For the purposes of risk assessment, valid dose-effect relationships are essential, and the research in the Faroes has endeavored to represent both ends of the exposure-outcome relationship with the highest possible precision and sensitivity. Because of the need to obtain precise measures of the causative exposure [7, 14], the mercury concentration has been measured in cord blood and other biological samples from exposed participants. Detailed comparison of various prenatal exposure indices within the cohort has shown that the cord-blood concentration is consistently the most precise predictor of nervous system deficits determined during postnatal follow-up [7, 14].

At the same time, valid outcome variables must be sensitive to MeHg neurotoxicity and relatively robust to impacts of confounders. In choosing feasible effect parameters, an important consideration includes suitability for the age group and culture. As confounder-independent outcomes, neurophysiological measurements have provided evidence that developmental MeHg neurotoxicity is detectable through age 14 years [26]. Clinical neurological tests have shown mercury-related adverse effects neonatally [29], but such tests may not be sensitive enough at school age [18]. Standardized neuropsychological tests are useful to gain insight into functional domains and overall cognitive functioning and have been widely applied as sensitive indicators of central nervous system (CNS) dysfunction associated with exposure to neurotoxicants in both environmental and occupational exposure settings [9,39]. In choosing such measures for the study of children, it is important to consider the developmental stage of the child at both exposure and at the time of testing as well as cultural and psychometric parameters of the tests being administered [2,39].

Risk assessments have so far been based on functional outcomes up to early school age obtained in three prospective epidemiological studies [11, 23, 28, 34]. Data from older children and adolescents have been available only from less informative cross-sectional studies. Developmental exposure to inorganic lead is known to cause permanent CNS damage, thus suggesting that deficits attributable to early developmental neurotoxicant exposure may be irreversible [2, 40]. However, long-term effects of prenatal and early childhood exposures to neurotoxicants may become more difficult to demonstrate with time, because an increasing number of other factors play a role in test performance and it may be difficult, if not impossible, to define and measure these covariates. Thus, the increasing complexity of confounder adjustment may cause problems in the statistical analysis of cognitive test data. Despite these potential problems, prospective assessment of cohorts with well defined exposures occurring early in development is essential to understanding how neuropsychological outcome measures can be utilized as long-term indicators of early neurotoxicant insults.

At age 14 years, a total of 878 of 1,010 living cohort members (86.9%) were examined. The 438 boys and 440 girls examined had an average age of 13.83 (S.D., 0.32) years. Most examinations took place at the National Hospital in Tórshavn, the capital of the Faroe Islands, during April–June of 2000 and 2001. For families who had moved to Denmark, examinations were also offered in Odense in November, 2000. Transportation costs and other expenses were reimbursed in accordance with rates approved by the Faroese ethical review committee. Each day, four participants were examined during the morning and four during the afternoon. Each group of participants rotated between four stations staffed by health professionals, who administered the clinical tests, and who had no access to information on individual exposure levels. These stations each lasted approximately one hour each and covered physical examination, neurophysiological tests, and two series of neuropsychological tests, each of which was administered by the same examiner in the same sequence throughout.

The thorough pediatric examination included otoscopy and assessment of neurological function. None of the adolescents had current middle ear infection. A total of 18 participants examined had neurological disorders thought to be independent of MeHg exposure and were excluded from the data analysis. The MeHg exposure of these participants did not differ from that of other cohort members. Of 860 included participants that completed the 14-year examination schedule, 815 were examined also at age 7 years.

Blood specimens and hair samples were again obtained, and the proximal 2-cm hair segment was used for mercury analysis [26]. Detailed quality assurance data showed acceptable analyses for both exposure biomarkers. Results in μg may be converted to nmol by multiplying by 5.0.

The study protocol was approved by the ethical review committee for the Faroe Islands and the U.S. institutional review boards, and written informed consent was obtained from all parents.

Table 3 shows associations with potential confounders. Continuous variables were trichotomized for the purpose of this table only. The expected pattern of associations reflects dietary habits and local differences in availability of whale meat during 1986–1987. Accordingly, participants from fishing villages had higher prenatal exposures than those living in towns or in Denmark. Participants tested in Denmark or in the Danish language had lower exposure. Lower exposure was also associated with a higher maternal Raven score. Adolescents from smaller primary schools had higher exposure, as had participants examined in the afternoon, at an older age, and those living with both parents.

The neuropsychological test results were generally within expected limits, based on Scandinavian and similar norms for this age range (Table 4). However, the Faroese adolescents performed above expectation given available norms for Block Designs, with a marked ceiling effect during the first year of examinations. Adding the three last designs from the adult version (WAIS-R) corrected for this imperfection in resolution at the high end, but was available for less than half of the participants.

