FDA's Midnight Mischief Heightens Mercury Risk to Pregnant Women, Infants : EPA Review of FDA action

EPA’s Comments on FDA’s Draft Report

An Evaluation of Risk to U.S. Consumers from Methylmercury in Commercial Fish Products, Including a Quantitative Assessment of Risk and Beneficial Health Effects from Fish
December 08, 2008

This document extends the written interim comments provided by EPA scientists and Program Experts on the Draft Report. This document of 12/08/08 is not meant to replace the 11/24/08 comments, but rather should be read with them. EPA has of this writing received only extremely limited revisions of the draft from FDA, and, thus, no way to note whether recommendations of 11/24/08 have been addressed. We include as an appendix to this document, a list of the issues identified in the 11/24/08.

EPA provided preliminary comments to FDA and OMB in written form on November 6, 2008, after a very short turnaround review. This 433 page Report was obtained by EPA late afternoon on October 30, 2008 and was supplemented by copies of the Peer Review materials on November 5, 2008. It did not appear to EPA that peer review comments had been fully addressed in the document EPA received on October 30, 2008. EPA’s preliminary review was completed in preparation for a joint EPA/OMB briefing on the Draft Report by FDA which was originally scheduled for November 3, 2008 and subsequently re-scheduled for November 6, 2008. Based on the issues identified in the preliminary review, EPA requested an additional 4-6 weeks review time prior to the public release of this document to complete a more thorough review. It was agreed at the meeting on November 6, 2008 that EPA would receive additional time. Interagency review meetings were subsequently set up for November 24 and December 2, 2008.

These comments of 12/08/08 summarize reviews made by EPA scientists in response to information available at a meeting of 12/02/08. This was intended to be a discussion among scientists to clarify and explain modeling procedures and inputs used by FDA in their draft report. Some clarification was gained. However, it is apparent to EPA scientists that much work needs to be done on the report to ensure a degree of transparency warranted by the complexity of the analyses, the controversial nature of quantification of mercury risk in the U.S., and the economic implications of conclusions based on the analysis. by and in preparation for the meeting of November 24, 2008. The Agency believes this is too important a public health issue to risk the publication of a Report that has not been thoroughly reviewed to ensure sufficient scientific rigor and transparency. The comments that follow are to be considered as interim; EPA still has not had sufficient time for the EPA scientists to thoroughly review a document of this size and complexity in the considered fashion that it warrants. These comments have not yet been vetted or approved by the senior EPA leadership.

EPA offers the comments below as well as additional marginal notes inserted to the draft text of the FDA report. Our comments do not constitute any concurrence with or approval of the document.

1. Sections on safety assessment were not deleted. Descriptions of the EPA RfD and ATSDR MRL are still inaccurate.

1. After the meetings of 11/24 and 12/02, we had assumed that Appendix E and most statements on safety assessment would be deleted. This would address our earlier comment that App. E requires a complete re-write as well as all those parts of the document that refer to the EPA RfD or ATSDR MRL. On 12/04/08 EPA received a revised App E and several pages other earlier sections of the document. With minor exceptions, this revision did not address our concerns of 11/24. We continue to recommend (as we did in the meeting of 11/24) that the sections on safety assessment be deleted. It is not necessary to trivialize existing, non-linear methylmercury risk assessments in order to establish a rationale for a predictive risk assessment. We have marked up a copy of the FDA draft indicating those sections (in addition to Appendices E and I) that should be deleted. We retain some comments below that have not been addressed, as well as additional comments on the misuse and misinterpretation of the EPA RfD and ATSDR MRL.




2. Please note marginal comments on the revised FDA pages on risk assessment. Note that we constructed a redlined copy of the FDA revision by comparison to the document EPA received 10/31/08. Some comments are new to this draft, but others were offered on 11/24, but have not been addressed.




