The content on this page is provided for reference purposes only. This content has not been altered or updated since it was archived.
Draft Report of Quantitative Risk and Benefit Assessment of Consumption of Commercial Fish, Focusing on Fetal Neurodevelopmental Effects (Measured by Verbal Development in Children) and on Coronary Heart Disease and Stroke in the General Population: Section IV, Quantitative Risk and Benefit Assessment Modeling for Selected Indicators of Fetal Neurodevelopment, Coronary Heart Disease, and Stroke
January 15, 2009
This information is distributed solely for the purpose of pre-dissemination peer and public review under applicable information quality guidelines. It has not been formally disseminated by FDA. It does not represent and should not be construed to represent any agency determination or policy.
(a) Conceptual Framework
This section provides an overview of the logic and design of the quantitative risk and benefit assessment. The assessment uses simulation modeling for uncertainty and variability to estimate:
- The net effect of eating commercial fish in the United States on early age verbal development in children as an indicator of neurodevelopment in the fetus. For purposes of this assessment, the net effect includes an adverse contribution from methylmercury and a beneficial contribution from fish, both of which are estimated in the assessment.
- The net effect of eating commercial fish on fatal coronary heart disease and stroke in the general population.
For neurodevelopment, our modeling is based on data on early age verbal comprehension from children who were prenatally exposed to methylmercury, or to nutrients from fish, or to both, as a result of their mothers' exposures while pregnant. Fetal exposure is not directly measured. The mother's exposure during pregnancy serves as a surrogate for fetal exposure without any adjustment.
For coronary heart disease (CHD) and stroke we model only fatal events. For this component of the analysis we divide the general population into four subpopulations and model them separately since the baseline risks and consumption patterns for each subpopulation are different. These are: (a) women of childbearing age (16-45); (b) women age 46 and older; (c) men age 16-45; and (d) men age 46 and older. Even for CHD and stroke, women were subdivided based on most likely childbearing years because the current fish advisory (thus consumption behavior) is driven by concern regarding neurodevelopmental effects on the fetus. Men are divided into the same age groups as women partly for ease of comparison as well as to capture differences in baseline risk by age.
The assessment is designed to estimate the consequences of current fish consumption and exposure to methylmercury (the "baseline") as well as the consequences of changes in fish consumption and exposures to methylmercury by U.S. consumers. Box IV-1 lists the questions about "baseline," that this assessment was designed to address. Results for these analyses are presented in Section V.
We modeled net effect three different ways:
- We estimate the likelihood and magnitude of effects at "baseline." We define "baseline" as being essentially commercial fish consumption and resulting exposure to methylmercury for women of childbearing age in accordance with the results of our exposure modeling of U.S. consumption and exposure. This modeling gives us a picture of consumption and exposure as of about the year 2005, since the data available for exposure modeling will always be subject to some time lag. Consumption at the "baseline" involves eating mostly fish that are at the lower end of the spectrum for methylmercury (since most commercial fish tend to be low in methylmercury, including most of the most popular commercial fish) but it also includes consumption of fish that are higher in methylmercury.
- In a separate analysis for fetal neurodevelopment effects, we estimate the likelihood and magnitude of effects at "baseline" U.S. levels of exposure to methylmercury but we assume that women of childbearing age eat only commercial fish that are at the low end of the spectrum for methylmercury. This modeling allows us to estimate whether maintaining current levels of exposure to methylmercury but only obtaining it through the consumption of lower methylmercury fish produces the same or different results as compared to the "baseline" results.
- We modeled various "what if" scenarios in which we estimate what would happen if women of childbearing age ate more or less fish or if the amount of methylmercury in the fish they ate were reduced (similar to the modeling in number two, above). The scenarios are listed in Box IV-2, below, and then discussed in detail later in this section.
We present the results in terms of the magnitude of the change on population-level effects. The simulations are based on two-dimensional population models that describe frequency of outcome in the population and the uncertainty associated with the estimates. The results are presented as population shifts above or below the "baseline." We described the "baseline" previously as recent levels of fish consumption and the resulting exposures to methylmercury experienced by women of childbearing age in the United States.
A potential limitation on the results from this "what if" modeling, however, is that for those scenarios involving increases or decreases in fish consumption, we were not able to take into account health consequences from corresponding increases or decreases in consumption of foods other than fish. Such modeling was beyond the scope and resources of this risk and benefit assessment. Analyses by the 2005 Dietary Guidelines Committee (2004) and by Mozaffarian and Rimm (2006) suggest, however, that the effects of food substitution might not have a significant impact on the outcomes (see the discussion on substitutionon substitution under "2005 Dietary Guidelines Committee" in Section A of the companion document entitled "Summary of Published Research on the Beneficial Effects of Fish Consumption and Omega-3 Fatty Acids for Certain Neurodevelopmental and Cardiovascular Endpoints").
Hypothetical scenarios involving changes in the baseline are posed in terms of "what if" questions shown in Box IV-2. As noted, the "what if" modeling does not take into account health effects from eating more or less of other foods as a consequence of eating more or less fish. Such modeling was beyond the limited purposes and resources of this project. The public health consequences of eating more or less of other foods can be relevant to the overall consequences of any risk management strategy that FDA might employ or contemplate, and may be worth considering for future analysis.
Box IV-1: Risk and Benefit Assessment Question Relating to Baseline (Current) Risk
For the health endpoints that we model (selected indicators of fetal neurodevelopment, fatal coronary heart disease and fatal stroke), what is the range of effects, from adverse to beneficial, that could be occurring in the U.S. population as a consequence of eating commercial fish? What are the uncertainties associated with these estimates, i.e., what is the range of possible effects in addition to the most likely effect estimated by the risk and benefit assessment (what are the confidence intervals surrounding the central estimates)?
Box IV-2: The Risk Assessment Questions We Posed that Involved "what if" scenarios.
What would be the effect on health endpoints if:
- Women of child-bearing age (age 15-45):
- Consume a maximum of 12 ounces per week? In other words, those who are consuming more than 12 ounces reduce their consumption to 12 ounces.
- Consume a maximum of 12 ounces per week of fish with relatively low concentrations of methylmercury? In other words, those who are consuming more than 12 ounces reduce their consumption to 12 ounces and those who are eating fish that average above "low" (as we define it in the scenarios) switch to only "low" fish.
- Consume any amount of fish with only relatively low concentrations of methylmercury?
- Consume exactly 12 ounces per week? In other words, those who are consuming less increase their consumption to 12 and those who are consuming more decrease their consumption to 12.
- Other subpopulations (women 46+ and all adult men):
- Decrease fish consumption across the board by 10 percent?
- One percent of fish eaters stop eating fish?
- All populations modeled:
- Increase fish consumption by 50 percent?
(b) Conceptual Model
For all endpoints we modeled the net effect from eating commercial fish. To assess the net effect of eating commercial fish on early age verbal development as an indicator of neurodevelopment, it was necessary to individually model components of net effect and then bring them together as a final step. We did so by combining the adverse methylmercury contribution and the beneficial fish contribution into a single dose-response function for net effect. For fatal coronary heart disease and fatal stroke we simply calculated a dose-response function from eating fish, thus these models involved fewer components. Figure IV-1 provides a visual description of the overall conceptual model. This section provides an overview of the modeling approach, with more detail provided in later sections.
The assessment was designed to estimate net effect of a range of U.S. exposures to a combination of: (1) methylmercury; and (2) nutrients in fish that could beneficially affect the endpoints we considered. In order to model these exposures we had to determine how much methylmercury is in each commercial species and how much of each species people appear to be eating. The major components of this modeling were:
- Estimating the amounts of fish that people eat. Amounts of fish eaten over time depend on the frequencies with which people eat fish and the serving sizes, i.e., the amount that people eat per meal.
