Long-term exposure to prenatal fluoride and social-emotional and behavioural outcomes in children in the New Hampshire Birth Cohort Study.

2.1. Study sample

This study included mother-child pairs from the New Hampshire Birth Cohort Study (NHBCS), originally designed to examine the impact of environmental contaminants found in unregulated water systems on the health of pregnant women and children living in rural New Hampshire (Gilbert-Diamond et al., 2011). Between 2009-2013, an unselected sample of pregnant women (18-45 years old) were recruited from prenatal clinics at approximately 28 weeks’ gestation. Eligibility criteria included receiving prenatal care at one of the study clinics in New Hampshire, using a private household water system, English literacy, carrying a singleton pregnancy, and having resided at the same address since their last menstrual period and not planning to move. Participants provided written informed consent in accordance with the guidelines from the Committee for the Protection of Human Subjects at Dartmouth College. The the analysis was restricted to mother-child pairs who had at least one maternal toenail fluoride measurement and/or a water fluoride measurement, a child behavioural assessment at age 5, and complete data on primary covariates described below (Figure 1).

1. Download: Download high-res image (321KB)
2. Download: Download full-size image

Fig. 1. Flow diagram of participant selection and analytic sample.

2.2. Maternal toenail fluoride

Women provided toenail clippings at approximately 28 weeks’ gestation and 5 weeks’ postpartum (collection procedures are reported elsewhere (Karagas et al., 2000)). Given their slow growth rate, toenail fluoride generally represents exposure approximately 3-6 months prior (Buzalaf et al., 2006; Whitford et al., 1999). Thus, collected toenails represent fluoride exposures occurring in early and mid-to-late-pregnancy (hereafter referred to as early- and late-pregnancy fluoride). The current study analyzed clippings from the lesser toes (i.e., excluding the big toe) based on the samples available in this cohort.

Toenail samples analyzed for fluoride were selected from a subsample of participants whose children completed intelligence testing at age 5 which was a primary outcome of interest (Goodman et al. In Preparation). Samples with a minimum mass of 1 mg were shipped to Indiana University School of Dentistry’s fluoride laboratory for analysis. Toenails were analyzed for inorganic fluoride using a modification of the hexamethyldisiloxane (HMDS; Sigma Chemical Co., St Louis, MO, USA) microdiffusion procedure, developed and validated for nails (Buzalaf et al., 2012; Whitford et al., 1999). The electrode’s limit of detection (LoD) was 0.019 µg/g fluoride. Following this analysis, toenail mass was found to influence the reliability of the HMDS-facilitated diffusion method, where samples <2.5 mg were not found to yield reliable results (Tamayo-Cabeza et al., 2026a). Therefore, participants whose toenail sample weights were below 2.5 mg were excluded from the early-pregnancy (n = 81) and late-pregnancy (n = 51) analyses.

2.3. Water fluoride

At enrollment, women received detailed instructions and materials to collect tap-water samples from their kitchen faucet (Karagas et al., 2000). If water filters were used, then one filtered and one unfiltered sample was requested. Water samples were analyzed for fluoride using a fluoride ion-selective electrode (Martínez-Mier et al., 2011), with an LoD of 0.02 mg/L.

At enrollment, women were asked: “Overall, what percentage of the water you use for drinking and cooking comes from your home tap water?”. Seventeen women reported “never,” and were excluded from the water analyses. For quality control purposes, duplicate samples were analyzed (n = 64; r = 0.97) to ensure reliability; the average water fluoride concentration was used for these participants. When both filtered and unfiltered samples were available (n = 42), filtered values were used (assuming these values more likely reflected their actual intake); when two filtered samples were provided (n = 3), their values were averaged.