3. Please note pages from the 10/31/08 document with sections on safety assessment that we recommend be deleted.




4. The RfDs, MRLs and similarly derived values are considered to be quantitative risk estimates by most risk assessors. They do not predict a risk at a particular exposure level, but are quantitative estimates of a population threshold for adverse effect.




5. While some changes have been made to the document there is still inappropriate mixing of concepts of uncertainty and safety factors. EPA and ATSDR use uncertainty factors to estimate a population threshold for adverse effect. When data permit (as was the case for the MeHg RfD), UFs are calculated as estimates of either variability or uncertainty or both. Safety factors for U.S. EPA involved risk management decisions rather than risk assessment. At this time the major use of safety factors in EPA are the application of FQPA factors when assessing pesticide exposure to children; this is not applicable to MeHg.




6. On p. 34 the document begins to describe safety assessments as designed to protect one or the other segment of a population (adults, fetus, etc.). The EPA RfD applies to the entire population, including sensitive subpopulations, but not excluding other populations. Thus, while it is based on endpoints measured consequent to fetal exposure, the U.S. EPA MeHg RfD is not restricted to protection against developmental neurotoxicity, but rather against all effects of MeHg exposure. P. 4-88 of WQC says the following: This RfD is applicable to lifetime daily exposure for all populations including sensitive subgroups. It is not a developmental RfD per se, and its use is not restricted to pregnancy or developmental periods”. This has not been adequately addressed in the FDA revised pages.




7. On p. 18, in the discussion of the NHANES data, it is stated that "All of the remaining 4.65 percent of women of childbearing age... retain some and typically most of the margin of safety that was built into the safety assessment level." This is an incorrect statement, predicated on mischaracterization of the RfD (e.g. not “margin of safety” in RfD). Note that several sections are recommended for deletion in EPA marked draft of 12/08/08.




8. Page 36 incorrectly characterizes the basis for the RfD. This was not a single benchmark dose (BMD). Rather RfDs were calculated on multiple endpoints, subsets of study populations, and studies as indicated in Table 4-8, p. 4-61, WQC, and on IRIS. The BMDL of 58 ppb mercury in cord blood was used as an example for calculations. But, as is clear from table 4-8, multiple BMDLs ranging from 24 – 103 ppb served as the basis for RfD calculations (ranging from 0.05 – 0.2 ug/kg bw/day). Note also the RfDs are published by EPA as ingested amounts; many scientists have chosen to use 58 ug / l cord blood as a comparison point, but EPA does not set RfDs as blood levels.




9. On page 36 is this statement:” Twelve ounces per week may be thought of as an approximate surrogate for the EPA Reference Dose with its 10-fold margin of safety.” EPA does not use the concept of margin of safety when referring to the RfD. EPA calculates margins of exposure (MOE). This is the difference between an observed no adverse effect level (NOAEL) lowest observed adverse effect level (LOAEL) or BMD. This is usually calculated as orders of magnitude the measured or estimated environmental exposure level is from the NOAEL, LOAEL, or BMD. The acceptability of the MOE is essentially a risk management choice that considers the factors that may be used to estimate uncertainty, as well as other factors (e.g. severity of the observed effect). A MOE of 10 would seldom be considered acceptable, with the possible exception of a reversible or adaptive effect.




10. Footnote 14 on p.35 is not accurate. It is the case that EPA generally defines the RfD as having uncertainty of about an order of magnitude. However, we have also stated that there is no evidence for a threshold for in the extant human studies. IRIS says the following: “It is also important to note that no evidence of a threshold arose for methylmercury-related neurotoxicity within the range of exposures in the Faroe Islands study. This lack is indicated by the fact that, of the K power models, K = 1 provided a better fit for the endpoint models than did higher values of K.“




11. Footnote 13 on p. 34 is not accurate. See discussions of safety factors vs. uncertainty factors




12. The extensive history of NAS peer review, and the additional independent external IRIS review of EPA's RfD should be presented: both in Appendix E and at least cited in appropriate places in the document. FDA is referred to the IRIS file, the Water Quality Criterion for the Protection of Human Health: Methylmercury (WQC), and to Rice et al.

