Population
The PELAGIE (Perturbateurs Endocriniens: Étude Longitudinale sur les Anomalies de la Grossesse, l’Infertilité et l’Enfance) mother-child cohort study has previously been described in detail [14]. Overall, 3421 pregnant women were included between 2002 and 2006 in Brittany, France, before the 19th week of gestation (median 10 weeks, interquartile range (IQR) 8–11). Mothers were asked at inclusion and when the child had turned 2 and 6 years old to complete self-administered questionnaires, focused on medical and social characteristics of the child and its family, including lifestyle, domestic habits, and occupation.
Among the 3323 women who gave birth to live-born singletons, 508 were lost to follow-up before the child turned 2. Among the 2815 remaining women, 1506 responded to the 2-year-old questionnaire, and 947 also responded to the 6-year-old questionnaire. Among these 947 women, we restricted our sample to the women who reported working at the beginning of the pregnancy (N = 866). We excluded families if the child was born before the 35th week of gestation (N = 8), had a chromosomal anomaly (N = 1), had severe head trauma during childhood (N = 1), or if a parent or sibling died during childhood (father: N = 2; sibling: N = 3), as well as children with missing behavior scores at age 2 and/or age 6 (N = 58). The eligible sample included 851 children (see Flowchart in Additional file 1: Figure S1).
All mothers provided written informed consent, and the appropriate French ethics committees approved the study procedures.
Assessment of child behavior
At the 2-year-old follow-up, parents assessed their children’s behavior using items derived from the Child Behavior Checklist (CBCL) [15, 16] and the Preschool Social Behavior Questionnaire (PSBQ) [17]. Items were 3-point Likert scales ranging from 0 to 2 points (0 = never, 1 = sometime, 2 = often), summed into four different subscales: a score for the attention deficit/hyperactivity subscale was computed from six items, and scores for the aggression, opposition, and emotionality subscales were derived from 3 items each. The detailed list of items is presented in Additional file 1: Table S1, further details about construction of the subscales are provided elsewhere [13].
The year the children turned 6, parents assessed their behavior using four subscales of the French version of the Strengths and Difficulties Questionnaire (SDQ) [18, 19]: emotional symptoms, peer relationship problems, conduct problems, and hyperactivity-inattention. These subscales were derived from 20 items scored from 0 (not true) to 2 (certainly true). When children had two missing SDQ items or more within a subscale (N = 3), no score was calculated. In 40 children with one missing item, the corresponding subscale score was extrapolated from the remaining available items. The sample included 793 children with available behavior subscales.
For all subscales at age 2 and 6, higher scores indicate more potential problems.
Exposure assessment
At inclusion (≤ 19 weeks of gestation), women were asked whether in their current job, they were using, producing, or exposed to one of 11 groups of products known to contain organic solvents (paints, strippers, varnishes, dyes, inks, glues, gasoline, grease remover, detergents and cleaning agents, textile treatment agents, and cosmetics) according to a 3-level scale (never, occasionally, regularly). Women reporting regular exposure to at least one group of products were considered “regularly exposed”; women reporting occasional exposure to at least one group of products were considered “occasionally exposed”; and finally, women were classified “not exposed” if they reported no exposure to any of these products [14]. The exposure status was available for 95.5% of the working women. We then excluded the 42 women who declared at the 2-year follow-up that they had changed job position in the late pregnancy. The final sample for analysis included 715 mother-child pairs.
Statistical analysis
We addressed the relations between the multidimensional and longitudinal child’s behavior data within the structural equation modeling (SEM) framework.
To deal with the multidimensional data, we defined latent variables representing internalizing, and externalizing behaviors at age 6, as suggested by Goodman et al. [20] for the SDQ, and we applied the same rationale at age 2 (Fig. 1). The internalizing behavior trait was thus defined at age 2 by the emotional symptoms subscale, at age 6 by the emotional symptoms and peer relationship problems subscales. The externalizing behavior trait was defined at age 2 by the attention deficit/hyperactivity, aggression and opposition subscales, at age 6 by the hyperactivity/inattention and conduct problems subscales. Covariance between the two latent traits at each age was estimated.
The longitudinal approach was considered by regressing each latent trait at age 6 linearly on the two latent traits at age 2, and by decomposing the associations between exposure and behavior latent traits at age 6 (total associations) into direct and indirect pathways. The indirect pathway represents the associations at age 6 that are determined by the associations between exposure and behavior traits at age 2 and the correlations between behavior traits at age 2 and 6. The direct pathway represents the marginal association at age 6, after adjusting for the associations at age 2.
First, a crude structural model representing the relations between latent behavior traits at ages 2 and 6 was fitted (Fig. 1). Each latent trait at age 6 was regressed linearly on all the two latent traits at age 2. All models’ parameters were estimated with a weighted least squares procedure (WLSM), which provides robust estimators and standard errors when the normality assumption is violated for observed variables [21]. Factor loadings and latent traits were standardized for easier interpretation. Parameters were considered significant if their 95% confidence intervals (95% CIs) did not include 0. A Satorra–Bentler scaled Chi-square statistic with P > 0.05, a RMSEA< 0.06, a CFI > 0.9, a GFI > 0.9 and a SRMR< 0.05 were considered to indicate good fit of the model [22].
In a second phase, we included the prenatal solvent exposure variable as a categorical variable. Both latent behavior traits at age 2 and age 6 were linearly regressed on the exposure variable. The association between exposure and each latent behavioral trait is interpreted as the mean change in the latent trait (expressed in number of standard deviations), expected for exposed compared to unexposed children.
All regression parameters were a priori adjusted for maternal age (continuous); maternal education (≤12 years, > 12 years); parity (0, ≥1); child’s sex; maternal tobacco consumption at the beginning of pregnancy (no, < 10 cig/day, ≥10 cig/day); breastfeeding (none, ≤16, > 16 weeks); mother-child interaction at age 2. This variable was based on five items collected at the 2-year follow-up on the activities shared with the child (e.g., singing, playing, and reading) (see Additional file 1: Table S2; see also Pelé et al. [13]). Missing values for covariates (< 3.6%) were imputed by the mode for categorical variables, and by the median for continuous variables.
Finally, we stratified our analyses by sex to explore possible differential pathways and associations in boys and girls. We tested the measurement invariance of the latent behavioral traits between boys and girls by comparing a restricted model constrained to estimate equal factor loadings in both sexes and the unrestricted model. The chi-square change (and associated p-value) was used to state whether measurement models were equivalent. SEM analyses were performed with the lavaan package, R software V3.3.0 [23].