2.4. Outcomes

Children’s social-emotional and behavioural functioning was assessed using the Behavior Assessment System for Children, Second Edition-Parent Rating Scale Preschool Form (BASC-2 PRS-P) (Reynolds and Kamphaus, 2004). The BASC-2 includes eight clinical scales of maladaptive behaviours. These scales are used to generate composite behaviour measures (Reynolds and Kamphaus, 2004): the Behavioural Symptoms Index (derived from the Aggression, Attention Problems, Atypicality, Depression, Hyperactivity, Withdrawal scales), Externalizing Problems (derived from the Aggression and Hyperactivity scales), and Internalizing Problems (derived from the Anxiety, Depression, and Somatization scales). The scales and composites are age-standardized using T-scores, with a mean of 50 and a standard deviation of 10. Higher scores on the maladaptive scales and composites indicate more problematic behaviour, with T-scores >60 considered at-risk. Participants were excluded from the analyses if any of the BASC-2 validity indices were scored as “extreme caution” (n = 5) (Reynolds and Kamphaus, 2004). The analyses focused on the age-5 assessment because ADHD-relevant behaviours, including those captured by the BASC Hyperactivity and Attention subscales, become more predictively valid for later clinical outcomes from approximately age 4 onward.

2.5. Potential covariates

A directed acyclic graph (DAG) was created to facilitate the selection of covariates (Supplemental Figure 1). We considered maternal age at enrollment, maternal education, child sex assigned at birth, recruitment clinic location, household income during pregnancy, maternal marital status, maternal smoking and alcohol consumption during pregnancy, prenatal dietary quality (alternate Healthy Eating Index [AHEI-2010]) (Chiuve et al., 2012), pre-pregnancy body mass index (BMI), parity, maternal depressive symptoms (Beck Depression Inventory, 2^nd edition (Beck et al., 1996)), breastfeeding duration, maternal race/ethnicity, parenting quality (Parenting Relationship Questionnaire [PRQ], measuring attachment, discipline practices, involvement, parenting confidence, and relational frustration (Rubinic and Schwickrath, 2010)), maternal intelligence (Kaufman Brief Intelligence Test, Second Edition(Kaufman, 2004)), and home environment quality [Home Observation for Measurement of the Environment Short Form (HOME-SF) (Mott, 2004).

Additional covariates considered include gestational age, birth weight, and concentrations of As, Pb, Cd, and Zn in toenails collected at 28 weeks’ gestation and 5 weeks’ postpartum and measured by inductively coupled plasma mass spectrometry (Agilent 7900x; detection limits: As and Pb, 0.02 µg/g; Cd, 0.005 µg/g; Zn, 5.0 µg/g (Doherty et al., 2020; White et al., 2018); values below the LoD imputed as LoD/?2).

2.6. Statistical analyses

The preregistration is available at https://osf.io/4reuh/. Deviations from the analytic plan are detailed alongside the plan and in Appendix A. The overall modeling strategy was to use the composite BASC-2 scores as primary outcomes due to their better psychometric properties compared with clinical scales (Reynolds and Kamphaus, 2004) and the BASC-2 clinical (subscales) as secondary outcomes.

In primary analyses, generalized estimating equations (GEE) were applied with an independent correlation structure (Farmus et al., 2021) to estimate associations between maternal toenail fluoride concentrations reflecting early- and late-pregnancy exposures, and BASC-2 composite scores, modeled dichotomously (T > 60, at-risk). According to the DAG, the minimum sufficient set of covariates to avoid overadjustment bias (Schisterman et al., 2009) – the primary covariates – included maternal age at enrollment, maternal education, child sex assigned at birth, and study site. These covariates were selected a priori from the full set of covariates. Each model included a main effect for toenail fluoride (per 1-µg/g increase), a two-way interaction between exposure window and fluoride concentration, a two-way interaction between sex and exposure window, a two-way interaction between sex and fluoride concentration, and a three-way interaction between sex, exposure window, and fluoride concentration. Interaction terms were included in the model regardless of their statistical significance to provide estimates by exposure window and sex in keeping with the study hypotheses. Multiple logistic regression were used to assess associations between water fluoride concentrations and BASC-2 composite outcomes, including effect modification by child sex.

Secondary analyses used GEE to estimate covariate-adjusted associations between maternal toenail fluoride concentrations reflecting early- and late-pregnancy exposures, and the individual BASC-2 clinical scores, modeled dichotomously (T ? 60, at-risk). Secondary analyses also used multiple logistic regression to assess associations between water fluoride concentrations and individual BASC-2 clinical scores.