2. Transparency is lacking in the approaches used to distinguish between low mercury containing fish and high mercury containing fish.
This could cause problems with risk communication if the document remains in its current form. The report tends to portray fish consumption simplistically and in a dichotomous way. This is relevant to all outcomes (e.g., neurotoxicity and cardiovascular outcomes).

3. Certain terms used throughout the document are concerning because they are not clearly defined (“background levels of Hg”’ and “high methylmercury-to-fish ratios”)
These terms should be defined carefully (i.e., <0.1 ppm) or replaced in the analysis with the actual concentrations.

4. Selection of endpoints used in modeling risk for adverse neurodevelopmental effects is flawed
The endpoints selected (age at first walking and age at first talking) are not the most sensitive measure of adverse neurodevelopmental effects seen in methylmercury exposed populations. More sensitive endpoints of cognitive and sensory function from the Faroe Island study were examined by the National Academy of Sciences (NAS) and recommended for use in risk assessment (which EPA adopted in its IRIS assessment for methyl mercury and in the Clean Air Mercury Rule). Additionally, there should be some discussion of the differences in baseline (i.e. unexposed) ‘age at talking’ and ‘age at walking’ for the Iraqi and Seychelles data, and how these differences might impact the results of the modeling (which includes both data sets).

5. There are concerns about the methodology used for the quantitative analysis
It is unclear whether this analysis was appropriately done, and whether the right questions were asked to support the conclusions being drawn.

6. There are many statements in the report that are not adequately supported or explained.
Some of these statements are critical to the conclusions of the report. For example, on p. 125-126 it states that the model may overstate the adverse effects of methyl mercury. The report then goes on to state that use of data from two sources (the Seychelles and the Daniels study) results in double-counting of both adverse effects of methyl mercury and benefits of fish (i.e. the data are confounded), but no explanation is provided as to why this would likely result in an overstatement of adverse effects (rather than an overstatement of benefits). To improve transparency, a clearer discussion of these tradeoffs should be provided in this and other similar discussions.

7. The current focus of FDA report on modeling data from for developmental milestones is seriously flawed in several ways.
By FDA admission they have not obtained reliable distributions/ confidence intervals for walking from the Iraqi data. The distributions developed are from the Seychelles data set; this presents some substantial uncertainties and concerns. These concerns are based in part on the advanced motoric developmental milestones in the Seychellois by comparison to U.S. norms FDA stated that they had not developed population distribution for these developmental milestones from population based samples in the U.S. population.

8. There are serious concerns about how z-scores were used in this analysis, both with respect to the ways in which they were derived and in the way they were used to compare across endpoints.

1. Z-score derivation: As described in section IV of the document (e.g., p. 116), and discussed at the meeting on 12/2/08, the s.d. from the Seychelles data on age of talking was used to derive the z-scores for this measure. It was stated by FDA that the slope of this model was driven mostly by the (high dose) Iraqi data. Thus, the appropriateness of using a z-score derived from the Seychelles data is questionable. Since (1) the size of the z-score depends entirely on the s.d., and (2) the argument supporting the validity of using these ‘walking and talking’ data to represent neurobehavioral effects of methyl mercury is based (at least in part) on a comparison of the ‘age at talking’ z-score to the IQ z-scores derived from other studies, inappropriate selection of the data (i.e. the s.d.) used to derive the z-score could undermine the validity of this analysis. This is even more critical since the benefit/risk analysis is apparently also based on combining z-scores from the ‘age at talking’ analysis with z-scores based on an analysis of data from the Daniels study. [We also note that the s.d. used to derive the z-scores for the ‘age at walking’ data was not provided in paper.] Discussion at the meeting on 12/02/08 resulted in a consensus that a rationale for selection of the s.d. used in calculating the z-scores should be provided, along with a discussion of the sensitivity of the conclusions to this choice. It was suggested that a sensitivity analysis be conducted using alternative s.d. values (e.g., from the Iraqi data or from the US population).
2. Comparison across neurobehavioral measures: It is unclear whether a z-score calculated using raw ‘age at talking’ data is appropriately compared to a z-score calculated on the basis of normalized population data for IQ. The comparability of these measures, both in terms of their distribution and the characteristics of the measures themselves, is questionable. A similar issue exists regarding comparing the ‘age at talking’ data to the data from the Daniels study, which uses yet another set of measures. The FDA analysis is predicated on combining and comparing these z-scores in various ways to support the risk/benefit analysis. Thus, it is essential that there be some further discussion regarding how the z-scores are influenced by the distributions and characteristics of the data, and ways in which the analysis and its conclusions might be influenced by these issues.
3. Results from the models should be presented as the original values (both medians and variance), in addition to the presentation of z-scores (e.g., in tables on p. 125).