- Estimating the species of fish that people eat. Different species of fish contain different average concentrations of methylmercury.
- Estimating how much methylmercury would likely be in each of these fish. In addition to variation among species, fish of the same species vary from one another in their methylmercury concentrations.
- Estimating dietary intake of methylmercury. This calculation is based on the previous three estimates.
- Estimating body levels of methylmercury. Over time, body levels are largely a result of dietary intake minus excretion. The average half life in the human body has been measured at about 50 days with a range of 42-70 days (Sherlock et al., 1984). We estimate body levels in terms of parts per million in hair. Many studies that have looked for associations between body levels of methylmercury and adverse effects have measured hair levels as the biomarker for body levels, although blood levels and other biomarkers have also been used. Hair is regarded as being a more reliable indicator of long term exposure than is blood. Blood is regarded as a good measure of current short-term exposure.
Adverse Effects from Methylmercury on Fetal Neurodevelopment
For fetal neurodevelopment we selected early age verbal development as an indicator of neurodevelopment and then developed dose-response functions for the adverse contribution that methylmercury could make to the net effect. This was the first dose-response function we modeled. In the United States, the fetal effect derives almost entirely from the methylmercury in the fish eaten by the mother and passed to the fetus. Available dose-response functions were then combined with information from the exposure assessment described above in order to estimate the size and likelihood of an adverse contribution through the range of U.S. exposures to methylmercury. In this report we present estimates for this contribution from the 10th percentile of exposure through the 99.9th percentile of exposure.
We did not model an adverse methylmercury contribution to the net effect for fatal coronary heart disease and fatal stroke. For these endpoints the potential for adverse effects from methylmercury exposure are not well enough understood and, furthermore, we did not have data on the concentration of methylmercury in the fish consumed. Thus we can only estimate whether the overall net effect from commercial fish is likely to be adverse, neutral, or beneficial.
Beneficial Effects from Commercial Fish on Fetal Neurodevelopment
For the chosen indicators of fetal neurodevelopment it was necessary to estimate a dose-response function for the beneficial contribution from the nutrients that can affect fetal neurodevelopment and that are passed to the fetus due to the mother's consumption of fish. Because estimating the contribution from individual nutrients is beyond the scope of this assessment, we modeled fish as a "package" of nutrients and assumed that all commercial fish are alike in terms of beneficial contribution.
Once a dose-response function was calculated, it was then combined with information from the exposure assessment to estimate the size and likelihood of a fish contribution independent of methylmercury attributable to a range of commercial fish consumptions in the United States. In this report we present estimates for this contribution from the 10th percentile of fish consumption through the 99.9th percentile of consumption.
(c) Criteria for Selecting Studies for Input into the Dose-Response Functions
The key challenge for modeling is identifying studies that can be used to inform the calculation of the dose-response functions.
Selected Indicators of Fetal Neurodevelopment: Methylmercury Adverse Contribution
- Methylmercury Effect Not Confounded: To estimate the effect from methylmercury alone, it was necessary to find data that measured an association between prenatal exposure to methylmercury and neurodevelopment where we could have reasonable confidence that the methylmercury effect was essentially not confounded (not offset or mitigated by) a beneficial effect from fish.
- Indicative of the Effect Magnitude: We could not model all aspects of neurodevelopment in a single assessment, i.e., all possible milestones and results from the myriad tests that exist for measuring all aspects of neurodevelopment, motor skills and verbal skills. Consequently, we had to model some aspects of neurodevelopment that we could assume to be reasonable indicators of at least part of the methylmercury's adverse effect on neurodevelopment as a whole.
- Individual Subject Data: We looked for studies from which individual subject data were available so that we could model individual variability into the assessment. Fetal neurodevelopmental endpoints are "continuous" in that the outcome in an individual is a matter of degree, e.g., the results on a test of neurodevelopment, or when an infant first talks (as opposed to whether an infant ever talks). Individual variability cannot be modeled from summaries of data because summaries presume a distribution (usually normal) that precludes the possibility of modeling individual variability as part of the dose-response function.
Another concern we had about using statistical summaries of data for a "continuous" endpoint is that the assessment would have to rely on how the investigators used and treated the data and the statistical techniques they used to evaluate them. We were especially reluctant to use statistical summaries that had been subject to a log(dose) transformation because the impact of the transformation on the secondary modeling results is difficult to determine.
Selected Indicators of Fetal Neurodevelopment: Fish Beneficial Contribution
- Fish Effect Not Confounded: To estimate the effect from fish independent of methylmercury, it was necessary to find data that measured an association between maternal fish consumption during pregnancy and neurodevelopment in their children where the beneficial fish effect was not significantly confounded by the methylmercury in the fish. ("Confounding" is an epidemiologic term that describes a variable that is associated with the health outcome of interest and is also associated with the exposure of interest). Since virtually all fish contain methylmercury if only in trace amounts, some confounding is probably inevitable but it can be minimized and taken into account in the modeling.
- Fish Effect Rather than Effects from Individual Nutrients: As mentioned previously, fish presents a "package' that includes lean protein, omegta-3 fatty acids, selenium, and other mineral and nutrients. We did not use data from studies that only measured the contribution from individual nutrients
- Comparability: We wanted fish contribution data that measured essentially the same underlying aspect of neurodevelopment as the methylmercury contribution data so that the dose-response functions from each of them could be combined into a single dose-response function for net effect.
- Individual Subject Data: We looked for studies form which individual subject data were available for the reasons described above for the methylmercury contribution.
Fatal Coronary Heart Disease and Fatal Stroke
Association Between Fish and Risk: We looked for studies that measured associations between fish consumption and risk of fatal coronary heart disease and fatal stroke. We concluded the literature supporting a direct link between methylmercury and these endpoints was not strong enough to support independently modeling that effect.
In four of the five studies that looked at methylmercury and CHD, data on exposure to methylmercury were obtained through methodologies that make comparison of exposures from one population to another, or to U.S. exposures, difficult. These methodologies involved measuring methylmercury levels in toenail clippings and blood serum (as opposed to whole blood). Without the ability to make such comparisons, it is not possible to know the methylmercury levels in the study participants as revealed by the established biomarkers, e.g., whole blood and hair. That knowledge would be essential for a quantitative assessment keyed to levels of exposure to methylmercury.
- Data Have Already been Subject to Meta-Analysis: In this context, a meta-analysis looks for an association between fish consumption and risk of coronary heart disease and stroke by combining the results of several studies that address the same question. Meta-analyses utilize their own criteria to determine whether individual studies are credible for inclusion in the analysis. We looked for meta-analyses with inclusion criteria that would be acceptable to us applying the criteria described below. We also looked for meta-analyses that calculated dose-response functions from the combined studies that they reviewed.
- Inclusion Criteria for Individual Studies: Our inclusion criteria, i.e., the characteristics that each study must possess, for the individual fish studies (also the inclusion criteria employed by the meta-analysis we selected for coronary heart disease (He et al., 2004a)) were:
- The study must have been a human study of clinical cardiovascular events. Therefore, studies that were in vitro or in animals do not meet this criterion. Similarly, studies that measured effects only in terms of biomarkers, rather than coronary events, do not meet this criterion.
- The study must have been conducted in adults with no history of heart disease (primary prevention). Studies in adults with existing heart disease (secondary prevention/intervention) will provide qualitative scientific support, but cannot be used quantitatively in the analysis.
- The study must have been an observational epidemiology study in populations. (There are no randomized clinical trials for primary prevention.) Randomized clinical trials for secondary prevention will provide qualitative scientific support.
- The study must have measures of exposure that are in terms of fish consumption and amount of fish eaten per unit of time (e.g., days, weeks). Studies based only on exposure to omega-3 fatty acids do not meet this criterion.
