II) (Jerrett et al. 2013; Krewski et al. 2000, 2009; Turner et al. 2011, 2014, 2016). For example, we estimated that a $10{-\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ increase in ${\text{PM}}_{2.5}$ concentrations was associated with a 9% (95% CI: 3, 16%) increase in lung cancer mortality (Turner et al. 2016, Table E4). The CPS-II is a large-scale prospective cohort in which the human health effects of ambient air pollution have been extensively investigated. The objective of this work was to clarify whether ambient air pollution is associated with specific types of cancer other than lung cancer by examining associations of ambient air pollution with nonlung cancer death among 623,048 CPS-II participants who were followed for 22 y.

Methods

Study Population

Approximately 1.2 million CPS-II participants were enrolled by 77,000 volunteers in 1982 throughout the entire United States, Washington, DC, and Puerto Rico. Participants, largely friends and family members of volunteers, were $\ge 30\mathrm{y}$ old with a family member $\ge 45\mathrm{y}$ old. A detailed self-administered questionnaire was completed at enrollment to obtain demographic, medical, and behavioral information (CPS-II questionnaires are available at https://www.cancer.org/research/we-conduct-cancer-research/epidemiology/cancer-prevention-questionnaires.html). Informed consent was implied by completion and return of the enrollment questionnaire. The Emory University School of Medicine Human Investigations Committee has approved all aspects of the CPS-II.

Volunteers ascertained the vital status of participants they had enrolled in 1984, 1986, and 1988, and death certificates, coded by trained nosologists, were used to determine the underlying cause of death, that is, “the disease or injury which initiated the train of morbid events leading directly to death” (WHO 1992). Since 1989, the National Death Index (NDI) has been used to determine vital status and underlying cause of death (Calle and Terrell 1993). Underlying cause of death was categorized according to the International Classification of Diseases (ICD) revisions 9 and 10 (WHO 1977, 1992). More than 99% of known deaths were assigned a cause.

Ambient Air Pollution Concentrations

${\text{PM}}_{2.5}$ concentrations were assigned to the geocoded participant residence at enrollment based on a national-level hybrid land use regression (LUR) and Bayesian maximum entropy (BME) interpolation model (Beckerman et al. 2013) used in previous work (Jerrett et al. 2016; Pope et al. 2015; Turner et al. 2014; 2016). Details on the geocoding of participant residences are presented elsewhere (Pope et al. 2015). A total of 104,172 observations were collected monthly from 1,464 monitoring sites between 1999 and 2008. Approximately 10% of observations were reserved for cross-validation. ${\text{PM}}_{2.5}$ concentrations were predicted in a base LUR model that included traffic within $1\mathrm{km}$ (based on modeled traffic counts) and the cube of green space within $100\mathrm{m}$. A BME interpolation model was then used to account for residual spatiotemporal variation in ${\text{PM}}_{2.5}$ concentrations, and the two estimates were combined. The cross-validation $R^2$ was 0.79, suggesting good spatiotemporal model prediction at locations other than those used to calibrate the model. Mean ${\text{PM}}_{2.5}$ concentrations at the enrollment residence for the period 1999–2004 were used in the present analysis to coincide with the cohort follow-up period.

${\text{NO}}_2$ concentrations were assigned at the census block group level to each participant enrollment residence. A national LUR model was used that incorporated both hourly monitoring data from 423 monitors and $\sim 4\text{million}$ satellite-based measurements in an approximate $10\mathrm{km}\times 10\mathrm{km}$ grid scale, including additional data on population density, land use, and distance to roadways for the year 2006 (model $R^2=0.78$) (Novotny et al. 2011; Turner et al. 2016).

Ozone (${\mathrm{O}}_3$) concentrations were obtained from the U.S. Environmental Protection Agency (EPA) and Centers for Disease Control and Prevention (CDC) Environmental Public Health Tracking Network Hierarchical Bayesian space time model (HBM) combining data from national air monitoring stations/state and local air monitoring stations (NAMS/SLAMS) and the Models-3/Community Multiscale Air Quality (CMAQ) photochemical model (U.S. EPA 2011). Daily 8-h maximum concentrations at a $36\mathrm{km}\times 36\mathrm{km}$ grid scale for the years 2002–2004 were assigned to each participant enrollment residence (Turner et al. 2016).

Statistical Methods

Both single- and multipollutant Cox proportional hazards regression models were used to estimate associations of ambient ${\text{PM}}_{2.5}$, ${\text{NO}}_2$, and ${\mathrm{O}}_3$ concentrations with death from 29 specific types of cancer according to an increment of the mean minus the fifth percentile for each pollutant ($4.4{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ for ${\text{PM}}_{2.5}$, $6.5\text{ppb}$ for ${\text{NO}}_2$, and $6.9\text{ppb}$ for ${\mathrm{O}}_3$) (Table 1). The fifth percentile–mean increment was used for comparability of results across pollutants and with previous work (Turner et al. 2016).

Table 1.

Distribution of air pollution concentrations, CPS-II cohort, United States ($\mathit{n}=\mathrm{623,048}$).

Air pollutant (units)	Time period	Mean (SD)	Minimum	Percentiles					Maximum	Increment
				5th	25th	50th	75th	95th		5th percentile–mean
${\text{PM}}_{2.5}$ (${\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$)	1999–2004	12.6 (2.8)	1.4	8.2	10.6	12.5	14.4	17.0	27.9	4.4
${\text{NO}}_2$ (ppb)	2006	11.6 (5.1)	1.0	5.1	8.1	10.8	14.1	21.2	37.6	6.5
${\mathrm{O}}_3$ (ppb)	2002–2004	38.2 (4.0)	26.7	31.3	36.2	38.1	40.1	44.9	59.3	6.9
Near-source ${\text{PM}}_{2.5}$ (${\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$)a	1999–2004	12.0 (0.9)	8.6	10.4	11.6	12.0	12.5	13.5	19.7	1.6
Regional ${\text{PM}}_{2.5}$ (${\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$)a	1999–2004	0.5 (2.7)	$-7.9$	$-4.0$	$-1.4$	0.5	2.5	4.6	13.0	4.5

Note: CPS-II, Cancer Prevention Study II; ${\text{NO}}_2$, nitrogen dioxide; ${\text{O}}_3$, ozone; ${\text{PM}}_{2.5}$, particulate matter with aerodynamic diameter $<2.5\mathrm{\mu }\mathrm{m}$.

^aThe land use regression Bayesian maximum entropy (LURBME) ${\text{PM}}_{2.5}$ model was created in three main steps: a) A base LUR model predicted ${\text{PM}}_{2.5}$ concentrations based on traffic within $1\mathrm{km}$ (based on modeled traffic counts) and the cube of green space within $100\mathrm{m}$; b) A BME interpolation model was then used to interpolate residual spatiotemporal variation in ${\text{PM}}_{2.5}$ concentrations; c) The two estimates were then combined. Regional ${\text{PM}}_{2.5}$ concentrations therefore represent a residual with some observations less than zero.

