Study participants
The Tohoku Study of Child Development is a prospective birth cohort study that comprises two birth cohorts from an urban area and a coastal area cohort of northeastern Japan. The protocol and research areas have been described previously [22, 23]. The subjects of this study were the participants of the urban area study.
Pregnant women were recruited through their obstetrical units in two hospitals in the Sendai city [22]. Women were eligible if 1) they did not have any serious illness that could impair the fetus’ development; 2) they did not suffer from pre-eclampsia or gestational diabetes mellitus; 3) in vitro fertilization wasn’t used; 4) their native language was Japanese. Children’s inclusion criteria were 1) the absence of congenital anomalies or serious illnesses; 2) singleton birth, born after 36 weeks of gestation; 3) birth weight more than 2400 g, or higher than 2500 g if gestational age at birth was between 36 and 37 weeks. Written consent was obtained from all women.
Of the 1500 women who were met between January 2001 and September 2003, 687 gave their consent to participate in the study (participation rate: 46%). During their pregnancy, 88 women had to withdraw from the study, either because their baby did not meet the inclusion criteria (n = 54), because they moved to another region (n = 18), or because they dropped out or weren’t reachable (n = 16) [22, 23]. This narrowed the final number of participants to 599. We then selected the participants who had their breast milk tested for p,p’-DDT and p,p’-DDE, had data on total breastfeeding duration (highly influential parameter in the toxicokinetic model), and had data on child weight and height at 42 months (used to calculate BMI, the health outcome of this study), leaving 290 mother-child pairs.
All the procedures of this study were approved by the Medical Ethics Committee of the Tohoku University Graduate School of Medicine and of the National Institute for Environmental Studies [22]. In addition, the protocol for the analyses presented in this paper was approved by the Université de Montréal Institutional Review Board.
Data collection
Data on demographics, lifestyle, medical history, and fish consumption was obtained through questionnaires administered four days after delivery [23]. Breastfeeding duration and anthropometric data such as children’s weight and height were assessed at the 7, 18, 30 and 42 months checkups through questionnaires. Only anthropometric measurements at 42 months were carried out on a large number of children by trained physicians or research coordinators using a certified electronic scale with 0.1 kg increments for weight and 0.1 cm increments for height (KS-110Mcp, Kansai-Seiki, Shiga, Japan). The measurement of body weight was performed without socks and heavy clothes.
Chemical analysis
Mothers were asked to send a breast milk sample of more than 50 ml in a clean glass bottle one month after delivery. Samples were aliquoted in 10 ml centrifuge tubes and immediately frozen at − 80 °C until analyses.
All manipulations were made by IDEA Consultants, Inc. laboratory in Tokyo [23]. A sample of 5 ml of breast milk was mixed with 2 ml of saturated sodium oxalate, 20 ml of 1:1 v/v ethanol/hexane, and 10 ml of diethyl ether. The first hexane layer was mixed for 30 min, separated, and the residue was extracted twice with hexane. The hexane layer was then dehydrated using anhydrous sodium sulfate and evaporated. The lipid extract was weighted, dissolved in hexane and purified using Florisil cartridge column chromatography [24]. p,p’-DDT and p,p’-DDE were then measured using a gas chromatography/high-resolution mass spectrometry (HRGS/HRMS) system. The quality control was performed in accordance with the German external quality assessment scheme (G-EQUAS). The method limit of detection (30 pg/g-lipid for p,p’-DDT and 10 pg/g-lipid for p,p’-DDE) was calculated using the Currie et al. method [25]. Because the toxicokinetic model uses levels expressed on a lipid basis, we expressed concentrations in ng/g-lipid.
