A prospective cohort study of ambient air pollution exposure and risk of uterine leiomyomata

Abstract

STUDY QUESTION

To what extent are ambient concentrations of particulate matter <2.5 microns (PM_2.5), nitrogen dioxide (NO₂) and ozone (O₃) associated with risk of self-reported physician-diagnosed uterine leiomyomata (UL)?

SUMMARY ANSWER

In this large prospective cohort study of Black women, ambient concentrations of O₃, but not PM_2.5 or NO₂, were associated with increased risk of UL.

WHAT IS KNOWN ALREADY

UL are benign tumors of the myometrium that are the leading cause of gynecologic inpatient care among reproductive-aged women. Black women are clinically diagnosed at two to three times the rate of white women and tend to exhibit earlier onset and more severe disease. Two epidemiologic studies have found positive associations between air pollution exposure and UL risk, but neither included large numbers of Black women.

STUDY DESIGN, SIZE, DURATION

We conducted a prospective cohort study of 21 998 premenopausal Black women residing in 56 US metropolitan areas from 1997 to 2011.

PARTICIPANTS/MATERIAL, SETTING, METHODS

Women reported incident UL diagnosis and method of confirmation (i.e. ultrasound, surgery) on biennial follow-up questionnaires. We modeled annual residential concentrations of PM_2.5, NO₂ and O₃ throughout the study period. We used Cox proportional hazards regression models to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for a one-interquartile range (IQR) increase in air pollutant concentrations, adjusting for confounders and co-pollutants.

MAIN RESULTS AND THE ROLE OF CHANCE

During 196 685 person-years of follow-up, 6238 participants (28.4%) reported physician-diagnosed UL confirmed by ultrasound or surgery. Although concentrations of PM_2.5 and NO₂ were not appreciably associated with UL (HRs for a one-IQR increase: 1.01 (95% CI: 0.93, 1.10) and 1.05 (95% CI: 0.95, 1.16), respectively), O₃ concentrations were associated with increased UL risk (HR for a one-IQR increase: 1.19, 95% CI: 1.07, 1.32). The association was stronger among women age <35 years (HR: 1.26, 95% CI: 0.98, 1.62) and parous women (HR: 1.28, 95% CI: 1.11, 1.48).

LIMITATIONS, REASONS FOR CAUTION

Our measurement of air pollution is subject to misclassification, as monitoring data are not equally spatially distributed and we did not account for time-activity patterns. Our outcome measure was based on self-report of a physician diagnosis, likely resulting in under-ascertainment of UL. Although we controlled for several individual- and neighborhood-level confounding variables, residual confounding remains a possibility.

WIDER IMPLICATIONS OF THE FINDINGS

Inequitable burden of air pollution exposure has important implications for racial health disparities, and may be related to disparities in UL. Our results emphasize the need for additional research focused on environmental causes of UL.

STUDY FUNDING/COMPETING INTEREST(S)

This research was funded by the National Cancer Institute (U01-CAA164974) and the National Institute of Environmental Health Sciences (R01-ES019573). L.A.W. is a fibroid consultant for AbbVie, Inc. and accepts in-kind donations from Swiss Precision Diagnostics, Sandstone Diagnostics, FertilityFriend.com and Kindara.com for primary data collection in Pregnancy Study Online (PRESTO). M.J. declares consultancy fees from the Health Effects Institute (as a member of the review committee). The remaining authors declare they have no actual or potential competing financial interests.

TRIAL REGISTRATION NUMBER

N/A

Introduction

Uterine leiomyomata (UL) are benign tumors of the myometrium that are an important cause of morbidity among premenopausal women. Symptoms can include heavy menstrual bleeding, anemia, pelvic pain and pregnancy complications (Hartmann et al., 2017). UL are the leading indication for gynecologic inpatient care in the USA (Whiteman et al., 2010) and account for up to $34 billion in annual health care costs (Cardozo et al., 2012). UL are clinically diagnosed in 25–30% of reproductive-aged women (Buttram and Reiter, 1981; Coronado et al., 2000; Stewart, 2001), but ultrasound studies indicate that the true incidence is much higher (i.e. lifetime incidence of 70–80%) as many cases are asymptomatic (Baird et al., 2003). Black women are diagnosed two to three times more often than white women and they tend to experience earlier onset of symptoms and more severe disease (Kjerulff et al., 1996; Marshall et al., 1997; Baird et al., 2003). Differential screening and prevalence of known risk factors do not fully explain this racial disparity (Wise and Laughlin-Tommaso, 2016).

