Associations of fine particulate matter pollution components with overall, prostate cancer, and cardiovascular disease mortality in men with prostate cancer: A cohort study

Abstract

Background

Over 3.5 million US men are living with prostate cancer (Pica), many with underlying cardiovascular disease (CVD). Fine particulate matter (PM_2.5) contributes to higher CVD mortality through inflammation and other mechanisms, and so may increase non-cancer mortality in men with PCa.

Methods

We conducted a retrospective cohort study of 886,876 men diagnosed with PCa between 2000 to 2015 and followed through 2018 across eight state cancer registries. Annual average predictions of five residential PM_2.5 component exposures (elemental carbon (EC), organic carbon (OC), nitrate, ammonium (NH₄⁺), sulfate (SO₄^2-)) were obtained from an ensemble-based machine learning model and assigned to geomasked addresses at diagnosis within 50 meters (urban) or 1 kilometer (rural). Adjusted hazard ratios (aHR) for associations of components separately and as a mixture with all-cause, PCa, and CVD mortality were estimated from covariate adjusted Cox models.

Results

There were 233,898 deaths over 5,836,741 person-years. Per interquartile range (IQR) increase, OC (aHR 1.03 [95% CI: 1.02–1.04]), NH₄⁺ (aHR 1.02 [1.01–1.03]), and SO₄^2- (aHR 1.11 [1.09–1.13]) were associated with all-cause mortality. CVD mortality was associated with higher EC (aHR: 1.04 [1.02–1.06]), OC (aHR 1.05 [1.03, 1.07]), NH₄⁺ (aHR 1.09 [1.06–1.12]) and SO₄^2- (aHR 1.19 [1.15–1.24]). There were no associations with nitrates or PCa mortality. Per IQR, PM_2.5 components mixtures were associated with higher all-cause (aHR 1.03 [1.02–1.04]), PCa (aHR 1.02 [0.99, 1.05]), and CVD mortality (aHR 1.08 [1.05, 1.11]).

Discussion

Certain PM_2.5 components were associated with higher all-cause and CVD mortality in men with PCa. Studies of air pollution in cancer survivors should consider impacts on non-cancer mortality.

Background

Prostate cancer (PCa) is a major contributor to cancer-related morbidity and mortality in US men, accounting for an estimated 35,250 cancer-related deaths each year¹. There are an estimated 3.5 million men are living with PCa (42% of all male cancer survivors)². Men with PCa experience excellent survival, with 15-year survival over 97%³. Therefore, similar to survivors of other cancers, men with PCa are often concerned about how environmental exposures may influence their prognosis and overall health, with an estimated 22% attributed their cancer diagnosis to environmental risk factors^4,5. In addition, many men with PCa are diagnosed with pre-existing cardiovascular disease (CVD)⁶, and CVD is a major competing risk for mortality, accounting for 23% of all deaths in PCa survivors^7,8. In addition, androgen deprivation therapy treatment for PCa is associated with higher risks of CVD⁹ and higher adiposity¹⁰ when used over long periods of time. These factors can contribute to increased risk of CVD death particularly among men diagnosed at older ages^11,12. In light of this, identifying environmental factors that may exacerbate CVD-related and PCa morbidity and mortality is a growing priority for PCa care and management⁶, and could inform policies to mitigate cancer-related harms associated with pollution¹³.

Fine particulate matter (PM_2.5) has been designated a causal risk factor for certain cancers¹⁴ and for CVD and death^15,16, and may exacerbate these risks in PCa survivors¹⁵. Mechanisms through which PM_2.5 may affect CVD include inflammation^17,18 and oxidative stress, which may contribute to atherosclerotic plaque formation¹⁹. Residents of neighborhoods with high levels of poverty or high proportion of non-White residents often experience a higher burden of air pollution²⁰ and may be more susceptible to its adverse effects due to compounding risks^20,21. Sources of PM_2.5 vary in different parts of the US and only a few studies have evaluated the effects of specific PM_2.5 components, which may help determine which sources are most strongly associated with higher CVD mortality in men with PCa^22,23. It is unknown whether men with PCa are more susceptible to risks of air pollution relative to the general population, but their older age, prevalence of comorbidities²⁴, and CVD risks associated with hormone therapy⁹ may increase their risks. Studies that examine associations of air pollution components, cause-specific mortality, and PCa survivorship in large, population-based databases could inform management of patients with CVD comorbidities and guide policies to reduce harmful pollution exposure^22,23.

We conducted a cohort study within the Multilevel Epidemiology Tumor Registry for Oncology (METRO) database²⁵, which has been used to investigate environmental and health care-access related drivers of survival disparities in men with PCa. We sought to characterize the association between PM_2.5 components with all-cause, PCa-specific, and CVD-specific mortality, evaluate effect modification by race and ethnicity, neighborhood socioeconomic status (nSES), and tumor stage, and estimate the impact of PM_2.5 component mixture on survival. We hypothesized that men residing in neighborhoods with higher PM_2.5 component exposure would experience higher all-cause and cause-specific mortality.

Methods

Study design and participants

We used data from METRO, a multistate registry cohort to examine environmental and health care access-related factors associated with PCa outcomes²⁵. The cohort included men diagnosed with PCa between 2000–2015, aged ≥40 years and followed until death, censoring on December 31, 2018, or 10 years, whichever came first. State cancer registries collected demographic, clinical, and treatment information from patients diagnosed within the state and applied quality control and completeness checks²⁶. For this study, we used data from the California, Detroit/Michigan, Massachusetts, New Mexico, New Jersey, Ohio, Pennsylvania, Utah and Seattle/Puget Sound/Washington population-based cancer registries. These states were selected to provide geographic and sociodemographic variation in cohort characteristics and PM_2.5 exposure. To comply with the required confidentiality procedures, we applied geomasking procedures to randomly displace geocoded addresses at diagnosis before appending exposures by 250m in urban areas and 400m in rural areas²⁷. This article adhered to STrengthening the Reporting of OBservational studies in Epidemiology (STROBE) guidelines for cohort studies.

Of 916,629 individuals meeting age criteria, we excluded those missing air pollution linkages (n=11,732), men without Census tract socioeconomic status data (n=449), and men who were diagnosed on autopsy (n=17,570). The institutional review boards of the Dana-Farber Cancer Institute and Rutgers University, The State University, approved this study and determined that, because existing data sources were used, no written consent was required for participation in the study. Access to cancer registry data additionally required state institutional review board approval from California, Massachusetts, New Jersey, New Mexico and Utah.

Fine particulate matter components

Outdoor PM_2.5 exposure was assessed from a novel ensemble-based model using machine learning, geographically-weighted averaging and super-learning model developed by Amini and colleagues that produced hyperlocal predictions of PM components²⁸. This model characterized elemental carbon (EC) and organic carbon (OC), primarily generated from tailpipe emissions and wood or other natural burning; ammonium (NH₄⁺), sulfates (SO₄^2-) and nitrates (NO₃⁻) which are secondary byproducts of coal from power plants²⁹. Some OC is also formed as secondary particles.