Models were stratified by 1-y age categories, gender, and race/ethnicity (white, black, other) to allow for separate baseline hazards according to these characteristics. In addition, in the multivariate model, we also adjusted for baseline values of education ($<\text{high}\text{school}$, high school, $>\text{high}\text{school}$); marital status (single, married, other); body mass index (BMI); BMI squared; smoking status (never cigarettes, pipes, or cigars; current cigarette smoker only; former cigarette smoker only; ever pipe/cigar); cigarettes per day and cigarettes per day squared (current and former cigarette smokers); years smoked and years smoked squared (current and former cigarette smokers); started smoking $<18\mathrm{y}\text{old}$ (yes/no) (current and former cigarette smokers); passive smoking (hours per day exposed to cigarette smoke of others at home, work, or other areas); usual dietary intake of vegetables, fruit, and fiber (combined, in quintiles) and fat (quintiles), plus a category for missing/bad dietary intake data; usual consumption of beer (yes, no, missing), wine (yes, no, missing), and liquor (yes, no, missing); occupational exposures (ever regular exposure to asbestos, chemicals/acids/solvents, coal or stone dusts, coal tar/pitch/asphalt, formaldehyde, or diesel engine exhaust) (yes/no); and an occupational dirtiness index to characterize workplace ${\text{PM}}_{2.5}$ exposure based on main lifetime occupation (six categories of exposure vs. a referent category, or missing) (Siemiatycki et al. 2003). In addition, we used 1990 census data to derive the following socioecomonic covariates defined at the ZIP-code level for each participant’s residence at enrollment: median household income, percentage of African American residents, percentage of Hispanic residents, percentage of adults with post-secondary education, percentage of unemployed residents $\ge 16\mathrm{y}$ old, and percentage of residents with household incomes $<125\%$ of the poverty level. For each characteristic, we modeled the value for the ZIP code of residence and for a second variable indicating the difference between that value and the county-level mean (Jerrett et al. 2016; Pope et al. 2015; Turner et al. 2016; U.S. Department of Commerce Bureau of the Census 1993). The time axis was follow-up time in days. Participants lost to follow-up or who were alive at the end of follow-up were censored.

For analysis of death from cancers of the uterus, cervix, or ovary, participants were excluded if they reported a previous hysterectomy or an artificial (vs. a natural) menopause ($n=\mathrm{108,956}$). An additional 7,143 participants were also excluded from analysis of ovarian cancer mortality if they reported having undergone an ovarian surgery because no information was available to distinguish partial from total oophorectomy.

A sensitivity analysis was performed to examine the influence of including additional baseline variables in the model including usual physical exercise (none, slight, moderate, heavy, missing), aspirin use in the past month (yes/no), and usual dietary intake of red meat (quintiles) for analysis of all cancer sites, as well as self-reported physician-diagnosed diabetes (yes/no). Detailed reproductive and hormonal variables, including age at menarche ($<12\mathrm{y}\text{old}$, 12–13 y old, $\ge 14\mathrm{y}\text{old}$, missing), parity (0, 1, 2, 3, $\ge 4$, missing), age at first birth (none, $<20\mathrm{y}\text{old}$, 20–29 y old, $\ge 30\mathrm{y}\text{old}$, missing), ever oral contraceptive use (yes, no, missing), ever postmenopausal hormone use (yes, no, missing), and postmenopausal status (yes, no, missing) were also included in models for sensitivity analysis of female reproductive cancers. Additional analyses were also performed in postmenopausal women only.

Further, we decomposed ${\text{PM}}_{2.5}$ concentrations into near-source and regional fractions to determine whether associations might vary for ${\text{PM}}_{2.5}$ from different air pollution sources (Turner et al. 2016). Near-source ${\text{PM}}_{2.5}$ concentrations were ${\text{PM}}_{2.5}$ concentrations predicted at each monitoring station based on land-use and traffic count data using the LUR model created in the first stage of the LURBME modeling process. Regional ${\text{PM}}_{2.5}$ was derived by subtracting the near-source ${\text{PM}}_{2.5}$ concentration from the total ${\text{PM}}_{2.5}$ concentrations from the LURBME model, thereby capturing spatial variation in ${\text{PM}}_{2.5}$ concentrations between monitoring stations.

Finally, where there were statistically significant positive associations observed, potential effect modification by gender, education ($<\text{high}\text{school}$, high school, $>\text{high}\text{school}$), and smoking status (never cigarettes, pipes, or cigars; current cigarette smoker only; former cigarette smoker only) was assessed on a multiplicative scale. Two-sided p-values were calculated according to the likelihood ratio statistic to compare models with and without multiplicative interaction terms between air pollutants and the potential modifiers. The proportional hazards assumption was tested by including interaction terms between ambient air pollution and follow-up time in the multivariate proportional hazards models. A p-value $<0.05$ was used to define statistical significance throughout this work.

SAS (version 9.2; SAS Institute Inc.) was used to conduct all analyses. Ethics approval for analysis was obtained from the Ottawa Hospital Research Ethics Board.

Results

From 1982 through 2004 there were a total of 743,543 (62.8%) participants who were alive and 438,123 (37.0%) who had died. Participants were excluded from analysis because of missing or invalid residence information ($n=\mathrm{385,422}$) or selected baseline covariate data ($n=\mathrm{130,119}$) (Jerrett et al. 2016; Pope et al. 2015; Turner et al. 2016) or having a prevalent cancer (except nonmelanoma skin cancer) at enrollment ($n=\mathrm{45,998}$). Follow-up was censored in September 1988 for 2,921 (0.2%) of participants who had insufficient information to link to the NDI. The present analysis is based on 623,048 CPS-II participants among whom 43,320 nonlung cancer deaths were observed during 11,936,799 person-years of follow-up. Follow-up ranged from 0.01 y to 22.5 y with a mean (SD) of 19.2 (5.6) y.

Mean (SD) ${\text{PM}}_{2.5}$, ${\text{NO}}_2$, and ${\mathrm{O}}_3$ concentrations at the participant residence at enrollment were $12.6\left(2.8\right){\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$, $11.6\left(5.1\right)\text{ppb}$, and $38.2\left(4.0\right)\text{ppb}$, respectively (Table 1). Correlations between these air pollutants ranged from a weak inverse correlation between ${\text{NO}}_2$ and ${\mathrm{O}}_3$ ($r=-0.09$) to a moderate positive correlation between ${\text{PM}}_{2.5}$ and ${\text{NO}}_2$ ($\mathit{r}=0.40$) (see Table S1).

Participant characteristics are presented in Table 2 and Table S2. The majority of participants were 40 y old to 69 y old at enrollment, 94.5% of participants were white, 55.3% were female, and 57.4% had a greater than high school education. A total of 44.7% of study participants were never smokers.

Table 2.

Distribution of selected participant characteristics at enrollment (1982) and air pollution concentrations, CPS-II cohort, United States ($\mathit{n}=\mathrm{623,048}$).