9. Section IV does not contain a separate discussion of results from the Daniels ‘benefits’ model, but only of the results of the combined model.
A separate discussion of the ‘benefits model’, similar to the discussions of the ‘adverse effects’ models, is needed.

10. A clearer discussion is needed of plateaus and thresholds, and how they are (or are not) accounted for in the models.

11. There are flaws with the assumptions and modeling of exposure
Substantial data on intake of seafood by US consumers are available. However, the data analyzed are insufficient in themselves to provide distributions of usual (that is individual average daily) intake of different fish species by pregnant women. As a result, the authors have made several significant and verified modeling assumptions. Significant among these is the LTSTCR variable, a “long term-to-short term consumer ratio”. The authors are correct in pointing out that short term fish consumption as measured over a 3 day period will not be a precise indicator of long term consumption patterns. They also appropriately make reference to data on consumer’s reported consumption over 30 days from NHANES. However, it is not statistically correct to assume that a parameter such as the LTSTCR can be used translate individual’s 3 day consumption (from USDA’s CSFII study) into their long term consumption. The 3 day survey does not provide enough data – regardless of the functional relationship used to transform it – to estimate long term consumption patterns for these individuals. A reader may point out that the analysis in the report (Figure AA-1) appears to reasonably reflect the 30 day NHANES consumption patterns. Indeed, the report indicates that the LTSTCR relationship was selected specifically for this purpose. However, the 3 day individual data records on grams of fish eaten and specific fish species consumed from CSFII are used to project longer term average consumption overall and for each species. Since the USDA data cannot be transformed to indicate usual fish intake for any individual, projected distributions of usual quantities of fish consumed cannot be relied on. The FDA report contains additional steps (“Variation in fish species consumed”) to try and project distributions of usual intake for individuals who reported consumption of more than one kind of fish within the 3 day survey. However, these further data manipulations do not overcome the limitation of not knowing usual intake for the individuals whose data is then being further modeled.

12. Data in Table AA-2 may not be appropriate for reflection of nationwide distribution of fish consumed in the U.S.
The issue of representativeness – whether data collected for other purposes can be taken provide a statistical distribution for the U.S. – is not discussed in the FDA report. Instead, the FDA analysis concentrates on conducting a highly complex analysis of concentration data for each species in which a “battery of 10 distributions was fit to each data set and the four that provided the best fit were used to construct a probability tree. This formal, mechanized statistical treatment does not lead to appreciation of the strengths and uncertainties of the database. In particular, if the data are not known to be statistically representative for the U.S., detailed distributional analysis is in the end misleading to the reader.

