Long-term exposure to ambient fine particulate matter and risk of liver cancer in the NIH-AARP Diet and Health Study

Abstract

Background:

Fine particulate matter (PM_2.5) exposure has been associated with liver cancer incidence and mortality in a limited number of studies. We sought to evaluate this relationship for the first time in a U.S. cohort with historical exposure assessment.

Methods:

We used spatiotemporal prediction models to estimate annual average historical PM_2.5 concentrations (1980–2015) at residential addresses of 499,729 participants in the NIH-AARP Diet and Health Study, a cohort in 6 states (California, Florida, Louisiana, New Jersey, North Carolina, and Pennsylvania) and 2 metropolitan areas (Atlanta, Georgia, and Detroit, Michigan) enrolled in 1995–1996 and followed up through 2017. We used a time-varying Cox model to estimate the association for liver cancer and the predominant histologic type, hepatocellular carcinoma (HCC), per 5 μg/m³ increase in estimated outdoor PM_2.5 levels, incorporating a 5-year average, lagged 10 years prior to cancer diagnosis and adjusting for age, sex, race/ethnicity, education level and catchment state. We also evaluated PM_2.5 interactions with hypothesized effect modifiers.

Results:

We observed a non-significantly increased risk of liver cancer associated with estimated PM_2.5 exposure (Hazard ratio [HR] = 1.05 [0.96–1.14], N = 1,625); associations were slightly stronger for HCC, (84 % of cases; HR = 1.08 [0.98–1.18]). Participants aged 70 or older at enrollment had an increased risk of liver cancer versus other age groups (HR = 1.50 [1.01–2.23]); p-interaction = 0.01) and risk was elevated among participants who did not exercise (HR = 1.81 [1.22–2.70]; p-interaction = 0.01). We found no evidence of effect modification by sex, smoking status, body mass index, diabetes status, or alcohol consumption (p-interaction > 0.05).

Conclusions:

Our findings in this large cohort suggest that residential ambient PM_2.5 levels may be associated with liver cancer risk. Further exploration of the variation in associations by age and physical activity are important areas for future research.

1. Introduction

Liver cancer is the sixth most commonly diagnosed malignancy and the third leading cause of cancer-related death worldwide (Rumgay et al., 2022a; Sung et al., 2021). Hepatitis B virus (HBV) and hepatitis C virus (HCV) are major causes (Rumgay et al., 2022b), although their prominence as liver cancer risk factors has declined due to increased availability of the HBV vaccine and antiviral therapies for both HBV and HCV (Rumgay et al., 2022a; Rumgay et al., 2022b). Other established risk factors include aflatoxin B₁, diabetes, obesity, non-alcoholic fatty liver disease (NAFLD), excessive alcohol consumption, and tobacco use (London et al., 2017). The liver is the predominant organ for metabolizing and detoxifying exogenous substances. While aflatoxin has been extensively investigated, other potential environmental risk factors for liver cancer are not well-studied.

The International Agency for Research on Cancer classified outdoor air pollution, and fine particulate matter less than 2.5 μm in diameter (PM_2.5) specifically, as carcinogenic to humans, primarily based on evidence for lung cancer (International Agency for Research on Cancer, 2016). Evidence is also accumulating that outdoor air pollution may be a risk factor for the development of cancers of other organ sites, including the liver. Because of its small diameter, inhaled PM_2.5 can pass the alveolar-capillary barrier and circulate to the liver through the bloodstream (Kim et al., 2014). PM_2.5 can also be swallowed and directly ingested, and inhaled PM_2.5 can proceed through mucociliary clearance processes in the airways, both leading to gastrointestinal exposure and potentially increasing the burden on the liver (Conklin, 2013; Kim et al., 2014). Mechanistic studies have shown that exposure to PM_2.5 can increase oxidative stress and lead to apoptosis, DNA damage, and dysregulation of DNA repair in the liver (Danielsen et al., 2010; Zheng et al., 2015).

Growing epidemiologic evidence supports an association between outdoor air pollution exposure and liver cancer risk. A recent systematic review and meta-analysis of studies of PM_2.5 and upper gastrointestinal cancers found the most consistent associations with liver cancer (Meta RR = 1.31, 95 % CI:1.07, 1.56) (Pritchett et al., 2022). This association was consistent across populations in different geographic locations and with varying exposure assessment approaches, although several studies lacked adjustment for individual level confounders (e.g., socioeconomic status) and most studies did not assess historical exposures. Few studies evaluated the association by tumor histology to assess potential etiologic heterogeneity.