Characteristic	%	${\text{PM}}_{2.5}$ (${\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$) Mean (SD)	${\text{NO}}_2$ (ppb) Mean (SD)	${\mathrm{O}}_3$ (ppb) Mean (SD)
Age (years)
$<40$	4.6	12.8 (2.9)	12.4 (5.7)	38.0 (3.9)
40–49	21.2	12.5 (2.8)	11.4 (5.1)	38.0 (3.8)
50–59	37.1	12.6 (2.8)	11.5 (5.0)	38.1 (3.9)
60–69	26.0	12.6 (2.9)	11.6 (5.1)	38.3 (4.1)
70–79	9.4	12.6 (2.9)	11.8 (5.0)	38.4 (4.2)
$\ge 80$	1.7	12.7 (2.9)	12.2 (5.1)	38.3 (4.2)
Race/ethnicity
White	94.5	12.5 (2.8)	11.5 (5.0)	38.2 (3.9)
Black	3.9	13.7 (2.5)	13.3 (5.2)	38.1 (3.3)
Other	1.6	12.9 (4.3)	15.6 (6.4)	38.3 (5.8)
Gender
Male	44.7	12.5 (2.8)	11.5 (5.1)	38.2 (4.0)
Female	55.3	12.6 (2.8)	11.7 (5.1)	38.2 (3.9)
Education
$<\text{High}\text{school}$	11.6	12.8 (2.8)	11.6 (5.3)	38.1 (3.8)
High school	31.1	12.6 (2.7)	11.4 (5.1)	38.1 (3.7)
$>\text{High}\text{school}$	57.4	12.5 (2.9)	11.7 (5.0)	38.2 (4.1)
Marital status
Single	3.3	13.0 (2.8)	13.1 (5.5)	37.5 (3.8)
Married	84.8	12.5 (2.8)	11.5 (5.0)	38.2 (3.9)
Other	11.9	12.9 (2.9)	12.4 (5.3)	38.1 (4.0)
BMI (${\mathrm{kg}/\mathrm{m}}^2$)
$<18.5$	1.7	12.6 (2.9)	11.7 (5.0)	38.5 (4.0)
18.5–24.9	50.3	12.5 (2.9)	11.6 (5.1)	38.2 (4.0)
25–29.9	36.5	12.6 (2.8)	11.5 (5.1)	38.1 (3.9)
$\ge 30$	11.5	12.8 (2.8)	11.7 (5.2)	38.1 (3.8)
Smoking status
Never	44.7	12.6 (2.9)	11.5 (5.1)	38.4 (4.0)
Current cigarette smoker	19.6	12.7 (2.8)	11.8 (5.1)	38.0 (3.8)
Former cigarette smoker	25.6	12.5 (2.9)	11.7 (5.1)	38.0 (4.0)
Ever pipe/cigar	10.2	12.5 (2.8)	11.5 (5.0)	38.0 (3.8)
Cigarettes per day
Current cigarette smoker
$<15$	4.9	12.7 (2.9)	12.0 (5.3)	37.9 (3.8)
15–19	1.5	12.7 (2.8)	11.8 (5.1)	37.9 (3.9)
20–29	7.2	12.6 (2.8)	11.6 (5.0)	38.0 (3.7)
$\ge 30$	6.0	12.7 (2.8)	11.7 (5.1)	38.0 (3.8)
Former cigarette smoker
$<10$	4.9	12.6 (2.9)	11.8 (5.1)	37.9 (4.0)
10–19	5.1	12.5 (2.9)	11.7 (5.1)	38.0 (4.1)
20–29	8.5	12.5 (2.9)	11.6 (5.1)	38.1 (4.0)
$\ge 30$	7.0	12.5 (2.9)	11.8 (5.2)	38.0 (4.0)
Duration of smoking (years)
Current cigarette smoker
$<26$	4.5	12.8 (2.8)	11.9 (5.3)	38.0 (3.7)
$<26$	4.7	12.7 (2.8)	11.7 (5.1)	37.9 (3.7)
26–32	4.7	12.7 (2.8)	11.7 (5.1)	38.0 (3.7)
33–39	5.6	12.6 (2.9)	11.8 (5.1)	38.1 (3.9)
$\ge 40$	6.1	12.5 (2.9)	11.6 (5.1)	38.0 (4.1)
Former cigarette smoker
$<12$	6.7	12.5 (2.8)	11.7 (5.1)	38.0 (4.0)
21–29	5.5	12.5 (2.9)	11.8 (5.1)	37.9 (4.0)
12–20	7.3	12.5 (2.9)	11.8 (5.2)	38.0 (4.0)
$\ge 30$
Age started smoking $<18\mathrm{y}\text{old}$
Current cigarette smoker	7.9	12.7 (2.8)	12.0 (5.3)	37.9 (3.8)
Former cigarette smoker	9.8	12.5 (2.8)	11.9 (5.3)	37.9 (4.0)
Passive smoke exposure (hours)
0	35.5	12.5 (2.9)	11.5 (5.1)	38.4 (4.2)
$>0–3$	33.8	12.5 (2.8)	11.6 (5.1)	38.0 (3.9)
$>3$	30.8	12.7 (2.8)	11.8 (5.2)	38.0 (3.7)
Beer consumption
Yes	24.9	12.6 (2.9)	11.8 (5.2)	38.3 (4.1)
No	16.3	12.5 (2.8)	11.6 (5.0)	37.9 (3.8)
Missing	58.8	12.6 (2.8)	11.6 (5.1)	38.1 (3.9)
Wine consumption
Yes	22.5	12.6 (2.8)	11.5 (5.1)	38.4 (3.9)
No	18.3	12.4 (3.0)	12.3 (5.2)	37.7 (4.3)
Missing	59.3	12.6 (2.8)	11.5 (5.0)	38.2 (3.8)
Liquor consumption
Yes	23.6	12.7 (2.9)	11.7 (5.2)	38.4 (4.0)
No	19.7	12.3 (2.9)	11.6 (5.0)	37.8 (4.0)
Missing	56.8	12.6 (2.8)	11.6 (5.1)	38.2 (3.9)
Vegetable/fruit/fiber consumption
1st quintile	17.2	12.8 (2.8)	11.6 (5.2)	38.1 (3.7)
2nd quintile	18.8	12.6 (2.8)	11.6 (5.1)	38.1 (3.8)
3rd quintile	18.1	12.5 (2.8)	11.6 (5.1)	38.1 (3.9)
4th quintile	18.9	12.4 (2.9)	11.6 (5.0)	38.1 (4.0)
5th quintile	18.9	12.4 (2.9)	11.6 (5.0)	38.2 (4.2)
Fat consumption
1st quintile	17.3	12.8 (2.9)	12.2 (5.3)	38.1 (4.0)
2nd quintile	18.3	12.6 (2.9)	11.8 (5.2)	38.0 (4.0)
3rd quintile	18.6	12.5 (2.8)	11.6 (5.1)	38.1 (4.0)
4th quintile	18.8	12.5 (2.8)	11.4 (5.0)	38.2 (3.9)
5th quintile	18.8	12.5 (2.8)	11.0 (4.8)	38.4 (3.8)
Missing/bad nutrition data	8.2	12.7 (2.9)	11.8 (5.2)	38.3 (4.0)
Industrial exposures
Yes	19.7	12.5 (2.9)	11.5 (5.1)	38.1 (4.0)
No	80.3	12.6 (2.8)	11.6 (5.1)	38.2 (3.9)
Occupational dirtiness index
Level 0	49.9	12.6 (2.8)	11.8 (5.1)	38.1 (4.0)
Level 1	13.7	12.5 (2.8)	11.4 (5.1)	38.2 (3.9)
Level 2	11.6	12.5 (2.9)	11.5 (4.9)	38.3 (4.0)
Level 3	4.8	12.5 (2.9)	11.6 (5.1)	38.2 (4.0)
Level 4	6.2	12.4 (2.7)	10.5 (5.0)	38.2 (3.8)
Level 5	4.3	12.5 (2.8)	11.4 (5.1)	38.2 (4.0)
Level 6	1.1	12.9 (2.9)	11.4 (4.8)	38.2 (4.0)
Missing	8.4	12.8 (2.9)	12.1 (5.3)	38.0 (3.9)
Median household income (thousands, USD)
$<25$	25.0	12.2 (2.8)	9.7 (4.2)	39.0 (3.3)
25–31	24.9	12.5 (2.8)	10.9 (4.4)	38.6 (3.8)
32–40	25.1	12.8 (3.0)	12.1 (4.9)	37.9 (4.2)
$\ge 41$	25.0	12.9 (2.8)	13.8 (5.8)	37.0 (4.2)
African American residents (%)
$<0.7$	25.0	11.6 (2.7)	10.0 (5.1)	37.7 (3.9)
0.7–2.4	25.0	12.3 (3.0)	12.4 (5.0)	38.1 (4.5)
2.5–8.8	24.5	13.0 (2.9)	12.5 (5.1)	38.5 (4.1)
$\ge 8.9$	25.5	13.4 (2.3)	11.6 (4.8)	38.3 (3.2)
Hispanic residents (%)
$<0.8$	25.0	13.2 (2.3)	9.1 (3.6)	38.5 (2.0)
0.8–1.7	25.0	12.7 (2.2)	10.4 (3.7)	37.9 (2.7)
1.8–5.8	24.6	11.9 (2.5)	12.5 (5.3)	37.1 (4.4)
$\ge 5.9$	25.4	12.5 (3.9)	14.5 (5.7)	39.1 (5.4)
Post-secondary education (%)
$<28.2$	25.0	13.1 (2.6)	10.1 (4.8)	38.8 (3.1)
28.2–37.0	25.0	12.5 (2.8)	11.3 (5.1)	38.0 (3.8)
37.1–48.1	25.0	12.3 (2.9)	12.3 (5.3)	38.1 (4.2)
$\ge 48.2$	25.0	12.5 (3.0)	12.8 (4.8)	37.7 (4.5)
Unemployment (%)
$<3.6$	25.0	12.7 (2.6)	11.8 (4.7)	37.2 (3.7)
3.6–4.9	25.0	12.5 (2.8)	12.1 (5.3)	38.1 (3.9)
5.0–6.6	25.0	12.3 (2.9)	11.2 (5.0)	38.7 (4.1)
$\ge 6.7$	25.0	12.8 (3.0)	11.3 (5.3)	38.7 (3.9)
Poverty (%)
$<4.6$	25.0	12.8 (2.4)	13.0 (5.3)	36.9 (3.5)
4.6–8.5	25.0	12.6 (3.0)	12.1 (5.0)	38.0 (4.4)
8.6–13.9	25.0	12.3 (2.9)	10.7 (4.7)	38.6 (3.9)
$\ge 14.0$	25.0	12.6 (3.0)	10.6 (4.9)	39.0 (3.7)

