Relationship of Perfluorooctanoic Acid Exposure to Pregnancy Outcome Based on Birth Records in the Mid-Ohio Valley

The study area in the mid-Ohio Valley was chosen to include PFOA-contaminated public water utilities in Mason and Wood counties in West Virginia and in Athens, Meigs, and Washington counties in Ohio, where levels ranged from near U.S. averages to orders of magnitude above typical levels (Frisbee et al. 2009). We obtained computerized live birth records with geographic identifiers from the Ohio and West Virginia health departments, and fetal death records from West Virginia only (unpublished data). We restricted the analysis to pregnancies that occurred from 1990 through 2004 because that is the period when the contrasts were highest between exposures resulting from contaminant releases and background sources (Shin et al. 2011a).

Because of the expense involved with refining address information and geocoding maternal residences, which is the basis for assigning exposure, we used case–control sampling to enhance precision for some of the birth outcomes of interest. We restricted eligibility to singleton births and excluded the small proportion (3%) of nonwhite births because there were too few to analyze separately. Comprehensively identified case groups consisted of stillbirths (fetal death ≥ 20 weeks completed gestation), preterm births (< 37 weeks completed gestation), and term low birth weight births (≥ 37 weeks completed gestation and birth weight < 2,500 g). A 10% sample of term births ≥ 2,500 g (n = 3,616) was selected and geocoded for assigning PFOA exposure.

To efficiently select control groups and expand the range of birth outcomes, we first constructed a stratified 10% random sample of the entire population of live births by combining the 10% sample of all term live births ≥ 2500 g with a 10% random sample of the comprehensively identified preterm cases and term low birth weight cases (the case groups described above) (n = 4063). Pregnancy-induced hypertension cases, term small-for-gestational-age (SGA) cases, and term birth weight births were identified exclusively from the 10% random sample of births. Control groups for each of the outcomes were derived as noncase subsets from the 10% random sample of the entire population of live births.

We conducted a series of case–control analyses: a) stillbirths (n = 106) compared with West Virginia live births (n = 1,844) because stillbirth information was available only from West Virginia; b) pregnancy-induced hypertension cases (n = 224) compared with births without pregnancy-induced hypertension (n = 3,828); c) preterm birth cases < 37 weeks (n = 3,613) and < 32 weeks (n = 491) corresponding to all preterm and very preterm births, respectively, compared with term births (n = 3,695), with assignment based on clinical assessment of gestational age; d) term low birth weight cases (n = 918) compared with term births ≥ 2,500 g (n = 3,616); and e) term SGA cases [births < 10th percentile by gestational age and sex (Oken et al. 2003)] (n = 353) compared with term appropriate-for-gestational-age births [births between the 10th and 90th percentiles, inclusive, by gestational age and sex (Oken et al. 2003)] (n = 2,990). In addition, we analyzed a continuous measure of birth weight among term births (n = 4,534).

The exposure metric used in the analysis is modeled serum PFOA concentration of the mother estimated for the early pregnancy period. The methods for estimating individual serum PFOA concentrations are described in detail elsewhere (Shin et al. 2011a, 2011b). Briefly, information on plant operations and chemical releases was combined with environmental characteristics of the region through a series of linked models to estimate air and water concentrations of PFOA from 1951 through 2008 (Shin et al. 2011a). Geocodes for birth residences were used to determine whether drinking water in participants’ homes was supplied by public water sources or private wells. When addresses could be geocoded only at the ZIP code level, an average exposure measure was calculated based on the proportion of the ZIP code population that was supplied by contaminated drinking water. With these estimated environmental levels of PFOA based on the geocoded maternal address, individual maternal serum levels for each pregnancy were estimated using age- and sex-based pharmacokinetic modeling with standard assumptions about water intake, body weights, and a PFOA half-life (Shin et al. 2011b).

Because we had only the residence listed on the birth certificate for the mother—not a residential history—and serum levels at a given time depend on exposure history, we assumed that she had lived in this location for the previous 6 years, which would account for approximately 63% of her serum concentration given a stable exposure rate and an estimated serum half-life of PFOA of 3.5 years (Olsen et al. 2007). Moving within the same public water supply district would have little or no impact on the estimated serum values, but moving across water districts or especially from more distant locations would lead to exposure misclassification.