Briefly, models were trained using pollution monitoring data from 987 unique locations with 230 initial predictors (e.g. time and location, remote sensing data, land cover, meteorology). Predictors were filtered based on collinearity (r ≥0.99) yielding 166 for training. Input data sources included component-specific monitoring data from the US EPA, state and national parks, and research studies^30–35. A 70% training and 30% test set were used, and different machine learning approaches were applied, including gradient bosting, random forest, and eXtreme Gradient Boosting (XGB). These predictions were then ensembled using an additional learner. The test set was used to reduce overfit of the prediction model derived in the training phase. Prediction model accuracy was assessed from the R² estimated from a linear regression with the PM_2.5 component monitoring value as the outcome and the predicted value as the independent variable using the test dataset. Test R² for components ranged from 0.86–0.96; for overall annual PM_2.5 it was 0.89. Further details regarding the modeling approach are available elsewhere^28,36–38.

Predicted mean PM_2.5 components (EC, OC, NH₄⁺, NO₃⁻, and SO₄^2-) were available at 50m in urban areas, and 1km in non-urban areas, from 2000–2019²⁸. Annual average exposure to EC, OC, NH₄⁺, NO₃⁻, SO₄^2-, was assigned to participants based on the grid cell of their residential address for their year of diagnosis. We assumed this to reflect residential PM_2.5 component exposure during the post-diagnosis period, assuming limited residential mobility. These exposure models have been used to examine associations of particulate matter pollution and risks of cardiovascular disease an hospitalizations in other administrative cohorts^39–42.

Outcomes

Mortality was ascertained by registries through linkages with state and national death records²⁵. Cause-specific mortality was assessed from records using ICD-09 and ICD-10 codes for the primary cause of death and categorized for PCa and CVD mortality. Surveillance Epidemiology and End Results algorithms were applied to address potential outcome misclassification⁴³. New Mexico cases were excluded from the analysis of CVD mortality because the state government did not authorize release of these data.

Covariates

We selected covariates based on expert knowledge drawn from prior studies of air pollution, green space and mortality^22,44–46, and use of VanderWeele’s “disjunctive cause criterion” which states one should control for covariates that are causes of the exposure, or outcome, or both, excluding instrumental variables⁴⁷. These approaches were supplemented by use of Directed Acyclic Graphs which ensured that adjustment sets were appropriate and did not lead to overadjustment or collider stratification bias⁴⁸. Directed Acyclic Graphs to guide covariate selection for analyses of PM components and all-cause mortality (Supplementary Figure 1), PCa mortality (Supplementary Figure 2), and CVD mortality (Supplementary Figure 3) depict the assumed causal relationships between sociodemographic factors, neighborhood context, unmeasured behavioral and occupational factors, and health care access. Although insurance and tumor stage are not causes of PM component exposure, they may mediate the relationship between unmeasured health behaviors and occupational factors that are true confounders of the effect of PM components on mortality and so we chose to adjust for them. We assumed that neighborhood contextual factors represent pre-diagnostic conditions over the same period as the PM components, rather than as mediators of the PM components to mortality relationship.

We considered neighborhood greenness as a confounder because it has been previously shown to be associated with mortality in men with PCa^49–51, and CVD mortality^52,53, and was associated with lower all-cause, PCa, and CVD mortality in METRO (Supplementary Table 1). Moreover leafy vegetation, particularly urban trees, may reduce air pollution through sequestration through leaves⁵⁴.

Registry data contributed to METRO included demographic measures (race and ethnicity abstracted by registrars from electronic medical records, age at diagnosis, diagnosis year), tumor characteristics, and socioeconomic measures (marital status, insurance). The Surveillance, Epidemiology and End Results (SEER) program classifies prostate tumors using summary stage, based on whether the tumor is restricted to the organ or origin (localized), or has metastasized to proximate organs or lymph nodes (regional), or further to distant lymph nodes or bone (distant)⁵⁵. We therefore categorized SEER summary as localized vs regional/distant as a marker of tumor severity. We adjusted for neighborhood socioeconomic status (quintiles) using Census tract-level z-score based measures of income, demographics, occupation, and educational attainment⁵⁶, outdoor nitrogen dioxide exposure within 1km of residence from an existing ensemble-based model developed by Di et al.^36,57, Census tract population density (<1000 people/mi², ≥1000 people/mi²) and neighborhood greenspace (quintiles) using Landsat satellite-derived Normalized Difference Vegetation Index (NDVI) within 1230m of address based on earlier studies⁴⁹, and state of residence to account for area-level confounding. For models with PCa-specific mortality, we further adjusted for receipt of definitive first course of treatment (surgery or radiation). All geospatial measures were aligned with calendar year of diagnosis, or decennial year for Census data.

Statistical analysis

For each component (EC, OC, NH₄⁺, NO₃⁻, and SO₄^2-), we fit Cox proportional hazards models with study follow up in months as the time scale for associations of PM_2.5 components with all-cause, PCa, and CVD mortality using the coxph function of the R survival package⁵⁸. We fit minimally adjusted models with just age and diagnosis year, and fully adjusted models including all covariates. We assessed the exposure-response relationship using linear terms scaled to the interquartile range (IQR), quintiles and a test for ordinal trend, and restricted cubic splines. Because we found that the association between the PM_2.5 components mixture and mortality was linear, we chose to present primary results assuming linear relationship between components and mortality. The linear model assumed a constant increase in risk associated with an incremental change in exposure, while in contrast, the models parameterizing exposure as quintiles assumed a constant risk within quintiles and change as a step function, making them less sensitive to outliers. Models using restricted cubic splines were fit with 3 degrees of freedom⁵⁹ for parsimonious modeling and because additional degrees of freedom did not improve model fit based on Akaike’s Information Criterion or likelihood ratio tests⁶⁰. Splines were fit using the pspline function in the R survival package⁵⁸. We evaluated effect modification by neighborhood socioeconomic status (tertiles), race and ethnicity, stage at diagnosis, population density, and NDVI (tertiles) using multiplicative interaction terms between the modifier and a continuous IQR increase in PM_2.5 component exposure. Likelihood ratio tests were used to compare nested models and evaluate statistical significance of non-linear associations and interaction terms.