In the USA, racial and economic disparities in environmental exposures have been recognized and documented for decades (Taylor, 2014). People of color are more highly exposed to criteria air pollutants and air toxins compared with white individuals (Gunier et al., 2003; Morello-Frosch and Jesdale, 2006; Morello-Frosch and Lopez, 2006; Miranda et al., 2011; Bell and Ebisu, 2012; Mikati et al., 2018; Rosofsky et al., 2018). Racial residential segregation (Morello-Frosch and Jesdale, 2006), disparate siting of polluting sources in communities of color (Mohai and Saha, 2015) and procedural inequities in remediation and regulatory action (Lavelle and Coyle, 1992) contribute to and sustain this disparity. The inequitable burden of air pollution exposure has important implications for racial disparities in several health outcomes, including asthma (Kravitz-Wirtz et al., 2018), hypertension (Song et al., 2020) and preterm birth (Benmarhnia et al., 2017).

Despite inequitable exposure to air pollution and the high burden of UL among Black women, there has been no assessment of the association between air pollution and UL in this population group. In the Nurses’ Health Study II, a large prospective cohort of nurses from the USA, of whom 95% identify as white, annual exposure to particulate matter <2.5 microns (PM_2.5) was associated with increased risk of self-reported physician-diagnosed UL, particularly among women aged ≤35 years (Mahalingaiah et al., 2014). In a registry-based case-control study of Taiwanese women, annual PM_2.5 and tropospheric ozone (O₃) exposures were associated with higher odds of ultrasound-confirmed UL ascertained via medical records, whereas sulfur dioxide (SO₂), carbon monoxide (CO) and nitrogen dioxide (NO₂) were not (Lin et al., 2019).

We examined the association between residential ambient concentrations of PM_2.5, NO₂ and O₃ with self-reported physician-diagnosed UL in a prospective cohort of Black women residing across the USA.

Materials and methods

Study population

The Black Women’s Health Study (BWHS) is an ongoing prospective cohort study of 59 000 Black women residing in the USA (Rosenberg et al., 1995). Participants were recruited beginning in 1995 primarily using mailing lists of Essence magazine, a general readership magazine targeted to Black women. The study also recruited participants through Black professional organizations and early respondents. Eligible women self-identified as Black or African American were aged 21–69 years and resided in the continental USA. Participation involved completion of biennial self-administered mailed and online questionnaires from 1995 through 2019. On the questionnaires, participants reported information on demographics, lifestyle factors and medical and reproductive histories. The institutional review board at the Boston University Medical Campus approved the study protocol. Participants indicated consent by completing and mailing back the questionnaires.

Estimation of air pollutants

We estimated ambient concentrations of three air pollutants for the present analysis: PM_2.5, NO₂ and O₃. Concentrations of each pollutant were estimated using different methods, at slightly different spatial resolutions, and with monitoring data at different time periods, but all were designed to capture long-term exposure around each participant’s residence. These models were applied to residential locations in 56 metropolitan areas across the USA.

For PM_2.5, we estimated concentrations at the level of the residential address using a hybrid modeling approach (Beckerman et al., 2013). Briefly, we began with construction of a deterministic land use regression model that regressed measures of traffic, land use and population on measured PM_2.5 as a dependent variable. We then applied Bayesian maximum entropy methods to monthly spatio-temporal residuals from the land use regression model. Models were developed using 104 172 monthly PM_2.5 measurements from January 1999 through December 2008 obtained from 1464 monitoring locations in the US Environmental Protection Agency’s (EPA) Air Quality System network. Cross-validation of the final hybrid model showed strong agreement between observed and predicted PM_2.5 levels with little evidence of bias (cross-validation R² = 0.79) (Coogan et al., 2016b).

We used a spatiotemporal land use regression model to estimate annual NO₂ levels at the census block group level (Novotny et al., 2011; Bechle et al., 2015). The spatial portion of the model incorporated annual average NO₂ concentrations from 2006 at 369 monitoring stations, 81 670 estimates of ground-level NO₂ concentrations derived from satellites, and both satellite- and ground-based land use data sets. The temporal portion of the model incorporated 48 886 monthly average NO₂ levels from monitoring stations (2000–2010). The final spatiotemporal model had an R² of 0.80.