We were able to geocode the maternal address listed on the birth certificate to the street level for 66% of births and to the ZIP code level for the remaining 34%, with little variation in these proportions by case or control status. We found a high correlation between the exposure estimate using street level geocode and ZIP code averages (r = 0.84), indicating modest loss of information from lack of street level assignment. To make optimal use of the available data, we used the ZIP code average exposure and other information from the birth certificate (maternal education, maternal age, parity, tobacco use, exposure year, residential state) to impute a new exposure estimate for the pregnancies with ZIP code geocodes only (n = 2,780). We used multiple imputation to generate 20 replications to obtain accurate information on the variability in the imputed values (Graham et al. 2007). Markov Chain Monte Carlo imputation was used with 200 burn-in iterations and 100 iterations between each imputation. The parameter estimates and standard errors determined from each imputed data set were pooled into a final parameter estimate and variance (Rubin 1987). Missing covariate data on maternal age (n = 2), maternal education (n = 61), parity (n = 6), and tobacco use (n = 64) were also imputed. The primary analyses incorporate the exposure estimates for the street-level geocode (66%) and the imputed exposure estimate when only ZIP code geocodes could be assigned (34%). In sensitivity analyses we considered the impact of restricting to births with exposure estimates based on street-level geocodes only and no missing covariate data.

Estimated serum PFOA concentrations were generated by calendar year, with the relevant exposure year for each pregnancy considered to be month 3 of gestation for stillbirths and month 4 for live births. The estimated serum PFOA level at the time of pregnancy was skewed, so we considered serum PFOA estimates as a continuous log-transformed measure, a continuous untransformed measure, and in quintiles of exposure. We aggregated the lowest two quintiles as the reference category for analysis because those quintiles covered a modest range typical of U.S. serum levels and our exposure estimates are best able to differentiate individual exposure at higher levels resulting from consumption of locally contaminated water.

We adjusted all analyses for maternal age (splines with 3 degrees of freedom), education (< 12, 12, 13–15, ≥ 16 years), parity (0, 1, ≥ 2), tobacco use (smoker, nonsmoker), exposure year (splines with 4 degrees of freedom), and state of residence (Ohio, West Virginia). In addition, we included gestational age (37, 38, 39, 40, 41, ≥ 42 completed weeks) as a covariate in the analysis of birth weight among term births.

For the binary outcomes, we used logistic regression to estimate the odds ratio (OR) and 95% confidence interval (CI) associated with PFOA exposure, using the log-transformed value to generate ORs for an interquartile shift from the 25th to 75th percentile, and the untransformed variable used to generate ORs per 100-ng/mL shift and for quintiles (comparing the 3rd, 4th, and 5th with the aggregated 1st and 2nd quintiles). Both crude and adjusted effect estimates were generated. We used multiple linear regression to estimate the association between PFOA and birth weight, using the same measures of estimated PFOA exposure described for the categorical outcomes. All statistical analyses were conducted using SAS version 9.2 (SAS Institute Inc., Cary, NC) and R version 2.12.1 (R Project for Statistical Computing, Vienna, Austria).

Births were nearly evenly divided by calendar period and by state (Table 1). Estimated exposure to PFOA tended to be higher for births that were more recent, among older mothers, more highly educated mothers, nonsmokers, and West Virginia residents (Table 1).

Stillbirth was unrelated to PFOA exposure based on either continuous or categorical exposure measures (Table 2). Although there was some suggestion of an increased risk based on the untransformed measure (OR = 1.20 per 100 ng/mL; 95% CI: 0.86, 1.68), the lack of gradient across quintiles does not support an association. Pregnancy-induced hypertension was unrelated to PFOA exposure regardless of metric (Table 2). Preterm birth was not associated with PFOA (Table 3).

PFOA was not related to risk of term low birth weight, with some suggestion of a reduced risk of SGA with higher exposure but no gradient across quintiles (Table 4). Birth weight among term births was largely unrelated to estimated PFOA (Table 4). There was a suggestion of a modest decrease in birth weight with higher exposure based on continuous exposure measures {somewhat stronger for male births [see Supplemental Material, Table 1 (https://doi.org/10.1289/ehp.1104752)]}, but no monotonic gradient across quintiles. In the sensitivity analysis restricted to the 66% of births with street-level geocodes and complete covariate information (see Supplemental Material, Tables 2–4), results were largely unchanged except for generating weak positive associations with both pregnancy-induced hypertension and late preterm births.