13. The discussion of benefits with respect to IQ is inconsistent with the measures modeled in the analysis
It seems inappropriate that much of the discussion of benefits is with respect to ‘IQ’ [and in fact several paragraphs refer to benefits in IQ points (e.g. p. 127), which were not evaluated in the model], rather than using the measures modeled in the analysis. In addition, presentation of figures plotting the actual values (i.e. test scores) versus the MeHg values, along with the modeled data, would allow the reader to better evaluate the model fit (which was not discussed).

14. The analysis used to predict maternal blood mercury levels as a function of estimated methylmercury ingestion my be flawed
The Sherlock et al. (1984) study relied on by FDA appears to be a valuable experimental investigation of methylmercury intake versus blood levels in 20 (presumably healthy) male volunteers. However, the elaborate “uncertainty” evaluation conducted by FDA – resulting in developing 120 probability models to describe the Sherlock data seems to miss a central point: to what extent are the experimental data on a “convenience” sample of 20 men representative of the population distribution of blood/intake relationship for methylmercury in the population of American women of child bearing age – and more specifically on relationships for pregnant women. However, instead of focusing on evaluating the degree to which these data are likely to be applicable to the population of concern, the FDA authors have focused their efforts on extensive computer simulations conducted on the assumption that the data are representative. A maternal hair to blood ratio for methylmercury is then needed to complete the exposure conversions. Here the authors had available a substantial and statistically based sample of women of child bearing age from the NHANES database. A population distribution for this variable was taken from the observed hair/blood ratios for the NHANES sample. However, the authors, while correctly recognizing the potential for “noise” (error related statistical variability) in a study of this nature, dealt with this concern by (arbitrarily) deleting the top and bottom 20% of the observed distribution of ratios. Where the observed ratios go from values on the order of 0.1 to 5, the truncated distribution extends only from 0.1 to 0.3. While the authors’ motivation may be understandable, a pruning of an empirical distribution in this manner does not provide a statistically (or scientifically) valid way to project the actual population distribution.

15. Many clarifications of and improvements to the cardiovascular modeling are needed.

1. Clear descriptions are needed for the modeling choices.
   In the ‘He analysis’ the selection of this data set – rather than others – needs to be logically described. The choice of modeling coronary heart disease in individuals above and below 45 years of age is rather crude. The report needs to address, with an explanation, whether or not the data are amenable to analyses of smaller age groups (e.g., 5 or 10 year spans). Addressing age differences is potentially important because coronary heart disease rates and fatal coronary disease risks change over time in the U.S. population. In the modeling based on the He et al data, the dose-response function is forced through the background risk rate at low doses. FDA should be able to show the reader if and how changes in background risks affect the model results. It seems that this would have a big effect on the predicted results. Further, in the FDA’s fatal coronary heart disease dose-response model based on He et al. (2004), the authors state that no model uncertainty was included in the analysis. While the authors note that this is consistent with data obtained from epidemiologic meta-regression/analysis, at a minimum this comment should be substantiated with a reference from the peer-review literature and some additional explanation. Given the importance of this analysis and the expertise of the FDA authors, it was surprising and disappointing that uncertainty analysis was not undertaken (but could have been). If the FDA has an opportunity to conduct sensitivity analyses, it would substantially increase the quality and potential usefulness of the analyses.
2. The report needs to provide some additional treatment and discussion of the confidence intervals in the ‘Carrington model’ results.
   The reported confidence intervals suggest that there is some likelihood that fish consumption increases coronary heart disease risks. The single sentence in the text (“However, the bulk of the probability distribution is less than zero, so it is more likely than not that increased fish consumption leads to a decrease in CHD death.”) was unsatisfying. The lack of discussion regarding this point leads to a broader concern regarding the general classification of the exposure as ‘fish.’ This is unsatisfying for a number of reasons. The authors do not review possible differences in the risk based on the type of fish consumed. Further, the authors do not review possible differences in risk based on the preparation of the fish (e.g., baked vs. fried). Individuals cannot use this report for decisions regarding the types of fish they choose to eat or preparation methods. Further, reduction in fatal heart attacks is an important benefit- perhaps the most important in the analyses. This needs additional analysis and discussion in the main text.
3. See additional comments from 11/24.
4. There needs to be further explanation about the studies describing cohorts that saw no association with mercury
   There are important details to discuss for two of the three cohorts that saw no association with mercury. If dental workers (who are likely exposed to inorganic mercury) are excluded from the Yoshizawa et al study, this study reports a positive association of heart attacks with toenail mercury exposures, although not statistically significant. The interpretation of the study by Ahlquist et al is difficult because they report plasma concentrations, which is a better marker for inorganic mercury exposures rather than methyl mercury. This leaves only the Halgren et al study as being negative. The document would be much improved if it provided better descriptions and analyses of each of these studies.
5. The authors should use the same outcomes for all coronary heart disease models (e.g., change in deaths could apply to both the He model and the Carrington model). Delete ‘averted deaths’; further should this be averted coronary heart disease deaths.