We estimated annual O₃ concentrations using the Downscaler, a Bayesian space-time fusion model developed by the US EPA (Berrocal et al., 2012). This model estimated daily 8-hour maximum O₃ concentrations at the centroid of each census tract in the contiguous USA by fusing data from the ground-based monitoring network with Community Model for Air Quality (CMAQ) model estimates with output on 12 × 12 km grids. We averaged daily estimates from 2007 to 2008 to approximate long-term average exposure at each residential location. When the air pollution exposure assessment in BWHS was conducted, O₃ data were only available from 2007 to 2008; subsequent data have shown a slight decline in O₃ over time, but very consistent spatial patterns across the USA. Validation of the Downscaler model using data from the summer of 2001 has been described in detail elsewhere (Berrocal et al., 2012); correlations with hold-out cross-validation locations for daily predictions were generally strong, ranging from 0.61 to 0.86.

Assessment of UL

We defined incident cases as report of a first diagnosis of UL confirmed by ultrasound or surgery from 1997 through 2011. On each follow-up questionnaire, women reported whether they were diagnosed with ‘uterine fibroids’ in the previous 2-year interval, the year in which they were first diagnosed, and the method of confirmation (‘pelvic exam’ and/or ‘ultrasound/hysterectomy’ on the 1999 and 2001 questionnaires or ‘ultrasound’ and/or ‘surgery’ on the 2003 questionnaire and later). Ultrasound has high sensitivity (99%) and specificity (91%) for detection of UL relative to histologic evidence (Dueholm et al., 2002). To maximize the specificity of our outcome definition, cases identified by pelvic exam only were considered noncases (Myers et al., 2006).

We previously conducted a validation study of self-reported UL in BWHS, the details of which have been described elsewhere (Wise et al., 2005). Briefly, we randomly selected 248 incident cases, mailed them supplemental questionnaires and requested permission to review their medical records. Self-reported UL were corroborated in 96% of the 126 women for whom we obtained medical records. The majority of cases (87%) reported that their UL were diagnosed because they sought treatment for symptoms or their UL were identified during a routine pelvic exam.

Covariate assessment

Women reported information on lifestyle factors (alcohol intake, smoking), gynecologic surgeries, anthropometric factors (height and weight), and reproductive history (parity, contraceptive use, menopausal status) on the 1995 and biennial follow-up questionnaires. Women reported educational attainment and menstrual history on the 1995 and 2003 questionnaires. On the 1997 and 2009 questionnaires, women reported their experiences of perceived racial discrimination in daily life (Williams et al., 1997; Wise et al., 2007), and on the 1999 and 2005 questionnaires, we measured depressive symptoms using the 20-item version of the Center for Epidemiologic Studies Depression Scale (CES-D) (Radloff, 1977). We obtained dietary data in 1995 and 2001 using a food frequency questionnaire modified from the 68-item Block instrument (Block et al., 1990; Kumanyika et al., 2003) and used factor analysis to identify two dietary patterns. The first was characterized by high intake of meat and fried food (i.e. ‘Western diet’) and the second was characterized by high intake of fruits and vegetables (i.e. ‘prudent diet’) (Boggs et al., 2011). We used factor analysis to develop a neighborhood socioeconomic status (SES) score based on census variables correlated with wealth, education and income (Coogan et al., 2015). Finally, on each questionnaire from 1997 on, we asked whether the women had undergone a Papanicolaou (Pap) smear in the past 2 years.

Statistical analysis

We first ascertained the method of UL confirmation on the 1999 questionnaire for the previous 2-year interval. Thus, for the present analysis, we began follow-up at the beginning of the previous cycle (1997) and ended in 2011. We excluded women who were lost to follow-up before 1999 (n = 2359), were postmenopausal (n = 17 011) or ≥50 years old (n = 1637) in 1997, reported a history of physician-diagnosed UL in 1995 or 1997 (n = 8355), resided outside of the 56 metropolitan areas in 1997 for which we had previously compiled air pollution data (n = 6903), or had missing data on UL year of diagnosis or method of confirmation. The final analytic sample included 21 998 women.