In sensitivity analyses, multiple logistic regression were used to estimate unadjusted, primary covariate-adjusted, and secondary covariate-adjusted associations between toenail fluoride concentrations at each time point and social-emotional and behavioural outcomes. Secondary covariates were from the remaining covariates considered in the DAG and retained if p < .20 or if inclusion altered the fluoride coefficient by > 10%. Retained secondary covariates included parity, maternal depression, HOME-SF score, household income, and PRQ subscales (attachment, involvement, parenting confidence, relational frustration). Sensitivity analyses were also performed using multiple imputation by chained equations for missing toenail data (n = 45% overall; n = 14% due to sample weight <2.5) and covariates (n = 6%). The imputed datasets were analyzed with multiple logistic regression models adjusted for primary covariates. Furthermore, sensitivity analyses were run with logistic regression models adding, one-at-a-time, gestational age, birth weight, and toenail concentrations of As, Pb, Cd, and Zn to the minimum (primary) set of covariates. Further sensitivity analyses were conducted with the outcomes modelled as continuous.

Standard errors for all analyses were derived from the regression models using default model-based (maximum likelihood) estimators-logistic regression for dichotomous outcomes and linear regression for continuous outcomes. Across all models, regression diagnostics were assessed as appropriate for the type of model (normality of errors, variance inflation factors, residual plots, nonlinearity checks, and influential observations) and no violations of regression assumptions were observed. In additional sensitivity analyses, analyses were run removing any potential influential observations, defined as those exceeding an absolute DFBETAS threshold of 0.2 for the toenail fluoride coefficient (Harrell, 2015). Statistical significance for main effects was set at p < .05, and interaction terms were set at p < .10. Analyses were conducted in R (version 4.4.3) and Stata/SE 18.5.

3.1. Fluoride measurements

Toenails were collected at a mean (SD) 28 (4) weeks’ gestation and 5 (5) weeks postpartum. Median (IQR) toenail fluoride concentrations reflecting early- and late-pregnancy exposures as well as water fluoride levels were 2.78 (1.92) µg/g, 2.15 (1.41) µg/g, and 0.20 (0.40) mg/L, respectively. Spearman rank correlations showed significant positive associations among toenail and water measures, ranging between 0.26 and 0.33 (Supplemental Table 4).

3.2. Fluoride and social-emotional and behavioural outcomes

Each 1-µg/g increase in toenail fluoride concentrations reflecting early pregnancy exposure was associated with higher odds of scoring in the “at-risk” range for Externalizing Problems (OR = 1.39, 95% CI, 1.15-1.67, P = .001). For late-pregnancy fluoride, the point estimate was similar, although the confidence interval was wider and included one (OR = 1.30, 95% CI, 0.99-1.71, P = .055; Fig. 2). No statistically significant two- or three-way interactions involving sex, exposure window, and fluoride concentration were identified (Supplemental Table 5). Water fluoride concentration did not show statistically significant associations with the composite scales, and there were no significant interactions by sex (Table 2).

1. Download: Download high-res image (203KB)
2. Download: Download full-size image

Fig. 2. Adjusted odds ratios (95% Confidence Intervals) from GEE models for BASC-2 composite scores per 1 µg/g increase in maternal toenail fluoride concentrations, by exposure window and child sex estimates from GEE models, adjusted for maternal age at enrollment, maternal education, child sex, and study site (N = 358). Outcomes are modeled dichotomously (T-score > 60, at risk). Points represent odds ratios; horizontal lines indicate 95% confidence intervals. Dashed vertical line marks the null value (OR = 1).

Table 2. Adjusted odds ratios (95% CI) from multiple logistic regression models of BASC-2 composite scores per 1 mg/L increase in water fluoride concentration, by sex (N = 381).

Outcome	Full Sample	Boys	Girls	Sex p-int
Empty Cell	OR (95% CI), p-value			Empty Cell
Composites
Behavioural Symptoms	0.99 (0.69, 1.43), .975	1.08 (0.71, 1.64), .707	0.92 (0.54, 1.55), .739	.620
Externalizing	1.12 (0.84, 1.50), .447	1.24 (0.92, 1.67), .162	0.89 (0.47, 1.66), .705	.345
Internalizing	1.00 (0.78, 1.28), .997	1.03 (0.76, 1.41), .847	0.96 (0.65, 1.41), .831	.773

Abbreviations: CI, confidence interval; p-int, P value for interaction.

Model covariates: maternal age; maternal education; child sex; study site.

Sex p-int for sex × fluoride interaction.