Note: BMI, body mass index; CPS-II, Cancer Prevention Study II; ${\text{NO}}_2$ , nitrogen dioxide; ${\mathrm{O}}_3$, ozone; ${\text{PM}}_{2.5}$, particulate matter with aerodynamic diameter $<2.5\mathrm{\mu }\mathrm{m}$; USD, U.S. dollar.

There was little variation in ambient air pollution concentrations by participant characteristics, although somewhat higher ${\text{PM}}_{2.5}$ and ${\text{NO}}_2$ concentrations were observed in the younger and older age groups and among participants who were black or “other” race/ethnicity (Table 2). ${\mathrm{O}}_3$ concentrations were slightly higher in older participants, in those with a low BMI ($<18.5{\mathrm{kg}/\mathrm{m}}^2$), and among never smokers.

The results from single-pollutant models per each fifth percentile–mean increment are presented in Table S3 according to the minimally adjusted model and in Table 3 according to the fully adjusted model. In the minimally adjusted model, there were several significant positive associations of ${\text{PM}}_{2.5}$ and colorectal, breast, cervical, and bladder cancer mortality and of ${\text{NO}}_2$ and stomach, colorectal, pancreatic, breast, and bladder cancer mortality as well as some significant inverse associations largely with ${\mathrm{O}}_3$. Fewer significant findings were observed in the fully adjusted models, with significant positive associations of ${\text{PM}}_{2.5}$ and mortality from kidney [HR per $4.4{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3=1.14$ (95% CI: 1.03, 1.27)] and bladder [$\text{HR}=1.13$ (95% CI: 1.03, 1.23)] cancer observed (Table 3) [or equivalently per each $10{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$, $\text{HRs}=1.36$ (95% CI: 1.06, 1.73) and 1.32 (95% CI: 1.07, 1.61), respectively]. There was a positive HR for colorectal cancer mortality [$\text{HR}=1.04$ (95% CI: 1.00, 1.08)] of borderline significance ($p=0.08$). The largest HR was observed for mortality from cervical cancer [$\text{HR}=1.34$ (95% CI: 0.98, 1.83)], although this result was no longer significant and was based on a small number of deaths ($n=115$). For ${\text{NO}}_2$, there was a weak positive association with mortality from colorectal cancer [HR per $6.5\text{ppb}=1.06$ (95% CI: 1.02, 1.10)] [or equivalently $\text{HR}=1.09$ (95% CI: 1.03, 1.16) per $10\text{ppb}$]. For ${\mathrm{O}}_3$, HRs were generally $<1$, with statistically significant inverse associations with mortality from stomach cancer [HR per $6.9\text{ppb}=0.90$ (95% CI: 0.81, 0.99)], pancreatic cancer [$\text{HR}=0.91$ (95% CI: 0.86, 0.97)], and leukemia [$\text{HR}=0.92$ (95% CI: 0.85, 0.99)].

Table 3.

Adjusted HRs (95% CIs)^a for nonlung cancer mortality per each fifth percentile–mean increment in air pollutant concentrations, single-pollutant models, CPS-II cohort, United States (1982–2004) ($\mathit{n}=\mathrm{623,048}$).