We assigned air pollution exposures to participants’ residential addresses collected at baseline and over follow-up. To capture long-term average exposures, for each residential address, we calculated the mean of modeled air pollutant data from all available years (1999–2008 for PM_2.5, 2000–2010 for NO₂ and 2007–2008 for O₃). Using these data, we calculated two primary exposure metrics. For the first (hereafter ‘cumulative grand mean’), for each year of follow-up, we averaged the mean air pollutant concentrations across all years from all residential addresses up to that year, weighted by time spent at each address. For example, in 1999, the cumulative grand mean for PM_2.5 is the average of mean PM_2.5 from 1999 to 2008 at the 1997 address, mean PM_2.5 from 1999 to 2008 at the 1998 address and mean PM_2.5 from 1999 to 2008 at the 1999 address. This metric is designed to capture long-term exposure (i.e. throughout the follow-up period). The second metric (hereafter ‘proximate grand mean’) simply updates the air pollutant concentrations based on the most recent address. For example, in 1999, the proximate grand mean for PM_2.5 is the mean PM_2.5 from 1999 to 2008 at the 1999 address. This metric is designed to assess exposure that is more proximate to diagnosis (i.e. at the most recent address). Because the etiologic time window for UL development is unknown and there is a lag between initiation and detection, we additionally calculated 1 year- and 2-year lagged exposures.

We used Cox proportional hazards regression models to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for the association between individual air pollutant concentrations and UL. We stratified models by age in 1-year intervals, 2-year questionnaire cycle and metropolitan area (n = 56). For pollutants with an approximately linear association with UL risk (assessed using restricted cubic splines (Durrleman and Simon, 1989)), we modeled air pollutants continuously and estimated HRs for a one-interquartile range (IQR) increase in exposure (2.9 µg/m³ for PM_2.5, 10.2 ppb for NO₂ and 6.8 ppb for O₃). We calculated person-time from 1997 until diagnosis of UL, loss to follow-up, death, menopause or end of follow-up (2011), whichever occurred first. We analyzed Schoenfeld residuals to assess the proportional hazards assumption.

We identified potential confounders using a directed acyclic graph (Supplementary Fig. S1). In addition to age, questionnaire cycle and metropolitan area (which were controlled for by inclusion in the strata statement), in the final models, we adjusted for education (≤high school, some college, college degree, graduate school), smoking history (never, former, current), body mass index (<20, 20–24, 25–29, 30–34, ≥35 kg/m²), neighborhood SES (quintiles), Western dietary pattern (quintiles), parity (0, ≥1 births), daily perceived racism score (quartiles) and CES-D score (<16, 16–24, ≥25). We additionally adjusted for co-pollutants (i.e. PM_2.5 models were adjusted for NO₂ and O₃).

In sensitivity analyses, we stratified our analyses by age (<35 vs. ≥35 years) because diagnostic specificity for UL is higher in younger women (Baird et al., 2003) and parity (nulliparous versus parous), given its strong protective effect on UL (Baird and Dunson, 2003; Wise et al., 2004). We conducted sensitivity analyses restricting to surgically confirmed cases, as a marker of disease severity, and restricting to women who reported a Pap smear within the past 2 years and women in the highest quintile of neighborhood SES, to reduce the possibility of detection bias. Finally, given previous reports in this cohort of interaction between air pollutants (Jerrett et al., 2017), we calculated goodness-of-fit chi-squared tests comparing models with and without linear interaction terms between pollutants.

We accounted for missing data on exposures and covariates in 1995 using multiple imputation. We generated five imputation data sets using fully conditional specification methods and statistically combined estimates across data sets using Rubin’s rule. For missing data during 1997–2011, we carried forward data from the most recent questionnaire cycle. NO₂ concentrations were missing for 0.9% of women, and covariate missingness ranged from 0% (age) to 6.5% (dietary scores).

Results

During 196 685 person-years of follow-up, 6238 participants (28.4%) reported physician-diagnosed UL confirmed by ultrasound or surgery. The remaining participants were censored at either menopause (28.4%), loss to follow-up (13.9%), death (0.1%) or end of follow-up (29.2%). Median concentrations of PM_2.5, NO₂ and O₃ at the 1997 address were 13.6 µg/m³ (range: 6.0–24.2), 17.8 ppb (range: 1.0–37.7) and 36.9 ppb (range: 25.4–55.0). PM_2.5 was weakly positively correlated with NO₂ (Spearman correlation coefficient (r) = 0.24) and O₃ (r = 0.16), whereas NO₂ and O₃ were moderately inversely correlated (r = −0.57).

