Study Participants. Study participants included 3448 infants (2280 singletons and 1168 twins; 1704 female and 1744 male) born to 2901 primarily white (83.4%), college-educated (85.3%), and older (62.4% over 30) women. 896 (32.4%) of the women conceived via fertility treatment and 1871 (67.6%) conceived naturally. Demographic and birth information for the infants and their mothers is found in Table 1.
Cytokine and Exposure Measurements. Summary statistics, including maximum, median and mean values for transformed and untransformed cytokines are shown on Table 2 along with number of missing and zero values, further summarized below. Median untransformed PFOA concentration was 2.075 ng/ml (IQR: 1.565), and median PFOS was slightly higher at 2.69 ng/ml (IQR: 1.61). The exposures and their missing and zero values are summarized in Table 2.
Missing Data. Missing data in the cytokine measurements ranges from a minimum of 7% (TARC) to a maximum of 52.8% (MIP-1a), with a median of about 14% missing, and with all measurements having at least some missing units. Cytokines with levels of missingness above 20% included MIP-1a (52.8%), SICAM (34%), MPO (26.5%) and 6-Ckine (20.4%). Missing units in the exposures PFOA and PFOS comprise 14.8%. See Table 2 for details.
Bivariate Analysis
Zero inflation. The study of cytokine distribution among the zero-inflated covariates TRAIL, TNF-α, IL-6, IL-5, IL-33 and SCF (Table 3) shows a stronger pattern of association of zero-valued cytokines with elevated mean PFAS compound value for PFOA than for PFOS (Table 3). Infants with zero-valued IL-6, IL-33, TRAIL, IL-5, and SCF measurements are all associated with significantly higher mean log PFOA values than infants with nonzero cytokine values. For PFOS, this is only true of SCF, with TNF-α marginal (p = 0.076).
ANOVA. There were fewer significant associations between cytokines and above median PFOS (10 cytokines), than for cytokines and above median PFOA (13 cytokines), especially after correction for multiple comparisons. After correction for multiple comparisons, cytokine mean value is significantly reduced for above median levels of PFOA in the cytokines 6-Ckine, Cathepsin, IL-16, IL-5, IL-6, SVCAM, and is marginal for reduced levels of TNF-α, IL-33 and SCF. Above median PFOA has significant and positive/increasing relationships with IL-1α, MIP-1d, MCP-1and CRP (Supplemental Table S2A). After correction for multiple comparisons, 6-Ckine, SCF, Cathepsin, TNF-α, IL-5 and IL-16 remain significantly reduced, and mean IL-1α and NCAM significantly elevated for above-median levels of PFOS (Supplemental Table S2B).
Adjusted Models
With the intention of selecting significant cytokines for further multivariate analysis, mixed effects regression models were adjusted for maternal age, race, BMI, fertility treatment and infant gestational age, and outcomes standardized to assist with effect size comparisons. Adjusted models for continuous (log-transformed) PFOA as exposure show results largely consistent with crude estimates, with significantly negative associations with IL-16, IL-5, IL-6, 6-Ckine, and TNF-α (Fig. 1A; Supplemental Table S3A). SCF, Cathepsin, and sVCAM are negative, but marginal (0.05 < p < 0.10). There are significantly positive associations for IL-1α, MCP-1 and MIP1-d. Estimates and standard errors from adjusted models for continuous (log-transformed) PFOS are slightly larger, and due to the nature of the FDR correction process and a reduction in the p-value of the most significant covariate following adjustment, associations remain significant after correction for fewer cytokines. Only 6-Ckine remains significant and is negatively associated with increasing PFOS (p < 0.05) after correction for multiplicity. TNF-α and IL-16 are significant and negatively associated with PFOS, while IL-1α and MIP-1d are significantly positive in association before FDR correction (Fig. 1B, Supplemental Table S3B).
Stratified models run on log-transformed continuous PFAS exposure data yield evidence both of consistency across the sexes and of differences by sex (Fig. 2); however the latter are not strong enough to constitute effect modification by sex. For both sexes, estimates for IL-1α and MIP-1d were consistent and significantly elevated, while IL-6, IL-16 and 6-Ckine were significantly reduced with increasing PFOA exposure (Fig. 2, panels A and B). For female neonates only, IL-5 was significantly reduced as a function of increasing PFOA, while for males, SCF (stem cell factor) and IL-33 were significantly reduced, though confidence intervals overlap for both. Males only also show significantly elevated inflammatory cytokines CRP and MCP-2 as a function of increasing PFOA. In both sexes, 6-Ckine is again reduced as a function of increasing PFOS. Among female infants only, IL-6, IL-16 and Cathepsin-D are also reduced; while for male infants only SCF is significantly reduced and MIP-1d significantly increased as a function of increasing PFOS (Fig. 2, panels C and D). In all of the above, confidence intervals overlap, so where there are differences, they are not substantial enough to indicate effect modification by sex. Finally, when interaction terms between sex and PFAS are included in models, they are at most just statistically marginal in significance.