Using a residential-level historical exposure assessment in a large prospective U.S. cohort, our main objective was to investigate the association between long-term exposure to ambient PM_2.5 at the residential address and the development of liver cancer. We also evaluated potential effect modification by a number of factors.

2. Methods

2.1. Study population and outcome ascertainment

The NIH-AARP Diet and Health Study is a large prospective cohort; major details of recruitment and data ascertainment are described elsewhere (Schatzkin et al., 2001). In brief, a baseline questionnaire about diet and lifestyle factors was mailed to 3.5 million members of the AARP who were aged 50–71 years and resided in one of six states (California, Florida, Pennsylvania, New Jersey, North Carolina, and Louisiana) or two metropolitan areas (Atlanta, Georgia and Detroit, Michigan) in 1995–1996. Incident cancer cases were identified through 2017 by probabilistic linkage to cancer registries in the 8 catchment states from recruitment and an additional 3 states where participants tended to move (Arizona, Nevada and Texas). A prior validation study showed that this approach detects about 90 % of incident cancer cases (Michaud et al., 2005). Registry data included the diagnosis date, grade, morphology, and stage for cancer cases. Incident liver cancer cases were classified according to the International Classification of Disease for Oncology (third edition), including hepatocellular carcinoma (C22.0) and intrahepatic bile duct carcinoma (C22.1). Vital status of study participants was identified by annual linkage with the Social Security Administration Death Master File, supplemented by linkage to the National Death Index and responses to study mailings. Person-years of follow-up were calculated from the date that the baseline questionnaire was received to the date that participants were diagnosed with liver cancer, died of any cause, moved out of a registry ascertainment area, were lost to contact tracing, or the end of the follow-up period (December 31, 2017).

Participants’ residential addresses at baseline have previously been geocoded; 91.5 % are at an exact address or street intersection level (i.e., a ‘high quality’ geocode). Of the 566,389 participants enrolled at baseline with complete questionnaires, we excluded participants who were proxy respondents (N = 15,760), reported a personal history of cancer at baseline (N = 49,325), were identified through the National Death Index with liver cancer as the underlying cause of death but were not found in a cancer registry (N = 1,075), did not have a follow-up time (N = 249), and did not have available exposure data (N = 251). Excluded participants were demographically similar to those included (Table S1). The final analytical cohort included 499,729 participants. Among these individuals, a total of 1,625 cases of primary liver cancer were diagnosed over a median of 21.5 years of follow-up. All participants provided consent and the study protocol was approved by the National Institutes of Health Institutional Review Board.

2.2. Exposure assessment

Annual residential-level outdoor air pollution exposures PM_2.5 from 1980 to 2015 were estimated for participants from nationwide spatiotemporal prediction models (Kim et al., 1999). Briefly, PM_2.5 measurements from the U.S. Environmental Protection Agency’s Federal Reference Method (FRM) network (1999–2010) and the Interagency Monitoring of Protected Visual Environment (IMPROVE) network (1999–2012) were used in the development of the model, which incorporated approximately 300 geographic predictors in universal kriging. Temporal trend estimation for 1980 to 2015 was determined using available data from FRM and IMPROVE, supplemented with data from the Clean Air Status and Trends Network, Weather Bureau Army Navy network visual ranges, and the Multi-Ethnic Study of Atherosclerosis and Air Pollution. NO₂ levels from 1990 to 1994 were estimated based on another spatio-temporal model which incorporated data from a national network of regulatory monitors, satellite data, and investigator-deployed monitors (Kirwa et al., 2021).

2.3. Statistical analysis

We used time-varying Cox proportional hazards models with calendar time as the time scale to estimate hazard ratios (HRs) and 95 % confidence intervals (CIs) of the association between PM_2.5 exposures and incident liver cancer, overall and in analyses restricted to hepatocellular carcinoma (HCC), the predominant subtype (84.1 % of cases, N = 1367). The risk sets were defined according to the year of diagnosis for each liver cancer case and included both cases and all uncensored participants at the time of the case diagnosis. For each set, we parameterized PM_2.5 exposures in 5-year moving averages lagged 10 years prior to cancer diagnosis and evaluated associations continuously (per 5 μg/m³ change), categorized in quartiles, and additionally with Q4 split at the 90th percentile. The reference group for categorical analyses was the lowest exposure quartile (Q1). The median of each exposure category was used to derive a continuous variable to test for linear trend. We also evaluated possible non-linear associations using restricted cubic splines, and the proportional hazards assumption with Schoenfeld residuals.