Cancer cause of death	ICD 9; ICD10	Number of deaths	${\text{PM}}_{2.5}$		${\text{NO}}_2$		${\mathrm{O}}_3$
			Per $4.4{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$		Per $6.5\text{ppb}$		Per $6.9\text{ppb}$
			HR	95% CI	HR	95% CI	HR	95% CI
Tongue and mouth	141, 143–145; C01–C06	262	1.03	0.84, 1.26	0.98	0.80, 1.19	0.93	0.74, 1.18
Salivary gland	142; C07–C08	58	0.93	0.61, 1.44	1.00	0.67, 1.48	1.44	0.91, 2.30
Pharynx	146–149; C09–C14	243	0.88	0.70, 1.10	0.86	0.69, 1.08	1.11	0.86, 1.43
Esophagus	150; C15	1,180	1.02	0.93, 1.13	1.02	0.92, 1.12	0.97	0.87, 1.08
Stomach	151; C16	1,340	1.00	0.91, 1.09	1.05	0.96, 1.15	0.90	0.81, 0.99
Colorectal	153–154; C18–C21	6,475	1.04	1.00, 1.08	1.06	1.02, 1.10	0.95	0.91, 1.00
Liver	155; C22	1,003	1.05	0.94, 1.16	1.03	0.93, 1.14	0.98	0.87, 1.10
Gallbladder	156; C23–C24	403	1.03	0.87, 1.22	1.00	0.85, 1.17	0.95	0.78, 1.14
Pancreas	157; C25	3,812	0.98	0.92, 1.03	1.04	0.98, 1.09	0.91	0.86, 0.97
Nose	160; C30–C31	46	0.65	0.37, 1.16	0.82	0.47, 1.42	1.18	0.69, 2.04
Larynx	161; C32	223	1.04	0.84, 1.30	1.07	0.86, 1.32	0.96	0.74, 1.24
Bone	170; C40–C41	81	0.81	0.55, 1.20	0.97	0.68, 1.38	0.95	0.63, 1.43
Connective tissue	171; C47, C49	377	1.06	0.89, 1.26	1.04	0.88, 1.22	0.89	0.74, 1.08
Melanoma	172; C43	862	1.05	0.94, 1.17	0.97	0.87, 1.08	1.02	0.90, 1.15
Other skin	173; C44, C46	195	1.10	0.88, 1.38	0.90	0.72, 1.14	0.97	0.75, 1.24
Breast (female)b	174–175; C50	3,844	1.03	0.97, 1.08	1.03	0.97, 1.08	0.99	0.93, 1.05
Uterusb,c	179, 182; C54–C55	611	1.04	0.91, 1.19	1.07	0.94, 1.22	0.98	0.84, 1.14
Cervixb,c	180; C53	115	1.34	0.98, 1.83	1.23	0.91, 1.66	1.01	0.69, 1.48
Ovaryb,c,d	183; C56	987	1.03	0.93, 1.15	1.03	0.93, 1.14	1.01	0.90, 1.14
Prostateb	185; C61	1,068	0.96	0.86, 1.06	1.01	0.92, 1.12	1.03	0.92, 1.15
Bladder	188; C67	1,324	1.13	1.03, 1.23	1.03	0.94, 1.12	1.03	0.93, 1.14
Kidney	189; C64–C66, C68	927	1.14	1.03, 1.27	0.99	0.89, 1.10	0.97	0.86, 1.09
Eye	190; C69	26	1.30	0.66, 2.53	1.14	0.57, 2.28	0.76	0.38, 1.53
Brain	191; C71	1,591	1.04	0.96, 1.14	0.99	0.92, 1.08	0.98	0.89, 1.08
Thyroid	193; C73	41	0.62	0.34, 1.12	1.24	0.75, 2.05	0.96	0.54, 1.71
Non-Hodgkin’s lymphoma	200, 202; C82–C85	2,840	1.00	0.94, 1.07	1.05	0.99, 1.12	0.98	0.92, 1.05
Hodgkin’s disease	201; C81	125	1.12	0.82, 1.54	1.16	0.87, 1.53	1.08	0.76, 1.53
Multiple myeloma	203; C88, C90	1,421	0.97	0.89, 1.07	1.00	0.92, 1.09	0.99	0.90, 1.09
Leukemia	204–208; C91–C95	2,584	1.01	0.94, 1.07	1.00	0.94, 1.07	0.92	0.85, 0.99

Note: CI, confidence interval; CPS-II, Cancer Prevention Study II; HR, hazard ratio; ICD-9, The International Classification of Diseases, 9th Revision; ICD-10, International Statistical Classification of Diseases and Related Health Problems, 10th Revision; ${\text{NO}}_2$, nitrogen dioxide; ${\mathrm{O}}_3$, ozone; ${\text{PM}}_{2.5}$, particulate matter with aerodynamic diameter $<2.5\mathrm{\mu }\mathrm{m}$.

^aAge, race/ethnicity, gender stratified and adjusted for baseline values of education; marital status; body mass index; body mass index squared; smoking status; cigarettes per day; cigarettes per day squared; duration of smoking; duration of smoking squared; age started smoking; passive smoking, vegetable/fruit/fiber consumption; fat consumption; beer, wine, liquor consumption; industrial exposures; occupation dirtiness index; and 1990 ecological covariates.

^bAs above but not gender stratified. Analysis limited to 344,593 women for analysis of breast cancer mortality, 235,637 women for analysis of uterine or cervical cancer mortality, and 228,494 women for analysis of ovarian cancer mortality. Analysis limited to 278,455 men for analysis of prostate cancer mortality.

^cWomen reporting a previous hysterectomy or an artificial menopause excluded here ($n=\mathrm{108,956}$).

^dWomen reporting having undergone an ovarian surgery also excluded here ($n=\mathrm{7,143}$).

The results were similar with further adjustment for physical exercise, aspirin use, and red meat intake for all cancer sites as well as with additional reproductive and hormonal variables for female reproductive cancers (see Tables S4 and S5). The results for female reproductive cancers were also similar upon restriction to postmenopausal women only. The results were similar in two-pollutant models including ${\text{PM}}_{2.5}$ and ${\text{NO}}_2$ (see Table S6) and in three-pollutant models with ${\mathrm{O}}_3$ (see Table S7).

Upon decomposition of total ${\text{PM}}_{2.5}$, there were statistically significant positive associations of both bladder and kidney cancer mortality per fifth percentile–mean increment with regional ${\text{PM}}_{2.5}$ [HRs per $4{.5\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3=1.14$ (95% CI: 1.04, 1.25) and 1.15 (95% CI: 1.02, 1.29), respectively] of similar magnitude per unit increment to those of total ${\text{PM}}_{2.5}$ but not with near-source ${\text{PM}}_{2.5}$ [HRs per $1.6{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3=1.02$ (95% CI: 0.92, 1.13) and 1.05 (95% CI: 0.92, 1.13), respectively] (see Table S8). There were also significant positive associations of near-source ${\text{PM}}_{2.5}$ and mortality from cancer of the stomach [HR per $1.6{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3=1.13$ (95% CI: 1.02, 1.26)] and colorectum [$\text{HR}=1.09$ (95% CI: 1.04, 1.15)] but not with regional ${\text{PM}}_{2.5}$ [HRs per $4.5{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3=0.97$ (95% CI: 0.88, 1.06) and 1.02 (95% CI: 0.97, 1.07), respectively]. There was also a positive HR for pancreatic cancer mortality with near-source ${\text{PM}}_{2.5}$ [$\text{HR}=1.06$ (95% CI: 1.00, 1.13)] of borderline significance ($p=0.06$), but not with regional ${\text{PM}}_{2.5}$ [$\text{HR}=0.96$ (95% CI: 0.90, 1.01)].