Regression with PFOA categorized by quartile yields a more nuanced result than regression with continuous data, with evidence of weak to moderate nonlinearity in associations for all significant cytokines except for IL-1α and IL-5 which show significant positive (IL-1α) and negative (IL-5) linear associations with PFOA by quartile, respectively. Notably, Cathepsin-D and sVCAM show U-shaped dose-response curves by quartile, and TNF-α has a weak inverted U-shaped response (Fig. 3A). See Supplemental Table S4A for PFOA quartile regression estimates, all cytokines. Note that by-quartile estimates are not adjusted for multiplicity.
Results from adjusted quartile regressions for PFOS are largely consistent with unadjusted bivariate models: 6-Ckine, IL-16, TNF-α, IL-6, SCF, cathepsin-D, SICAM and IL-8 show significant and negative association in one or more quartiles, and MIP-1d, NCAM and IL1-a elevated for one or more PFOS quartiles (Fig. 3B, Supplemental Table S4B). Plots by quartile of these estimates, with confidence intervals, show weak to strong nonlinearity in response for all, with Cathepsin, MIP-1d, SICAM and NCAM displaying a U-shaped dose-response response and IL-8 showing an inverted U-shaped response.
We then selected subsets of cytokines for further analysis based on the adjusted continuous and by-quartile regression results. For PFOA, the subset included Th1 family cytokine TNF-α; Th2 family cytokines IL-5, IL-6, and IL-33; IL-1 family cytokine IL-1α, chemokines MIP-1d, 6-Ckine; pleotropic inflammatory cytokine IL-16; SCF, Cathepsin-D, and cell adhesion molecule sVCAM. The subset for PFOS was largely similar, including Th1 family cytokine TNF-α and IL-8; Th2 family cytokine IL-6; IL-1 family cytokine IL-1α, chemokines MIP-1d, 6-Ckine; pleotropic inflammatory cytokine IL-16; and cell adhesion molecules NCAM and SICAM (see Supplemental Table 1 for a listing of cytokine family and function).
Exploratory Factor Analysis
We performed factor analysis for both subsets with the full N = 3448 sample, using a varimax rotation for the PCA, and extracting loadings for three factors/components for the PFOA cytokine subset and two for the PFOS subset. For the PFOA subset, four factors account for 54% of the variance. See Fig. 3A for the loadings visualized as a heatmap array by factor and cytokine.
-
Factor 1 primary loadings: sVCAM, Cathepsin-D, 6-Ckine, MIP-1d; 15% of variance explained.
-
Factor 2 primary loadings: IL-33, SCF; 14% of variance explained.
-
Factor 3 primary loadings: IL-16, IL-1α, 6-Ckine; 13% of variance.
-
Factor 4 primary loadings: TNF-α, IL-5, IL-6; 12% of variance.
For the PFOS subset, three factors were identified by parallel analysis, and accounted for 42% of the variance; see Fig. 3B for a heatmap visualization of the loadings by cytokine and factor.
-
Factor 1 primary loadings: Cathepsin; 6-Ckine, MIP-1d, NCAM; 16% of variance explained.
-
Factor 2 primary loadings: TNF-α, IL-6, IL-8, Cathepsin; 13% of variance explained.
-
Factor 3 primary loadings: IL-16, IL-1α, 6-Ckine; 13% of variance explained.
Imputation level regression-based factor scores were then computed for each PFAS-specific subset of cytokines using the loadings above. While not all component cytokines are the same, note the similarity in loadings between the PFAS subsets.
Factor Analysis Regressions. Factor analysis regressions are most successful when they reveal significant underlying structure or associations in a dataset that may not be revealed by standard multivariable regressions. In this case, multiple regression models using factor scores computed from pooled loadings do show unexpected structure, in that variables significant by themselves in standard bivariate and adjusted regressions can form groupings that are not statistically significant. We performed mixed effects regressions on both continuous exposure data and exposure data categorized by quartile, with consistent results, though the quartile regressions again reveal more nuance.
For continuous (log transformed) PFOA, all axes had negative associations with the exposure, and all except for the third component (primarily loaded by IL-16, IL-1α, and 6-Ckine) were statistically significant (Table 4A). Estimates from quartile regressions were consistent: for the first and second factors, estimates were significant and negative only for the third and fourth quartiles of PFOA (reference level is significant and is quartile 1, Fig. 4A). Estimates were significant and negative across all exposure quartiles for the fourth factor (loaded with Th1 and Th2 family cytokines TNF-α, IL-5, and IL-6). Again estimates for the third component are not significant (Table 4B).
Estimates from regressions with continuous (log-transformed) PFOS were significant and negative for the second (Th1 and Th2 family cytokines) and third (primarily IL-16, IL-1α, and 6-Ckine) axes (Table 5A). Note that the first component (Cathepsin; 6-Ckine, MIP-1d + CAM) is significant for PFOA but not PFOS, while the third (primarily IL-16, IL-1α, and 6-Ckine) is significant and negative for PFOS but not for PFOA. There is at least one significant quartile estimate for all factors in quartile regressions with PFOS (reference level is significant and is quartile 1, Fig. 4B). Estimates for the second factor (primarily Th1 and Th2 cytokines plus cathepsin-D) are negative and significant across all PFOS exposure quartiles (Table 5B).