We used multiple imputation with chained equations to model the exposure-outcome associations that would have been observed had no covariate data been missing, assuming that the available covariate data are sufficient to explain patterns of missingness (Rubin’s “Missing at Random” assumption)⁶¹. Multiple imputation with chained equations leverages all available covariate information to estimate predicted values for missing covariates, and because simpler approaches (complete case, mean) can be biased⁶². Further details have been reviewed elsewhere⁶³. We used predictive mean matching for continuous variables and linear discriminant analysis for discrete variables. Covariates for the imputation model were the same as used in the main analysis models. Five imputed datasets were created and Rubin’s Rules were applied to calculate standard errors⁶¹. The multiple imputation procedure was implemented using the R mice package⁶⁴.

Because men with PCa often die from causes other than PCa, we wished to analyze associations of PM components with PCa mortality and CVD mortality accounting for the presence of these competing risks⁶⁵. We fit stabilized inverse probability of censoring weights (IPCW) to account for competing risks of other causes in the PCa mortality and CVD mortality analyses⁶⁶. Applying IPCW for the competing risks of death allowed us to estimate the “direct effect”, or association of PM components with the rate of cause-specific mortality that does not flow through other competing risks⁵⁵. This is referred to as the “marginal hazard” in the statistical literature⁵⁵. IPCW were calculated using logistic regression models using the glm function in the R stats package⁶⁷ for deaths of other causes, conditional on covariates. Further details regarding this methodology are available elsewhere⁶⁸.

In addition to assessing PM_2.5 components individually, we sought to analyze them jointly as a single “mixture”. For this purpose, we used quantile g-computation⁶⁹, a method that applies causal inference based modeling, to assess PM_2.5 components as a mixture for each quantile increase across all components of the mixture (here, the five PM_2.5 components). We specified five quintiles, and because we found no evidence of non-linearity for the PM_2.5 components mixture effect, we assumed additive, non-linear effects of the component and presented results for the combined effect of all five components on all-cause mortality, PCa mortality, and CVD mortality. We specified the same Cox models for all-cause, PCa and CVD mortality as described previously for the outcome models used for quantile g-computation. These methods have been used elsewhere to examine associations of pollutant components and PCa risk⁷⁰. Quantile g-computation was implemented with the R qgcomp package⁷¹.

Unmeasured confounding poses a threat to validity because our registry data did not capture information on important behavioral and occupational factors (Supplementary Figures 1–3). Sensitivity analyses based on statistical relationships between exposures, outcomes, and unmeasured confounders provide quantitative information to assess whether unmeasured confounding could explain away estimated exposure-outcome associations⁷². We selected and calculated E-Values to assess sensitivity of findings to unmeasured confounding by behavioral risk factors (mainly smoking)⁷³ due to the widespread use of this method in the epidemiologic literature^74,75, and because unlike other forms of sensitivity analysis which require the investigator to provide estimates of strength of association between the confounder, exposure, and outcome which may be subjective, the E-Value simply requires the point estimate and confidence interval. The E-Value captures the minimum association between an unmeasured confounder and either the exposure (PM_2.5 component) or outcome (all-cause, PCa, CVD mortality) that would be required to attenuate the point estimate to 1, or to shift the confidence interval to include 1. We calculated E-Values using formulae provided by VanderWeele and Ding⁷³, and used the formula for HR for common outcomes when estimating E-Values treated for all-cause mortality due to >15% cumulative incidence over follow-up. Based on previously reported associations of smoking with air pollution and mortality in PCa survivors⁷⁶, E-Values of 1.5 or higher for the lower confidence interval bound would indicate that the reported association is robust to confounding by smoking or similar factors. All statistical analyses were performed using R version 4.4.2⁶⁷.

Role of the funding source

The funder of the study had no role in study design, data collection, data interpretation or writing of the report.

Results

Cohort characteristics

There were 886,878 participants in METRO meeting eligibility criteria, with mean age (SD) 67.1 (9.6) years (Table 1). Cohort characteristics were presented stratified by EC as an illustrative example of the frequencies and other summary statistics for covariates across levels of PM components. Of these, 76% were Non-Hispanic White, 11% were Non-Hispanic Black, and 13% were Hispanic or other race. Most participants (65%) were diagnosed with localized disease and had private (30%) or Medicare (28%) insurance. Participants in the highest compared to lowest quintile of EC were less likely to be Non-Hispanic White (57% vs 90%), reside in the most deprived quintile of nSES (17% vs 31%), and more likely to reside in neighborhoods with lower mean (SD) NDVI (0.22 (0.10) vs 0.39 (0.12)). Correlations between PM_2.5 components ranged from −0.08 (OC with SO₄) to 0.86 (NH₄ with SO₄) (Supplementary Figure 4). The range of values for each PM component within each quintile is provided in Supplementary Table 2.

Table 1.

Characteristics of men diagnosed with prostate cancer in the Multilevel Tumor Registry for Oncology (METRO) cohort stratified by Elemental Carbon (n=886,876)