In stratified analyses of associations of ${\text{PM}}_{2.5}$ with kidney and bladder cancer mortality and of ${\text{NO}}_2$ with colorectal cancer mortality, associations of ${\text{PM}}_{2.5}$ with kidney and bladder cancer mortality appeared to be limited to men [$\text{HR}=1.23$ (95% CI: 1.08, 1.40) and $\text{HR}=1.16$ (95% CI: 1.05, 1.29), respectively], with little or no evidence of associations in women [$\text{HR}=1.00$ (95% CI: 0.83, 1.20) and $\text{HR}=1.03$ (95% CI: 0.86, 1.23), respectively], although differences between men and women were not significant (interaction p-values of 0.13 and 0.25, respectively) (see Table S9). There was no clear evidence of effect modification according to categories of smoking status (interaction p-values $\ge 0.30$). Associations of ${\text{NO}}_2$ with colorectal cancer mortality were somewhat stronger among those with a $<\text{high}\text{school}$ or high school level of education [$\text{HR}=1.12$ (95% CI: 1.01, 1.24) and $\text{HR}=1.15$ (95% CI: 1.07, 1.23), respectively] compared with a $>\text{high}\text{school}$ level of education [$\text{HR}=0.99$ (95% CI: 0.94, 1.05); $p\text{for}\text{interaction}=0.03$]. There was no evidence that the proportional hazards assumption was violated for associations at these sites ($p>0.05$) (data not shown).

Discussion

Ambient air pollution was not associated with death from most nonlung cancers in our large prospective study population. However, there were statistically significant positive associations between ${\text{PM}}_{2.5}$ and death from bladder and kidney cancer, ranging from 13–14% increases in risk per each fifth percentile–mean increment ($4.4{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$). For ${\text{NO}}_2$, there was a statistically significant positive association with colorectal cancer mortality of 6% per $6.5\text{ppb}$. The results were similar in two-pollutant models including both ${\text{PM}}_{2.5}$ and ${\text{NO}}_2$ as well as in three-pollutant models with ${\mathrm{O}}_3$. The magnitudes of the associations here were somewhat stronger than, although compatible with, those of ${\text{PM}}_{2.5}$ and lung cancer mortality observed in previous work [9% (95% CI: 3, 16%) per $10{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ or equivalently, 4% (95% CI: 1, 7%) per $4.4{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$] (Turner et al. 2016).

A small but growing body of literature has examined associations between ambient air pollution and nonlung cancer risk; however, the evidence for such associations is limited. For bladder cancer, a Spanish hospital-based case–control study including 1,219 incident cases and 1,271 controls reported that living $>40\text{y}$ in a city of $>\mathrm{100,000}$ inhabitants was associated with a significantly higher risk of the disease [$\text{odds}\text{ratio}\left(\text{OR}\right)=1.30$ (95% CI: 1.04, 1.63)] (Castaño-Vinyals et al. 2008). In a death certificate–based case–control study in Taiwan, including 680 bladder cancer deaths and 680 matched noncancer or nongenitourinary death controls, there were significant positive trends with increasing tertiles of particulate matter with aerodynamic diameter $<10\mathrm{\mu }\mathrm{m}$ (${\text{PM}}_{10}$), ${\text{NO}}_2$, and sulfur dioxide (${\text{SO}}_2$), but not with carbon monoxide (CO) or ${\mathrm{O}}_3$ concentrations (Liu et al. 2009). There was a positive, nonsignificant association between nitrogen oxides (${\text{NO}}_{\mathrm{x}}$) concentrations and bladder cancer incidence ($n=221$) [incidence rate ratio (IRR) per $100{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3=1.32$ (95% CI: 0.80, 2.19)] in an analysis of 54,304 participants in the Danish Diet Cancer and Health cohort (Raaschou-Nielsen et al. 2011). In a retrospective cohort study of 85,559 individuals in Malagrotta (Rome, Italy), there was a positive association between hydrogen sulfide (${\mathrm{H}}_2\mathrm{S}$) concentrations from a municipal waste landfill and bladder cancer mortality in women [HR per $0.043{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3=1.35$ (95% CI: 1.00, 1.82), $n=12$], but not in men [$\text{HR}=0.88$ (95% CI: 0.51, 1.52), $n=61$] (Ancona et al. 2015). Taken together, the results from our study and those from most previous studies are generally consistent with the recent IARC evaluation noting a positive association in studies of outdoor air pollution and bladder cancer risk (IARC 2013).

There was no association [HR per $10{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3=0.98$ (95% CI: 0.58, 1.64)] between ${\text{PM}}_{2.5}$ and urinary cancer mortality (including 155 bladder and kidney cancer deaths combined) in a cohort of 66,820 elderly Hong Kong residents (Wong et al. 2016). The Danish Diet Cancer and Health cohort reported a positive, although imprecise, association of ${\text{NO}}_{\mathrm{x}}$ with kidney cancer incidence [$\text{IRR}=1.73$ (95% CI: 0.89, 3.73) per $100{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ (Raaschou-Nielsen et al. 2011). A total of 95 kidney cancer cases were observed in that study.

Few studies of colorectal cancer have been reported to date. Although our study showed a significant positive association between ${\text{NO}}_2$ and death from colorectal cancer and a borderline association with ${\text{PM}}_{2.5}$, a Hong Kong cohort reported no association [HR per $10{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3=1.01$ (95% CI: 0.79, 1.30)] between ${\text{PM}}_{2.5}$ concentrations and mortality from cancers of the lower digestive tract ($n=719$) (Wong et al. 2016). There were nonsignificant inverse associations in men and women for ${\text{PM}}_{10}$ concentrations from a waste incinerator and colorectal cancer mortality ($n=149$ deaths in men and 95 deaths in women) in Malagrotta (Ancona et al. 2015). There were also nonsignificant inverse associations for ${\text{NO}}_{\mathrm{x}}$ concentrations and both colon and rectal cancer incidence in Denmark ($n=414$ and 246 cases, respectively) (Raaschou-Nielsen et al. 2011). The number of included colorectal cancer deaths ($>\mathrm{6,000}$) in CPS-II here is substantially larger than the numbers of cases/deaths in previous studies.