Table I shows the means or percentages of baseline characteristics within the lowest and highest quintiles of PM_2.5, NO₂ and O₃ concentrations. PM_2.5 concentrations were positively associated with the Western dietary pattern score and were inversely correlated with neighborhood SES, current smoking and parity. NO₂ concentrations were positively associated with current smoking and were inversely correlated with age, education, neighborhood SES and parity. Finally, O₃ concentrations were associated with similar variables as NO₂ but in opposite directions (which is expected, given the inverse correlation between these two pollutants).

Table I.

Distribution of characteristics by exposure to fine particulate matter (PM_2.5), nitrogen dioxide (NO₂) and ozone (O₃), Black Women’s Health Study, 1997–2011.

	PM2.5 (µg/m3)		NO2 (ppb)		O3 (ppb)
	Quintile 1	Quintile 5	Quintile 1	Quintile 5	Quintile 1	Quintile 5
Characteristica	<12.1	≥15.8	<12.7	≥25.3	<33.5	≥41.9
No. of women	4400	4399	4402	4391	4398	4401
Age (years), mean	36.0	35.6	36.0	34.5	34.8	35.7
College degree, %	27.9	29.0	30.7	26.4	27.0	30.7
Lowest quintile of neighborhood SES, %	10.5	18.7	8.8	31.9	30.3	10.4
Current smoker, %	15.5	13.7	11.7	17.9	18.2	10.7
Current drinker, %	29.1	28.3	25.5	28.6	30.0	25.7
Body mass index ≥30 kg/m2, %	28.6	27.3	27.0	27.9	29.0	25.7
Highest quintile of vegetable/fruit diet pattern score, %	15.2	16.2	13.2	17.7	16.6	14.7
Highest quintile of Western diet pattern score, %	18.8	22.6	20.7	17.3	17.6	21.1
Parous, %	61.0	54.4	60.5	55.4	55.2	57.9
Birth within past 2 years, %	9.6	8.8	10.5	7.4	7.4	10.0
Depot medroxyprogesterone acetate use, %	4.3	3.6	4.0	3.6	3.3	3.4
Age at menarche <11 years, %	11.0	10.4	10.4	11.3	11.1	10.0
Highest quartile of perceived racism, %	29.0	23.7	27.7	24.3	26.3	24.6
CES-D score ≥16, %	31.6	29.8	27.4	31.9	32.6	27.7
Pap smear within last 2 years, %	79.1	78.3	80.0	75.6	76.6	80.1

^aAll characteristics standardized to the age of the cohort in 1997.

CES-D, Center for Epidemiologic Studies Depression Scale.

Restricted cubic spline analyses demonstrated that the association of all three pollutants with UL was approximately monotonic (Supplementary Fig. S2). When we modeled exposures continuously (Fig. 1), we found no appreciable association between concentrations of PM_2.5 and NO₂ with UL, but a positive association between O₃ concentrations and UL. The adjusted HRs for a one-IQR increase in PM_2.5 (2.9 ppb) and NO₂ concentrations (10.2 ppb) were 1.01 (95% CI: 0.93, 1.10) and 1.05 (95% CI: 0.95, 1.16), respectively, whereas the adjusted HR for a one-IQR increase in O₃ concentrations (6.8 ppb) was 1.19 (95% CI: 1.07, 1.32). We observed little difference in HRs when using the 1-year and 2-year lagged exposures.

Figure 1.

Exposure to PM_2.5, NO₂ and O₃ (modeled continuously) and rate of uterine leiomyomata, Black Women’s Health Study, 1997–2011. Hazard ratios (HRs) and 95% confidence intervals (CIs) are for a one-interquartile range increase in exposure and are adjusted for education, smoking history, body mass index, neighborhood socioeconomic status, Western diet score, parity, perceived daily racism score, Center for Epidemiologic Studies Depression Scale (CES-D) score and other pollutants (i.e. PM_2.5 models are adjusted for NO₂ and O₃). Cumulative grand mean exposure is the average of the air pollutant concentration from all residential addresses up to each year, weighted by time spent at each address. Proximate grand mean exposure is the air pollutant concentration from the most recent address.

Note: Figure Replacement Requested.

