Study design
We used data from a prospective pregnancy and birth cohort study, the Health Outcomes and Measures of the Environment (HOME) study, in Cincinnati, Ohio. The original study was designed to test and quantify the association between low-level prenatal and postnatal exposure to various environmental toxicants with the development of cognitive and behavioral problems in children. We enrolled pregnant women in the study from March 2003 to January 2006. Eligibility criteria at enrollment were: a) age ≥ 18 years, b) between 13 to 19 weeks of gestation, c) residing in a house built before 1978, d) not on a medication for seizure or thyroid disorders, and e) negative HIV status. Of 1263 eligible pregnant women, 468 (37%) agreed to participate in the study and 389 women remained in the study until delivery of a liveborn singleton infant. The details about the study cohort are described elsewhere [24].
Measurement of mercury exposure
We collected whole blood samples from mothers at approximately 16 - and 26 weeks of gestation, and at the time of delivery. We also collected whole blood samples from children at birth (cord blood) and at the 2, 3, 4, 5, and 8 year visits. The whole blood samples were collected in EDTA vacutainer tubes that were lot tested for metal contamination, and stored at − 80 °C until shipment to the United States Centers for Disease Control and Prevention (CDC) for analysis. Total whole blood mercury was quantified using inductively coupled plasma mass spectrometry (ICP-MS) [25]. This multi-element analytical technique is based on quadruple ICP-MS technology. All analytical results were referenced to the National Institute of Standards and Technology (NIST) Standard Reference Materials (SRMs) along with two levels of quality control materials in each analytical run. The limit of detection (LOD) for total mercury was 0.2 μg/L. For results below the LOD (16, 19, 14, 16, and 8% for 16-week, 26-week, maternal blood samples at delivery, cord blood), we imputed a value of the LOD divided by square root of 2 [26].
Assessment of children’s behaviors
Children’s behavior was assessed using the second edition of the Behavioral Assessment System for Children (BASC-2) Parent Rating Scale during follow-up visits when children were 2, 3, 4, 5, and 8 years. The BASC-2 is a reliable and valid assessment of a child’s adaptive and problem behaviors in community and home settings [27, 28]. The scale has four composite scores comprised of 12 clinical subscales. We examined the externalizing problems composite (hyperactivity and aggression subscales), internalizing problems composite (anxiety, depression, and somatization), the behavior symptoms index (BSI) composite (aggression, hyperactivity, depression, atypicality, withdrawal, and attention problems); and adaptive skills composite (activity of daily living, adaptability, social skills, and functional communication). We used T-scores based on combined-sex norms for analysis that are normalized to a mean of 50 and standard deviation (SD) of 10. A score > 60 is considered “at-risk behavior,” with the exception for adaptive skills and corresponding subscales, in which a score < 40 is non-optimal. We were primarily focused on examining the relationship of prenatal mercury with composite behaviors (i.e., externalizing problems, internalizing problems, BSI, and adaptive skills), but we also examined the associations with ADHD (i.e., aggression, hyperactivity, attention) and anxiety subscales as prior studies found that prenatal mercury was associated with these specific behaviors [8, 12, 16, 18].
We also assessed anxiety in children using the self-reported Spence Children’s Anxiety Scale (SCAS) when the children were about 8 years of age. The SCAS assesses six domains including generalized anxiety, panic/agoraphobia, social phobia, separation anxiety, obsessive-compulsive disorder, and physical injury fears [29]. A trained examiner oriented the child to the use of a Likert scale before reading each question to the child and allowing them to circle their answer on the score sheet. We used age- and sex-standardized scores for the analysis.
All assessments for the study were conducted in a standardized clinic setting to minimize distractions. The examiners were blinded to maternal and child mercury concentrations.
Covariates
Our selection of potential confounders was guided by other research studies [16, 18, 30]. We included variables as potential confounders that were associated with both mercury exposure and behavior changes in children but were not known to be on the causal pathway. Fish intake, a source of methylmercury exposure as well as nutrients that are beneficial for brain development, can be an important confounder of the effect of mercury on child behavior [5, 31]. At the baseline visit (about 20 week of gestation), we asked the mothers about fish or shellfish consumption during pregnancy from the estimated date of conception until the survey was taken. We categorized the frequency of reported fish intake as: none, < once per month, 1 to 3 times per month, and > once per week.
