The Study of Women’s Health Across the Nation (SWAN) recruited participants aged 42 to 52 years from 1996 to 1997 in seven sites across the United States (Santoro et al., 2011). Eligibility criteria included presence of an intact uterus who were not pregnant and had had at least one menstrual period and were not taking hormone medications within the prior three months. Race/ethnicity was based on participant self-identification. Investigators recruited Black women from Boston, MA, Chicago, IL, Pittsburgh, PA, and southeast MI, Hispanic women from Newark, NJ, Chinese women from Oakland, CA, and Japanese women from Los Angeles, CA and White women at all seven sites. The SWAN Multi-Pollutant Study (SWAN MPS) enrolled a subsample of 1,400 participants using repository biospecimens-collected at the SWAN visit 03 (the SWAN MPS baseline, 1999–2000) to assess multiple environmental pollutants in White, Black, Chinese, and Japanese women (Ding et al., 2020; Park et al., 2019a; Wang et al., 2019a). Hispanic women were not included due to a lack of biological samples at the Newark site. Of the 1,400 SWAN MPS participants with available PFAS measurements, the current analysis excluded women with missing information on hypertension (n=18), those who had prevalent hypertension (n=306) at the SWAN MPS baseline, and those with missing values for other covariates (n=18), yielding a final analytic sample of 1,058 women followed from 1999 to 2017.

The research protocols were approved by the institutional review boards at each of the collaborating institutions. Written informed consent was obtained from all participants at each visit.

Serum samples were analyzed at the Division of Laboratory Sciences, National Center for Environmental Health, Centers for Disease Control and Prevention (CDC). The analytic methods used to estimate serum PFAS concentrations has been described previously (Kato et al., 2011a). In brief, an online solid phase extraction-high performance liquid chromatography-isotope dilution-tandem mass spectrometry were developed for quantification of perfluorohexane sulfonate (PFHxS), linear perfluorooctane sulfonate (n-PFOS), sum of branched isomers of PFOS (Sm-PFOS), linear perfluorooctanoate (n-PFOA), sum of branched PFOA (Sb-PFOA), perfluorononanoate (PFNA), perfluorodecanoate (PFDA), perfluoroundecanoate (PFUnDA), perfluorododecanoic acid (PFDoDA), and 2-(N-ethyl-perfluorooctane sulfonamido) acetate (EtFOSAA), and 2-(N-methyl-perfluorooctane sulfonamido) acetate (MeFOSAA) in 0.1 mL of serum. The coefficient of variation of low- and high-quality controls ranged from 6% to 12%, depending on the analyte. The limit of detection (LOD) was 0.1 ng/mL for all the analytes. Detection frequencies, medians (interquartile ranges, IQR), geometric means (geometric standard deviation), and ranges are listed in Table S1. Sb-PFOA (%>LOD: 17.1%), PFDA (%>LOD: 41.3%), PFUnDA (%>LOD: 32.9%), and PFDoDA (%>LOD: 3.7%) were not included in statistical analyses due to low detection frequencies. N-PFOS, Sm-PFOS, n-PFOA, PFHxS, PFNA, EtFOSAA and MeFOSAA were detected in >97% serum samples and thus were included in further analyses. Concentrations below the LODs were substituted with LOD/2 (Hornung and Reed, 1990). Total PFOS (PFOS) was computed as the sum of n-PFOS and Sm-PFOS.

A standard mercury sphygmomanometer was used to record systolic blood pressure (SBP) and diastolic blood pressure (DBP) at the approximately annual visits . Blood pressure was measured according to a standardized protocol, with readings taken on the right arm after 5 min in the seated position. Three readings were taken with a minimum two-minute rest period between measures. The average of two sequential blood pressure values was employed in these analyses. Hypertension was defined as SBP≥140 mmHg or DBP≥90 mmHg or use of antihypertensive medications, in accordance with the definitions used in the sixth and seventh reports of the Joint National Committee on Prevention, Detection, Treatment and Control of High Blood Pressure (JNC-6, JNC-7) (Carey et al., 1985; Chobanian et al., 2003).