We identified an initial set of potential confounders based on the literature, including demographic characteristics, diet, lifestyle and other health-related behaviors collected in the baseline questionnaire: age, sex, race and ethnicity, education level, body mass index (BMI), smoking status and frequency, alcohol consumption, physical activity, daily caloric intake, self-reported history of diabetes, and catchment state. From these variables, we then used a directed acyclic graph (DAG) to identify a minimum sufficient adjustment set, and further evaluated the remaining factors in the DAG (e.g., smoking, caloric intake) based on a 10 % change-in-estimate criterion using forward selection. The final multivariable models were ultimately adjusted only for the minimum sufficient set, which included age, sex, race and ethnicity, education level, and catchment state.

We evaluated effect modification via models of continuous exposure by age (50–54, 55–59, 60–64, 65–69, and ≥ 70 years), sex (male or female), BMI (underweight and normal [<25 kg/m²], overweight [25.0–29.9 kg/m²] or obese [≥ 30 kg/m²]) (Centers for Disease Control and Prevention, 2022), physical activity (none, < 3 times/week or ≥ 3 times/week), self-reported history of diabetes (yes, no), smoking (non-smoker, former smoker, current smoker), and catchment state. We also evaluated potential interaction by alcohol consumption, categorized into three groups (non-drinkers, moderate drinkers, heavy drinkers) for analyses, defined based on the Departments of Agriculture and Health and Human Services Dietary Guidelines for Americans 2020–2025 (Snetselaar et al., 2021). Moderate drinkers were those whose alcohol intake included two drinks or fewer per day for men, and one drink or fewer per day for women. We defined heavy drinkers as those who consumed larger amounts. Statistical interactions were evaluated using a log likelihood test comparing models with and without interaction terms between the potential modifier and PM_2.5. We also evaluated the independent effect of a common co-pollutant, NO₂, in a single pollutant model. Due to the high correlation between NO₂ and PM_2.5 in 1990–1994 (r = 0.73), we did not mutually adjust but instead assessed their joint effects in separate models of cross-classified exposure categories (PM_2.5 levels in quartiles, and NO₂ levels split at the median).

We conducted sensitivity analyses restricted to participants with high quality geocoded addresses (N = 457, 252, 91.5 %). We additionally conducted sensitivity analyses using attained age as the time scale for models, adjusting for time trend. All statistical analyses were performed in SAS v. 9.4 and R (v4.2.2, Boston, MA). We used an alpha level of 0.05 for determining statistical significance.

3. Results

Over 90.0 % of the participants in the analysis were of non-Hispanic White ancestry; however, although Black participants comprised less than 4 % of the cohort, they made up nearly 8 % of those in the highest PM_2.5 exposure quartile (Table 1). We saw few differences in the proportions of participants across exposure categories by education level, BMI, smoking status, physical activity, or self-reported history of diabetes. Five-year average PM_2.5 concentrations prior to enrollment (1990–1994) differed substantially across the catchment states, with the lowest occurring in Florida (Table S2, median = 12.14 μg/m³) and the highest in Atlanta, Georgia (median = 18.35 μg/m³). California residents comprised the majority of participants and also made up the greatest proportion of the highest exposure quartile (46.5 %). In contrast, Florida comprised 21.3 % of the cohort but less than 1 % of the most highly exposed.

Table 1.

Baseline characteristics of study population by ambient PM_2.5 levels at the residence (1990–1994) ^{a, b}

