Methods
Study Population
The parent birth cohort consisted of a total of 2,246 women, 18–50 years of age, who were recruited at three tertiary care academic centers: Brigham and Women's Hospital and Beth Israel Deaconess Medical Center in Boston, Massachusetts, and The Hospital of the University of Pennsylvania in Philadelphia, Pennsylvania, between October 2006 and September 2008. Eligibility included those women who sought out routine prenatal care before 15 weeks completed gestation, who were > 18 years of age, and planned to deliver at the enrolling institutes. Exclusion criteria included higher-order multiple gestations (triplets or greater). Written informed consent was obtained from all participating women, and the study protocol was approved by each institutional review board. In 2011, a nested case–control study of singleton PTBs was selected from the Brigham and Women's Hospital participant pool (n = 1,648), who were originally enrolled as part of this larger prospective birth cohort. These women were originally recruited at two Brigham and Women's Hospital clinical facilities and one private practice facility. This nested case–control study consisted of 130 women who delivered before 37 weeks of gestation and 352 randomly selected women who delivered at or after 37 weeks. The study was approved by the institutional review boards of Brigham and Women's Hospital and the University of Michigan.
Maternal urine samples were obtained at four visits during pregnancy. Initial visit samples were collected at median 9.7 weeks gestation (range, 4.7–16.1 weeks), visit 2 at median 17.9 weeks (range, 14.9–21.9 weeks), visit 3 at median 26.0 weeks (range, 22.9–29.3 weeks), and visit 4 at median 35.1 weeks (range, 33.1–38.3 weeks). All specimens were stored at –80°C until analysis. Demographic information was collected at the initial visit. Clinically relevant pregnancy characteristics were collected at the initial visit and subsequently at three additional time points throughout pregnancy. Gestational age was confirmed by ultrasound scanning at < 15 weeks gestation if inconsistent with last menstrual period dating.
We further classified preterm birth by clinical presentation (McElrath et al. 2008). In this analysis there were 56 "spontaneous" PTB [arising from clinical presentation of spontaneous preterm labor and/or preterm premature rupture of membranes (PPROM)] and 35 "placental" preterm births (comprising PTB following preeclampsia or intrauterine growth restriction). Additionally, 39 cases were excluded from this subset analysis because they were delivered because medical protocol required their elective delivery before 37 weeks. The deliveries that were performed due to obstetrical protocol included the management of prior classical cesarean section, abdominal cerclage, prior term intrauterine demise, and suspected uterine wall thinning (due to prior surgery). These cases were not analyzed separately, as they have no known unifying etiology.
Urinary BPA Concentrations
Total BPA (free + conjugated) was measured in all available urine samples (n = 1,695) by NSF International in Ann Arbor, Michigan, based on methods developed by the Centers for Disease Control and Prevention (CDC) (Lewis et al. 2013; Silva et al. 2007). Levels below the limit of detection (LOD) were kept if a numerical value was reported or replaced by dividing the LOD (0.4 ng/mL) by the square root of two if no value was reported (Hornung and Reed 1990). Urinary specific gravity (SG) was measured in all samples as an indicator of urine dilution using a digital handheld refractrometer (ATAGO Company Ltd., Tokyo, Japan). Urinary BPA concentrations were corrected for SG using the following formula: Pc = P[(1.015 – 1)/(SG – 1)], where Pc represents the SG-corrected BPA concentration (nanograms per milliliter), P represents the measured concentration in urine, 1.015 is the median SG of all samples measured, and SG represents the specific gravity of the individual sample (Meeker et al. 2009). Both uncorrected and SG-corrected metabolite levels were log-normally distributed and were ln-transformed for statistical analysis to more closely approximate normality and to reduce the likelihood of influential values given the skewed distribution.
Statistical Analysis
Analysis was performed using SAS version 9.2 (SAS Institute Inc., Cary, NC) and R version 2.15.2 (R Core Team 2014). p-Values < 0.05 were used to define statistical significance. Sociodemographic characteristics of the participating women have been previously described (Ferguson et al. 2014), and associations with cases/control status were examined using chi-square tests. To assess variability in BPA levels across pregnancy, we examined differences in levels by visit both in the population overall and in cases and controls separately and tested the differences by Wilcoxon rank sum test. Geometric means and standard deviations of SG-corrected BPA levels at individual visits were calculated and differences by visit were tested using linear mixed models with random intercepts for subject to adjust for intraindividual correlation. Spearman correlations between measures of BPA across subjects were calculated using SG-corrected values. To examine temporal variability in BPA levels by subject, ICCs and 95% CIs were calculated using uncorrected and SG-corrected BPA.
Geometric average BPA concentrations were calculated using the visit 1–visit 3 time point measurements. Visit 4 measurements were excluded from the average because of a relatively small proportion of cases with samples available from that time point. Crude logistic regression models, where preterm birth was the outcome, included average urinary specific gravity and BPA concentrations. In full models, maternal age, race/ethnicity (white/African American/other), and education level were included a priori, and additional covariates were added in a forward step-wise model selection procedure with inclusion in final models if they altered log-transformed BPA concentration effect estimates by > 10%. Additional variables that were considered included health insurance category (private/HMO/self-pay vs. Medicaid/SSI/MassHealth), prepregnancy body mass index (BMI), smoking status during pregnancy (yes/no), parity (nulliparous/parous), prior history of preterm birth (yes/no), and use of assisted reproductive technology (ART) (yes/no). As a sensitivity analysis, we explored whether addition of either time point specific summed di(2-ethyheyxl) phthalate (ΣDEHP) or average urinary ΣDEHP, which were shown earlier in this population to be associated with an increased odds of PTB, altered any effect estimates (Ferguson et al. 2014). Adjustment for mono-n-butyl phthalate (MBP), mono(2-ethylhexyl) phthalate (MEHP), and mono(2-ethyl-5-carboxypentyl) phthalate (MECPP) was also explored, but results were not included because separate adjustment did not alter the results seen with adjustment for ΣDEHP. ΣDEHP, MEHP, MECPP, and MBP metabolites were associated with preterm birth in a previous study (Ferguson et al. 2014) in this population, and levels were weakly but significantly correlated with BPA in this study (Spearman rho = 0.17, 0.11, 0.12, and 0.28, respectively). Phthalate measurement methods and primary associations with PTB are provided in detail by Ferguson et al. (2014).
Windows of vulnerability to BPA exposure were then assessed by fitting separate logistic regression models with preterm birth as the outcome to calculate odds ratios (ORs) corresponding to BPA levels from each individual visit and, due to low correlation between time points, together in the same model. We did not control for multiple comparisons.
To explore the longitudinal nature of the relationship between BPA exposure and risk of PTB, we initially used linear mixed-effects models to generate random slopes and intercepts for BPA exposure over time. Random intercepts and slopes were then used as predictors in the adjusted logistic regression models. We also fit generalized additive mixed models (GAMMs) and linear mixed-effects models using BPA concentrations as the response variable, and preterm birth, gestational age at sample collection, and other covariates as the explanatory variables, to explore the interaction between preterm birth and gestational age at time of urine sample collection.
For additional secondary analyses, we repeated the above steps for subtypes of preterm birth, including placental and spontaneous preterm birth. PTB cases that did not fit into the subtype were excluded from analysis instead of being recoded as controls. We additionally recreated models of overall preterm birth stratified by infant sex. Covariates from full models were kept the same for sensitivity analyses for comparison.