We examined other variables as potential confounders, including maternal age at delivery (≤30 years and > 30 years), maternal ethnicity (white and nonwhite), annual household income (≤ $40,000 and > $40,000), maternal education (completed and not completed bachelor’s degree), marital status (married/living with partner and unmarried/living alone), and child sex. Maternal age at delivery and annual household income were categorized based on the frequency distribution of the variables. We also adjusted for the caregiving environment using the Home Observation for Measurement of the Environment (HOME) Inventory conducted during a 12-month home visit [32]. We categorized HOME Inventory scores as < 40 and ≥ 40 based on the distribution of the scores. We also measured maternal depression using the second edition of the Beck Depression Inventory (BDI-II) at 20 weeks of gestation [33]. We further classified the scores as minimally depressed (BDI-II score ≤ 13) and greater than minimally depressed (BDI-II score > 13). In addition, we examined whether children’s blood lead and mercury concentrations, and maternal pregnancy concentrations of serum cotinine, polychlorinated biphenyls (PCBs), urinary bisphenol A (BPA) were confounders or effect modifiers of the associations between prenatal mercury concentrations and child behavior.
Statistical analysis
We defined mean prenatal mercury concentrations as the mean of mercury measured in whole blood collected at 16 and 26 weeks gestation, at delivery from the mothers, and in cord blood from the infants. We used average of the log2-transformed mercury concentrations to reduce the influence of outliers. We compared 389 women and their children’s (original sample) demographic characteristics with 320 dyads who had at least one blood mercury and BASC-2 data. We compared the 320 dyads with at least one mercury and behavior assessment with 69 dyads not included in the analyses because of missing either mercury and behavior assessment. We also stratified mean prenatal mercury concentrations and child BASC-2 scores of 226 dyads who were followed at the 8-year visit by demographics and by maternal and child characteristics.
We examined the associations of prenatal mercury with BASC-2 composite externalizing problems, internalizing problems, and BSI, and adaptive skills using linear mixed models with an unstructured covariance matrix. The linear mixed models account for the correlated nature of the repeated observations. Thus, the coefficient from a linear mixed model would represent an average increase in mean BASC-2 scores for every 2-fold increase in mercury concentrations.
We constructed linear mixed models for composite scores. First, we constructed unadjusted models. Then we added fish intake in the unadjusted models to test for confounding with whole blood mercury. Next, we added participants’ demographics and characteristics, child blood lead concentrations, maternal serum cotinine concentrations with and without child mercury concentrations to the unadjusted models and called this multivariable model 1. We added fish intake to the multivariable model 1 and call this model multivariable model 2. As mentioned before, the choice of the inclusion of the variables in the models was based on the criteria of confounders (associated with both mercury exposure and behavior changes in children but were not known to be on the causal pathway) and also guided by previous research [16, 18, 30]. The findings of the different regression models are presented in Fig. 2. Multivariable model 2 had the lowest Akaike’s information criterion (AIC) values compared to multivariable model 1 and considered this our primary regression model. We then constructed primary regression model for all corresponding BASC-2 subscales. We used locally weighted scatterplot smoothing (LOESS) analysis to examine the shape of the relationship for blood mercury and behaviors in adjusted models. The relationships appeared to be linear (Fig. 3). We used statistical software SAS 9.3 and R for all analyses and GraphPad Prism to prepare some graphs for the manuscript.
Secondary analysis
We conducted secondary analyses to test for evidence of a developmental window of vulnerability for behaviors that were associated with mean prenatal mercury concentrations. We constructed a series of regression models for 16- and 26 weeks gestation, at delivery, cord and mean prenatal mercury concentrations with and without including cord blood concentrations. Not only anxiety, we also examined the association of prenatal mercury concentrations and other composite scores/subscales at 8-years of age. We examined anxiety at 8-years of age in more depth because there is some evidence suggesting an association between prenatal mercury exposure and anxiety [16, 34]. Although this was at higher prenatal mercury levels, it was important for us to examine if such association exists at very low – yet representative – levels of prenatal mercury exposure as we found some evidence of positive association between mean prenatal mercury concentrations and BASC-2 anxiety subscale. So, we also examined the Pearson correlation coefficient between 8-year BASC-2 anxiety subscale and 8-yr SCAS anxiety total scores. Finally, we examined the associations of prenatal mercury concentrations with parent reported BASC-2 anxiety subscale and self-reported SCAS total scores at 8-years using multivariable linear regression models.
Additionally, we examined models stratified by child sex and maternal race/ethnicity. We checked for two-way interactions by adding interaction terms for child age, child sex, maternal race/ethnicity, child blood lead concentrations, maternal cotinine concentrations, and PCBs with mercury at each prenatal visit. Finally, we examined the association between mean of post-natal child blood mercury concentrations and behavior scores in children.