Characteristics	All participants	PM2.5 concentration (1990–1994; μg/m3)
		Q1: 3.5–13.0	Q2: 13.1–15.4	Q3: 15.5–17.4	Q4: 17.5–27.0
Participants (N)	499,729	124,940	124,929	124,929	124,931
Age at baseline (mean ± SD)	62.0 ± 5.4	62.5 ± 5.3	61.8 ± 5.4	61.8 ± 5.4	62.0 ± 5.4
Age groups (%)
50–54 years	13.4	11.6	14.2	14.3	13.4
55–59 years	22.6	20.6	23.5	23.4	23.0
60–64 years	28.1	27.9	27.9	28.1	28.3
65–69 years	32.1	35.5	30.7	30.6	31.6
≥ 70 years	3.9	4.4	3.7	3.7	3.8
Sex (Male, %)	59.7	60.6	60.4	60.8	56.8
Race and ethnicity (%)
Non-Hispanic White	91.0	94.2	92.7	91.9	85.4
Non-Hispanic Black	4.0	1.6	2.5	4.0	7.8
Hispanic	1.9	2.1	1.7	1.2	2.6
Asian, PI, or AI/AK Native	1.7	0.8	1.8	1.6	2.5
Unknown	1.4	1.3	1.3	1.3	1.7
Education level (%)
High school or less	25.7	25.8	25.0	25.5	26.4
Less than college	33.0	35.9	32.7	30.8	32.6
College graduate or post-graduate	36.4	35.2	39.5	41.0	37.8
Unknown	3.0	3.2	2.9	2.8	3.2
BMI (mean ± SD)	27.1 ± 4.8	27.0 ± 4.7	27.1 ± 4.8	27.1 ± 4.8	27.3 ± 5.0
BMI category (%) c
Underweight and normal	34.4	35.0	34.6	34.2	33.9
Overweight	41.5	42.1	41.8	41.7	40.4
Obese	21.5	20.4	21.2	21.7	22.7
Unknown	2.6	2.5	2.5	2.5	2.8
Smoking status (%)
Non-smoker	35.0	31.4	35.6	36.6	36.4
Former smoker	49.1	52.0	49.2	48.2	46.9
Current smoker	12.0	12.8	11.4	11.3	12.6
Unknown	3.9	3.8	3.8	3.8	4.1
State of residence (%)
California	30.9	24.1	30.0	23.0	46.5
Florida	21.3	64.0	19.2	2.0	0
Georgia	2.8	0	0.2	3.5	7.7
Louisiana	3.8	2.6	7.1	5.4	0.1
Michigan	5.1	0.4	3.7	8.8	7.5
North Carolina	8.3	3.7	11.5	14.2	3.6
New Jersey	12.7	3.6	20.2	21	6
Pennsylvania	15.1	1.7	8.1	22.1	28.7
Total energy intake (kcal, mean ± SD)	1870.5 ± 996.5	1886.9 ± 957.1	1870.8 ± 1001.4	1865.3 ± 953.2	1859.0 ± 1069.5
Total drinks per day (mean ± SD)	1.0 ± 2.8	1.2 ± 3.1	1.0 ± 2.8	0.9 ± 2.6	0.9 ± 2.8
Physical activity at least 20 min (past 12 months, %)
Never	4.5	4.2	4.2	4.4	5.2
< 3 times/week	48.5	45.4	48.1	50.1	50.5
≥ 3 times/week	45.8	49.2	46.7	44.4	43.1
Unknown	1.2	1.1	1.1	1.1	1.3
Self-reported history of diabetes (%)
Yes	9.1	8.8	8.9	9.0	9.6
No	90.9	91.2	91.1	91.0	90.5

^a:Percentages may not add up to 100% due to rounding.

^b:Abbreviations: PM: particulate matter; SD: standard deviation; BMI: body mass index.

^c:Underweight and normal: < 25.0 kg/m²; overweight: 25.0–29.9 kg/m²; obese: ≥ 30.0 kg/m².

In multivariable models, we observed a 5 % suggested increase in the risk of liver cancer per each 5 μg/m³ increase in PM_2.5 (HR = 1.05 [0.96–1.14]; Table 2). The positive association was evident across quartiles, without monotonic trend (p-trend = 0.44). When restricted to the cases with HCC, we found a modestly stronger association (HR = 1.08 [0.98–1.18] per 5 μg/m³) that was similarly elevated across quartiles and greatest for participants in the highest exposure category (HR_Q4vsQ1 = 1.15 [0.95–1.39]; p-trend = 0.13). Using age as the time scale resulted in only nominal changes to these associations (Table S3). When we additionally split Q4 at the 90th percentile, we found non-monotonic increases in risk overall and for HCC (Table S4). The risk was highest in the 75th-90th percentile both overall and for HCC (HR = 1.10 [0.90–1.34] and HR = 1.19 [0.96–1.48], respectively) and dropped somewhat above the 90th percentile (overall: HR = 1.06 [0.75–1.31], HCC:HR = 1.10 [0.88–1.39]). The spline analysis did not indicate statistically significant departure from linearity (data not shown).

Table 2.

