The Avon Longitudinal Study of Parents and Children (ALSPAC) enrolled pregnant women from three health districts of the county of Avon, Great Britain, with an expected delivery date between April 1991 and December 1992. A total of 14,610 children joined the cohort at birth. Details of recruitment methods are described elsewhere (Golding et al. 2001
In 2004–2005, when enrolled offspring were 13 years old, there were 11,820 singleton active participants in the ALSPAC, of whom 5,756 (49%) were girls. Of these girls, 3,682 had returned at least two valid assessments of pubertal status between the ages of 8 and 13 years. We obtained a sample of 448 girls from the group of girls with at least two valid pubertal assessments for a study of maternal serum concentrations of PFCs and menarche (Christensen et al. 2011
). Cases (n
= 218) consisted of all the girls who attained menarche < 11.5 years, and controls (n
= 230) were a random sample of noncases who attained menarche ≥ 11.5 years of age. We used this sample to study associations of maternal exposures to PFCs, estimated from the serum concentrations, with fetal and postnatal growth. Median age at menarche in the full cohort was 12.93 years (Rubin et al. 2009
Birth weight (in grams), length (in centimeters), and gestational age (in weeks) were abstracted from medical records. Weight and height at 2 (mean ± SD = 1.7 ± 0.3), 9 (9.2 ± 0.9), and 20 (19.7 ± 2.8) months were obtained by health professionals as part of the routine infant health surveillance program. We converted the infant’s weights to weight-for-age SD scores (z
-scores) using the 1990 British growth reference curves for girls (Cole et al. 1995
). Ponderal index was calculated using the following formula: (weight in grams/length in cubic centimeters) × 100. Self-reported data on maternal prenatal characteristics and behaviors were obtained from the mothers during pregnancy. Breast-feeding data were obtained from questionnaires administered when the girls were 4 weeks old. Data collection instruments and methods have been described in detail elsewhere (Golding et al. 2001
). Potential confounders assessed include gestational age (weeks); maternal educational level (< O level/O level/> O level); maternal prepregnancy body mass index (BMI); maternal age at delivery (years); previous live births (0/≥ 1); maternal smoking during pregnancy (yes/no); mother’s ethnic background (white/nonwhite); breast-feeding of child between birth and 4 weeks (yes/no); and gestational age when maternal serum sample was obtained. An O-level education is the qualification obtained at age 16 years when obligatory schooling ends. We assigned maternal prepregnancy weight status according to the Centers for Disease Control and Prevention (CDC) adult BMI classification: underweight (< 18.5), normal (18.5–24.9), overweight (25.0–29.9), and obese (≥ 30.0).
PFOS, PFOA, and PFHxS were measured in 448 stored maternal serum samples collected during 1991–1992 at pregnancy. The median gestational age when serum samples were obtained was 15 weeks, and the interquartile range was 10–28 weeks. Valid results were available for 447 girls. Serum samples were analyzed at the National Center for Environmental Health of the CDC (Atlanta, GA). Analytical methods used have been described elsewhere (Kuklenyik et al. 2005
). Limits of detection were 0.2 ng/mL for PFOS and 0.1 ng/mL for PFOA and PFHxS. We prepared low-concentration (~ 2–9 ng/mL) and high-concentration (~ 6–25 ng/mL) quality control materials with pooled human serum that was analyzed with standards, reagent blanks, and study samples. Depending on the analyte, the precision of measurements, expressed as the relative SD, was 8–13%.
Our analyses were conducted on a sample of girls and their mothers previously selected for a nested case–control study to test associations of PFCs with earlier age at menarche. Because ignoring the sampling scheme can lead to biased results, we accounted for sampling selection probabilities by constructing stratum-weighted linear and longitudinal regression models to estimate the associations between PFC concentrations and fetal and postnatal growth (Richardson et al. 2007
). The weight for the cases (all girls who attained menarche < 11.5 years) was 1, and the weight for the controls (a random sample of girls who attained menarche ≥ 11.5 years of age) was 15.1.
The markers of fetal growth studied were birth weight, birth length, gestational age, and ponderal index. We constructed stratum-weighted adjusted linear regression models to explore associations of each birth outcome with maternal serum concentrations of PFOA, PFOS, and PFHxS. We used the same linear regression methods to model outcomes at 20 months. The single-pollutant stratum-weighted models for each study outcome were constructed using backward elimination to identify potential confounders, with p < 0.20 as the cutoff for retention. Trend tests for analyses of associations with outcomes at birth and at 20 months were done where the exposure was modeled as an ordinal variable and coded as 0, 1, and 2. We tested whether exposure–response trends were significantly different from the null using a p < 0.05.
We examined the influence of missing values for covariates missing values for ≥ 10% of girls. The covariates missing values in this range were maternal smoking during pregnancy and maternal prepregnancy BMI. Comparisons of bivariate associations between the exposures and study outcomes before and after exclusion of girls missing data on either maternal prepregnancy BMI or smoking during pregnancy were conducted to verify for consistency. Results of such analyses suggested that exclusion of subjects missing values for theses covariates had the potential to introduce bias because changes in the magnitude and direction of bivariate associations were observed after these exclusions. Because maternal smoking and maternal pre-pregnancy BMI are predictors of the daughter’s growth, categories for missing values were included in the analyses of fetal and postnatal growth.
Linear mixed models were used for the longitudinal analyses of weight (Verbeke and Molenberghs 1997
). We included girls who had 3 or 4 weight-for-age SD scores using weights obtained at birth or 2, 9, or 20 months (n
= 410). The “time” variable represents the month of age when weight measurements were taken. These models were also constructed with sampling weights to account for the nested case–control sampling.
First, we created exploratory graphs to observe weight-for-age patterns over time by exposure and potential predictors (maternal prepregnancy BMI, maternal age at delivery, maternal educational level, previous live births, maternal smoking during pregnancy, gestational age, gestational age when maternal serum sample was obtained, and breast-feeding in the first 4 weeks). Weight-for-age patterns between 2 and 20 months appeared to be quadratic rather than linear. As described previously, a category for missing data was included for maternal smoking during pregnancy and maternal prepregnancy BMI. Then we fit saturated linear mixed-effects models for each of the three PFC exposures (exposures categorized as tertiles), and included intercept, age, and age-squared parameters for the exposures and potential predictors. To assess whether exposures or predictors influenced weight patterns, we included interactions between age and each covariate and age-squared and each covariate. We used the saturated model to test for random-effects and various random-effects covariance structures. It was appropriate to include a random intercept in addition to random slopes (for age and age-squared), with an unstructured random-effects covariance structure. We removed covariates in a hierarchical way, using backward elimination, to obtain a reduced model. The exposure was retained regardless of significance. Nested models were compared with likelihood ratio tests and potential predictors that were significant at the p < 0.20 were retained. Finally, we constructed graphs to depict the predicted mean weight-for-age SD scores, based on the final linear mixed-effects models, which were calculated at the 4 time points (birth and 2, 9 and 20 months), with PFC exposure as tertiles and all other covariates at their reference. All analyses were conducted using SAS version 9.2 (SAS Institute Inc., Cary, NC). Analyses of markers of fetal growth were performed in PROC GLM and the longitudinal mixed models were performed in PROC MIXED.
Human subject protection was assessed and approved by the ALSPAC Law and Ethics Committee, the local research ethics committees, and the CDC Institutional Review Board.
Informed consent was provided by the mothers at the time of enrollment. The daughters did not provide consent as such because they were children.