All multiple regression analyses were based on complete cases (Table 5). Thus, the number of participants included depended upon the choice of exposure biomarker and confounders. Higher prenatal MeHg exposure was associated with lower finger tapping scores, increased reaction time, and lower cued naming scores; the cord-blood mercury concentration was a significant predictor of outcomes of all three types. The maternal hair mercury concentration showed significant or near-significant associations with deficits only on the three conditions of finger tapping, and the two measures of reaction time. The cord tissue mercury concentration showed no clear relation to finger tapping and reaction time results, but was associated with deficits on the naming and verbal learning results. One surprising find was that higher MeHg exposure was associated with better scores on WMS-III Spatial Span.

The estimated mercury effects remained virtually unchanged after adjustment for the extended set of confounders, except that, in this reduced data set, the p-value for the effect on the cued naming score was slightly above a 5% significance level. Some outcomes were transformed to obtain normally distributed residuals, but this change barely affected the mercury effects. Table 5 therefore shows the effect on the untransformed variables. The logarithmic transformation of the exposure variables assume a linear effect of each doubling of the exposure. The appropriateness of this assumption was confirmed in non-parametric models (Figs 2 and ​and33).

In contrast to the prenatal exposure variables, markers of postnatal exposures were generally only weakly related to cognitive test scores at 14 years, with many coefficients suggesting effects in the direction opposite to expectation. Only one significant association was seen, i.e., for the hair mercury concentration at age 14 years on the NES2 finger tapping score on the preferred hand (effect, −5.18; p = 0.019). This effect appeared less clear when corrected for the full set of confounders (effect, −3.18; p = 0.17) and decreased further when adjusted for the prenatal exposure level (effect, −2.74; p = 0.26). On the other hand, prenatal effects changed only slightly after adjustments for postnatal exposure markers. However, due to limited power in a regression model with both prenatal and postnatal exposure variables, the p-values generally increased slightly. Thus, the p values for the three outcomes significantly associated with the cord blood mercury concentration without adjustment for postnatal exposure increased to levels ranging from 0.061 (CPT reaction time) to 0.11 (finger tapping), after adjustment for postnatal exposure. On the other hand, the Copying test (5 most difficult designs) now showed a significant p-value of 0.03 for the effect of prenatal exposure effects after adjustment for postnatal exposure.

Likewise, PCB exposure, although only available for half of the subjects, showed only weak associations with the outcomes, none of them reaching statistical significance. The p values for the three outcomes showing significant deficits associated with the cord blood mercury concentration (Table 5) varied from 0.15 (CPT reaction time) to 0.98 (finger tapping with both hands).

In general, a high percentage of the outcome variation could not be explained in the regression models and most covariate associations were weaker than at age 7 years [18]. Age was not a general predictor of test performance at this developmental stage and within the limited age range covered, and was only weakly related to scores on verbal abstraction, attention, and manual motor speed tasks. Sex was a stronger predictor of test performance, where boys performed faster on tasks assessing manual motor speed, while girls earned higher scores on most other tasks. No sex difference was seen in vocabulary, nonverbal abstraction, and visuospatial memory. The maternal Raven score was again an important predictor for most outcomes. Residence in towns was positively associated with naming ability and, to some degree, with other abilities. Computer experience was associated with better performance on the continuous performance tasks and finger tapping. However, the full complement of predictors explained only about 10% of the total variation (Table 6). Although prenatal MeHg exposure seemed a less important predictor of neuropsychological performance at age 14 than at age 7 years, this risk factor appeared to represent about the same proportion of the total variance explained by the regression model at the two occasions.

Assuming that prenatal MeHg neurotoxicity causes permanent changes, the mercury-associated effects observed at age 14 years could represent the deficits previously observed at age 7 years, as modified by learning and maturation. An analysis was therefore conducted based on outcomes available from both examinations and with a significant mercury effect at age 7 or 14 years (Table 7). For most outcomes, prenatal MeHg exposure showed only weak effects on the test score difference. The sign of the coefficients generally suggested that highly exposed participants may have been able to compensate in part for deficits seen at age 7 years. However, this effect was weak and only statistically significant for one of the CVLT scores. For two finger tapping scores (non-preferred hand and both hands), this analysis indicated that, between ages 7 and 14 years, highly exposed participants had fallen even further behind those less exposed, but again this tendency was not statistically significant. A supplementary analysis of the relative difference in test scores yielded similar results, but these models provided a poorer fit to the data.

