Study population
The TIMOUN birth cohort included 1068 pregnant women during their second- or third-trimester prenatal visit recruited at public hospitals and antenatal care dispensary between 2004 and 2007 in Guadeloupe archipelago (FWI). Eligible participants resided in Guadeloupe for more than 3 years. Around 7% refused to participate mainly because of refusal of the spouse, not wishing to participate in the follow-up, and not wishing to provide biological samples (Kadhel et al. 2014). Mothers answered a standardized questionnaire during a face-to-face interview at enrollment. This questionnaire included sociodemographic characteristics, occupational, medical, and obstetrical information. Gestational age (in weeks of amenorrhea) was estimated by obstetricians in charge of the follow up of the pregnancy. At delivery, dietary habits and alcohol consumption during pregnancy and newborn’s health data (including birth weight) were collected and a cord blood sample was obtained to document prenatal exposure to chlordecone and other environmental contaminants. At 7 years-old, the whole cohort of liveborn singleton children (N = 1033) were invited to participate in a clinical examination, 444 families could not be contacted, refused to participate, or were excluded for other reasons (Supplementary material; Figure S1). At 7-years, 576 of the participating children underwent a neuropsychological evaluation. A maternal interview provided information concerning current health and past medical history, lifestyle, duration of breastfeeding, child behavior and other characteristics. Children’s blood samples were also obtained at age 7. The study was approved by the relevant ethical committee for studies involving human subjects (Comité de Protection des Personnes Sud-Ouest et Outremer III; n° 2011-AOOSSI--40). Each parent provided written informed consent.
Assessment of intellectual abilities
We assessed children’s intellectual abilities with the French validated version for France of the Wechsler Intelligence Scale for Children, fourth edition (WISC-IV; Wechsler 2003). The eight core subtests required to compute the Full-Scale IQ score (FSIQ) were performed. The administration of these core subtests provides scaled scores standardized for age, which are combined to obtain four composite scores in domains of verbal comprehension (Similarities and Vocabulary), processing speed (Coding and Symbols), working memory (Letter-Number Sequencing and Digit Span), and perceptive reasoning (Block Design and Matrix Reasoning). The sum of the 8 scaled scores provides the FSIQ score, a measure of global intellectual functioning, which was our primary outcome of intellectual abilities.
Behavior assessment
Child behavior problems were documented with the French version of the Strengths and Difficulties Questionnaire (SDQ), a 25-items screening questionnaire completed by the child’s parent. Items are representing attributes, some positives and other negatives, scored on a 3-point Likert scale: 0 (“not true”), 1 (“somewhat true”), and 2 (“certainly true”), documenting emotional symptoms, conduct problems, hyperactivity/inattention, peer relationship problems and prosocial behavior. There are strong theoretical and empirical supports for classifying behavior problems according internalizing and externalizing-types of problems. The SDQ’s emotional symptoms (5 items) and peer relationship subscales (5 items) are usually combined into an internalizing problems subscale, as are the conduct (5 items) and hyperactivity/inattention (5 items) subscales into an externalizing problems subscale (Goodman et al. 2010), and higher scores indicate higher difficulties. The SDQ is adapted for many different cultures and languages and has demonstrated excellent psychometric properties (http://www.sdqinfo.com/).
Biomarkers of exposure to chlordecone and other contaminants
Blood samples from the umbilical cord and the child at the 7-year visit were collected in EDTA tubes to document respectively prenatal and childhood exposure to chlordecone and other environmental contaminants. Plasma samples were stored at -30 °C in Polypropylene Nunc® tubes following centrifugation. Chlordecone, polychlorinated biphenyl congener 153 (PCB-153), dichlorodiphenyl dichloroethene (DDE) and lipids were measured in plasma. Total mercury (Hg), lead (Pb), and cadmium (Cd) were quantified in whole blood using inductively coupled plasma mass spectrometry. Determination of chlordecone and PCB-153 concentrations were done by the Center for Analytical and Research Technology at Liege University (Belgium). Contaminant concentrations were analyzed by high-resolution gas chromatography (Thermo Quest Trace 2000, Milan, Italy) equipped with a Ni63 electron capture detection. Preparation of samples and quantification method were previously described (Multigner et al., 2010). The limit of detection (LOD) was 0.06 μg/L for chlordecone, PCB-153, and DDE in cord blood, and 0.02 μg/L for child chlordecone, PCB-153, and DDE. Total cholesterol and triglyceride in plasma were determined by standard enzymatic procedures (DiaSys Diagnostoc Systems GmbH; Holzheim, Germany) and total lipid concentrations were calculated as described by Bernert et al. (2007). Blood Hg, Pb, and Cd concentrations were measured by inductively coupled plasma mass spectrometry (ICP-MS) at the laboratory of the Centre de toxicologie du Québec. The LOD for Hg, Pb and Cd were 0.4 μg/L, 2 μg/L and 0.1 μg/L respectively, and each run of samples included a standard.