	Quintiles of Elemental Carbon (EC)
Characteristic	Q1	Q2	Q3	Q4	Q5	Total
N	177376	177375	177375	177375	177375	886876
Patient characteristics
Age at diagnosis (mean (SD)	66.88 (9.24)	66.86 (9.50)	67.00 (9.64)	67.24 (9.74)	67.49 (9.73)	67.09 (9.58)
Race (%)
Non-Hispanic White	159659 (90.0)	147537 (83.2)	136935 (77.2)	128710 (72.6)	100698 (56.8)	673539 (75.9)
Non-Hispanic Black	5445 (3.1)	13015 (7.3)	20789 (11.7)	25880 (14.6)	31852 (18.0)	96981 (10.9)
Other	12272 (6.9)	16823 (9.5)	19651 (11.1)	22785 (12.8)	44825 (25.3)	116356 (13.1)
Insurance (%)
Private	59202 (33.4)	57903 (32.6)	55337 (31.2)	50524 (28.5)	44961 (25.3)	267927 (30.2)
Uninsured	4367 (2.5)	6126 (3.5)	8663 (4.9)	9712 (5.5)	10813 (6.1)	39681 (4.5)
Medicaid	3066 (1.7)	3249 (1.8)	3354 (1.9)	3382 (1.9)	5685 (3.2)	18736 (2.1)
Medicare	55435 (31.3)	52529 (29.6)	48857 (27.5)	42695 (24.1)	34139 (19.2)	233655 (26.3)
Other Government	1042 (0.6)	1068 (0.6)	1367 (0.8)	1296 (0.7)	794 (0.4)	5567 (0.6)
Missing	54264 (30.6)	56500 (31.9)	59797 (33.7)	69766 (39.3)	80983 (45.7)	321310 (36.2)
Year of diagnosis (%)
2000–2004	41130 (23.2)	47039 (26.5)	53833 (30.3)	67257 (37.9)	80075 (45.1)	289334 (32.6)
2005–2009	50794 (28.6)	51359 (29.0)	55123 (31.1)	70977 (40.0)	69015 (38.9)	297268 (33.5)
2010–2015	85452 (48.2)	78977 (44.5)	68419 (38.6)	39141 (22.1)	28285 (15.9)	300274 (33.9)
Marital status (%)
Not married	33991 (19.2)	39043 (22.0)	46569 (26.3)	44927 (25.3)	47409 (26.7)	211939 (23.9)
Married	84383 (47.6)	101598 (57.3)	105489 (59.5)	110470 (62.3)	109229 (61.6)	511169 (57.6)
Missing	59002 (33.3)	36734 (20.7)	25317 (14.3)	21978 (12.4)	20737 (11.7)	163768 (18.5)
Neighborhood socioeconomic status (%)
Quintile 1	55189 (31.1)	32797 (18.5)	28852 (16.3)	29950 (16.9)	29945 (16.9)	176733 (19.9)
Quintile 2	33648 (19.0)	35728 (20.1)	34563 (19.5)	34273 (19.3)	38892 (21.9)	177104 (20.0)
Quintile 3	31168 (17.6)	35068 (19.8)	34998 (19.7)	35305 (19.9)	40711 (23.0)	177250 (20.0)
Quintile 4	29906 (16.9)	34148 (19.3)	37463 (21.1)	38447 (21.7)	37641 (21.2)	177605 (20.0)
Quintile 5	27465 (15.5)	39634 (22.3)	41499 (23.4)	39400 (22.2)	30186 (17.0)	178184 (20.1)
SEER Summary Stage (%)
Localized	125140 (70.6)	116573 (65.7)	112914 (63.7)	112554 (63.5)	112452 (63.4)	579633 (65.4)
Regional	18185 (10.3)	15058 (8.5)	14279 (8.1)	14359 (8.1)	16359 (9.2)	78240 (8.8)
Distant	6629 (3.7)	6168 (3.5)	5970 (3.4)	5705 (3.2)	6942 (3.9)	31414 (3.5)
Missing	27422 (15.5)	39576 (22.3)	44212 (24.9)	44757 (25.2)	41622 (23.5)	197589 (22.3)
State (%)
California	47417 (26.7)	49967 (28.2)	52676 (29.7)	60050 (33.9)	103002 (58.1)	313112 (35.3)
Massachusetts	42648 (24.0)	17042 (9.6)	4301 (2.4)	2177 (1.2)	528 (0.3)	66696 (7.5)
Michigan (Detroit)	3690 (2.1)	13507 (7.6)	18312 (10.3)	14029 (7.9)	5296 (3.0)	54834 (6.2)
New Jersey	8074 (4.6)	18169 (10.2)	24274 (13.7)	26774 (15.1)	24112 (13.6)	101403 (11.4)
New Mexico	9113 (5.1)	2285 (1.3)	1922 (1.1)	1773 (1.0)	3013 (1.7)	18106 (2.0)
Ohio	21375 (12.1)	33839 (19.1)	28046 (15.8)	20709 (11.7)	7271 (4.1)	111240 (12.5)
Pennsylvania	24920 (14.0)	29283 (16.5)	33636 (19.0)	35717 (20.1)	25540 (14.4)	149096 (16.8)
Utah	6423 (3.6)	4270 (2.4)	4506 (2.5)	4790 (2.7)	2360 (1.3)	22349 (2.5)
Washington (Seattle)	13716 (7.7)	9013 (5.1)	9702 (5.5)	11356 (6.4)	6253 (3.5)	50040 (5.6)
Environmental factors
PM2.5, μg/m3 (mean, SD))	8.42 (2.99)	10.55 (2.78)	11.26 (2.71)	12.27 (2.80)	14.46 (3.83)	11.40 (3.64)
NO2, ppb (mean (SD))	16.53 (6.85)	22.01 (6.20)	25.59 (6.22)	29.32 (6.61)	36.37 (8.23)	25.96 (9.59)
NDVI 1230m (mean (SD))	0.39 (0.12)	0.35 (0.10)	0.33 (0.10)	0.30 (0.10)	0.22 (0.10)	0.32 (0.12)
EC, μg/m3 (mean (SD))	0.36 (0.09)	0.56 (0.04)	0.68 (0.03)	0.81 (0.05)	1.13 (0.21)	0.71 (0.28)
OC, μg/m3 (mean (SD))	1.51 (0.41)	1.92 (0.50)	2.13 (0.58)	2.37 (0.66)	3.10 (0.94)	2.21 (0.83)
NO3, μg/m3 (mean (SD))	0.94 (0.46)	1.36 (0.47)	1.55 (0.48)	1.71 (0.53)	2.20 (0.76)	1.55 (0.69)
NH4, μg/m3 (mean (SD))	0.71 (0.45)	0.95 (0.51)	1.05 (0.52)	1.20 (0.53)	1.40 (0.48)	1.06 (0.55)
SO4, μg/m3 (mean (SD))	1.79 (1.10)	2.23 (1.24)	2.33 (1.24)	2.54 (1.30)	2.61 (1.18)	2.30 (1.25)

Abbreviations: Elemental Carbon (EC), Ammonium (NH₄), Nitrate (NO₃), Organic Carbon (OC), Sulfate (SO₄), Particulate matter ≤2.5 microns (PM_2.5), Nitrogen dioxide NO₂), outdoor Light at Night (LAN), Normalized Difference Vegetation Index (NDVI), neighborhood Socioeconomic Status (nSES)

Associations of fine particulate matter components with cause-specific mortality

The cohort experienced 249,959 all-cause deaths, 64,594 PCa deaths, and 61,242 CVD deaths over 6,226,017 person-years of follow-up. Comparing Kaplan-Meier survival curves for cause of death and quintiles of pollution components revealed higher mortality associated with higher air pollution (log-rank P <.001) (Supplementary Figure 5). Hazard ratios for associations of PM_2.5 components with all-cause mortality are presented in Table 2. Per interquartile range increase in PM_2.5, in fully adjusted models, there was higher rate of all-cause mortality associated with EC (aHR: 1.01, 95% CI: 1.00, 1.02), OC (aHR: 1.03, 95% CI: 1.02, 1.04), NH₄⁺ (aHR: 1.02, 95% CI: 1.01, 1.03), and SO₄^2- (aHR: 1.11, 95% CI: 1.10, 1.13). Higher NO₃⁻ exposure was associated with lower mortality (aHR: 0.97, 95% CI: 0.96, 0.98). Supplementary Figure 6 presents spline results, revealing higher risks in the lower end of the distribution of OC and NO₃⁻, but in the higher end of the distribution for NH₄⁺ and SO₄^2-.

Table 2.