When ${\text{PM}}_{2.5}$ concentrations were decomposed into near-source and regional components, there were stronger associations for near-source ${\text{PM}}_{2.5}$ with colorectal (and stomach) cancer mortality, supporting similar findings for ${\text{NO}}_2$, and for regional ${\text{PM}}_{2.5}$ with both bladder and kidney cancer mortality. Little is known regarding the role of different air pollution sources or components in cancer, including at sites other than the lung. We have not specifically examined what ${\text{PM}}_{2.5}$ sources were further related to this decomposition.

Studies of occupational exposure to diesel engine exhaust and polycyclic aromatic hydrocarbons (PAHs), which are formed during incomplete combustion processes, have suggested positive associations with both bladder and kidney cancer (Brown et al. 2012; IARC 2005; Siemiatycki et al. 2004). Recent population-based case–control studies of men in Canada reported positive associations between any exposure to high concentrations of occupational diesel exhaust and both bladder [$\text{OR}=1.64$ (95% CI: 0.87, 3.08), $n=658$ cases and 1,360 controls] and rectal [$\text{OR}=1.98$ (95% CI: 1.09, 3.60), $n=840$ cases and 1,360 controls] cancer incidence, which increased among those with $>10\text{years}$ of exposure (Kachuri et al. 2016; Latifovic et al. 2015). Studies of workers in dusty occupations (e.g., mineral dust, wood dust), as well as in steel and iron processing, have also noted positive associations with colorectal and stomach cancer (Kreuzer et al. 2012; Oddone et al. 2014; Raj et al. 2003; Santibañez et al. 2012).

The largest HR was for ${\text{PM}}_{2.5}$ in relation to death from cervical cancer, although the findings were based on only 115 deaths and were not statistically significant. The positive association with ${\text{NO}}_2$ was weaker. In the Danish Diet Cancer and Health study, the strongest association was between ${\text{NO}}_{\mathrm{x}}$ and cervical cancer incidence [i.e., $\text{IRR}=2.45$ (95% CI: 1.01, 5.93) per $100{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$], but it was based on only 35 cases (Raaschou-Nielsen et al. 2011). In both studies, there were no data on human papillomavirus (HPV) infection (or on other potentially relevant infections for cancer mortality at other sites, e.g., Helicobacter pylori for stomach cancer or hepatitis B or C for liver cancer), and confounding cannot be ruled out. There was also no information on access to or compliance with cervical cancer screening programs, which may differ in areas with differing levels of ambient air pollution. However, we adjusted our models for a range of area-level socioecomonic covariates and screening rates in a subset of CPS-II participants for both breast ($>90\%$) and colorectal ($>65\%$) cancer were high (Patel et al. 2003; Stevens et al. 2011). A recent cross-sectional study of women from a clinical trial of cervical disease diagnostic techniques in Texas reported a positive association of cervical dysplasia with residential census-tract level estimates of ambient benzene, diesel particulate matter, and PAH concentrations (Scheurer et al. 2014).

Although we found no significant associations with death from breast cancer based on fully adjusted models, other studies reported positive associations between ambient air pollution and breast cancer incidence (Crouse et al. 2010; Mordukhovich et al. 2016; Reding et al. 2015). Crouse et al. (2010) reported a positive association between ${\text{NO}}_2$ concentrations and postmenopausal breast cancer incidence [OR per $5\text{ppb}=1.31$ (95% CI: 1.00, 1.71)] in a hospital-based case–control study including 383 case and 416 control participants. The prospective U.S. Sister Study reported no association between ${\text{PM}}_{10}$, ${\text{PM}}_{2.5}$, or ${\text{NO}}_2$ concentrations and incident breast cancer overall in an analysis of 47,591 participants including 1,749 breast cancer cases, but a positive association was reported between ${\text{NO}}_2$ and $\text{estrogen}\text{receptor}\left(\text{ER}\right)+/\text{progesterone}\text{receptor}\left(\text{PR}\right)+\text{disease}$ [HR per $5.8\text{ppb}=1.10$ (95% CI: 1.02, 1.19)] (Reding et al. 2015). In our mortality-based study, no information on ER or PR status was available.

There was no positive association of any ambient air pollutant and total leukemia mortality in the present study. There was a positive association of ${\text{NO}}_2$ concentrations at the residence with incident adult acute myeloid leukemia [OR per $10{\mathrm{mg}/\mathrm{m}}^3=1.31$ (95% CI: 1.02, 1.65), $n=531$] but not with other leukemia subtypes in a Danish case-control study (Raaschou-Nielsen et al. 2016). There was no positive association of ${\text{PM}}_{2.5}$ with total incident leukemia or with chronic lymphocytic leukemia in a Canadian case–control study of 1,064 total case and 5,039 control participants (Winters et al. 2015). There was no information on leukemia subtype in CPS-II.

There was no association of any ambient air pollutant with total brain cancer mortality in this analysis, which included 1,591 brain cancer deaths. The Danish Diet Cancer and Health cohort reported a positive association of ${\text{NO}}_{\mathrm{x}}$ with incident brain tumor risk [IRR per $100{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3{\text{NO}}_{\mathrm{x}}=2.28$ (95% CI: 1.25, 4.19)] (Raaschou-Nielsen et al. 2011), although there were few brain tumor cases ($n=95$). A subsequent national case–control study in Denmark with a larger number of incident brain tumor cases ($n=\mathrm{4,183}$) reported a significant positive association of ${\text{NO}}_{\mathrm{x}}$ concentrations $\ge 100{\mathrm{\mu }\mathrm{g}/\mathrm{m}}^3$ with nongliomas [$\text{OR}=2.30$ (95% CI: 1.15, 4.59)] but not with gliomas [$\text{OR}=0.89$ (95% CI: 0.44, 1.77)] (Poulsen et al. 2016). There was no information on tumor histology or cranial location in CPS-II.

In the present study, there were no clear positive, and some cases there were significant inverse, associations observed with ${\mathrm{O}}_3$ in both single- and multipollutant models possibly owing to broader spatial patterns in ambient air pollution concentrations, to negative correlations with other air pollutants, and to the larger spatial scale of ${\mathrm{O}}_3$ concentrations, which were unable to capture fine-scale variation and scavenging effects in urban areas (Williams et al. 2014). ${\mathrm{O}}_3$ is thought to increase DNA damage (U.S. EPA 2013), which plays a role in several types of cancer. However, in previous studies of the CPS-II cohort, ${\mathrm{O}}_3$ was not associated with lung cancer death (Jerrett et al. 2013; Krewski et al. 2009; Pope et al. 2002; Turner et al. 2016). A positive association between ambient ${\mathrm{O}}_3$ and incident male lung cancer was reported in the Adventist Health Study on Smog (AHSMOG), although few lung cancers were observed ($n=16$) (Beeson et al. 1998).

Strengths of this study include a large-scale, well-established cohort design with large numbers of nonlung cancer deaths observed at many cancer sites. Air pollution exposures were estimated at each participant’s residence using national-level exposure surfaces that have previously been used to examine mortality associations in CPS-II (Jerrett et al. 2016; Pope et al. 2015; Turner et al. 2014; 2016). Detailed data were collected at enrollment on a variety of cancer risk factors including cigarette smoking, occupation, diet, and various hormonal and reproductive factors.