Covariates and potential confounders
We collected socio-demographic and lifestyle factors and medical history at enrollment, after delivery, and at subsequent follow-ups via administered questionnaires. We used directed acyclic graphs (Supplemental Material; Figure S2) to identify a set of covariates to include in the models from the following potential covariates: child’s age at testing (in months), sex (male; female), birth weight (grams), maternal age at pregnancy, parity (nulliparous; primipara and multiparous), duration of breastfeeding (no breastfeeding; ≤ 6 months; 7-18 months; ≥ 18 months), maternal marital status (married or in a couple; single; living with own family), socio-economic status (SES) indicators during pregnancy based on maternal education (none or some elementary school; some high school; high school diploma; college/university studies) and monthly household income (≤ 800 Euros; 800 – 2300 Euros; > 2300 Euros), maternal nonverbal intellectual abilities (Raven score) assessed with the Raven’s Progressive Matrices (Raven et al. 1998), maternal alcohol consumption (no; yes) and smoking during pregnancy (no; yes). We included in multivariable models, child’s age and sex, maternal age, parity, Raven score, education, marital status, monthly household income, and alcohol and smoking during pregnancy.
Statistical analysis
Chlordecone and other chemicals’ concentrations were log-transformed (base 2) to address skewness and limit the influence of outliers. Values below limit of detection were replaced by LOD/√2 (Helsel & Hirsh 1991). Initial exploratory analyses included descriptive statistics and univariate associations between exposures and measured outcomes, and potential covariates of interest. We also investigated correlations between log-transformed concentrations of environmental exposures using Pearson correlations.
First, we used structural equations modeling (SEM) adjusting for the same set of covariates to investigate simultaneously the associations between cord- and 7-years exposures, with WISC-IV and SDQ scores. We ran two separate analyses for the intellectual and behavioral functions. For the WISC-IV scores, we employed a simple path analysis to incorporate all outcomes in a single model. The results are presented as the absolute change in the test score associated with a twofold increase in chlordecone concentrations. For the SDQ test, we used a two stages confirmatory factor analysis (CFA) model with first order latent functions of conduct problems, hyperactivity/inattention, emotional symptoms, and peer problems indicated by their subsequent measured SDQ item scores (Figure 1). We added two second order latent functions for externalizing problems (indicated by conduct problems and hyperactivity/inattention first order latent functions) and internalizing problems (indicated by emotional symptoms, and peer problems first order latent functions) (Figure 1). We adopted this two-stage CFA model since it has previously been found to show superior model fits and because it was considered theoretically meaningful (Niclasen et al. 2013). All observed indicators showed significant correlations to their latent functions (Figure 1), and the models exhibited an excellent fit to the data (Comparative fit index = 0.92, root mean square error of approximation = 0.04, Standardized Root Mean Square Residual = 0.035, and χ2 p-value: 0.06). For the SDQ SEM model, the results for behavioral functions are presented as a standard deviation (SD) change in the latent functions associated with a twofold increase in chlordecone concentrations. The SEM approach allows to reduce both multiple comparisons testing and measurement errors. Because there were missing data for some covariates and prenatal chlordecone exposures, we used the Full Information Maximum Likelihood estimation, which utilizes all available information and avoids list-wise deletion due to missing data.
We also examined potential effect modification by sex using multi-group SEM analyses. In these analyses, we constrained the loadings, variances, co-variances, and intercepts of the latent variables to be equal across the two groups of females and males. Differences in the associations in the two exposure groups were tested by comparing the value of d/SEd to the standard normal distribution, where d is the difference between the two estimates, and
is the standard error of the difference (Altman and Bland 2003; Oulhote et al. 2017).
In additional analyses, we conducted multivariable linear regression analyses to investigate the associations between cord- and 7-years chlordecone concentrations in relation to the WISC-IV (FSIQ and four composite scores). The results are presented as the adjusted mean difference (aMD) in scores with 95% confidence interval (95% CI) for a two-fold increase in chlordecone concentrations. For the SDQ internalizing and externalizing measured scores, and because scores on these tests exhibit an over-dispersed distribution, we used negative binomial regressions. Negative binomial regressions model the ratio of the mean SDQ scores among exposed and non-exposed. The results are presented as adjusted mean ratios (aMR) of the SDQ externalizing and internalizing problems with corresponding 95% CIs. In all these models, the exposure estimate was reported for a two-fold increase in chlordecone concentrations—for example, an aMR of 1.1 for a two-fold increase in chlordecone suggests that the mean SDQ subscale score was 10% higher for each doubling of the chlordecone concentrations. We also examined sex-specific associations by including a product interaction term (sex x chlordecone concentrations) in all analyses.
We finally explored potential nonlinear relations of chlordecone concentrations (log2-transformed) with WISC-IV and SDQ scores using generalized additive models (GAMs) with penalized smoothing regression splines (Wood 2011) and visually inspected plots of the smoothed data. For the behavioral function (externalizing problems and internalizing problems), we extracted the latent function scores from the factor-analytic component of the SEM and used them as outcome variables, adjusting for covariates (Oulhote et al. 2014). Cord-blood and 7-years child chlordecone concentrations were introduced as spline functions. We assessed departure from linearity (at p<0.1) by comparing the model with chlordecone concentrations introduced as a spline function with chlordecone concentrations introduced as a linear term. None of the dose-response relationships significantly deviated from linearity, all at a p>0.2 (Supplemental Material; Figures S3 and S4).
In sensitivity analyses, we adjusted simultaneously for additional exposures measured in cord blood that have been shown to exert neurotoxic effects, namely, blood Pb, Cd, and Hg, and lipid-standardized PCB-153 and pp’-DDE concentrations.
All significance tests were two-sided, and the level of significance was set at a p-value<0.05 for main estimates and interactions. However, we provide all the estimates, and base our conclusions on the effect estimates in addition to concordant trends rather than solely on hypothesis testing. All statistical analyses were conducted using R version 4.0.3 (R Core Team 2020).