16. There are limitations to using ‘unit fish’ as an exposure measure in the modeling
While it may be a useful initial modeling construct, the insights that can be obtained from such a model, particularly given the wide range of types of seafood consumed and the differences in nutrient and pollutant concentrations, are relatively limited. The development of models that examine the effects of differing levels of nutrients and pollutants in seafood and different consumption patterns would allow the FDA to model with more insight as to health risks/benefits associated with different choices regarding the types fish consumers choose to consume and rates at which they consume commercial seafood. The authors do not review possible differences in risk based on the type of fish consumed. Further, the authors do not review possible differences in risk based on the preparation of the fish (e.g., baked vs. fried). Risk communicators will not be able to use this report to help consumers make better decisions regarding the types of fish that consumers should choose to eat, their rates of seafood consumption, or preparation methods.

17. Lack of full documentation of decision and assumptions underlying model development and selection of input data:
Documentation of the overall design of the risk assessment, particularly the 2-stage probabilistic simulation, is not detailed enough to allow a full critical review of the approach. The documentation needs to be expanded in two areas. First, the modeling options and input data associated with each modeling node in the simulation needs to be clearly presented in a table, including the rationale for selection and whether each modeling option or data input represents coverage for variability or uncertainty. Second, the 2-stage probabilistic simulation needs to be more clearly described in a step-wise fashion (e.g., a clearer graphical pseudocode needs to be provided) so that the model’s looping structure can be fully understood.

18. More complete discussion of alternate modeling options considered in designing the analysis. Although the exposure model developed for this analysis is innovative and may be potentially defensible, I would expect that alternate approaches were considered for specific steps of the analysis (some allusion to this was given by the FDA modelers during our discussion clarifying the modeling approach – the modelers noted that they had considered multiple modeling approaches for specific steps and dropped certain options in the design phase). If this is the case and potential approaches were considered and rejected, then the write-up should more fully document those decisions (perhaps in an appendix). Certainly, if the literature provides multiple ways of addressing a specific modeling step, then those options should be considered and the rationale provided for why a specific approach was selected and others dropped. Further more, alternate modeling approaches, if credible, should be included as part of uncertainty analysis, or at the very least, sensitivity analysis. This would represent part of an expanded discussion of the rationale behind specific modeling choices. Examples of where this discussion of alternate approaches might be warranted include the approach used to estimate annual servings based on 3 day serving data. I would expect that there are alternate approaches for completing this particular extrapolation/adjustment.

Stepping back for a second, this comment really speaks to the fact that an initial design-phase document was not first developed and shared such that the modelers could have received feedback from peer reviewers on key design elements before finalizing the design and completing the analysis. Such an initial review of the model might have resulted in modifications or additional model strategies included in the analysis to deal with model uncertainty. Given the innovative and extremely complex nature of the risk modeling, this design-phase peer review was critical.