Figure 2 shows stratified and sensitivity analyses for the association between O₃ and UL. Our findings for O₃ were slightly stronger, although less precise, among women aged <35 years (HR for a one-IQR increase = 1.26, 95% CI: 0.98, 1.62). Post hoc analyses restricted to women <30 years of age were even stronger, but less precise (HR = 1.53, 95% CI: 0.84, 2.77). Results were also stronger among parous women (HR = 1.28, 95% CI: 1.11, 1.48), who are generally at lower risk of UL, compared with nulliparous women (HR = 1.07, 95% CI: 0.90, 1.28). Our findings were slightly stronger when we restricted to surgically confirmed cases at all ages (HR = 1.23, 95% CI: 0.96, 1.57) and surgically confirmed cases among women age <35 years (HR = 1.54, 95% CI: 0.75, 3.14). Results were similar when restricted to women who reported a Pap smear within the last 2 years (HR = 1.25, 95% CI: 1.11, 1.40) and women in the highest quintile of neighborhood SES (HR = 1.34, 95% CI: 1.06, 1.70).

Figure 2.

Exposure to O₃ and rate of uterine leiomyomata, overall and restricted to women age <35 years or ≥35 years, parous women, surgically confirmed cases (hysterectomy or myomectomy), or women with a Papanicolaou smear within the past 2 years, Black Women’s Health Study, 1997–2011. Hazard ratios (HRs) and 95% confidence intervals (CIs) are for a one-interquartile range increase in cumulative grand mean O₃ exposure and are adjusted for education, smoking history, body mass index, neighborhood socioeconomic status, Western diet score, parity, perceived daily racism score, Center for Epidemiologic Studies Depression Scale (CES-D) score and other pollutants (e.g. PM_2.5 models are adjusted for NO₂ and O₃).

We observed little evidence of multiplicative interaction between pollutants. Likelihood ratio tests comparing models with and without linear interaction terms were not statistically significant (P-value > 0.9).

Discussion

In this large prospective study of Black women from across the USA followed from 1997 through 2011, higher residential concentrations of O₃ were associated with increased risk of UL, with an estimated increase of 19% for a one-IQR increase in exposure. Residential PM_2.5 and NO₂ concentrations, on the other hand, were not appreciably associated with UL. There was little evidence of interaction between air pollutants.

This is the third epidemiologic study to examine the association between air pollution and physician-diagnosed UL. While all three studies have found positive associations between air pollution and UL, the associations have varied in magnitude and by the pollutant estimated. In a nested case-control study of women from Taiwan (including 11 028 women with ultrasound-confirmed physician-diagnosed UL ascertained through a health insurance database and 44 112 with no UL diagnosis), concentrations of PM_2.5, O₃, CO, NO₂ and SO₂ were modeled at the residential zip code level. A 10 µg/m³ increase in PM_2.5 concentrations was associated with 9% increased odds of UL (95% CI: 1.05, 1.12) and a 10 ppb increase in O₃ concentrations was associated with 6% increased odds of UL (95% CI: 1.03, 1.10); other pollutants were not associated with UL. In the Nurses’ Health Study II, a prospective cohort study of 85 251 nurses followed for close to 20 years (Mahalingaiah et al., 2014), a 10 µg/m³ increase in residential PM_2.5 concentrations was associated with 1.11 times the risk of self-reported physician-diagnosed UL confirmed by ultrasound or surgery (95% CI: 1.03, 1.19). Larger PM fractions and living close to major roadways (correlated with NO₂ exposure) were not appreciably associated with UL, and the authors did not assess exposure to O₃. PM_2.5 concentrations were substantially higher in the Taiwanese study and slightly higher in NHS II compared with the BWHS (median = 33.1 vs. 15.3 vs. 13.6 µg/m³, respectively), which could partially explain the discrepant results. Median O₃ concentrations were slightly higher in the Taiwanese study (42.1 ppb) compared with the BWHS (36.9 ppb).

The association between O₃ and UL was stronger among BWHS participants aged <35 years, and even more so among women aged <30 years, although the results were imprecise. This suggests that O₃ exposure may be associated with early-onset disease. Similarly, our findings were stronger when restricting to surgically confirmed cases, a subgroup of cases that are more likely to be symptomatic and severe. Taken together, these two sensitivity analyses indicate that O₃ exposure could be related to both early-onset and symptomatic UL. However, stratified estimates were imprecise, particularly when examining risk among women aged <30 years.