Given careful review of the literature for important variables associated with both PFOS exposure and hypertension risk (Ding et al., 2022; Park et al., 2019b; Wassertheil-Smoller et al., 2000), we selected a comprehensive set of confounders. Demographic and socioeconomic characteristics assessed at baseline included age (in years), self-defined race (Black, Chinese, Japanese, White), study site (Southeast Michigan, Boston, Pittsburgh, Oakland, Los Angeles), education (high school or less, some college, college degree or higher), and difficulty paying for basics (very hard, somewhat hard, not hard at all). Lifestyle risk factors collected at the MPS baseline included cigarette smoking (current, former, or never) (Ferris, 1978), secondhand smoking (0, 1–4, ≥5 person-hours) (Coghlin et al., 1989), total calorie intake (in kcal) (Block et al., 1986), alcohol intake (<1 drink/month, >1 drink/month, and ≤1/week, >1 drink/week) (Block et al., 1986), and body mass index (BMI). BMI (kg/m²) was calculated using measured height and weight. Physical activity was assessed in various domains, including sports/exercise, household/caregiving, and daily routine, with a score ranging from 3 to 15 (15 indicating the highest level of activity) (Sternfeld et al., 1999). Menopausal status was categorized into surgical postmenopause, natural postmenopause, late perimenopause, early perimenopause, premenopause, or unknown due to hormone therapy (HT) using bleeding patterns and information about HT use (Sowers et al., 2007).

Descriptive analyses were conducted to examine participant characteristics at the MPS baseline in the total population and by racial/ethnic groups. Chi-square tests or Fisher’s exact tests were implemented to compare differences in race/ethnicity for categorical variables; analysis of variance (ANOVA) or Kruskal-Wallis tests were used for continuous variables. Serum concentrations of PFAS were log-transformed with base 2 to ensure normality.

The principal interest of the study was to assess the extent to which PFAS could account for racial/ethnic disparities in incident hypertension. The conceptual model is shown in Figure S1. Causal mediation analysis was conducted to examine the mediating role of PFAS on racial/ethnic disparities in hypertension. These associations were estimated using accelerated failure time (AFT) models with a Weibull distribution (Gelfand et al., 2016; VanderWeele, 2011). The Weibull distribution was selected by comparing Akaike information criterion (AIC) values. The outcome AFT model initially took the following general form: log TRace,PFAS,Covariates=θ0+θ1Race+θ2PFAS+θ3Race×PFAS+θ4Covariates+σε, Where T is time to hypertension, Race represents racial/ethnic groups (including Black or Asian participants compared with White participants), PFAS represents log2-transformed serum concentrations of PFAS, Race × PFAS is the interaction term between race/ethnicity and log2-transformed serum concentrations of PFAS Covariate, are baseline confounders, σ describes the Weibull distribution scale and shape parameters, and ε symbolizes the errors which are independently and identically distributed.

The racial/ethnic disparity was calculated and expressed as relative survival and interpreted as a ratio of time to develop hypertension comparing Black or Asian participants to White. If the relative survival equals to 1, then there is a null association between race/ethnicity and incident hypertension; if the relative survival is less than 1, Black or Asian race/ethnicity compared with White is associated with shorter time to the development of hypertension; and if the relative survival is greater than 1, Black or Asian race/ethnicity is associated with the longer time to the development of hypertension. We then fit a linear regression model for the mediator, log2-transformed PFAS concentrations log PFASRace,Covariates=β0+β1Race+β2Covariates+ε.

Direct effects are equal to (VanderWeele and Robinson, 2014), EYG0Race=1,G0,Covariates−EYRace=0,PFAS,Covariates=exp {[θ1+θ3(β0+β2Covariates+θ2σ2]+0.5θ32σ2}; and mediated effects equal to (VanderWeele and Robinson, 2014), EYG1Race=1,G1,Covariates−EYG0Race=1,G0,Covariates=exp (θ2β1+θ3β1);

Assuming G(0) is a random draw of PFAS concentrations in the White population and G(1) is a random draw of PFAS concentrations in the Black population, Race = 1 represents the Black population and Race = 0 stands for the White population. Similar formula can be derived for the Asian population. The direct effects obtained for race/ethnicity not through PFAS can be interpreted as the disparity in incident hypertension that would remain for participants if the PFAS distributions of the Black population were set equal to those of the White population, assuming we have controlled for sufficient variables such that the associations between PFAS and hypertension actually reflects the effects of PFAS on hypertension. The mediated effects can be interpreted as how much of racial/ethnic disparities in hypertension is due to racial/ethnic differences in serum PFAS concentrations. Note that the assumptions required here for identification are much weaker than those for natural direct and natural indirect effects because it is nearly impossible to capture the effects of physical phenotype, parental physical phenotype, genetic background, and culture context, all of which are associated with race/ethnicity (VanderWeele and Robinson, 2014).

Given the presence of exposure-mediator interactions, we also calculated the controlled direct effects and the percentage of the racial/ethnic disparities in incident hypertension that could potentially be eliminated by intervening to set PFAS concentrations to the 10^th percentiles of exposure in the SWAN MPS study population (Valeri and VanderWeele, 2013; Vanderweele, 2013). The controlled direct effects are disparity in incident hypertension that would remain for participants if the PFAS concentrations are fixed to a specific level, in this case, the 10^th percentiles in the study population. The percentage eliminated is a more policy-relevant measure, which captures the extent to which racial/ethnic disparities in hypertension could be eliminated by intervening on serum PFAS concentrations.