Hazard ratios for associations of fine particulate matter components with all-cause mortality among men with prostate cancer in METRO (n=886,876)

	Continuousa	Quintile 1	Quintile 2	Quintile 3	Quintile 4	Quintile 5	P trend
PM2.5 component	aHR (95% CI)		aHR (95% CI)	aHR (95% CI)	aHR (95% CI)	aHR (95% CI)
Elemental Carbon
Cases/Person-years		42,766/1,171,618	45,792/1,205,830	48,889/1,261,352	53,510/1,357,442	59,002/1,363,004
Ageb	1.04 (1.03, 1.04)	Ref	1.01 (1.00, 1.03)	1.00 (0.99, 1.02)	1.01 (0.99, 1.02)	1.10 (1.08, 1.11)	<.001
Confoundingc	1.01 (1.00, 1.01)	Ref	1.02 (1.01, 1.04)	1.00 (0.98, 1.02)	0.99 (0.98, 1.01)	1.02 (1.00, 1.04)	0.14
Organic Carbon
Cases/Person-years		41,275/1,181,877	46,826/1,271,585	51,680/1,308,081	55,482/1,330,787	54,696/1,266,918
Ageb	1.05 (1.05, 1.06)	Ref	1.03 (1.02, 1.05)	1.09 (1.08, 1.11)	1.11 (1.10, 1.13)	1.16 (1.15, 1.18)	<.001
Confoundingc	1.03 (1.02, 1.04)	Ref	1.05 (1.04, 1.07)	1.11 (1.10, 1.13)	1.15 (1.13, 1.17)	1.14 (1.11, 1.16)	<.001
NO 3 −
Cases/Person-years		44,634/1,227,606	45,677/1,249,068	47,788/1,218,037	52,350/1,300,895	59,510/1,363,640
Ageb	1.06 (1.05, 1.06)	Ref	0.97 (0.96, 0.98)	1.06 (1.05, 1.08)	1.09 (1.07, 1.10)	1.13 (1.12, 1.14)	<.001
Confoundingc	0.97 (0.96, 0.98)	Ref	1.04 (1.02, 1.05)	1.08 (1.06, 1.10)	1.04 (1.02, 1.06)	1.00 (0.98, 1.02)	<.001
NH 4 +
Cases/Person-years		45,923/1,146,702	41,832/1,161,624	43,033/1,190,715	56,838/1,415,692	62,333/1,444,515
Ageb	1.05 (1.04, 1.06)	Ref	0.90 (0.88, 0.91)	0.91 (0.9, 0.92)	1.00 (0.99, 1.01)	1.05 (1.04, 1.06)	<.001
Confoundingc	1.02 (1.01, 1.03)	Ref	0.92 (0.90, 0.94)	0.92 (0.9, 0.93)	0.96 (0.94, 0.98)	0.98 (0.96, 1.00)	0.015
SO 4 2-
Cases/Person-years		46,056/1,115,466	45,363/1,194,618	42,031/1,190,369	53,694/1,413,880	62,815/1,444,912
Ageb	1.03 (1.02, 1.03)	Ref	0.91 (0.90, 0.93)	0.83 (0.82, 0.84)	0.93 (0.92, 0.95)	1.00 (0.99, 1.02)	<.001
Confoundingc	1.11 (1.10, 1.13)	Ref	0.92 (0.91, 0.94)	0.85 (0.83, 0.87)	0.95 (0.93, 0.98)	1.03 (1.00, 1.06)	<.001

^aPer interquartile range increase, Cox models ^badjusted for age, diagnosis year, ^cfurther adjusted for race and ethnicity, neighborhood socioeconomic status, insurance, marital status, stage at diagnosis, Normalized Difference Vegetation Index, nitrogen dioxide, population density, and state

Competing risk-adjusted hazard ratios for associations of PM_2.5 components with PCa mortality are presented in Table 3. In fully adjusted models, an IQR increase in NH₄⁺ (aHR: 1.03, 95% CI: 1.00, 1.07) and SO₄^2- (aHR: 1.10, 95% CI: 1.06, 1.15) was associated with higher PCa mortality. There was no association between EC, OC, or NO₃⁻. Results from models with splines are presented in Supplementary Figure 7, which revealed highest risks in the middle of the distribution of OC and NO₃⁻.

Table 3.

Marginal hazard ratios for associations of fine particulate matter components with prostate cancer-specific mortality among men with prostate cancer in METRO (n=701,511)

	Continuousa	Quintile 1	Quintile 2	Quintile 3	Quintile 4	Quintile 5	P trend
PM2.5 component	aHR (95% CI)		aHR (95% CI)	aHR (95% CI)	aHR (95% CI)	aHR (95% CI)
Elemental Carbon
Cases/Person-years		11,935/1,171,618	12,069/1,205,830	12,404/1,261,352	12,938/1,357,442	15,248/1,363,004
Ageb	1.04 (1.03, 1.05)	Ref	0.98 (0.95, 1.01)	0.97 (0.95, 1.00)	0.97 (0.94, 0.99)	1.10 (1.07, 1.13)	<.001
Confoundingc	0.99 (0.97, 1.01)	Ref	1.01 (0.98, 1.05)	0.98 (0.94, 1.02)	0.97 (0.93, 1.01)	0.99 (0.94, 1.04)	0.54
Organic Carbon
Cases/Person-years		11,293/1,181,877	11,677/1,271,585	12,601/1,308,081	13,723/1,330,787	15,300/1,266,918
Ageb	1.08 (1.06, 1.09)	Ref	0.99 (0.97, 1.02)	1.05 (1.03, 1.08)	1.11 (1.08, 1.14)	1.20 (1.17, 1.24)	<.001
Confoundingc	0.98 (0.96, 1.00)	Ref	1.03 (0.99, 1.06)	1.05 (1.02, 1.09)	1.07 (1.03, 1.12)	1.04 (0.99, 1.10)	0.38
NO 3 −
Cases/Person-years		12,559/1,227,606	11,310/1,249,068	12,373/1,218,037	13,466/1,300,895	14,886/1,363,640
Ageb	1.06 (1.04, 1.07)	Ref	0.90 (0.88, 0.93)	1.01 (0.98, 1.03)	1.06 (1.03, 1.09)	1.10 (1.07, 1.12)	<.001
Confoundingc	0.97 (0.95, 0.99)	Ref	1.02 (0.98, 1.06)	1.06 (1.02, 1.11)	1.04 (0.99, 1.09)	0.99 (0.95, 1.05)	0.29
NH 4 +
Cases/Person-years		13,598/1,146,702	11,652/1,161,624	11,963/1,190,715	13,051/1,415,692	14,330/1,444,515
Ageb	1.00 (0.98, 1.01)	Ref	0.85 (0.83, 0.88)	0.87 (0.85, 0.89)	0.89 (0.87, 0.92)	0.97 (0.95, 1.00)	0.52
Confoundingc	1.03 (1.00, 1.07)	Ref	0.92 (0.89, 0.95)	0.94 (0.90, 0.98)	0.93 (0.89, 0.98)	1.00 (0.95, 1.05)	0.14
SO 4 2-
Cases/Person-years		13,868/1,115,466	13,055/1,194,618	11,450/1,190,369	12,280/1,413,880	13,941/1,444,912
Ageb	0.95 (0.94, 0.97)	Ref	0.87 (0.85, 0.89)	0.78 (0.76, 0.80)	0.82 (0.79, 0.84)	0.90 (0.88, 0.92)	<.001
Confoundingc	1.10 (1.06, 1.15)	Ref	0.91 (0.88, 0.94)	0.85 (0.81, 0.89)	0.94 (0.89, 1.00)	1.01 (0.95, 1.08)	<.001

^aPer interquartile range increase, Cox models ^badjusted for age, diagnosis year, ^cfurther adjusted for race and ethnicity, neighborhood socioeconomic status, insurance, marital status, stage at diagnosis, Normalized Difference Vegetation Index, nitrogen dioxide, population density, and state and receipt of definitive treatment. Models weighted using inverse probability of censoring weights for competing events.