To take into account multiple statistical testing, exposure imprecision and missing data, we then applied structural equation models. The crude, initial models fitted the data poorly, but after adjustment for local dependence and differential item functioning, models with an acceptable fit were obtained for all five test groups (Table 8). The strongest mercury effect was found in regard to the motor and attention groupings. The effect for the verbal grouping is close to being significant at the 5% level, while the effect on spatial and memory groups is far from significant.

Maternal fish intake during pregnancy was then added to the structural equation models (Table 8). Although fish intake appeared to exert a weak beneficial effect, it was not statistically significant. Inclusion of this adjustment changed the mercury coefficients in the direction anticipated.

Significant mercury effects corresponded to a decrease corresponding to 6.5–8.5% of the standard deviation for each doubling in the exposure (Table 8). These results are in accordance with the strongest regression coefficients in Table 5. For motor function, a doubling of the true MeHg exposure corresponded to 1.3 fewer taps on NES2 finger tapping with the preferred hand. As a result, at age 14 years, this effect is slightly stronger than at 7 years. Other effects were about the same or less. In the regression analysis, a strong positive mercury effect had been obtained for Spatial Span, and this tendency led to poorer fit of the structural equation for attention. A better fitting model with a stronger effect would be obtained by excluding this score, but this possibility was not further pursued.

While the effect at age 14 years appeared to be somewhat weaker than seven years before, the overall explanatory power of the statistical models was also more limited than it was previously. A multitude of events and incidents, beneficial and adverse, may have occurred during the time between the two examinations and may have attenuated the correlation between two sets of tests and decreased the association with early-life covariates. For example, differences in time of onset of puberty, together with very high growth rates in this phase of development, may contribute unadjusted variability that overshadows the association with defined predictors, including MeHg exposure. In addition, the covariates and methylmercury exposure might particularly affect the rate at which skills are acquired at the younger age, while innate ability might be a more important predictor of performance at 14 years. Thus, although the neuropsychological performance results at age 14 were more affected by variability that could not be captured by adjustments using known covariates, the relative importance of MeHg exposure was about the same at the two examinations. In addition, analyses that incorporated test results at both examinations showed that the mercury-associated deficits had changed only little between age 7 and 14 years.

When considering the estimated effect size, comparison of regression coefficients with the standard deviations of the test scores is appropriate [19]. Because of the logarithmic transformation of the exposure scale, effects must be expressed in terms of log-scale increases. We found that a doubling of the prenatal MeHg exposure at age 7 years corresponded to a decrease in performance in the range of about 5–10% of the standard deviation [18,19]. The effect at age 14 years was similar in size (Table 8). The regression coefficients obtained for mercury may be compared to those obtained for age. At age 7 years, a doubling in MeHg exposure corresponded to a loss in development by about 1–2 months [18]. At age 14, age relationships for the neuropsychological tests are less steep, and a doubling in the MeHg exposure therefore corresponded to greater delays in development.

While most tests showed a mercury association in the direction predicted, Spatial Span showed an unanticipated positive effect of increased prenatal MeHg exposures. This tendency caused a poor fit of one of the structural equation models. Although a chance finding cannot be excluded, one must also consider the possibility that MeHg exposure may be associated with intake of essential nutrients from seafood, and that, for example, long-chain n-3 fatty acids may have beneficial effects on brain development [1,22]. Partial adjustment for this factor was provided by adding maternal fish intake during pregnancy as a covariate. This addition changed the mercury regression coefficients in the direction anticipated to better reflect the true adverse effect of methylmercury per se. This confounding had previously been considered as an explanation of the weak tendency of decreased (improved) visual evoked potential latencies at higher prenatal MeHg exposure levels [18]. It therefore seems likely that the outcomes of the present study are affected by both contaminant toxicity and beneficial nutrient effects in the opposite direction, although the different seafood components may not affect different brain functions to the same degree. Furthermore, it is apparent that any beneficial nutrients would have compensated only partially for the neurotoxic effects of MeHg in this cohort. However, adjustment for fish intake was based only on maternal recall of the frequency of fish dinners during pregnancy, thus not allowing any precise assessment of essential nutrient intakes. If more precise indicators of nutrient supply had been available, statistical control for their beneficial effects might have revealed an even stronger MeHg effect.