To evaluate the overall mediating effects of PFAS mixtures on incident hypertension, we constructed an integrative index health risk of exposure to multiple chemicals in epidemiological research (Park et al., 2017; Wang et al., 2019b, 2018). The underlying idea behind the index is to build a risk score as a weighted sum of the chemical concentrations from the simultaneous assessment of multiple chemicals. Weights are determined by the magnitudes of the associations between chemicals and health outcomes of interest from the same regression model. To achieve this goal, we first used quantile g-computation to evaluate the associations between multipollutant PFAS and incident hypertension (Keil et al., 2020). Weights represented the association between each PFAS and incident hypertension per one quantile increase in PFAS concentrations. Quantile g-computation was implemented using the “qgcomp” package in R statistical computing environment. An index was then computed as a weighted sum of all seven PFAS by PFASMixturesi=∑j=1Pβ^jEij, where Eij (j = 1, …, P) is the log-transformed concentrations of the jth PFAS for the ith subject; β^j is the beta coefficients for the jth selected PFAS. As a final step, direct and mediated effects were estimated under the framework of causal mediation analysis, similar to individual PFAS. We also considered interactions between race/ethnicity and the index and calculated the racial/ethnic disparities in the development of hypertension that could be eliminated by fixing all PFAS concentrations simultaneously their 10^th percentiles in our study population

Several sensitivity analyses were performed to confirm the robustness of our findings. PFAS are artificial chemicals believed to contribute to obesity (Ding et al., 2021; Liu et al., 2018), thus, we additionally adjusted for BMI at the MPS baseline to exclude the possibility of confounding. Second, we evaluated percent mediated risk assuming no exposure-mediator interactions to assess the additional impact of those interactions. Finally, in addition to standard approaches for causal mediation analysis, we also considered more general approaches in health disparity research i.e., counterfactual disparity measures, or the disparity that would remain in the Black or Asian population assuming the same distribution of PFAS concentrations as the White population (Naimi et al., 2016). We calculated the percent of reduction in racial/ethnic disparities in the development of hypertension when setting the same distributions of PFAS across racial/ethnic groups. This method assumes no exposure-mediator interaction and thus interaction between race/ethnicity and PFAS were not included in the analysis. The definition of hypertension was based on SBP≥140 mmHg or DBP≥90 mmHg or use of antihypertensive medications, in accordance with the widely accepted definitions used in the seventh report of the JNC-7. We have also conducted sensitivity analysis using guideline per ACC/AHA 2017 recommendations (Whelton et al., 2018), i.e., stage 1 hypertension defined as blood pressure at or above 130/80 mmHg and stage 2 hypertension as blood pressure 140/90 mmHg.

Table 1 shows the median (IQR) or frequency of participant characteristics and each mediator in the total population and by racial/ethnic group. The study sample included 577 (54.6%) White, 161 (15.2%) Black, and 320 (30.2%) Asian. The median age of participants was 49.2 years at the MPS baseline (1999–2000). The median follow-up was 5.9 years for 470 participants with incident hypertension, and 16.2 years for 588 participants without hypertension cases. The majority of the study population had college education or higher (53.3%), did not have difficulty paying for basics (70.2%), were never smokers (64.4%), were not exposed to environmental tobacco smoke (62.2%), did not drink alcohol (50.8%), and were pre- or early perimenopausal (72.2%).

Participant characteristics differed significantly by racial/ethnic groups. Black participants were less likely to have a college education or higher (31.4%), compared with White (60.4%) and Asian (51.0%) (P<0.0001). Furthermore, Black participants were more likely to have difficulty paying for basics (P<0.0001), be current smokers (P<0.0001), be exposed to environmental tobacco smoke (P<0.0001) and have had a hysterectomy or oophorectomy (P<0.0001) than White and Asian women. Black and Asian participants tended to report no or low alcohol use (P<0.0001), and low intensity of physical activity (P<0.0001). Asian participants had the lowest BMI among racial/ethnic groups (P<0.0001).

Serum concentrations of PFOS, n-PFOS, Sm-PFOS, EtFOSAA and MeFOSAA were significantly higher among Black participants compared with other racial/ethnic groups (P<0.0001, Table 1). Serum concentrations of n-PFOA were higher among White participants, intermediate among Black, and lowest among Asian participants (P<0.0001). Asian women had lower concentrations of PFHxS compared to White and Black women (P<0.0001). Asian and Black participants had higher concentrations of PFNA than White participants (P<0.0001).