Association between long-term exposure to PM_2.5 and incident liver cancer in the NIH-AARP cohort ^a

	Cases (n)	HR (95 % CI) b
Liver cancer
Per 5 μg/m3	1625	1.05 (0.96–1.14)
Q1: 3.5–13.0	371	1.00 (referent)
Q2: 13.1–15.4	399	1.05 (0.90–1.23)
Q3: 15.5–17.4	439	1.05 (0.88–1.25)
Q4: 17.5–27.0	416	1.08 (0.91–1.29)
p for trend		0.44
Hepatocellular carcinoma
Per 5 μg/m3	1367	1.08 (0.98–1.18)
Q1: 3.5–13.0	312	1.00 (referent)
Q2: 13.1–15.4	332	1.04 (0.88–1.24)
Q3: 15.5–17.4	367	1.08 (0.89–1.31)
Q4: 17.5–27.0	356	1.15 (0.95–1.39)
p for trend		0.13

^a:Abbreviations: PM: particulate matter; HR: hazardous ratio; CI: confidence interval.

^b:Adjusted for age, sex, race and ethnicity, education level, and state of residence.

Our stratified models did not show clear evidence of a difference in the association between women (HR = 1.08 [0.91–1.28] per 5 μg/m³) and men (HR = 1.03 [0.93–1.14]; p-interaction = 0.89; Table 3). However, the association was notably stronger in participants over 70 years old at baseline (HR = 1.50 [1.01–1.23]) versus participants in the other age groups (p-interaction = 0.01). The associations varied by catchment state, with HRs ranging from 0.51 (North Carolina, 95 % CI: 0.32–0.81) to 1.17 (Detroit, Michigan, 95 % CI: 0.52–2.63, p-interaction = 0.01). We found the strongest association among participants reporting no physical activity (HR = 1.81 [1.22–2.70] versus those reporting exercising < 3 times per week (HR = 1.07 [0.95–1.21] or ≥ 3 times per week (HR = 0.95 [0.83–1.09]; p-interaction = 0.01). Alcohol intake, smoking status, BMI, and diabetes status did not appear to modify the association (p-interaction all > 0.5). Associations did not change substantially when we restricted the analysis to participants with high quality geocodes (changes in HRs < 5 %, data not shown). In a single-pollutant model for NO_2, we found a suggestion of increased risk of liver cancer per interquartile range (IQR; 10.43 ppb) change (Table S5, HR = 1.04 [0.97–1.11]) and a slightly stronger association for HCC (HR = 1.06 [0.99–1.14]). In cross-classified models, an association with higher levels (Q3 and Q4) of PM_2.5 was observed in the group with ≥ median NO₂ exposure, although no associations were statistically significant (Table S6).

Table 3.

Association between long-term exposure to PM_2.5 and incident liver cancer in the NIH-AARP cohort, by demographic characteristics and lifestyle risk factors ^a

	Cases (n)	Mean ± SD (μg/m3)	HR (95 % CI) b	p for interaction
Sex c
Male	1250	15.4 ± 3.2	1.03 (0.93–1.14)	0.89
Female	375	15.6 ± 3.3	1.08 (0.91–1.28)
Age groups
50–54 years	160	15.6 ± 3.0	1.17 (0.87–1.57)	0.01
55–59 years	336	15.6 ± 3.2	1.02 (0.84–1.25)
60–64 years	510	15.5 ± 3.2	1.16 (0.99–1.34)
65–69 years	546	15.3 ± 3.3	0.97 (0.83–1.13)
≥ 70 years	73	15.3 ± 3.3	1.50 (1.01–2.23)
State of residence
California	508	16.6 ± 4.0	1.07 (0.97–1.18)	0.01
Florida	306	12.2 ± 1.4	0.86 (0.56–1.30)
Georgia	41	18.2 ± 1.3	1.15 (0.32–4.17)
Louisiana	66	14.6 ± 1.5	1.00 (0.42–2.40)
Michigan	70	16.9 ± 1.7	1.17 (0.52–2.63)
North Carolina	131	15.4 ± 1.9	0.51 (0.32–0.81)
New Jersey	241	15.5 ± 1.6	1.12 (0.74–1.70)
Pennsylvania	262	17.1 ± 1.7	1.14 (0.78–1.68)
Physical activity d
None	85	15.7 ± 3.3	1.81 (1.22–2.70)	0.01
< 3 times/week	853	15.6 ± 3.2	1.07 (0.95–1.21)
≥ 3 times/week	669	15.3 ± 3.2	0.95 (0.83–1.09)
BMI category c, f
Underweight and normal	381	15.5 ± 3.2	0.95 (0.80–1.13)	0.24
Overweight	666	15.4 ± 3.2	1.10 (0.95–1.26)
Obese	541	15.6 ± 3.2	1.01 (0.86–1.19)
Alcohol consumption c
Non-drinkers	469	15.6 ± 3.2	1.05 (0.89–1.22)	0.42
Moderate drinkers	841	15.6 ± 3.2	1.02 (0.89–1.16)
Heavy drinkers	315	15.2 ± 3.2	1.11 (0.93–1.34)
Self-reported diabetes c
Yes	330	15.6 ± 3.2	1.03 (0.85–1.24)	0.55
No	1295	15.5 ± 3.2	1.04 (0.94–1.15)
Smoking status e
Non-smoker	409	15.7 ± 3.2	1.02 (0.86–1.21)	0.29
Former smoker	929	15.4 ± 3.2	1.06 (0.95–1.19)
Current smoker	221	15.5 ± 3.2	1.13 (0.88–1.44)