The main limitation of this study is the use of cancer mortality rather than cancer incidence end points, with inferences of associations of ambient air pollution here reflecting both disease incidence and survival following diagnosis. Because lung cancer is rapidly fatal, with 5-y survival rates ranging from $\sim 13–18\%$ for the periods 1987–1989 and 2005–2011 respectively, the use of mortality data reasonably approximates disease incidence (American Cancer Society 2016). Other rapidly fatal cancer sites include the pancreas ($5-\text{year}\text{survival}=4–8\%$), the liver and intrahepatic bile duct (5–18%), the esophagus (10–20%), the stomach (20–30%), and the brain and other nervous system sites (29–35%) (American Cancer Society 2016). In contrast, survival is greater for cancer at other sites, including the urinary bladder, the kidney and renal pelvis, and the colorectum, of interest here (5-y survival ranging from 57–79%), with survival from disease playing an increasing role in associations with ambient air pollution observed here (American Cancer Society 2016).

There is little research on whether ambient air pollution may be related to cancer progression or survival. One recent study of $>\mathrm{350,000}$ California lung cancer patients reported that higher residential ambient air pollution concentrations (${\text{NO}}_2$, ${\text{PM}}_{2.5}$, ${\text{PM}}_{10}$) were associated with poorer survival, particularly among patients diagnosed in earlier disease states (i.e., with localized disease) (Eckel et al. 2016). Mean ambient air pollution concentrations were also somewhat higher among those diagnosed with more advanced disease and among those with an unknown stage at diagnosis (i.e., either patients who were dying before stage information was obtained or who had limited workup performed), possibly reflecting differences in access to medical care, which may vary by levels of ambient air pollution. Little is known regarding possible impacts at other cancer sites, although reduced breast cancer survival was associated with higher ${\text{PM}}_{2.5}$ and ${\text{PM}}_{10}$ concentrations in another study (Hu et al. 2013). Further studies of nonlung cancer incidence are needed to disentangle the observed associations with ambient air pollution.

Covariate data were only available at enrollment and were not updated over the follow-up period in CPS-II. It is unlikely that CPS-II participants would begin smoking cigarettes over the follow-up time, given the mean age of $>55\text{years}$ at enrollment, although participants may increasingly become former smokers. There were also limited data on occupational exposure history. Participant residence data were also only available at enrollment. Changes in participant residence after enrollment as well as changes in coding of the underlying cause of death over the follow-up period would likely be nondifferential and would result in attenuation of the magnitude of the associations observed. A Canadian study observed little impact of accounting for residential mobility on ${\text{PM}}_{2.5}$ or ${\mathrm{O}}_3$ mortality associations; however, associations with more spatially resolved ${\text{NO}}_2$ strengthened somewhat (Crouse et al. 2015).

There is also a lack of historical ambient air pollution data, although correlations between ${\text{PM}}_{2.5}$ and ${\mathrm{O}}_3$ concentrations assigned to CPS-II participants over recent decades were moderately strong, ranging from approximately 0.6 to $>0.8$, indicating that the use of more recent ambient air pollution estimates may be reflective of longer-term exposure patterns (Krewski et al. 2009; Pope et al. 2002; Turner et al. 2016). The rank ordering of U.S. cities was also similar over time in the context of generally declining ambient air pollution concentrations. The use of recent ambient air pollution estimates may result in somewhat inflated HRs because increments of recent concentrations represent greater contrasts of historically higher concentrations (Krewski et al. 2009; Pope et al. 2002; Turner et al. 2016). Little is known regarding potential latency periods for cancer development in relation with ambient air pollution concentrations, which may also differ from those of lung cancer.

Correlations among pollutants were generally weak. Ambient air pollution concentrations were estimated using different approaches at different time periods and different geographic units of scale, possibly complicating interpretation of the correlation structure among pollutants. The LURBME model outperformed a range of other geostatistical and remote sensing ${\text{PM}}_{2.5}$ models in the CPS-II (Jerrett et al. 2016). Similar positive cardiorespiratory mortality–${\mathrm{O}}_3$ associations were observed in recent work with ${\mathrm{O}}_3$ concentrations estimated at either $12\mathrm{km}\times 12\mathrm{km}$ or $36\mathrm{km}\times 36\mathrm{km}$ scales in the Eastern United States (Turner et al. 2016).

Owing to multiple testing and to the large number of cancer sites evaluated ($n=29$), it is possible that some of the significant associations observed may be due to chance. As such, the results of this study should be replicated in other studies, particularly in studies of cancer incidence. Finally, our findings may not be entirely generalizable because CPS-II participants are of generally higher socioeconomic status and more limited racial/ethnic composition than the broader U.S. population.

Potential mechanisms through which ambient air pollution may be associated with other nonlung cancers remain to be fully elucidated; however, ambient air pollution represents a complex mixture of exposure to a broad range of carcinogenic and mutagenic substances, including PAHs and other aromatic hydrocarbons, benzene, metals, and xenoestrogens, which may be transported and metabolized in the body (Crouse et al. 2010; IARC 2012, 2013). Populations exposed to outdoor air pollutants and to diesel engine exhaust have elevated urinary 1-hydroxypyrene and hemoglobin adducts of nitro-PAHs and low-molecular-weight alkenes (Ciarrocca et al. 2014; Duan et al. 2016; IARC 2012, 2013). A recent study of 23,820 participants in Taiwan, including 464 incident hepatocellular carcinoma (HCC) cases, reported a positive association of ${\text{PM}}_{2.5}$ with HCC incidence, which may have been mediated by alanine transaminase levels, suggesting that ${\text{PM}}_{2.5}$ exposure may lead to HCC via chronic inflammation (Pan et al. 2015).

Active cigarette smoking has also been associated with a range of other nonlung cancers, including cancers in various urinary and digestive sites such as the bladder, the kidney, and the colorectum, among others (Carter et al. 2015; IARC 2009). However, levels of exposure to ${\text{PM}}_{2.5}$ from ambient air pollution are substantially lower and of differing chemical composition and toxicity compared with those from active cigarette smoking or occupational exposure (Pope et al. 2011).

Patients with a history of diabetes, which has also been linked with ambient air pollution (Eze et al. 2015; Pope et al. 2015), have also been observed to be at increased risk for bladder, kidney, and colorectal cancers, although it is unclear whether this is due to shared underlying risk factors (such as obesity) or to other metabolic features of the disease (such as hyperinsulinemia, hyperglycemia, or chronic inflammation) (Campbell et al. 2012; Giovannucci et al. 2010). The results herein were virtually unchanged with further adjustment for prevalent diabetes at enrollment (not shown).

Conclusion

The results from this large prospective study suggest that ambient air pollution was not associated with most nonlung cancer causes of death. Nonetheless, observed associations with mortality from kidney, bladder, and colorectal cancer merit further research, particularly in studies of cancer incidence.