The biologic mechanisms through which O₃ exposure could increase UL risk are unclear. O₃ is a secondary pollutant formed by atmospheric transformation of precursors, including NO₂, in the presence of sunlight. It is highly unstable and reactive and thus does not penetrate beyond the lungs. O₃ is a powerful oxidant, and exposure induces an immune-inflammatory response in the lung, causing oxidative damage to the lungs and the airway lining fluids via involvement of the innate immune system (Zhang et al., 2019). However, extrapulmonary effects of O₃ have also been observed (Martinez-Lazcano et al., 2013; Shah et al., 2015; Turner et al., 2016; Croze and Zimmer, 2018), indicating that the immune-inflammatory response can extend into the circulatory system. Although most studies on the pathophysiology of UL have focused on hormonal mechanisms, non-hormonal pathways exist, including both oxidative stress and inflammation, which may play a role in UL pathology (Fletcher et al., 2017; Cetin et al., 2020). The observed association between O₃ exposure and UL could also be mediated by hypertension. Ambient O₃ concentrations have been associated with increased risk of hypertension (Coogan et al., 2017; Weaver et al., 2021), which could lead to a proatherogenic state involving smooth muscle cell injury and cytokine release (Faerstein et al., 2001). Supporting this hypothesis, several studies have found associations between hypertension and UL risk (Emembolu, 1987; Faerstein et al., 2001; Luoto et al., 2001; Aboyeji and Ijaiya, 2002; Boynton-Jarrett et al., 2005). In addition, studies in both rats and humans have shown that O₃ exposure can cause increased concentrations of circulating cortisol and corticosterone (Miller et al., 2015, 2016), indicating the potential for an effect on the stress response. Activation of the hypothalamic pituitary adrenal axis as a result of chronic stress has been associated with luteinizing hormone secretion and adrenal progesterone production (Xiao et al., 2000; Nepomnaschy et al., 2004), which have been associated with higher UL risk (Rein et al., 1995; Baird et al., 2006). Chronic stress itself has also been associated with UL risk (Wise et al., 2007, 2013; Boynton-Jarrett et al., 2011; Vines et al., 2011; Qin et al., 2019). Of note, both particulate matter and NO₂ have been linked with these biologic processes, so it is unclear why we observed an association with O₃ only; however, this finding has some consistency with our observation of stronger associations of diabetes with O₃ than with NO₂ or PM_2.5 (Coogan et al., 2016a,b; Jerrett et al., 2017).

We used monitoring data from a range of years in our modeling approach (1999–2008 for PM_2.5, 2000–2010 for NO₂ and 2007–2008 for O₃). These years were selected based on which years of data were available at the time exposure assessment was conducted. Therefore, we primarily accounted for spatial but not temporal variation in air pollutant concentrations, particularly for O₃ where we only had 2 years of data. However, in general, the spatial patterns of the exposures tend to remain relatively stable over periods of 10 years or so (Jerrett et al., 2005). In our data, we observed little variability in concentrations from year to year, and we hypothesized that inclusion of data from all years would better estimate long-term exposure. However, a consequence of this approach is that exposures occurring after the outcome contribute to the exposure assessment, which could result in some degree of exposure misclassification. Previous analyses in this cohort (Coogan et al., 2016b) have found that exposures including monitoring data from all years are similar to exposures including monitoring data only from years preceding the risk set.

We used validated national statistical models to estimate air pollutant concentrations at participants’ residences. Although our models had high R² values and low levels of error and bias, there are limitations to our approach. The models relied on data from monitors, which are not equally spatially distributed. In areas where monitors are sparse, there could have been over-smoothing of exposure estimates, although the majority of BWHS participants lived within 50 km of a monitor. We also estimated pollutants at three different spatial scales (residential address for PM_2.5, census block group for NO₂ and census tract for O₃); to the extent that there is spatial variability at smaller scales, this may have introduced exposure misclassification. In addition, we estimated pollutant concentrations at residential locations only. We did not have information on indoor sources of air pollution or pollutant exposures during transit or at the workplace. Both of these limitations likely led to exposure misclassification that could be quite substantial, depending on the time spent at home and the available monitoring data around the home. Employed women spend more time away from home and are more likely to have health insurance and access to regular healthcare, and therefore may be more likely to report a UL diagnosis due to more opportunities for detection. In addition, women experiencing UL symptoms (e.g. heavy menstrual bleeding) likely spend more time at home, given the association between UL symptoms and work absenteeism (Downes et al., 2010; Borah et al., 2013; Fortin et al., 2018; Hasselrot et al., 2018; Marsh et al., 2018). This could lead to differential exposure misclassification, where women with UL are more likely to be misclassified with respect to their pollutant exposures compared to women without UL. It is unclear if true exposure would be systematically higher or lower among these women, and therefore, it is difficult to predict the direction of the expected bias.