The results of the multiple regression analyses must be interpreted in light of various uncertainties, in particular exposure misclassification [7, 14]. Thus, while our previous studies have shown that the cord-blood mercury concentration is the exposure parameter that best predicts test performance, closely followed by the mercury concentration in cord tissue [17], both parameters are participant to substantial variability that exceeds the normal laboratory error. The mercury concentration in maternal hair is substantially more imprecise and may be severely affected by, e.g., permanent hair treatment [41]. Such exposure imprecision is likely to be random and would therefore bias the regression coefficients toward null. However, the imprecision of the exposure markers was taken into account in the structural equation models by viewing the observed parameters as reflections of a true underlying exposure variable. These more advanced statistical analyses revealed that the grouped test score outcomes, labeled as motor, attention, and verbal functions, had the clearest dose-effect relationships for developmental MeHg exposure.

In both the multiple regression and structural equation analyses, tests or groups of tests that showed the best dose-effect relationships with regard to prenatal MeHg exposure tended to be highly sensitive tests with wide score ranges or extremely accurate outcome measurements. For example, reaction time and manual motor speed as measured by computer are highly accurate. The Boston Naming Test has 60 items, allowing much better discrimination of the subject’s abilities than tests with fewer items. Therefore, the mercury-associations of the tests with superior psychometric properties strengthen the argument that MeHg has a long-term adverse effect on brain development and function. The tests assessing visuospatial skills appeared to have a low ceiling in this population at 14 years, while responses on tests of short-term memory generally had a narrow range, thereby limiting their ability to reveal exposure-effect relationships. Even given these provisos, the confluent finding of effects on tasks assessing attention, manual motor speed, and language/semantic memory suggests that the effects of MeHg at these levels of brain development are multi-focal or diffuse, affecting widely divergent brain structures. The existence of sustained neurotoxic effects in adolescents is also supported by the neurophysiological examinations that were carried out in parallel [26]. Thus, auditory evoked potential latencies at age 14 years increased at higher prenatal MeHg exposures, and the heart rate variability decreased. The notion of permanent effects is also consistent with the known neuropathological effects of prenatal MeHg exposure at much higher levels producing clinical Minamata disease [15,28], but clearly as a much more subtle picture.

Postnatal exposures were generally weakly related to cognitive test scores at 14 years, although postnatal exposures were poorly characterized and averaged only about one-fifth of the prenatal exposures. Significant effects were seen on the finger tapping scores, but these effects could be due to chance and were less certain, when corrected for the full set of confounders. Because the neuropsychological outcomes may be affected by both prenatal and postnatal exposures, and the exposures at different times were interrelated [26], these effects would be difficult to separate. In addition, using the postnatal exposure to adjust the regression coefficients for prenatal exposure is problematic, as this approach will involve two error-prone and associated predictors, thereby causing complex biases they may be difficult to predict [12]. Given these caveats, postnatal exposure in this population did not appear to affect the neurobehavioural outcomes to any appreciable degree. However, our results on peak V of the evoked potentials showed a possibly specific effect of postnatal (recent) MeHg exposures [26], thus supporting the finding that recent exposure could conceivably cause a decrease in motor speed. This effect is perhaps surprising, because the average MeHg exposure of the 14-year-olds was reflected by a hair-mercury concentration of about 1 μg/g, a level that corresponds to the Reference Dose used by the U.S. Environmental Protection Agency [34]. However, these findings are in accordance with several cross-sectional studies carried out in 7-to-12-year-old children, who also showed decreased motor speed associated with increased hair-mercury concentrations [15]. While the hair-mercury concentration in the latter studies reflected the child’s own recent exposure, an association with prenatal exposures could not be ascertained.

PCB is another neurotoxicant present in seafood, and increased exposures to methylmercury and PCBs are correlated [16]. At age 7 years, prenatal PCB exposure was associated with deficits on CPT reaction time, naming, and long-term verbal memory, although these tendencies all but disappeared after adjustment for mercury exposure [16,18]. This result was confirmed in a structural equation analysis that allowed for measurement errors in biomarkers of both MeHg and PCB [6]. No association with PCB exposure was seen at age 14 years.

The findings of the present study add to the substantial evidence on developmental methylmercury neurotoxicity [11,15,23,28]. One previous study carried out in the Seychelles [27] found no significant association between prenatal methylmercury and performance on neurobehavioral tests at age 8 years. However, the two studies are not contradictory, because the Seychelles study cannot exclude an effect at the magnitude observed at the Faroes [24]. In addition, the mercury effect was assessed only from analyses of maternal hair collected six months after parturition and after adjustment for the child’s postnatal mercury exposure [27]. Joint analyses based on three major prospective studies are in excellent agreement with the findings in the Faroes [23,28].