19. Need for additional sensitivity analysis to examine behavior of model:
Because the exposure model is attempting to characterize behavior across a population displaying significant variation in behavior (in terms of the number of meals consumed in a year and the types of fish species involved), the model may be extremely sensitivity to non-linearities and potential associations (e.g., correlations) between variables. This is particularly true in predicting exposure for individuals further from “typical” fish consumption behavior (e.g., those who consume a larger amount of fish species with higher methylmercury). If the model does not accurately capture non-linearities in underlying input datasets, or reflect underlying associations between parameters, then significant bias could be introduced into the results. It is clear that data on specific aspects of fish consumption is lacking, necessitating some of the innovative modeling presented in the analysis (e.g., estimation of repeat fish species consumed for consumers with different meal frequencies). But, this lack of data, also points to an inability to accurately capture potential associations between variables. When professional judgment, or available empirical data suggest an association between input parameters, but that association can not be estimated for inclusion in the formal model, a sensitivity analysis should be conducted examining the potential impact of that association (should it exist) on exposure results. In the absence of available data characterizing the suspected association, a reasonable surrogate value could be used as a bound. The results of this kind of sensitivity analysis could be used as part of an overall analysis of the model’s ability to generate reasonable estimates of exposure, particularly towards the “tails” of the consumption (or methylmercury intake) distribution. It may well be that overall confidence in the model’s ability to capture exposure for specific subsets of the population comprising the upper tail of the distribution is significant lower than confidence associated with the central-tendency or “more typical” consumer.

An example of a potential association between input parameters that could be examined through sensitivity analysis involves the species consumed and the repetition ratio variables. The potential exists for a non-random association between the species consumed (as represented in 3-day CSF II study) and the repetition ratio (derived from 30 day NHANES study). As I understand the exposure modeling, a repetition ratio is sampled from a constructed distribution (based on the NHANES data) and assigned to a simulated individual drawn from the CSF II survey. That repetition ratio is then used to determine what fraction of the fish consumption days for that simulated individual are to be modeled using their CSF II-reported species (the remainder of their consumption is modeled using species drawn from the NMFS data). My concern relates to a potential linkage between the CSF II species and the selection of the repetition ratio. There is the potential that the repetition ratio data in NHANES is non-randomly related to fish species (i.e., certain fish species are favored in those consumers who repeat consume the same species – tuna being an example in my mind). If this is the case, then linking the CSF II simulated individual (with their specific reported species) to a repetition ratio should not be random and could be guided conceivably by some correlation factor, or species-related linkage in the selection of the repetition ratio. If this relationship between species and repetition ratio is strong and favors higher MeHg fish (i.e., those consuming the same species tend to favor a high MeHg fish), then consideration for this factor could impact both the central-tendency MeHg exposure estimates as well as (and to a greater extent) the high-end MeHg exposure estimates.

This issue could be examined fist, by looking at the NHANES data to see if the repetition ratios are random with regard to fish species, or if they display a trend/correlation. If there is a trend, then the exposure model could be refined to explicitly reflect this linkage. Conversely, if it is not possible to resolve whether this correlation exists, it would be advantageous to at least conduct a sensitivity analysis where the FDA considers varying degrees of correlation between the species consumed and the repletion ratio. This sensitivity analysis would at least reveal whether this potential limitation in the model (source of uncertainty) could impact the results in the model such that a different policy-conclusion could be reached.

A close review of the modeling approach as implemented could identify additional instances of potential associations between exposure-related variables. I believe that these should be examined as part of a sensitivity analysis, unless it is possible to use empirical data to characterize these associations, thereby incorporating them into the formal exposure model.