- The study must have included at least three levels of fish consumption (that is, the study cannot just have compared no fish to some fish but must have included at least three levels of fish consumption), in order to be able to develop a quantitative dose-response function.
- The study must have reported relative risk and corresponding 95 percent confidence intervals of CHD mortality relating to each exposure level (that is, amount of fish consumed).
- The study must have been a prospective cohort study design that was published in an English language journal.
- Summary Data Were Acceptable: The availability of individual subject data for fatal coronary heart disease and stroke was not a criterion. Unlike fetal neurodevelopment, where effects can involve subtle variations in test scores, fatal coronary heart disease and stroke have clearer criteria for diagnosis. We determined that summary data would be adequate for modeling under such circumstances.
(d) Exposure Modeling Overview
The following flow diagram and table provide an overview of the exposure modeling and the key input parameters. Each of the model components and the associated input data are described in detail below. Table IV-1 presents a summary the knowledge gaps, assumptions used to fill those gaps, and the implications of those assumptions. The assumptions primarily address how the available data are used and adjusted to provide a national picture of exposure for both commercial fish consumption and methylmercury. This study is based on previously published work by Carrington and Bolger. A discussion about the modeling is provided after the flow diagram and table.
|1||Consumer Survey Data:
How much and what types of commercial fish do people eat over a one year period? There is no consumer survey that covers an entire year.
|To the extent that it is used (other sources of data are used as well), the CSFII 3-day survey is presumed to be nationally representative for: 3 day frequency of intake; % of U.S. consumers eating seafood over a 3 day period; Characterization (in part) of the variety of fish people eat; Serving size||Although newer NHANES have similar average fish consumption for most adults, there is some indication that fish consumption in women of childbearing age may have decreased since the CSFII survey was conducted. If this is so, then the implication for the risk & benefit assessment results would be a slight overestimation of fish consumption and thus a slight overestimation of net effect.|
|2||Short-to-Long Term Frequency Extrapolation:
How much and what types of commercial fish do people eat over a one year period? There is no consumer survey that covers an entire year.
|For those individuals consuming fish, the 30 day survey is presumed to also represent annual (365-day) frequency
An exponential function is used to map short term frequency of consumption (CSFII) to the 30 day frequency (NHANES). While the model itself is well grounded empirically, there is an uncertainty in the extent to which the relative position of individuals in the short term survey corresponds to the long-term survey (i.e. a 90th percentile short-term consumer may be higher or lower than 90th percentile long-term consumer).
|The extent to which relative position varies is treated as a source of uncertainty in the model. Persons who consume seafood very rarely (less than once per month) are not well characterized. The implication for the risk & benefit assessment is that it may mischaracterize small effects in those consumers who eat fish less than once per month.|
|3||% of Consumers Eating Fish Over an Entire Year:
How much and what types of commercial fish do people eat over a one year period? There is no consumer survey that covers an entire year.
|As part of the long-term correction, an adjustment is made to account for the fact the number of fish consumers increased as the length of the survey period increases. A range of 85-95% consumers who eat fish was presumed for annual intake, with the lower bound being the percentage that ate fish in the 30-day survey.||The percentage of consumers eating fish over a year is a very minor source of uncertainty in the modeling.|
|4||Long-term Species Consumption Patterns:
How much and what types of commercial fish do people eat over a one year period? There is no consumer survey that covers an entire year.
|Data from the 30 Day Survey can be used to reasonably determine the extent to which each individual in the CSFII varies their pattern of fish consumption. The CSFII data associated with the individual can be used to reasonably determine repeated consumption, whereas market share data can be used to reasonably determine varied consumption.||There are fairly substantial changes in the composition of the seafood market since the CSFII survey was conducted. Although newer data are employed for the majority of the meals consumed, the estimates for individuals who consistently eat the same species are dominated by older data. Therefore species with greatly increased market share (e.g. shrimp and tilapia) are underrepresented while tuna is over represented. The implication is that the methylmercury adverse contribution to net effect may be slightly overstated for some repeat eaters.|
|5||Mercury concentration distributions in commercial species are known from years of sampling, but not known with 100% accuracy.||Three different approaches were taken to generating estimates for the range of mercury concentrations in each species. 1) Empirical distributions of FDA survey data with no uncertainty, 2) modeled FDA survey data with model uncertainty, 3) surrogate distributions based on older NMFS data with model uncertainty. Which is assumed to still be representative||The greatest source of uncertainty involves mercury concentrations of a small (<10%) portion of the market, which might not be current. The uncertainty is minimized by the fact that no clear trend toward increased methylmercury concentrations in commercial species can be seen in the data (see Section III of this report). The implication for the risk and benefit assessment results appears to be negligible.|
|6||Mercury Speciation Factor:
Most of the mercury in fish is methylmercury, but for ease of lab analysis the amount of total mercury in fish is typically measured, rather than the methylmercury. How much of the total mercury is methylmercury?
|Fixed conversion factors were used to adjust for mercury content. While most of the total mercury is methylmercury in most seafood species, there is good evidence that shellfish is much less. The conversion factors are based on a study published by Height and Cheng (2006) in which they estimated how much mercury was methylmercury in finfish and shellfish. The assumption is that these conversion factors enable us to correctly estimate the amount of methylmercury in fish based on the previously measured amount of total mercury.||Although there are minor variations among and between species in the inorganic contributions to the total content, these variations are considered negligible.|
|7||Serving Size Adjustment:
Are serving sizes the same as they were when measured in the CSFII survey?
|Because current per capita consumption is more accurately measured by market share disappearance, we applied a correction factor of 11% to make the CSFII-derived serving estimates consistent with market data.||Using CSFII serving sizes without a correction factor would generate slightly lower estimates of exposure to both fish and mercury.|
Estimating Species and Amounts of Fish that People Eat
The objective for the stage of the exposure assessment was to estimate commercial fish consumption, i.e., the amounts and species that people consume, for the U.S. population over a period of time long enough to capture infrequent fish consumption and to characterize chronic (i.e., steady state) exposure. We chose a one year time period for this purpose.
In order to estimate amounts and species consumed over a period of one year, we extrapolated average daily fish consumption over a one year period from the results of short term food consumption surveys in which people were asked to recall what they ate on three days. We assume that this extrapolation yields a distribution that is reasonably representative of amounts and species of commercial fish consumed in the United States over a one year period.
We estimated U.S. fish consumption, i.e., amounts and species, using three sources of data:
- The U.S. Department of Agriculture's Continuing Survey of Food Intake by Individuals (CSFII) survey conducted between 1989 and 1991 (three day survey)
- The NHANES survey data from 1999-2002 (30 day survey)
- National Marine Fisheries Service market share data on consumable commercial fish (2005).
The three-day survey was the U.S. Department of Agriculture Continuing Survey of Food Intake by Individuals (CSFII) (USDA 1993). It surveyed both men and women and obtained information about portion sizes that they ate. These data were statistically representative of the U. S population.
The 30-day survey was a fish and shellfish consumption frequency questionnaire that had been administered as part of the NHANES survey during 1999-2000. It captured information about frequency and various categories of fish type, e.g., clams, tuna, swordfish, and salmon. However, this survey only involved women of childbearing age and children up to 11 years of age and did not obtain information about serving size. These omissions made it impossible for us to rely solely on the 30-day survey for our exposure assessment. Since the three-day survey provided information lacking in the 30-day survey, and vice versa, we used the two surveys together.