^a:Abbreviations: PM: particulate matter; SD: standard deviation; HR: hazard ratio; CI: confidence interval; BMI: body mass index.

^b:Base model adjusted for age, sex, race and ethnicity, and state of residence; additionally adjusted for

^c:education level;

^d:education level and BMI category;

^e:education level and smoking frequency.

^f:Underweight and normal: <25 kg/m²; overweight: 25.0–29.9 kg/m²; obese: ≥ 30 kg/m².

4. Discussion

In this large prospective cohort, we found a suggestive positive association between long-term ambient PM_2.5 at the residence and incident liver cancer. Each 5 μg/m³ increase in PM_2.5 was associated with a 5 % increase in the risk of liver cancer and an 8 % increase in the risk of HCC specifically. Associations were most apparent among participants who were older than 70 years at study enrollment or who were not physically active.

Our finding of an increased risk of liver cancer with higher PM_2.5 exposure is consistent with previous studies on ambient air pollution and liver cancer incidence in the U.S., although none have been conducted within a cohort and most lacked individual-level historical exposure assessment. A study of 56,245 HCC cases between 2000 and 2014 from Surveillance, Epidemiology, and End Results (SEER) found a positive association between county-level ambient PM_2.5 levels and HCC incidence (incidence rate ratio (IRR): 1.26 [1.08–1.47] per 10 μg/m³) (VoPham et al., 2018). Another ecologic study in SEER with 185,012 liver cancer cases diagnosed in 1992–2016 and estimated county-level PM_2.5 exposure observed a positive association of similar magnitude with the additional years of data (IRR per 10 μg/m³ = 1.32 [1.11–1.57]) (Coleman et al., 2020).

Studies outside the U.S. have similarly observed a positive association, although many have been limited by small numbers of cases. An investigation within the European Study of Cohorts for Air Pollution Effects (ESCAPE) that included participants from Denmark, Austria, and Italy observed an increased risk of liver cancer associated with residential level PM_2.5 (n = 279 cases, meta-HR per 5 μg/m³ = 1.34 [95 %CI: 0.76–2.35] (Pedersen et al., 2017). Another pooled prospective cohort of participants from six European countries, the Effects of Low-Level Air Pollution (ELAPSE) study, found an association with liver cancer similar in magnitude to that observed in our study (N = 512 cases, HR = 1.12 [0.92–1.36] per 5 μg/m³) (So et al., 2021). These two European studies both evaluated exposures at residential addresses and had comparable PM_2.5 levels (ESCAPE_2008–2011: 11.3–13.6 μg/m³; ELAPSE₂₀₁₀:15.0 μg/m³) to our study (mean_1990–1994 = 15.49 μg/m³). However, exposures in both ESCAPE and ELAPSE were estimated after diagnosis of some liver cancer cases. A prospective cohort in Taiwan found a positive association between long-term PM_2.5 exposure and HCC risk on the main island (N = 203 cases; HR = 1.20 [0.95–1.52] per 13.1 μg/m³) and in the Penghu Islets (N = 261 cases; HR = 1.22 [1.02–1.47] per 0.73 μg/m³). Their exposure levels estimated close to the end of follow-up were notably higher than those in our study and in the European studies (median_2006–2009: Main Island: 36.0 μg/m³; Penghu Islets: 24.1 μg/m³) (Pan et al., 2016).