Regarding the individual clinical scales, each 1-µg/g increase in toenail fluoride reflecting early pregnancy exposure was associated with higher odds of scoring in the “at-risk” range for Hyperactivity (OR = 1.20, 95% CI, 1.01-1.43, P = .04). For late-pregnancy fluoride, estimates were imprecise but suggested elevated odds of Hyperactivity (OR = 1.24, 95% CI, 0.97-1.58, P = .09). There was also a statistically significant sex difference in the association of fluoride with Depression (sex p-interaction = .04): among girls, each 1 µg/g increase in toenail fluoride concentration reflecting early pregnancy exposure was associated with 1.37 higher odds of scoring in the “at-risk” range for Depression (95% CI 1.09-1.71, P = .006), whereas no association was observed in boys (OR = 0.99, 95% CI, 0.79-1.23, p = .90). Furthermore, there was a sex difference in the relation between late-pregnancy fluoride and both Anxiety and Withdrawal (p-interaction < .10), with increased but imprecise odds of being “at risk” in girls (respective OR = 1.24 and 1.35), but no association in boys (Supplemental Table 6).

No statistically significant differences in the association of toenail fluoride with outcomes by exposure window were found (Supplemental Tables 5 and 6). Water fluoride concentration did not show statistically significant associations with the clinical scales, and there were no significant interactions by sex (Supplemental Table 7).

3.3. Sensitivity analyses

As shown in Supplemental Fig. 2 and Supplemental Table 8, the results were similar across analysis strategies (using GEE or logistic regression to assess dichotomous outcomes); covariate inclusion (unadjusted, adjusted for primary covariates or primary covariates plus birth weight, gestational age and metals); or when using multiple imputation of missing data. However, when the full set of secondary covariates was included, effect estimates for Externalizing Problems increased, although confidence intervals widened likely due to the decreased sample size. To evaluate whether these differences reflected sample restriction, post-hoc analyses controlling for primary covariates only were conducted among the subsample of participants with complete data on all secondary covariates. In these analyses, effect estimates for associations between toenail fluoride concentrations and behavioural composite scores were similar in magnitude to those from the primary models using the larger sample.

Across primary models, between 8 and 13 observations were identified as potentially influential. Sensitivity analyses excluding these observations preserved the direction and statistical significance of all primary findings (Supplemental Table 9).

When outcomes were modeled continuously, toenail fluoride concentrations reflecting early pregnancy exposure demonstrated positive associations with all the composite scores, not just Externalizing Problems, although only Internalizing Problems reached statistical significance (Supplemental Tables 10 and 11).

4.1. Strengths and limitations

Strengths of this study include the use of maternal toenails, a biomarker reflecting cumulative exposure over the preceding 3-6-months of exposure), measured at two timepoints in pregnancy; a well-characterized cohort with rich covariate data; and extensive sensitivity analyses that yielded consistent results. Effect estimates were larger in models adjusted for secondary covariate than models adjusted for primary covariates suggesting that residual confounding is likely to have biased results toward the null, and that our primary estimates may underestimate the true associations.

Several limitations warrant consideration. First, variability in toenail growth rates could cause introduce exposure misclassification. Second, we did not measure postnatal fluoride exposure from diet (e.g., drinking water, formula prepared with fluoridated water), or from ingestion of fluoridated toothpaste. If prenatal and postnatal exposures are correlated, residual confounding is possible, and we cannot exclude the contribution of postnatal exposure to the observed associations. Third, selection into the analytic sample may have introduced bias, as mothers of children who completed the BASC-2 differed modestly on some demographic characteristics. We adjusted for maternal age, education, and infant sex in all analyses, though we cannot exclude residual bias from unmeasured factors associated with both exposure and outcome. Fourth, participants were rural New England residents who were predominantly White and college-educated and who relied on private water systems. Whether these findings extend to populations with optimally fluoridated water, or to populations differing in demographic compositions, or in sociocultural contexts that shape both exposure and child development, requires further investigation. Fifth, although we tested multiple hypotheses, all analyses were specified a priori, and associations observed concentrated in the domains and directions hypothesized in advance. Finally, social-emotional and behavioural outcomes were assessed at a single preschool timepoint, and the small proportion of children in the at-risk range (e.g., 6% for Externalizing Problems) limited statistical power. Nevertheless, associations were robust across sensitivity analyses.