We used data from the National Marine Fisheries Service of the U.S. Department of Commerce (NMFS 2007) on "edible (for human use) meat weight" for individual commercial fish species that are imported into, or landed in, the United States in order to develop a rank order of popularity for commercial fish. We used these data to help estimate the types of fish consumed over a year. These data were used to supplement the short term survey data for characterization of long-term variation in species consumed over an entire year. NMFS market share data were also used to adjust portion sizes to reflect current levels of consumption. Since the NMFS data are more recent, they more accurately reflect current national patterns of fish consumption.
Variations in the Species that People Consume
In order to estimate the species of fish that people eat, we developed and implemented the following process:
Using the 30-day survey: For each individual in the survey who ate at least four fish meals during the survey period, we developed a "repetition ratio" to reflect the extent to which the individual ate the same fish or ate a variety of fish. The mathematics of the "repetition ratio" are provided in Appendix A. We assume that the distribution of "repetition ratios" from this survey is representative of the entire U.S. population, even though the survey only involved women of childbearing age. This distribution is described in detail in Appendix A. We then applied the ‘repetition ratios" to the results of the three-day survey since that survey was representative of the U.S. population rather than just women of childbearing age.
Using the three-day survey and the NMFS market share data: The individuals in the three-day survey reported eating fish from zero to four times during the survey period. For each of the 3,525 individuals in the survey who ate at least one fish meal during the period of the survey, we randomly selected one of the "repetition ratios" developed from the 30-day survey. On the basis of the "repetition ratio" that was selected for this individual, we would either assume that the types of fish reported for that person in the three-day survey were the only types of fish eaten by that person all year or that the individual ate other types of fish during the year in addition to the fish he or she reported eating during the three days, with the proportion of other fish determined by the repetition ratio. For example, if the "repetition ratio" were 0.5, we would assume that half of the person's fish meals consisted of the fish he or she reported in the survey. We would fill in the other half with fish selected randomly from the NMFS market share data after "weighting" those fish based on popularity.
Amounts of Fish that People Consume
Estimating amount of fish consumed in a year involved extrapolating the data on frequency and serving size from the three and 30 day surveys to (a) the entire U.S. population; and (b) a year's worth of fish consumption. We developed formulas for this purpose as described in Appendix A.
Estimating Levels of Methylmercury in Commercial Fish
Data:Total mercury concentrations in most commercial fish species are available from FDA surveillance data (1990-2004) (FDA 2006). Data for a small number of minor species were obtained from reports from a National Marines Fisheries Survey (NMFS 1978) and the EPA (EPA 2000, page 59). These data are summarized in Table AA-2 in Appendix A.
Method: A realistic estimate of exposure to methylmercury requires consideration of the variations in concentrations of methylmercury that occur across and within commercial fish species. Variations in methylmercury concentrations from fish to fish are generally attributed to differences in size (Barber et al., 1972, page 638; Kraepiel et al., 2003, page 5,554) and age of the fish as well as differences in the concentrations of methylmercury in what the fish consumed.
The primary source of data for this part of the assessment was FDA's database of mercury concentrations in commercial species of fish. For many species in the database, FDA provides a mean, median, high-low range, and standard deviation based on all the samples in the database for the species in question. These values are for the total mercury in the fish, rather than for methylmercury, because the standard laboratory analysis is for total mercury. Recent analysis by FDA scientists has shown that for finfish, methylmercury constitutes about 95 percent of the total mercury in the fish, and about 45 percent of the total mercury in molluscan bivalve shellfish (e.g., clams, oysters, mussels) (Hight & Cheng 2006). Consequently, for purposes of this exposure assessment, we reduced the mercury values in the FDA database by five percent for finfish and 55 percent for bivalve molluscs. The methylmercury concentrations in bivalve molluscs tend to be low to the point of being essentially nondetectable, so the actual reductions for these species had a minimal impact even though the percentage was relatively high.
Rather than using only one number, like an average or another type of "best estimate," to represent this variation, we used a statistical simulation approach that allowed for the inclusion of a range of concentrations for individual fish in each species. Approaches for developing distributions of mercury in fish are described in "Mercury Concentrations in Individual Species" in Appendix A.
Estimating Methylmercury Intake from Consumption of Commercial Fish
The model for the fetal neurodevelopmental endpoint requires that we calculate the exposure to methylmercury for women of childbearing age. Such a calculation is not essential to the modeling for fatal coronary heart disease or stroke, however, since the data employed in those models come from studies of fish consumption in which the exposures to methylmercury were not measured.
The modeling involved extending our statistical simulation modeling for amounts and types of fish by selecting a value for the concentration of methylmercury in each type of fish from the distribution of methylmercury values for that fish. A new value was randomly selected for each iteration of the model.
Converting Dietary Methylmercury Intake to Hair Levels of Methylmercury
The next step involved estimating the actual level of methylmercury in the body on the basis of dietary intake. As indicated previously, methylmercury is excreted with a half life of around 50 days; consequently, the level of methylmercury in a person's body would not be identical to their accumulated daily intake.
As also indicated previously, mercury concentration in scalp hair is the most commonly used biomarker of a person's body level of methylmercury. Much of the data from scientific studies that we use in the assessment of neurodevelopmental risk to the fetus measure the "dose" of methylmercury to the fetus in terms of the concentrations of methylmercury in the mother's hair. We retain this measure of dose in the fetal neurodevelopment assessment.
In order to do so, however, we first had to convert dietary intake to mercury blood levels and then convert from blood levels to hair levels. We converted to blood levels by using the results from a study (Sherlock et al., 1984) with controlled exposures to fish that related dietary mercury to blood levels. Estimations of hair levels from given methylmercury blood levels were calculated with a distribution developed from the 1999-2000 NHANES survey. The impact of body weight on blood mercury was calculated using a function of body weight to the power of 0.44. The data and methodology we used for converting dietary intake into blood levels and then into hair levels are described in Appendix A.
Also, for purposes of conveying information in Table V-7, we wanted to estimate what hair levels of methylmercury would be from eating certain amounts of commercial fish over time. The main purpose of the table is to show the size of a beneficial effect on fetal neurodevelopment from eating various amounts of commercial fish (10th percentile of consumption through the 99.9th percentile of consumption). For the sake of context, we wanted to show what the methylmercury hair levels were likely to be for most people at each level of consumption. In order to do that, we had to assume how much methylmercury there was likely to be in the fish. We chose the average concentration of methylmercury in commercial fish weighted for consumption, i.e., popularity, 0.086 ppm. We then estimated the exposure to methylmercury by using the following equations:
Blood (µg / L) = Hair ppm ÷ 0.303 (Weighted average ratio from NHANES 1999-2000 data)
Diet (µg / day) = Blood µg/L ÷ 0.85 (from Sherlock et al., 1984)
Fish (g / day) = Diet (µg/day) ÷ 0.086 ppm (0.086 ppm is average methylmercury concentration in U.S. fish, weighted for consumption.)
Differentiating Between "Mercury" and "Methylmercury" For Purposes of Exposure Assessment
Much of the data available to us on exposure to methylmercury is actually reported as exposure to total mercury that includes both inorganic and organic forms of mercury. Inorganic mercury in the body primarily comes from sources other than fish. An important issue for our quantitative risk and benefit assessment, therefore, is estimating how much mercury in a person's hair or blood is likely to be methylmercury from eating fish. In summary:
- We know that most mercury in fish is methylmercury. As stated previously in this report, methylmercury constitutes between 93-98 percent of total mercury in finfish and 38-48 percent in molluscan shellfish (Hight and Cheng, 2006). (Molluscan shellfish, e.g., clams and oysters, have such small amounts of total mercury in them per FDA's monitoring program that the quantity of mercury that is not methylmercury in those species is tiny.) We took these percentages into account when calculating methylmercury exposure from fish.
- Most methylmercury in the U.S. diet comes from fish. Small exposures are possible, however, from eating other animals that were fed fish meal (Lindberg et al., 2004). As described in Appendix A, we calculate that about 0.1 ppb of methylmercury in the diet is from sources other than fish. We took this amount into account in our exposure assessment.