21. Modeling fish consumption-related benefits with a simpler total fish approach rather than a species-differentiated approach analogous to the approach used in modeling MeHg risk:
The FDA analysis models neurological benefits of fish consumption based on a simple consideration of the total amount of fish consumed (i.e., benefits modeling based on unit of non-species differentiated fish consumption). Given that fish species vary in the amount of beneficial agents such as omega-3 fatty acids that they contain, this simplified benefits modeling approach introduces significant uncertainty (and possibly measurement error) into the analysis. It is possible that an approach that estimates neurological benefits based directly on consumption of these beneficial agents found in fish and not simply on total fish consumption, could produce a different overall mean population benefit. In addition, several of the FDA scenarios modeled consider various substitution scenarios whereby public health education initiatives produce changes in the types of fish consumed (favoring lower MeHg species). At least in the case of one of these FDA scenarios, the neurological benefits related to fish consumption are assumed to remain unchanged due to the simple unit-fish consumption approach used in modeling benefits. However, if substitution of low MeHg for high MeHg fish results in consumption of fish also possessing different concentrations of the beneficial agents (e.g., omega-3 fatty acid), then the overall benefit-disbenefit tradeoff estimated for these scenarios could be different than what is currently presented in the FDA analysis. Again, without examining this issue as part of a sensitivity analysis, the FDA does not know how significant a source of uncertainty this issue represents and what the implications it might have for conclusion drawn from the analysis.

Ideally, this issue would be addressed by refining the benefits component of the risk model to work with species-specific modeling (or age least groupings of species – oily versus white meat for example), in which case the exposure results generated in estimating MeHg (in terms of amounts of different fish species concerned for each simulated consumer) could also be used on the benefits side of the modeling. A key fist step would be pulling together the key input parameters and datasets for this refined species-specific benefits modeling approach including (a) data on the concentrations of key benefit agents (e.g., omega-3 fatty acids) in different fish species and (b) the concentration-response functions linking these various agents to neurological endpoint (benefits). In the absence of readily available studies characterizing “b” (i.e., neurological benefits linked directly to omega-3 fatty acids for example), it might be possible to use some simple assumptions to derive benefits functions specified in terms of omega-3 fatty acid intake rather than fish intake which could be suitable for a sensitivity analysis

APPENDIX A: LIST OF ISSUES IDENTIFIED IN 11/24/08 COMMENTS DOCUMENT.

1. FDA criteria for inclusion of neurodevelopmental studies of the model of mercury effects are overly restrictive, excluding several studies that may be useful.

2. A key component of the FDA analysis is to estimate the negative effect of mercury on neurodevelopment independent of any countervailing positive effect of fish consumption. Use of the Iraqi data is not the only way to obtain such an estimate.

3. The summary statements are imprecise and, therefore, tend to trivialize or overstate results.

4. The definition of the Reference Dose for MeHg needs to be properly described and placed in context. It is marginalized as an important value in the risk assessment of methylmercury.

5. The report includes several incomplete characterizations of the literature. [Note EPA has included many examples-- 57 instances at this writing -- of this in the comments added to the Draft FDA Report].

6. Many components of the FDA models are not adequately described.

7. Methodology used for the quantitative analysis of MeHg effects.

8. Transparency of the procedures and assumptions in the assessment is lacking. In particular, many vague terms are used in the summaries and to a lesser extent in the main text.

9. Human studies in general, are not well characterized.

10. Vulnerable populations were not adequately discussed.

11. There are many concerns with inputs to applications of the models

12. Information regarding the extent to which the various models
referenced and applied in the report have been validated and/or peer reviewed is not provided.

13. Sensitivity analysis is needed to determine whether these additional sources of uncertainty are potentially of concern (i.e., do they significant impact the risk assessment results such that they could potentially produce a different policy decision).

14. Transparency is lacking in the approaches used to distinguish between low mercury containing fish and high mercury containing fish.

15. The Report needs to clearly describe the choices regarding coronary heart disease modeling.

16. The report should analyze non-fatal coronary heart disease risks.

17. The report needs to provide some additional treatment and discussion of the ‘Carrington model’ and its results.

18. We do not understand how the coronary heart disease models treat changes in risk over time (pg 144).

19. Alternative risk benefit analyses were not discussed.