Toxicokinetic model
We used the toxicokinetic model of prenatal and lactational exposure to lipophilic persistent organic pollutants (including p,p’-DDT and p,p’-DDE) developed by Verner et al. [18] to estimate children’s levels. The model has two compartments: one for the mother’s lipids and one for the child’s lipids. The mother is exposed through diet and the child is exposed in utero through the placenta, and postnatally through breastfeeding (Fig. 1). The model assumes that lipophilic persistent organic pollutants are completely absorbed through the gastrointestinal tract and distribute evenly in lipids [18]. The percentage of lipids in fetal tissues was set to values at birth, and fetal and child growth was calculated using average growth curves [26]. The increase in maternal body fat mass during pregnancy was calculated as the difference between weight gain and increase in lean tissues. After delivery, the difference between pre-pregnancy weight and postpartum weight was attributed only to adipose tissue, which was assumed to be composed of 75% lipids [18]. Elimination rates were based on published half-lives (p,p’-DDT: ≈ 5 years [27] and p,p’-DDE: 13 years [28]). Exposure through breast milk in children was calculated based on average age-specific hourly milk intake and milk lipid content. The hourly breast milk intake for the first year of life was calculated using a formula based on data from Arcus-Arth et al. [29], and from Kent et al. [30] for the second year of life. We generated profiles of maternal, breast milk and children’s p,p’-DDT and p,p’-DDE levels for each participating mother-child pair. For each pair, the model incorporated individual-specific data on mother’s age at delivery, the total breastfeeding duration, mother’s body weight before pregnancy and one month after delivery, weight gain during pregnancy, and child’s sex. Available information did not allow estimating the duration of exclusive breastfeeding precisely. For that reason, only the total duration of breastfeeding was used, and daily breast milk intake was based on equations derived from studies including exclusively and partially breastfed children. For each pair and each chemical, iterative model simulations were performed to calibrate maternal lifetime daily oral intake until the breast milk level estimated with the model matched the breast milk level measured in the study. This maternal daily dose was then used in the model to estimate children’s areas under the concentration vs. time curve (AUCs) for three periods: 0 to 6 months, 6 to 12 months, and 12 to 24 months of age. Average children’s levels during these periods were calculated by dividing AUCs by the integration period. We performed toxicokinetic model simulations using acslX (Aegis Technologies Inc., Huntsville, AL, USA).
Statistical analysis
Data was missing for 11% of the weight gain during pregnancy and 26% of the maternal weights one month after delivery. Missing data was due to non-response to the questionnaires. We tested the Missing at random assumption with t-tests for numeric variables, chi-squared test for categorical variables and an overall test of randomness (Little’s MCAR test). Most results for t-tests showed statistical significance (p-value < 0.05) but most of the chi-squared tests did not. Little’s MCAR test showed statistical significance. Thus, we concluded that the missing data was not missing completely at random (MCAR), but was missing at random (MAR) as there might be systematic differences between the missing and observed values, but these can be entirely explained by other observed variables [31]. Missing data related to the mothers’ body weight was imputed using the MICE multiple imputation package in R with predictive mean matching (pmm), and we set the number of imputed datasets to 100 [32, 33]. Toxicokinetic model simulations were performed for each mother-child pair, for each chemical, and for each of the imputed datasets.
The covariates were selected using a directed acyclic graph (DAG), an approach that reduces the degree of bias that can occur when measuring causal relationships between exposure and effect [34] (see Fig. S1 in Supplementary Material). Potential confounders included total breastfeeding duration (months, continuous [35];), weight gain during pregnancy (kg, continuous [36];), pre-pregnancy BMI (kg, continuous [36];), and fish intake (g/year, continuous [37];).
We calculated sex and age-specific BMI z-scores using the World Health Organization’s growth charts, which is an international standard taking into account children from different ethnic and cultural backgrounds [38]. Summary statistics were computed for our study outcome, exposures and covariates of interest. We performed correlation analyses to evaluate the relationship between measured (breast milk levels) and estimated (children’s levels) exposure metrics. Multiple linear regression analyses were performed to examine the relationship between exposure metrics (i.e., breast milk levels, estimated children’s levels for the 0–6, 6–12, and 12–24 months periods) and BMI z-scores at 42 months of age. Exposure metrics were ln-transformed prior to regression analyses. Two statistical models were evaluated. In Model 1, potential covariates identified in the DAG were kept in the regression models if they were correlated with the BMI z-scores and exposure metrics with a p-value < 0.2. In Model 2, all potential covariates identified in the DAG were included in the regression model. Analyses for both models were carried out including all children, and stratified by sex. We then looked at variance inflation factors (VIFs) to assess multicollinearity, and verified that the residual plots did not display unwanted patterns. Statistical analyses were performed in R, version 1.1.456 [32].