The median age of diagnosis of liver cancer in the U.S. is approximately 66 years (National Cancer Institute, 2022). We found that the PM_2.5 association with liver cancer was strongest among participants who were over 70 years old at study enrollment. Given that most prior studies did not stratify by age, likely due to small case numbers, our ability to directly contrast this finding with other evaluations is limited. We speculate that one explanation is that the oldest cohort participants would have experienced higher levels of air pollution in the past, such as before the implementation of air quality controls in many parts of the U. S., which began largely in the 1970’s. We also acknowledge that this finding is based only on 73 cases and could be due to chance.

We observed variation in PM_2.5 exposure levels and heterogeneity in the association between PM_2.5 exposure and liver cancer across the catchment states. The two previous U.S.-based ecologic studies did not investigate between-state IRRs (Coleman et al., 2020; VoPham et al., 2018). The exposure levels in our study were relatively higher in two metropolitan areas (median_1990–1994: Atlanta, Georgia: 18.35 μg/m³; Detroit, Michigan: 16.98 μg/m³) and Pennsylvania (median_1990–1994: 17.42 μg/m³) and risks for participants living in these areas were higher compared with other states, although we were limited by small case numbers to observe statistically significant differences. PM_2.5 is a complex aerosol mixture, and the variation in the association we observed across states could be a result of different distributions of its constituent chemical components (Bell et al., 2007). Geographic variation in the PM_2.5 association within the U.S. has been observed with other cancers, including for breast cancer in this cohort (White et al., 2023), and in the Sister Study, which included participants from all 50 states (White et al., 2019). Associations with specific PM_2.5 components were observed between copper and lung cancer (HR = 1.25 [1.01–1.53] per 5 ng/m³) and sulfate and gastric cancer (HR = 1.93 [1.13, 3.27] per 200 ng/m³) in ESCAPE, where PM_2.5 components were assessed directly (Raaschou-Nielsen et al., 2016; Weinmayr et al., 2018). We found an inverse association between PM_2.5 and liver cancer in North Carolina (HR = 0.51 [0.32–0.81], N = 131), which was not expected. We assessed but did not find many differences in demographics and lifestyle factors between North Carolina participants and the cohort as a whole. Given the limited plausibility of PM_2.5 exhibiting a true protective effect on liver cancer, we interpret this result as either a chance finding or due to confounding by unknown factors not addressed in our analyses.

Previous studies have observed a protective effect of physical activity on the risk of liver cancer, independent of BMI (Behrens et al., 2013; Zelber-Sagi et al., 2021), and physical activity may take place outdoors, which could influence an individual’s exposure to PM_2.5. However, few studies have evaluated whether physical activity may modify the association between PM_2.5 and liver cancer. We found that participants who did not exercise had a significantly increased risk of liver cancer associated with PM_2.5. We also observed a suggestive, weaker association among moderate exercisers but found no association among individuals reporting the most frequent physical activity (≥3 times per week). These findings may indicate that active physical activity could decrease the risk of developing PM_2.5-associated liver cancer. However, the increased risk we observed among inactive individuals could have other explanations. These participants may have spent more time at home, thus their residential PM_2.5 exposures could be less misclassified compared to more active individuals. These findings could also reflect other, unmeasured factors such as unknown health problems that may have preceded their liver cancer diagnosis; for example, participants who were less physically active due to poor health could be more vulnerable to the carcinogenic effects of PM_2.5. However, we found that physical activity was not correlated with self-reported health status at enrollment (r = 0.21). It will be important for other studies with better characterization of both factors to explore this relationship further.

Smoking is an established risk factor for liver cancer (London et al., 2017) and some constituents in tobacco smoke have similar carcinogenic effects as the chemical components of PM_2.5 (Zhou, 2019). We did not find that smoking adjustment influenced our associations, likely due to its weak correlation with ambient PM_2.5 levels. Although we did not detect a significant statistical interaction, we found the strongest PM_2.5-liver cancer association among current smokers. This finding is consistent with observations in the ELAPSE (So et al., 2021) and ESCAPE (Pedersen et al., 2017) studies, supporting the hypothesis that smoking might exacerbate the effects of ambient PM_2.5 on liver cancer risk. Also similar to these studies, we did not observe an interaction between PM_2.5 exposure and alcohol consumption on liver cancer risk. In contrast, the study in Taiwan found an increased risk for PM_2.5-associated HCC among drinkers but not among non-drinkers (Pan et al., 2016). One possible explanation for this inconsistency is that the studies have used varying definitions of alcohol intake to define strata. For instance, we used U.S. dietary guidelines to define groups based on daily intake, similar to ELAPSE and ESCAPE, while the study in Taiwan had only a binary categorization of drinking status (yes vs. no) and no information on the frequency or amount of consumption.