- People have mercury in their bodies in addition to methylmercury. We excluded mercury other than methylmercury. To do this, we used data from NHANES, described in Section II, that show both the total mercury and the inorganic mercury in each person surveyed. We can calculate the amount of methylmercury (i.e., organic mercury) in an individual by subtracting the inorganic mercury from the total mercury. This calculation also tells us what the ratio is between total mercury and methylmercury.
- People have mercury in their bodies even though they eat no fish. In NHANES there are respondents who reported eating no fish but whose hair or blood showed the presence of mercury. These people can be found through the 15th percentile of mercury exposure per NHANES.
(e) Data Selections and Dose-Response Models
Dose-Response Modeling Flow Diagram and Table
The following flow diagram and table provide an overview of the dose-response modeling and the key knowledge gaps, assumptions, and implications at each step. The assumptions primarily address how the available data are used and adjusted to provide a national picture of health effects associated with both commercial fish consumption and methylmercury. A discussion about the modeling is provided after the flow diagram and table.
Figure IV-3: Flow Diagram for the Dose-Response Modeling. The numbers correspond to numbers in Table IV-6 that describe knowledge gaps, assumptions and implications. The numbers start with "8", Figure IV-2. Box "8" here carries the results of the exposure modeling over to this flow diagram.
|8||Exposure Assessment Integration||Individual estimates of average daily intake of methylmercury and fish are carried over into the dose-response assessment, along with the body weight of the individual and the number of assumed eaters for each uncertainty estimate. As is widely assumed in the scientific literature, long-term average exposure from fish consumption is considered to be the best dose metric; other options are not considered. All dose-response data available to us for modeling have measured long-term exposure.||Implications for health of, higher exposures for shorter periods of time are not addressed in this modeling exercise.|
|9||Blood methylmercury concentrations were estimated based on dietary exposure.||The distribution of blood levels developed for this estimation is based on a 90-day human study with controlled exposures to methylmercury. We assume that this estimation provides reasonably accurate blood methylmercury concentrations.||Model uncertainty and sampling error are represented in the model. Again average, or steady-state, chronic exposure is presumed to be the best measure of dose Since the confidence intervals are relatively narrow, this is likely to be a minor source of uncertainty.|
|10||Non fish exposures to methylmercury . What percentage is coming from other sources?||We assume a very small contribution from other sources, based on studies that have shown that animals such as chickens that have been fed fish meal will contain very small amounts of methylmercury.||This part of the model has no impact on assessing the health impact of consuming fish - it is included in order to make the model consistent with the NHANES survey values at the low end of the population distribution.|
|11||Blood-Hair Relationship:||This relationship is characterized with an empirical distribution constructed from U.S. survey values (NHANES). As a result of potential environmental contamination at high concentrations and measurement error at low concentrations, However, for several reasons, some of the observed variability in hair/blood ratios may not be attributable to actual pharmacokinetic variation. Therefore, we assume:
This is significant source of uncertainty at the tails of the population distribution for the methylmercury-neurobehavioral effect, which uses hair-mercury as a measure of exposure. The implication is that the model has wider confidence intervals for the methylmercury effect than would otherwise have been the case.
Regarding assumption (2): Sherlock et al. (1984) employed multiple dose levels of methylmercury in fish and demonstrated that the relationship is approximately linear. Consequently, this assumption is not a source of significant uncertainty.
|12||Shape of relationship between the fish consumption and CHD mortality||Two different dose-response functions were used (the "meta-analysis" model, herein referred to as "CM," and the "pooled analysis" model, herein referred to as "CP." They were based on different assumptions in three key areas: 1:
"CM": Assumes a linear dose-response function.
"CP": Assumes the possibility that it is not linear.
"CM": Assumes that the underlying mean value estimated from the different studies that were incorporated into the dose-response function represents the true value.
"CP": Assumes that each study that was incorporated into the dose-response function represents a plausible true value.
"CM": When dose = zero, there is no uncertainty about the rate of disease.
"CP": When dose = zero, there is uncertainty about the rate of disease.
|1: The "CM" function does not predict a ceiling on benefits (more fish always = greater benefits) but the "CP" function contains a ceiling at 20 grams per day of fish. Implication: the baseline central estimates are similar but beyond 20 grams per day the predictions would diverge. The "CM" function probably overestimates benefits at high levels of consumption and under represents the uncertainty associated with lower levels of exposure.
2: While the central estimates are similar for both functions, the confidence intervals for "CP" are notably wider, to the point where they include a small possibility of deaths caused by fish consumption.
3: The "CM" function expresses less uncertainty at lower levels of exposure while the "CP" function expresses greater uncertainty.
|The health benefits of individual nutrients in fish.||Assumed that all commercial fish are alike for purposes of benefits. (The dosimetry for both dose-response functions treats all forms of fish equally).||The modeling exercise is not able to provide advice on the mix of fish types needed to maximize the benefits from fish consumption. However, the modeling for fetal neurodevelopment indicates that avoidance of higher methylmercury fish can reduce the attenuation of these benefits and minimize the likelihood of an adverse effect.|
|13||Shape of relationship between the fish consumption and stroke mortality.||Two different dose-response functions were used (the "meta-analysis" model, herein referred to as "SM," and the "pooled analysis" model, herein referred to as "SP." They were based on different assumptions in three key areas:
"SM": Assumes a low dose slope that is different from a high dose slope, with most of the effect (and less of the uncertainty) being at low doses and less effect (but more uncertainty) at the high doses.
"SP": Assumes less certainty at the low doses but is similar to SM at the high doses.
"M": Assumes that the underlying mean value estimated from the different studies that were incorporated into the dose-response function represents the true value.
"P": Assumes that each study that was incorporated into the dose-response function represents a plausible true value.
"SM": When dose = zero, there is no uncertainty about the rate of disease.
"SP": When dose = zero, there is uncertainty about the rate of disease.
|A significant source of uncertainty in the simulation model. While the central estimates are similar for both functions, the low-dose confidence intervals are notably wider in the "SP" analysis.
1: The "SP" confidence intervals being wider to account for more uncertainty at low doses.
2: The confidence intervals for "SP" are notably wider, to the point where they include a small possibility of deaths caused by fish consumption.
3: The "SM" function expresses less uncertainty at lower levels of exposure while the "SP" function expresses greater uncertainty.
The implications for all of the above are that while central estimates are similar, the size of the confidence intervals are dependent on some of the modeling assumptions made.
|14||Choice of indicator for neurobehavioral benefits||
Assume that early age verbal comprehension is a useful indicator of neurobehavioral development. The specific tests were MacArthur Communicative Development Inventory at 15 months of age and the language component of the Denver Developmental Screening Test at 18 months of age. Verbal comprehension, as measured by these tests, was selected because it matched criteria we developed for the inclusion of data for the
; (4) comparability, i.e., same general aspect of neurodevelopment, with age of talking endpoint. We lacked data on any other beneficial aspect of neurodevelopment that met these criteria. ….Endpoint choice; reason
This is a significant source of uncertainty in the
model because we used data from one large study only, and could
not develop dose-response functions for other beneficial
fish effects for purposes of comparison. However, the results from the modeling are generally consistent with results from studies that have been published in the scientific literature involving other aspects of neurodevelopment measured at later ages in life.
|14||Shape of the relationship between fish consumption and neurobehavioral benefits||
One function with a linear slope derived from two measures of verbal comprehension performance in a single cohort. A linear function does not include a "plateau" above which greater fish consumption does not lead to greater benefits. Uncertainty bounds were placed on the linear slope to reflect that inclusion of other studies or endpoints would provide a wider range of plausible interpretation. Linear only – no model uncertainty, The central estimate is based on verbal measures in one cohort. No maximum effect parameter in the model.