Competing risk-adjusted hazard ratios for associations of PM_2.5 components with CVD mortality are presented in Table 4. An IQR increase in EC (aHR: 1.03, 95% CI: 1.02, 1.05), OC (aHR: 1.05, 95% CI: 1.03, 1.07), NH₄⁺ (aHR: 1.08, 95% CI: 1.05, 1.11) and SO₄^2- (aHR: 1.19, 95% CI: 1.15, 1.24) were associated with higher CVD mortality. Some evidence of non-linearity was observed for OC, NO₃⁻, NH₄⁺, SO₄^2-, but this was consistent with higher CVD risks at higher levels of component pollutants (Supplementary Figure 8).

Table 4.

Marginal hazard ratios for associations of fine particulate matter components with cardiovascular disease-specific mortality among men with prostate cancer in METRO (n=685,281)

	Continuousa	Quintile 1	Quintile 2	Quintile 3	Quintile 4	Quintile 5	P trend
PM2.5 component	aHR (95% CI)		aHR (95% CI)	aHR (95% CI)	aHR (95% CI)	aHR (95% CI)
Elemental Carbon
Cases/Person-years		9,651/1,144,755	10,718/1,179,852	11,879/1,237,243	13,520/1,330,233	15,475/1,333,934
Ageb	1.05 (1.04, 1.06)	Ref	1.03 (1.00, 1.06)	1.04 (1.01, 1.07)	1.04 (1.01, 1.07)	1.13 (1.10, 1.16)	<.001
Confoundingc	1.03 (1.02, 1.05)	Ref	1.02 (0.99, 1.06)	1.00 (0.96, 1.04)	1.00 (0.96, 1.04)	1.06 (1.01, 1.10)	0.01
Organic Carbon
Cases/Person-years		9,334/1,153,576	11,213/1,246,056	12,768/1,283,154	14,010/1,302,518	13,918/1,240,713
Ageb	1.05 (1.04, 1.06)	Ref	1.04 (1.01, 1.07)	1.10 (1.07, 1.14)	1.10 (1.07, 1.13)	1.15 (1.11, 1.18)	<.001
Confoundingc	1.05 (1.03, 1.07)	Ref	1.03 (1.00, 1.07)	1.11 (1.08, 1.15)	1.14 (1.10, 1.18)	1.17 (1.12, 1.22)	<.001
NO 3 −
Cases/Person-years		9,886/1,202,816	10,840/1,214,322	11,529/1,196,047	12,909/1,275,903	16,079/1,336,928
Ageb	1.08 (1.06, 1.09)	Ref	1.04 (1.01, 1.07)	1.13 (1.10, 1.16)	1.14 (1.10, 1.17)	1.22 (1.19, 1.26)	<.001
Confoundingc	0.99 (0.97, 1.01)	Ref	1.05 (1.01, 1.09)	1.08 (1.04, 1.13)	1.02 (0.98, 1.06)	1.03 (0.98, 1.07)	0.58
NH 4 +
Cases/Person-years		10,094/1,112,574	9,268/1,129,295	10,030/1,179,678	15,244/1,389,728	16,607/1,414,742
Ageb	1.11 (1.09, 1.13)	Ref	0.90 (0.88, 0.93)	0.93 (0.91, 0.96)	1.10 (1.07, 1.13)	1.12 (1.09, 1.15)	<.001
Confoundingc	1.08 (1.05, 1.11)	Ref	0.94 (0.91, 0.98)	0.96 (0.92, 1.00)	1.06 (1.01, 1.10)	1.07 (1.02, 1.12)	<.001
SO 4 2-
Cases/Person-years		10,170/1,084,678	10,351/1,160,043	9,889/1,177,727	14,164/1,388,368	16,669/1,415,200
Ageb	1.07 (1.06, 1.09)	Ref	0.89 (0.87, 0.92)	0.83 (0.81, 0.86)	1.00 (0.97, 1.03)	1.04 (1.01, 1.07)	<.001
Confoundingc	1.19 (1.15, 1.24)	Ref	0.92 (0.89, 0.95)	0.86 (0.83, 0.90)	1.06 (1.01, 1.12)	1.14 (1.08, 1.21)	<.001

^aPer interquartile range increase, Cox models ^badjusted for age, diagnosis year, ^c further adjusted for race and ethnicity, neighborhood socioeconomic status, insurance, marital status, stage at diagnosis, Normalized Difference Vegetation Index, Nitrogen dioxide, population density, and state. Models weighted using inverse probability of censoring weights for competing events. Models excluded New Mexico due to lack of available cause-specific data for CVD.

Quantile g-computation

A one quintile increase in the PM_2.5 components mixture was associated with higher all-cause mortality (aHR: 1.03, 95% CI: 1.02, 1.04) and CVD mortality (aHR: 1.08, 95% CI: 1.05, 1.11), but not PCa mortality (aHR: 1.02, 95% CI: 0.99, 1.05). For all-cause mortality and CVD mortality, the PM_2.5 components mixture weights for the quantile g-computation analysis for OC and SO₄^2- were positive, while NO₃⁻ and EC were negative; NH₄⁺ was also positive for CVD mortality (Supplementary Figure 9).