Table 2 shows the total effects, direct effects, and mediated effects of individual PFAS on racial/ethnic disparities in incident hypertension, after adjusting for age, study site, education, difficulty paying for basics, smoking status, environmental tobacco smoke, alcohol intake, menopausal status, physical activity, and total calorie intake at the MPS baseline. Black participants had significantly higher risks of developing hypertension (relative survival: 0.58, 95% CI: 0.45, 0.76) compared to White participants. The direct effects of Black versus White race/ethnicity on time to hypertension were significant in each PFAS model. The mediated effects through PFAS were significant for PFOS, n-PFOS, and two precursors (i.e., EtFOSAA and MeFOSAA). Specifically, the direct-effect of Black versus White race/ethnicity was 0.62 (95% CI: 0.47, 0.82) and the mediated-effect for PFOS was 0.95 (95% CI: 0.89, 1.00). From these data, 8.2% (95% CI: 0.7, 15.3) of the racial/ethnic disparities in hypertension would be explained if PFOS distributions in the Black and White populations were equivalent. Similarly, the percent mediated by n-PFOS was 10.0% (95% CI: 0.3, 17.8), EtFOSAA was 6.9% (95% CI: 0.2, 13.8) , and MeFOSAA was 12.7% (95% CI: 1.4, 22.6). By contrast, no significant mediation was observed for n-PFOA (percent mediated: −3.2%, 95% CI: −10.0, 2.6), PFNA (percent mediated: 2.0%, 95% CI: −3.2, 6.5), or PFHxS (percent mediated: −0.3%, 95% CI: −3.3, 2.4).

Given the significant interactions between race/ethnicity and PFAS on time to hypertension, we ran models fixing each PFAS at the concentration at the 10^th percentile of exposure. In these models the controlled direct effects represent the racial/ethnic disparities in time to hypertension that would remain in the SWAN MPS study population. In these models, the percent of racial/ethnic disparities eliminated was significant for PFOS (10.2%, 95% CI: 0.9, 18.6), n-PFOS (12.7%, 95% CI: 0.4, 22.1), Sm-PFOS (6.3%, 95% CI: 0.3, 13.6), EtFOSAA (7.5%, 95% CI: 0.2, 14.9), and MeFOSAA (17.5%, 95% CI: 2.1, 29.8).

The total effects, direct effects, and mediated effects of racial/ethnicity on time to hypertension comparing Asian with White participants are displayed in Table 3. Asian participants did not have higher risks of earlier onset of hypertension, with the relative survival of 0.84 (95% CI: 0.63, 1.11) and in the mediation analysis neither the direct nor the mediated effects were significant.

Sensitivity analyses further adjusting for BMI at the MPS baseline did not change our study results (Tables S2–S3). Ignoring interactions between race/ethnicity and PFAS, yielded a lower but significant percent of mediation of 5.0% (95% CI: 0.2, 9.3) for PFOS, 5.7% (95% CI: 0.4, 10.1) for n-PFOS, 5.7% (95% CI: 0.1, 10.2) for EtFOSAA, and 6.0% (95% CI: 0.6, 11.1) for MeFOSAA for racial/ethnic differences in time to hypertension comparing Black versus White participants (Table S4), but not for Asian compared to White participants (Table S5). Results of the counterfactual models did not change the overall conclusions of this study (Tables S6–S7). Similar results were observed when defining hypertension as blood pressure at or greater than 130/80 mmHg (Table S8).

For PFAS mixtures, the quantile g-computation estimated β coefficients for all PFAS, as shown in Table S9, including n-PFOS (β=0.071), Sm-PFOS (β=0.011), EtFOSAA (β=0.058), MeFOSAA (β=0.086), n-PFOA (β=−0.003), PFNA (β=−0.001), and PFHxS (β=−0.071). These β coefficients were then incorporated as the weights into the construction of quantile g-computation estimator. The direct and mediated effects of racial/ethnic disparities in hypertension comparing Black with White participants were statistically significant (Figure 1). The direct effect had a relative survival of 0.66 (95% CI: 0.47, 0.92), which means the differences of 34% shorter time to hypertension that would remain if we were able to set the PFAS distributions in the Black population equal to that of the White population. The mediated effect was 0.88% (95% CI: 0.79, 0.98), which resulted in percent mediated of 19.1% (95% CI: 4.2, 29.0). In other words, 19.1% of the racial/ethnic disparities in hypertension could be explained by PFAS mixtures. With an interaction between race/ethnicity and PFAS mixtures, the percent eliminated was 22.3% (95% CI: 5.1, 33.2) if intervening overall PFAS concentrations to the 10^th percentiles in our study population simultaneously. No significant mediation by PFAS mixtures was observed for differences in the development of hypertension comparing Asian with White participants.