The uncertainty represented in the model is the primary source of uncertainty in the simulation model estimates.
(1): Although the omission of a maximum effect is not likely to have an impact on the scenarios presented here, it could result ins overestimation of benefits were seafood consumption greatly increased.
|15||Choice of indicator for neurobehavioral deficits from methylmercury exposure||Assume: a) age of talking is an useful indicator of neurodevelopment;
b) the magnitude and size of the methylmercury effect on age of talking is similar to the methylmercury effect on a wide range of other neurodevelopmental endpoints measured when children were older.
c) age of talking and early age verbal comprehension are similar enough to allow combining in a net benefits model
Regarding assumption a): From the scientific literature: age of talking requires the effective integration of a large number of complex sensory neural mechanisms (Marsh et al., 1995b).
Regarding assumption c): from response to comment (per yesterday's discussion, needs rewrite)They both are in the same domain (verbal) of neurodevelopment. Moreover, as incorporated into the modeling, they were measured at essentially the same ages. The tests administered in the U.K. were at ages 15 & 18 months. The age of talking data from Iraq involved children who talked both sooner and later than these ages.
|15||Shape of the relationship between methylmercury exposure and neurobehavioral deficits||
Three different functions. One extrapolates high dose (Iraq) effects developmental milestones using a combination of linear and nonlinear dose-response relationships. . A second produces a linear slope for IQ with the SEM serving as the uncertainty characterization, based on data from Faroes, Seychelles, and NZ. The third analyses, based on the same three studies, generated a distribution of Z-score slopes that reflect a wide range of different neurobehavioral performance measures. When converted to Z-scores, all three functions yield very similar central estimates. Consequently, there was no need to rely on any one function.
The uncertainty represented in the model is the primary source of uncertainty in the simulation model estimates. The three analyses taken together pretty much cover the landscape of possible sources of uncertainty. However, since each analysis taken on its own has some additional uncertainty, there may be some additional uncertainties associated with each function. The models may not be ideal for characterizing some specific neurobehavioral effects (e.g. effects motor or cognitive development).
Regarding assumption c): the measures used were age of first talking (which was later combined with data on beneficial fish effects in order to measure overall net effect), age of first talking; IQ from Seychelles, Faroe Islands, and New Zealand; and a wide battery of test scores from Seychelles, Faroe Islands, and New Zealand. Since only age of talking was later incorporated into the net effect modeling, the other results (age of walking, IQ, and wide battery of test scores) were used for purposes of comparison and support. The results from the age of talking model are similar to the results from these other models. Collectively they reflect a wide range of neurodevelopmental performance.
|16||Metric for combining beneficial and adverse effects||Net effects are estimated through the use of Z-scores that covert the original measures to relative measures that scale each effect by the amount of variation that occurs in a normal population. This conversion is somewhat dependent on what population is taken to represent a normal population.||The standard deviations used to calculate Z-scores for the analysis of the pooled Iraq and Seychelles data are taken from the Seychelles (SD= 2.57 months for talking and 1.97 for walking). Iraq data were not used because there are very few data at low levels of exposure that can be used to characterize normal variation in the milestones. We believe that these that values are close to those for the U.S. and elsewhere. For age of walking, this assumption is supported by a study of children in six different countries, including the U.S., which yielded a standard deviation of 1.8 (WHO, 2006). We are unable to find reports of statistical descriptions or raw data in the U.S. or elsewhere on the verbal milestone (three words) used in Iraq and Seychelles.|
Fetal Neurodevelopment: Adverse Methylmercury Contribution to Net Effect – Age of First Talking and Walking
The study of the poisoning event in Iraq, where methylmercury was ingested in bread, provides data on an association between prenatal exposure to methylmercury independent of fish and neurodevelopment. Unlike most studies of the effects of methylmercury, the Iraq data do not involve fish consumption. These data are probably the least ambiguous data on methylmercury toxicity in humans currently available, with effects unlikely to reflect methylmercury from fish consumption. If we were to use data from other studies, it would be necessary to statistically separate two closely correlated variables – methylmercury and fish – in order to estimate the methylmercury contribution. It is highly uncertain whether separation of such highly correlated variables can be done.
The researchers in Iraq collected data on ages of first walking and talking that showed dose-response relationships between delays and prenatal exposures to methylmercury (Marsh et al., 1987). Moreover, they published individual results from each of the study participants. For these reasons, we used data on the attainment of early age milestones from Iraq to measure the methylmercury effect independent of fish.
We modeled dose-response functions for both age of first talking and age of first walking. However, as explained later in this section, we only used the function for age of first talking to represent the methylmercury contribution to the net effect. The results for age of walking were included for purposes of comparison along with IQ results and results from a range of neurodevelopmental tests.
One source of uncertainty from the data on age of first talking and walking from Iraq is the exact age of the children when they first walked and talked, since birthdays were not recorded in Iraq. The mothers provided the ages of their children within six month time frames. We believe these estimates are sufficiently accurate. Likely errors were no larger than a few months either way, which would be within a range of normal variation for these milestones. Moreover, we would not expect errors in recollection to be biased in favor of the children being either older or younger than estimated. Finally, at high doses, the adverse effects were larger than the six month time frames and could span years.
We were concerned, however, that a model that only used the data from Iraq would produce results of limited utility due to the small size of the study population (81 mother-infant pairs) and the fact that few subjects within this population experienced relatively low levels of exposure (Marsh et al., 1987). Because the Iraq data come from one of the most extreme exposures ever to occur with methylmercury, they might be viewed as anchoring the model at the upper end of observed effect but are much less robust at the low end. For these reasons we looked for additional sources of data where the endpoints measured were ages of first talking and walking.
Another potential source of data for the consumption of methylmercury other than from fish was the study conducted in the Faroe Islands. The primary source of methylmercury in that study population was from pilot whale, although fish was also a source of methylmercury. The Faroe Islands study obtained early age developmental milestone data on age of first creeping, sitting, and standing (Grandjean et al., 1995), rather than walking and talking as were measured in Iraq. Because the milestones measured in Iraq and the Faroe Islands were not identical, we concluded that we could not combine them into the single dose-response function model that underlies this assessment. If we had used the milestone data from the Faroe Islands, we would have had to do so in lieu of the Iraq data. An additional impediment was that individual scores on developmental milestones (or on neurodevelopmental tests that were administered when the children were older) have not been made available from the Faroe Islands study. We would have had to use summary data that had undergone log(dose) transformation. For these reasons we did not incorporate data from the Faroe Islands for this aspect of the modeling (although data from the Faroe Islands were employed in IQ and other modeling described below).
The only other studies that measured age of first talking and walking were the studies in Peru (Marsh et al., 1995b) and in the Seychelles Islands (Myers et al., 1997). The individual subject data from Peru were never published and are not available primarily due to the age of the study (conducted between 1981 and 1984).
On the other hand, we have obtained the individual subject data on age of talking and walking from the Seychelles Islands. A potential problem with these data is that they derive from exposure to methylmercury solely from maternal consumption of fish ((Shamlaye et al., 1995, page 601). Nonetheless, we combined these data with data from Iraq in order to model a methylmercury effect independent of fish for the following reasons:
- Adding data from the Seychelles helps characterize the variation in the response at low doses where the contribution of methylmercury to the overall variation is relatively small. Also, adding data from 680 mother-infant pairs from the Seychelles produces a more robust assessment.
