National Toxicology Program

Iraq & Seychelles Study Group Comments

Thank you for the copy of the Executive Summary of the workshop. Overall it is an excellent account of the main issues raised in the workshop. We compliment you on a job well done under challenging circumstances. We have included below our comments on specific points in the summary.

Page 5, lines 5-7
"...the Seychelles was predicated on the occurrence of severe clinical

neurological findings...." In fact the Seychelles is a confirmatory study based on findings of development delays in apparently healthy children.

Severe neurological effects stemming from prenatal exposures had already been known since the early 1960s based on the Minamata outbreak.

Page 6, lines 32-33
"Cord tissue may be a better matrix for measurement of lipophilic substances such as polychlorinated biphenyls (PCBs)." The statement appears to contradict what is stated in other parts of the executive summary and in the reports from the panels. It is advised that PCBs in cord tissue be tested to see if they correlate with levels in traditional matrices such as serum or breast milk (see pages 11 & 21).

Page 7, lines 12-13
"... analyses of trimester-specific exposure would be a valuable

addition...." Trimester measurements were made in the Seychelles and already reported (Cernichiari et al., 1995). A high degree of correlation was found between mercury levels in each trimester such that correlation with mercury levels would be the same whether based on a specific trimester or on average levels for the whole of gestation.

Page 9, section 4.4 Influence of age
This section discusses the issue of the preferable age at testing. We respect the views of the panel but this area is subject to disagreement even among experts. The main point about the Seychelles study was absent from this discussion, that is to say that the Seychelles is a longitudinal study where the children are followed up regularly in order to keep track of any environmental factors that might operate to affect performance in the neurobehavioral tests. We believe a longitudinal study is preferable to a single cross-sectional set of observations made at only one point in the age of the child. The question of past influence of child performance and on the permanence or reversibility of any observed effects can only be answered by a continuing longitudinal follow-up of the population under study.

Page 10, lines 1-8
This discussion of methods for adjusting for age differences at testing is moot as far as the Seychelles is concerned. As stated on line 1, "The

Seychelles study controlled for age by testing all children at the same age...."

Pages 15 and 16
In the discussion of covariates, the possibility of "overfitting" is suggested in the analysis strategy employed in the SCDS. The term has two different interpretations that should be distinguished in discussing this question. The first and most important for the SCDS is the idea of including what are sometimes called mediating variables in the list of covariates. Such variables are really intermediate outcomes, which are "on the causal pathway" between the exposure and the final outcome, and adjustment for such variables can eliminate an association which is actually present in the data.

Two possible such variables are mentioned on page 16, birthweight and length of breast feeding.

At the workshop the Seychelles team was asked about the correlation between mercury levels in maternal hair and the child's birthweight. This question prompted us to examine birthweight as an endpoint. The same full and reduced regression models were used as in previous analyses of data from the SCDS Main Study, each with and without an interaction between mercury level and gender. For the full model the interaction was only marginally significant (p = 0.10) while for the reduced model the interaction was statistically significant (p = 0.05). Estimation of separate slopes for males and females showed that both of the individual slopes were positive. In the reduced model, the slope (SE) for males was 0.015 (0.005) and for females 0.0008 (0.005)), although only the slope for males was significant (p = 0.0038 for males and p = 0.88 for females). This result is consistent with previous findings of what appear to be beneficial effects in males in the SCDS Main Study. Thus. if anything. birthweight would appear to be mediating a beneficial effect, and inclusion of this variable as a covariate may have masked other beneficial effects presumably associated with the nutritional benefits of fish consumption. With regard to breast feeding we would argue that this is another example of a variable that represents both mercury exposure (postnatal in this instance) and nutritional benefits. The issue is not overfitting, but the confounding of these beneficial and toxic factors.

The second possible occurrence of "overfitting" is when covariates are included in the regression analysis whose coefficients in the model are exactly zero. This can lead to inflation of the estimated standard errors of the other coefficients in the model and consequently to a loss of power. We would argue that our use of full and reduced models with prespecified covariates avoids this problem and provides the strongest possible confirmatory evidence. Furthermore, although the report mentions the problem of overfitting with its consequent loss of power, the more serious problem of underfitting, which results in biased estimates of association, receives scant emphasis. Of the two problems, bias would seem to be the more serious. On the basis of the hypotheses generated from the Iraq study, we viewed the SCDS as a confirmatory rather than an exploratory study, and we continue to believe that a prespecified analysis plan results in the strongest possible inference.

We especially appreciate the opportunity you have given us to respond to the Executive Summary.

Return to Workshop Table of Contents