Assessment of effect modification

Effect modification of associations of PM_2.5 components with all-cause mortality, PCa mortality, and CVD mortality stratified by nSES, race and ethnicity, and stage are presented in Figure 1, with numeric estimates in Supplementary Table 3. For all-cause mortality, associations of PM_2.5 components were often strongest among those with the most deprived (tertile 1) nSES (EC P_het <.001, NO₃⁻ P_het = .041, NH₄⁺ P_het <.001). For example, in tertile 1, an IQR increase in EC (aHR: 1.02, 95% CI: 1.01, 1.02) and NH₄⁺ (aHR: 1.04, 95% CI: 1.03, 1.06) was associated with higher all-cause mortality, but associations were generally weaker in more advantaged tertiles of nSES. Higher exposure to PM_2.5 components was associated with significantly lower mortality in Hispanic or Other men compared to Non-Hispanic White or Non-Hispanic Black men (P_het <.001 for all PM_2.5 components). For example, an IQR increase in SO₄^2- was associated with higher all-cause mortality in non-Hispanic White (aHR: 1.13, 95% CI: 1.11, 1.15) and non-Hispanic Black (aHR: 1.11, 95% CI: 1.08, 1.14) men, but not Hispanic or Other (aHR: 0.94, 95% CI: 0.90, 0.97) men. There was evidence of effect modification for associations of PM_2.5 components by tumor stage at diagnosis for EC, NO₃⁻, NH₄⁺, and SO₄^2- (P_het < .001 for all comparisons). For example, an IQR increase in SO₄^2- was associated with higher mortality among men with localized disease (aHR: 1.17, 95% CI: 1.15, 1.19), but not regional or distant disease (aHR: 0.86, 95% CI: 0.84, 0.88). Stronger associations of PM_2.5 with all-cause mortality were observed among those in low compared to high population density areas, although the magnitude was similar (P_het ≤ .004). There was significant evidence of effect modification across tertiles of NDVI for OC, NH₄⁺, and SO₄^2- (P_het <.001 for all comparisons), with strongest associations in the highest tertile of greenness.

Figure 1. Hazard ratios for PM_2.5 components with all-cause, prostate cancer, and cardiovascular mortality stratified by neighborhood socioeconomic status, race and ethnicity, stage at diagnosis, population density, and greenness.

Cox models adjusted for age, diagnosis year, race and ethnicity, neighborhood socioeconomic status, insurance, marital status, stage at diagnosis, Normalized Difference Vegetation Index, nitrogen dioxide, population density, and state (receipt of definitive treatment for prostate cancer mortality). Models for prostate cancer-specific mortality and cardiovascular disease-specific mortality used inverse probability weights for competing risks. Models for cardiovascular disease mortality excluded New Mexico participants due to missing cause of death data. EC=elemental carbon, OC=organic carbon, NO₃=nitrate, NH₄=ammonium, SO₄=sulfate, nSES=neighborhood socioeconomic status. P_het=p-for-heterogeneity from likelihood ratio tests (four degrees of freedom for nSES, two degrees of freedom for race, one degree of freedom for stage and population density). The 95% confidence intervals for stratum-specific estimates represent precision of the estimate for association of the PM_2.5 component with mortality, and can be used to test the null hypothesis that the stratum-specific hazard ratio is equal to 1 (no difference in hazard per interquartile range increment in PM_2.5 component).

These patterns were generally consistent for PCa mortality and CVD mortality. For PCa mortality, there was strong effect modification observed for stage with stronger associations with NH₄⁺ (P_het <.001) and SO₄^2- (P_het <.001). An IQR increment in NH₄⁺ (aHR: 1.29, 95% CI: 1.25, 1.33) and SO₄^2- (aHR: 1.37, 95% CI: 1.32, 1.42) were each associated with higher PCa mortality. For CVD mortality, the strongest effect modification for nSES was observed for NH₄⁺ (tertile 1: aHR: 1.09, 95% CI: 1.05, 1.13 vs tertile 3: aHR: 1.12, 95% CI: 1.09, 1.16), with statistically significant effect modification (P_het = .018). Associations of PM_2.5 components with CVD mortality were weaker in Hispanic or other compared to Non-Hispanic White or Non-Hispanic Black men for EC, NO₃⁻, SO₄^2- (P_het ≤ 0.013). In results stratified by stage, there was evidence of effect modification for NH₄⁺ (P_het < .001) and SO₄^2- (P_het < .001), but no statistically significant effect modification for associations of EC, OC, and NO₃⁻. Stronger associations of NH₄⁺ (aHR: 1.11, 95% CI: 1.08, 1.14) and SO₄^2- (aHR: 1.23, 95% CI: 1.19, 1.27) were observed among those with localized compared to regional/distant disease. Stronger associations were observed among those residing in low population density areas for NH₄⁺ (P_het = .036). Among those residing in low population density areas, an IQR increment was associated with higher CVD mortality for NH₄⁺ (aHR: 1.12, 95% CI: 1.08, 1.16). There was significant effect modification observed for NDVI for the association between SO₄^2- and CVD mortality (P_het <.001), with the strongest association observed in the lowest tertile of NDVI.

Sensitivity to unmeasured confounding

E-Values for sensitivity to unmeasured confounding by smoking are presented in Table 5. Associations of smoking with mortality in men with PCa range from 3.26 for all-cause mortality, 1.82 for PCa mortality, and 3.53 for CVD mortality⁷⁶. However, confounding also depends on the prevalence ratio for smoking across high vs low levels of PM_2.5 components. These estimates ranged from 1.04–1.18 in studies with available data^17,77,78. Because confounding depends on the weaker of the associations between confounding and exposure vs outcome⁷³, we can interpret E-Values smaller than 1.04–1.18 in Table 5 as potentially biased by unmeasured confounding by smoking.

Table 5.

E-values for unmeasured confounding for associations of PM_2.5 components with all-cause, prostate cancer, and cardiovascular mortality

	Per IQR		Quintile 5 vs 1
	E-value Point Estimate	E-value CI Lower Bound	E-value Point Estimate	E-value CI Lower Bound
All-cause Mortality
EC	1.07	1.00	1.14	1.03
OC	1.16	1.13	1.41	1.36
NO3−	1.17	1.14	1.04	1.00
NH4+	1.13	1.07	1.14	1.01
SO42-	1.37	1.33	1.16	1.00
Prostate Cancer Mortality
EC	1.12	1.00	1.10	1.00
OC	1.17	1.04	1.25	1.00
NO3−	1.22	1.12	1.08	1.00
NH4+	1.22	1.06	1.06	1.00
SO42-	1.44	1.32	1.13	1.00
CVD Mortality
EC	1.22	1.14	1.30	1.12
OC	1.28	1.22	1.61	1.48
NO3−	1.13	1.00	1.19	1.00
NH4+	1.38	1.29	1.35	1.18
SO42-	1.67	1.57	1.54	1.36

EC=elemental carbon, OC=organic carbon, NO₃⁻=nitrate, NH₄⁺=ammonium, SO₄^2-=sulfate. Note: E-Value calculations directly from results presented in Tables 2–4 may not match due to rounding.