- Offsetting effects from fish were minimized. The model's characterization of the dose-response relationship (adverse) was still driven primarily by the Iraq data because the effects attributable to methylmercury were much larger in Iraq. As a consequence, the dose-response function from the Iraq-Seychelles was not substantially different from a dose-response function we calculated solely from the Iraq data. If we were to model solely from the Iraq data, the median estimate would be a delay of 0.048 months for each additional part per million of mercury in maternal hair. The Iraq and Seychelles combined median is a delay of 0.045 months for each additional part per million of mercury in maternal hair. (A general description of an "Iraq only" analysis can be found in Carrington et al., 1997.)
The results from this modeling reflect several major assumptions. The first is that the predicted methylmercury effects have not been offset to any substantial degree by benefits obtained elsewhere in the diet. For example, we assume that the predicted effect has not been reduced by selenium obtained from vegetables or another source. The second assumption is that methylmercury might have a threshold of effect, i.e., that methylmercury might not produce an adverse effect below a certain level of exposure. Because we do not know what a threshold level might be for methylmercury, the probabilistic modeling that we employed included simulations of various possible thresholds, including no threshold. Third, we assume that the standard deviations we used to calculate Z-scores are close to those for the United States and elsewhere. We used a standard deviation from the Seychelles of 2.57 months for talking and 1.97 for walking. For age of walking, this assumption is supported by a study of children in six countries, including the United States, which yielded a standard deviation of 1.8 (WHO, 2006). Any difference between the standard deviation we use from the Seychelles and the standard deviation for the United States would not affect the original estimates for the adverse methylmercury contribution but it could slightly affect the estimates for net effect in the United States. The results from the adverse methylmercury contribution were converted to Z-Scores so that they could be incorporated in the net effect modeling. In any case, since milestone standard deviations do not vary greatly among populations, the choice of reference population represents a minor source of uncertainty for the Z-Score estimates.
A fourth assumption is that ages of first talking and walking are useful measures for neurological health. As stated by Marsh et al. (1995b):
"Age at which an infant talks, stands alone and walks without assistance may appear to be crude indices of development. However, they all require the effective integration of a large number of complex motor and sensory neural mechanisms, and when supported by neurological observations of behavior, vocalization, understanding, motor and sensory functions, they provide very good standards for comparisons on an individual infant or group basis."
Both early speech and motor development have been associated with greater IQ at eight years of age; early speech development has been associated with reading comprehension at 26 years of age (Murray et al., 2007).
There is another perspective on these endpoints, however, as expressed by Crump et al. (1998): "The measures of effect in the Iraqi study (late walking, late talking, and neurological score) are relatively crude measures of neurological deficit and may not be as sensitive to methylmercury as more subtle but equally important effects that could be occurring, such as effects upon IQ."
To address this concern, we included in this risk and benefit assessment two analyses that were developed outside of FDA, one of which was on the effect of prenatal exposure to methylmercury on IQ (Axelrad et al., 2007) in the Seychelles, Faroe Islands and New Zealand; while the other was on the effect of prenatal exposure to methylmercury on a battery of tests administered in these three locations (Cohen et al., 2005b). We used these results in a comparative analysis against the results from age of talking. This comparative analysis reveals a consistency of outcome in certain respects.
Fetal Neurodevelopment: Comparative Analysis on the Adverse Methylmercury Contribution to Net Effect – IQ and Battery of Tests
One of these dose-response models used the single metric of IQ (Axelrad et al., 2007). Although there are some uncertainties associated with this metric, one advantage is that it incorporates a range of sub-tests in several "domains" of neurodevelopment, each of which increases the likelihood that it includes tests that could be sensitive to effects of methylmercury at low doses. Moreover, IQ's predictive value for achievement throughout life has been studied extensively. There is a body of literature that can provide insight into the potential consequences for achievement later in life of very small changes in IQ that modeling might predict. Another advantage provided by the IQ model is that it addresses neurodevelopmental results that were measured from ages six through nine. If, as been hypothesized, effects from prenatal exposure to methylmercury are difficult to detect until a child becomes older, they could be more likely to appear at ages six though nine than at ages of first talking and walking.
As stated above, this model calculated changes in IQ as the response to methylmercury exposure using test results from the Seychelles Islands, Faroe Islands, and New Zealand studies. In the Seychelles and New Zealand studies, the researchers looked for an association between IQ score and prenatal exposure to methylmercury. The Faroe Islands study did not test for IQ per se, but Axelrad et al. incorporated results from some tests administered at age seven because these were regarded as being significant components of IQ. The three slopes were weighted and averaged into one linear IQ slope for methylmercury exposure. The model predicts a loss of 0.18 of an IQ point for each part per million of methylmercury in maternal hair.
Another dose-response model, published by Cohen et al. (2005b), calculated dose-response slopes from a wide battery of neurodevelopmental tests from Seychelles, New Zealand, and the Faroe Islands. These three slopes were combined into one linear slope, using weighted averages.
Cohen et al. (2005b) conducted two analyses with data from the Faroe Islands. The first analysis linearized the published log linear function in the range of U.S. exposures while the second analysis linearized in the range of exposures in the original study. Because the first analysis was based on a model in which effects become larger as doses become smaller (log(dose) transformation), we regard the second analysis as being the more plausible of the two. The authors calculated their dose-response function as if the effect involved IQ points. The second, more plausible analysis, predicts a loss equivalent to 0.2 of an IQ point for each part per million of methylmercury in maternal hair (Cohen et al., 2005b, page 362).
An uncertainty associated with these dose-response functions is the extent to which they reflect methylmercury's contribution to the net effect without being significantly confounded by fish. As explained previously, we interpret both our age-of-first talking and age-of-first walking models as roughly indicative of a methylmercury effect independent of any offsetting benefits from fish consumption. We interpret these results similarly.
The majority of methylmercury in the Faroe Islands was from the consumption of pilot whale rather than fish (Grandjean et al., 1999). The nutritional profiles of pilot whales and fish are not the same (Julshamn et al., 1987) so there was less opportunity for confounding by nutrients in fish than there would have been if the source of the methylmercury had only been fish. The IQ results from New Zealand also appear to reflect high exposures to methylmercury relative to the amounts of fish consumed, i.e., exposures that derive from consumption of shark. The Seychelles results involved lower methylmercury fish than were consumed in New Zealand but the IQ slope calculated from these data by Axelrad et al. (2007) was adverse, suggesting that fish confounding was not substantial. In total, we assume that confounding by fish did not significantly alter these results, although it probably did occur to some degree.
Fetal Neurodevelopment: Beneficial Fish Contribution to Net Effect – Early Age Verbal Comprehension
In order to develop a dose-response function for the beneficial fish effect we looked for data that showed an association between maternal consumption of fish and beneficial neurodevelopmental outcomes. Because we wanted to calculate a dose-response function that we could then combine with the adverse dose-response function for methylmercury, we looked for endpoints that were either identical or reasonably comparable to early age milestone data on first talking (early age verbal) and/or first walking (early age motor). We also looked for individual subject data rather than data summaries. We wanted to use an association from fish rather than from nutritional supplements, but with only minimal confounding from methylmercury.
The study that met these criteria was that of 7,421 mother-infant pairs in the ALSPAC study in the United Kingdom (Daniels et al., 2004). The study measured associations between maternal fish consumption and subsequent test scores. For a subset of its cohort it also measured associations with prenatal methylmercury exposure and test scores but found none.(11) For that reason we concluded that the fish consumption results were not confounded by methylmercury to any significant degree. Although the neurodevelopmental outcomes measured in the children did not include age of talking, the tests did include verbal comprehension at young ages, i.e., vocabulary comprehension on the MacArthur Communicative Development Inventory (MCDI) at 15 months of age and the language component of the Denver Developmental Screening Test (DDST) at 18 months of age. Furthermore the children were of the same age as children who first talk. In light of the similarity in age to age of first talking, we assume that these results were comparable -- even though not identical -- to the milestone results on age of first talking from Iraq and the Seychelles Islands. Moreover, individual subject results from these tests were available to us.