Ultrasound studies of unselected populations estimate that around 40% of UL are symptomatic (Baird et al., 2003). Women who seek care for their symptoms generally undergo a pelvic exam as an initial diagnostic measure, followed by ultrasound confirmation. The 60% of UL that do not cause symptoms are either detected incidentally (e.g. during a routine pelvic exam or an ultrasound or surgical procedure for an unrelated condition) or remain undiagnosed. Because our outcome measure was based on self-report of a physician diagnosis, we likely under-ascertained UL in our study. If detection was related to air pollution exposures, this could have led to a detection bias (differential outcome misclassification), which would lead to upward bias. For example, individual- or neighborhood-level SES was strongly related to air pollutant concentrations (inversely related to PM_2.5 and NO₂ and positively related to O₃) and women of higher socioeconomic position may have greater access to medical care and willingness to seek care. However, we attempted to minimize this bias in a sensitivity analysis by restricting to women who reported a recent Pap smear and restricting to women in the highest quintile of neighborhood SES and found that our results were similar. In addition, our findings persisted when we restricted to women younger than 35 years, among whom the specificity of UL classification is higher (Baird et al., 2003). Finally, almost all BWHS participants had health insurance and access to regular care, thus lessening the concern over outcome misclassification.

Although we controlled for several individual- and neighborhood-level confounding variables, residual confounding remains a possibility. One key factor we did not account for is vitamin D exposure. Vitamin D deficiency is a suspected risk factor for UL: a recent meta-analysis that pooled results across nine studies collectively consisting of 1730 participants found that serum vitamin D concentrations were lower in UL cases compared with controls (Mohammadi et al., 2020). In addition, several studies show that air pollution is related to vitamin D deficiency, either by blocking ultraviolet B photons or reducing time outdoors (Mousavi et al., 2019). Therefore, it is possible that our observed finding of a positive association between O₃ concentrations and UL is confounded by vitamin D.

Close to 14% of the cohort was lost to follow-up during the study period. However, the median concentrations of PM_2.5, NO₂ and O₃ were similar among those lost to follow-up and those followed until experiencing the outcome or a censoring event. This diminishes concerns about differential loss to follow-up as an important source of bias in this analysis.

This prospective study, with a 14-year follow-up period, a large number of incident cases and extensive data on potential confounders, provides compelling evidence that long-term exposure to air pollution, particularly O₃, could be associated with increased risk of UL. This finding is consistent with results from the one previous study to examine the association between O₃ and UL and is supported by biologic evidence that air pollution increases oxidative stress and inflammation, which may be related to UL development. Because of inequitable exposure to air pollution and the higher burden of UL among Black women, these findings have important implications for racial disparity in UL and support further assessment of air pollution and other environmental factors that could contribute to UL burden.

Data availability

The data underlying this article cannot be shared publicly for the privacy of individuals that participated in the study. The data will be shared on reasonable request to the corresponding author.

Authors’ roles

A.K.W. took the lead on data analysis and drafting the manuscript. All authors made substantial contributions to the conception or design of the work, or to the acquisition, analysis or interpretation of data, critically revised the work for important intellectual content, approved the final submitted version and agreed to be accountable for all aspects of the work.

Funding

This research was funded by the National Cancer Institute (U01-CAA164974) and the National Institute of Environmental Health Sciences (R01-ES019573). The funding sources had no role in study design, collection, analysis or interpretation of data, writing of the report or the decision to submit the article for publication.

Conflict of interest

L.A.W. is a fibroid consultant for AbbVie, Inc. and accepts in-kind donations from Swiss Precision Diagnostics, Sandstone Diagnostics, FertilityFriend.com and Kindara.com for primary data collection in Pregnancy Study Online (PRESTO). M.J. declares consultancy fees from the Health Effects Institute (as a member of the review committee). The remaining authors declare they have no actual or potential competing financial interests.