For all-cause mortality, the strongest E-Values for point estimates and confidence intervals were for the association of Quintile 5 vs 1 of OC with all-cause mortality (1.41, 1.36), implying that smoking would have to be at least 37% more prevalent among those in quintile 5 vs 1 of OC to change our inference. Given that the published estimates suggest excess prevalence of only 4–18%, it does not seem likely that confounding by smoking alone could explain this finding. Confidence Interval Lower Bound E-Values for associations of PCa mortality with an IQR increase in PM_2.5 components were between 1.00–1.12, which are comparable in magnitude to the strength of association between smoking and air pollution reported above and so imply that confounding by smoking could explain away these results. However, for the per IQR change in SO₄^2-, (1.44, 1.32), we observed that confounding bias stronger than reported in the literature for the association between smoking and air pollution would be needed to explain away this finding. For CVD mortality, E-Values for confidence intervals comparing quintile 5 to 1 for OC (1.48) and SO₄^2- (1.36), and per IQR for NH₄⁺ (1.29) and SO₄^2- (1.57) suggested that smoking would have to be 29–57% higher in quintile 5 vs 1 to change inferences, far greater than the reported estimates of prevalence ratios for smoking and air pollution and implying that confounding by smoking alone is insufficient to explain these results.

Discussion

In this population-based study of outdoor PM_2.5 components and cause-specific mortality in men with PCa, exposure to PM_2.5 components was associated with higher all-cause mortality. This was predominantly driven by association with higher CVD mortality, among the most common causes of death in men with PCa⁷. The components that contributed most to harms were OC and SO₄^2-, with non-linear exposure-responses observed for most components. Sensitivity analyses revealed that reported associations with CVD mortality were most robust to assumptions of no unmeasured confounding. In general, associations with PM components and mortality were strongest in the lowest (most deprived) tertiles of nSES, areas with high greenspace, and men with localized tumors. Associations were stronger in Non-Hispanic White and Non-Hispanic Black men compared to Hispanic or other men with PCa.

Multiple epidemiologic studies have examined associations of outdoor air pollution with PCa risk^70,82–84, with mixed results^85,86 possibly due to differences in exposure assessment and timing with respect to diagnosis date. In studies of cancer survival, most have assessed cancer-specific mortality rather than CVD mortality, finding no association between exposure to outdoor PM_2.5 and PCa mortality⁸⁷. For example, Coleman et al. studied cancer and cardiopulmonary mortality using SEER Medicare data, with PM_2.5 assigned to Census tracts of residence^22,23. These studies found no association between PM_2.5 and PCa-specific mortality. Additional studies using individual-level data have evaluated associations of census tract-level PM_2.5 with mortality in cancer survivors, again finding limited evidence for associations with PCa mortality^88,89. Our study extends this evidence by including CVD mortality as an outcome, using residential address-level linkages, and evaluating a more recent follow-up period than earlier studies. The links between air pollution and PCa progression and mortality are less clear than for CVD, and so the lack of consistent association with PCa mortality could serve as a negative control⁹⁰.

Prior studies of air pollution and cancer mortality did not examine contributions of individual PM components, but emerging evidence suggests that different geographic regions of the US have different PM component mixtures that lead to different health impacts^39,41. US cohort studies examining sources and components of PM_2.5 in relation to all-cause mortality have found that sulfur linked to coal combustion, and EC, linked to traffic were most strongly associated with overall mortality⁹¹. Other cohorts in different geographic regions found associations of OC with mortality and cerebrovascular disease than other components^92,93. Recent nation-wide population-based analyses using Centers for Medicare & Medicaid Services found that strongest associations of PM_2.5 with atherosclerotic CVD mortality were associated with coal and biomass burning, followed by industrial pollution, oil combustion, and then motor vehicle pollution, though the range of relative rate ratio estimates was small (1.04 to 1.07)⁴¹. Another US Medicare study using the same PM components model as that in our study found that higher EC, OC, and NH₄⁺ were associated with higher mortality in a linear exposure-response fashion³⁹. Comparing adjusted hazard ratios for all-cause mortality per IQR with each component in single pollutant models, we found similar associations for EC and OC, weaker associations for NO₃⁻ and NH₄⁺, and stronger associations for SO₄^2- (Supplementary Figure 10).

Biological mechanisms underpinning these relationships include inflammation, immune suppression, and epigenetic alterations^13,15,16. Studies examining atherosclerotic lesions based on carotid intima-media thickness and coronary artery calcium in relation to PM components found that OC and sulfur-related components may increase CVD risks⁹⁴. A recent experimental study comparing exposure to PM_2.5 nitrate, sulfate and ammonium in mice showed that exposure to all pollutants led to declines in respiratory function, although the strongest influences were seen from nitrate exposure⁹⁵. These impacts were shown to be mediated by neutrophil infiltration, supporting a role of inflammation in mediating influences on CVD mortality. However, further research is needed to establish PM component-specific biological mechanisms.

These findings may have implications for clinical management of patients with PCa due to the high burden of CVD morbidity in this population and high risks of non-PCa death^6,96. Given that men with PCa are often diagnosed with CVD and respiratory illness, which are sensitive to air pollution, they may benefit from information about health risks associated with air pollution, ways to monitor air quality, and ways to protect themselves when air quality is poor (masks, air purifiers, reducing intense outdoor physical activity)⁹⁷. Furthermore, given that long-term androgen deprivation therapy has been linked to higher CVD risk^9,98, understanding patients’ home environment and exposure to air pollution may be important for mitigating these risks. Efforts to limit exposures of sulfate air pollution from coal plants through closures or regulatory standards may also improve CVD outcomes⁹⁹.

This study has some limitations. Unmeasured confounding by behavioral factors, comorbidities, ozone, and healthcare access is a threat to validity. We quantified the magnitude of confounding that would be required to change our inference using E-Values and found that associations of air pollutant components with CVD mortality would persist even under moderate confounding by smoking. Air pollution was assessed at a single address, and so we were unable to assess potential exposures that occurred prior to diagnosis or at non-residential locations or use time-varying Cox models. However, our address-level exposure linkages are more precise compared to earlier studies of air pollution in men with PCa which relied on single Census tract-level linkages. Future studies which have residential mobility data on participants after diagnosis, along with daily or monthly pollutant estimates could provide more precise assessments of short- vs long-term impacts of pollution on mortality in PCa survivors. Misclassification of cause of death has been reported in large registry-based studies¹⁰⁰, but because air pollution is assessed independently of mortality, this misclassification is likely non-differential and so would be expected to attenuate associations towards the null. Study strengths include a large, multistate population capturing diverse geographic and sociodemographic characteristics which enhances external validity of findings.

Conclusion

In this large, population-based study of men with PCa, we report elevated risks of all-cause and CVD mortality in men with PCa associated with higher levels of PM_2.5 components. Exposure to OC and SO₄^2- appeared to drive these relationships. Future studies should assess potential variation in effect estimates based on treatments with adverse CVD profiles, behavioral factors, and tumor characteristics. Our findings can encourage research on environmental modifiers of treatment complications in men with PCa.