NO₂ is a non-carcinogenic pollutant, which is considered a proxy for traffic-related air pollution and is moderately to highly correlated with ambient PM_2.5 depending on region (Hamra et al., 2015; VoPham and Jones, 2023). Previous studies have noted consistent evidence of an association between NO₂ and lung cancer (Hamra et al., 2015), but few have evaluated its role in liver cancer. We found a suggestive increased risk associated with NO₂ in single pollutant models, consistent with the findings in ELAPSE (So et al., 2021) and ESCAPE (Pedersen et al., 2017). In ELAPSE, the association with NO₂ was robust to adjustment for PM_2.5. The ESCAPE evaluation did not conduct two-pollutant models. Our cross-classified analyses suggest that the observed PM_2.5 associations may have been driven in part by high levels of NO₂, although confidence intervals overlapped and no point estimates reached statistical significance, making interpretations challenging.

We accounted for socioeconomic status (SES) in our analyses by model adjustment for education level (the only available individual-level proxy for participants) because heterogeneity in PM_2.5 exposure by SES has been demonstrated consistently; in North America, lower SES populations are exposed to ambient air pollutants at the highest levels (Hajat et al., 2015). A recent study confirmed that in the U.S. specifically, census tract level PM_2.5 concentrations are highest in low SES populations (Boing et al., 2022) Even though our cohort is of moderate-to-high SES, the greatest proportion of participants with less than a high school education was in the top quartile of PM_2.5 concentrations, consistent with this pattern. The evidence in Europe appears more mixed, with inequities by SES varying in direction (Hajat et al., 2015). Additionally, liver cancer incidence is also greater in populations with low SES in the U.S (Flores et al., 2021). Both the pattern of higher exposure and greater incidence of the disease among those with low SES in the U.S. make a reasonable argument for us to consider SES as a confounder in our analyses. However, while some prior studies may not have similarly accounted for SES, we note that the geographic location of the study is relevant as to whether this factor is important for consideration.

Our study had several strengths, including its investigation of the PM_2.5-liver cancer association within a large, prospective cohort with geographic variability and long-term follow-up. We used a validated spatiotemporal model to estimate historical exposure levels 10 years prior to cancer diagnosis at the residential level, which allowed for a reasonable latency period for the development of this solid tumor. We were able to account for individual risk factors such as physical activity, smoking habits, and alcohol consumption. However, we acknowledge some limitations. The cohort is mainly comprised of non-Hispanic White individuals, which potentially reduces the generalizability of our findings. Although we observed some interesting patterns of association in stratified analyses, we had less than optimal power to fully explore effect modification across some subgroups. We lacked information to evaluate potential confounding by HBV and HCV infection status, although it is not clear that HBV and HCV status would be associated with PM_2.5 exposure. Previous studies have shown associations between outdoor air pollution and both NAFLD and cirrhosis (Li et al., 2023; Orioli et al., 2020), and patients with NAFLD and/or cirrhosis have a higher risk of developing liver cancer (Shah et al., 2023). We did not have data on these conditions in our study to allow for exploration of their potential role in the association between PM_2.5 exposure and liver cancer. Lastly, we did not have residence histories for study participants and assessed exposures based only on the enrollment address, assuming the participants resided at the enrollment address in the past. Residence history reconstruction for a subset of the cohort in California indicated that the median pre-enrollment duration at the address was 13 years, and 91 % of participants remained at their enrollment address 10 years after the study began (Medgyesi et al., 2021). These data offer some support for both historical and prospective residential stability, assuaging concerns about mobility as a source of exposure misclassification and contributing to errors in case ascertainment.

5. Conclusion

Our results suggest that long-term ambient PM_2.5 exposure at the home may be associated with the development of liver cancer. We found the greatest risk in certain subgroups, including individuals 70 years and older and those reporting no physical activity. There was also some suggestion of geographic variation in the relationship. Future studies in more diverse populations could consider assessing PM_2.5 chemical constituency to extend this work.

Data availability

Data can be requested to the AARP Executive Committee with an approved scientific protocol. More information can be found here: https://www.nihaarpstars.com.