The C8 Health Project enrolled > 69,000 residents of the mid-Ohio Valley who were part of a class action lawsuit, administered a questionnaire eliciting health and residential history information, and collected blood to measure serum levels of perfluoroalkyl acids and clinical health markers. Geocoding had already been attempted for every residential address in the C8 Health Project (Shin et al. 2011b).

We sought to link births from 1990–2004 birth certificates in a 13-county region in West Virginia (Cabell, Jackson, Mason, Pleasants, Putnam, Ritchie, Wirt, and Wood counties) and Ohio (Athens, Gallie, Meigs, Morgan, and Washington counties) to self-reported births in the C8 Health Project. This 13-county region encompasses the 5 counties included in study I. We considered a match to be accurate based on consistency between the birth certificate and C8 Health Project information for one of three combinations: a) mother’s county and state of residence + mother’s date of birth + month and year of infant’s birth + infant’s sex; b) mother’s county and state of residence + mother’s age + mother’s first name, middle initial, last name + month and year of infant’s birth + infant sex; or c) mother’s county and state of residence + mother’s address + mother’s age + month and year of infant’s birth + infant sex. This process yielded 3,323 matched births. We extended the matching through the West Virginia Birth Score project, a program of the West Virginia Department of Health (2012) to identify infants at risk of developmental problems. Using the mother’s Social Security number as well as other birth certificate items, Birth Score was able to identify an additional 1,224 matched births through 2004, yielding a total of 4,547 birth certificates linked to the C8 Health Project for analysis.

We used the same environmental/pharmacokinetic models from study I to generate annual exposure estimates, except that the link to the C8 Health Project enabled us to incorporate lifetime residential history, thus reducing exposure misclassification (Shin et al. 2011a, 2011b). We examined the model-based exposure estimates, a Bayesian time-dependent calibration in which the measured 2005–2006 serum concentration was used for updating estimates (Savitz et al. 2012), and a traditional calibration method assuming that a higher-than-expected 2005–2006 serum concentration reflects an entire lifetime of higher-than-expected exposure. As in study I, we considered a log-transformed continuous measure, an untransformed continuous measure, and quintiles, aggregating the lowest two quintiles as the referent. Pregnancy outcomes and covariates were defined as described for study I. We used generalized estimating equations with an exchangeable correlation structure to account for correlation across a woman’s multiple pregnancies.

As in study I, PFOA exposures were higher in the more recent time period, among older and more highly educated mothers, and among nonsmokers; however, in study II exposures were higher for Ohio births than West Virginia births, presumably reflecting the different geographic scope of the two studies (Table 5).

Pregnancy-induced hypertension showed essentially no association with estimated serum PFOA based on the uncalibrated estimates (Table 6). With Bayesian calibration, the log-transformed and categorical measures suggested a positive association, with an increased risk primarily in the 4th quintile. Traditional calibration showed elevated risk in the 3rd and 4th quintiles, but the continuous exposure measures did not indicate an association.

Preterm birth < 37 weeks was weakly related to PFOA levels based on continuous measures in the uncalibrated analyses, but with no gradient across quintiles (Table 7). Bayesian calibration yielded a similar pattern and traditional calibration yielded no association. Preterm birth < 32 weeks, with only 40 cases, showed some suggestion of an association without calibration, enhanced somewhat with traditional calibration, and markedly strengthened with Bayesian calibration.

Term low birth weight was largely unrelated to PFOA based on exposure measures without calibration, except for an isolated increased risk in the 4th quintile (Table 8). With both calibration methods, the log-transformed continuous measure showed a stronger association than the untransformed continuous measure. There were also irregular findings across quintiles, but none suggestive of a monotonic gradient. Term SGA was weakly associated with increasing PFOA levels based on continuous and categorical exposures across all exposure estimates (Table 8). The continuous birth weight analysis yielded some support for an association (Table 8), with a predicted decrement of 22 g for an interquartile shift in exposure (95% CI: –46, 2.1 g), and decrements of 25 g (95% CI: –64, 13 g) and 33 g (95% CI: –73, 6.5 g) in the fourth and fifth quintiles of exposure based on the uncalibrated exposure estimates. This pattern was similar for the Bayesian calibrated estimates and somewhat reduced with traditional calibration. No sex differences were apparent [see Supplemental Material, Table 5 (https://doi.org/10.1289/ehp.1104752)].