By contrast, the data available to us from that study did not include individual subject data on early age motor skills that would be comparable to age of first walking from Iraq and Seychelles. The DDST total scores included a motor component, but it could not be separated from total score. Consequently, for the fish contribution to the net effect, we developed a dose-response function based on early age verbal comprehension results from the United Kingdom. As an aside, because we now had data on early age verbal development for both the methylmercury contribution to the net effect and the fish contribution to the net effect, we elected to use the dose-response function for age of walking as part of the comparative analysis.
We chose a linear dose-response function for this effect. A linear function does not include a "plateau" above which greater fish consumption does not lead to greater benefits. We did not use a model with a maximum effect parameter or other non-linear models because we have not yet discerned a shape to the dose-response function (the "fish" effect is small relative to the other sources of variance to allow model discrimination). We assume that such a plateau must exist but the results of several studies suggest that it must exist somewhere above the 95th percentile of consumption (12 ounces per week). Future modeling efforts may be in a better position to model a plateau, assuming it is not so high as to be irrelevant to U.S. consumption patterns. Daniels et al. (2004), which modeled dose response using quartiles for fish consumption, suggests a plateau but no other studies have investigated the existence of such a plateau.
The following table provides a study-by-study review from the standpoint of whether a study was included into, or excluded from, the modeling for fetal neurodevelopment. The table shows which studies were used and how they were used. For the studies that were not used, the table briefly explains why. Decisions about inclusion/exclusion essentially followed the criteria for selecting studies as provided in section (c), previously.
It is worth noting that these studies involve exposures to methylmercury that are longer than "episodic, " e.g., single meal or a few meals clustered together. Most involve long term exposure. Consequently, the data on associations between methylmercury and fetal neurodevelopment provide a basis for evaluating risk from long term consumption of fish over time, but not from isolated meals that might cause a shorter term elevation in methylmercury. Whether exposure from a single meal or a series of meals eaten over a short period of time has the same impact on risk as a sustained exposure over time at identical levels cannot be determined from these data. The risk to the fetus from shorter term exposure to methylmercury, e.g., from a single meal, remains an untested question.
|Study (location, authors, year)||Size of Study Pop.||Source of MeHg||Outcome Measures||Availability of individual subject data||Application to the Risk and Benefit Assessment|
(Marsh et al., 1987)
|81||Mother's consumption of bread||
--Age of first talking
--Age of first walking
|Yes||Age of talking & walking data used in modeling performed in FDA for MeHg effect independent of any countervailing effect from fish.|
(Myers et al., 1995)
|Approx. 700||Mother's consumption of fish||
--Age of first talking
--Age of first walking
|Yes||Age of talking & walking data were combined with similar data from Iraq in modeling a MeHg effect described above.|
(Marsh et al., 1995b)
|131||Mother's consumption of fish||
--Age of first talking
--Age of first walking
|No||Not used. Individual age of talking data not available.|
(Grandjean et al., 1995)
|583||Mother's consumption of fish and pilot whale||
-- Age of first sitting
-- Age of first creeping
-- Age of first standing
|No||Not used. Individual subject data not available. Also, the developmental milestones that were measured (sitting, creeping, standing) are different from ages of first talking & walking.|
Quebec: Cree Native Americans
(McKeown-Essen et al., 1983
Ages 12-30 months:
--Denver Developmental Scale
|No||Not used. Individual subject data not available. Also, whether exposure to MeHg was solely from fish or also from marine mammals was not published.|
(Kjellström et al., 1986 & 1988)
|38 at age 4; 61 at age 6 ("high exposure" part of study pop.)||Mother's consumption of fish||
--Denver Developmental Screening Test
--Neurological Screening Tests
--battery of tests including IQ
--Not used in modeling performed in FDA. Individual subject data not available. Also, data not comparable to early age verbal.
-- IQ data were used in IQ modeling performed outside FDA and these results are included in this assessment.
(Myers et al., 1997 & 2003; Davidson et al., 1995a & 1998)
|Approx 700||Mother's consumption of fish||Battery of neurodevelopmental tests at ages 6.5 mo., 19 mo., 29 mo., 66 mo., & 9 yrs. IQ at age 9 yrs.||No||
--Not used in modeling performed in FDA. Individual subject data not available.
-- IQ data and other test results were used in modeling performed outside FDA and these results are included in this assessment.
(Grandjean et al., 1998; Debes et al., 2006)
|900+||Mother's consumption of fish and pilot whale||Battery of neurodevelopmental tests at ages 7 & 14.||No||
--Not used in modeling performed in FDA. Individual subject data not available. Summary data would be subject to log(dose) transform. Also, data not comparable to early age verbal.
-- 9-yr data that constitute aspects of IQ were used in modeling performed outside FDA and these results are included in this assessment.
(Daniels et al., 2004)
|7,421||Mother's consumption of fish||
--MacArthur Communicative Development Inventory
--Denver Developmental Screening Test
|Yes||Data on verbal skills at 15 & 18 months used in modeling performed in FDA of net effect from fish consumption.|
(Hibbeln et al., 2007a)
|Approx. 9,000||Mother's consumption of fish||
Ages 6 mo. through 8 yrs:
--various neurodevelopmental tests including IQ
|No||To the extent that Hibbeln et al. used the same data from the U.K. as Daniels et al., above, the data were used through the use of the Daniels et al., data.|
(Oken et al., 2005)
|135||Mother's consumption of fish||
Ages 5.5 – 8.4 mos:
--visual recognition memory test
|No||Not used. Individual subject data not available. Also, data not comparable to early age verbal.|
(Oken et al., 2008)
|341||Mother's consumption of fish||
Age 3 yrs:
--Wide Range Assessment of Visual Motor Abilities test
--Peabody Picture Vocabulary Test
|No||Not used. Individual subject data not available. Also, data not comparable to early age verbal.|
(Lederman et al., 2008)
|329||Mother's consumption of fish||
Age 3 yrs:
--Bayley Scales of Infant Development psychomotor score
Age 4 yrs:
|No||Not used. Study was published after completion of our assessment. Also: (1) individual subject data not available; (2) summary data would have been subject to log(dose) transform; and (3) the outcomes that were measured were not comparable with early age communication.|
(Jedrychowski et al., 2006)
|233||Mother's consumption of fish||
Age 1 yr:
--Bayley Scales Mental and Motor
|No (with qualifications)||We have the Bayley Scales Mental Scores but the verbal component is not distinguishable from the total. (We did model dose-response from the Bayley scores separately in order to determine the size of the dose-response function.)|
(Jedrychowski et al., 2007)
|374||Mother's consumption of fish||
Ages 2 & 3 yrs:
--Bayley Scales Mental and Motor
|No||Not used. Individual scores not available.|
(Oken et al., 2008a)
|25,446||Mother's consumption of fish||
Ages 6 & 18 mos.:
--range of neurodevelopmental milestones
|No||Not used. Study was published after completion of our assessment. Also: (1) individual subject data not available; and (2) the developmental milestones that were measured were different from ages of first talking & walking.|
Fetal Neurodevelopment: The Net Effect from Eating Commercial Fish
In order to estimate the net effect on fetal neurodevelopment from maternal consumption of commercial fish, we developed this model by combining the results from age of talking in Iraq and the Seychelles (representing "methylmercury") with early age verbal comprehension results from the United Kingdom (representing "fish"), using Z-Scores (as described in Section V). We assume that this combination includes a certain amount of double counting but not to the extent that it would skew the results significantly one way or the other. That is, presumably it double counts methylmercury slightly from both the Iraq-Seychelles data and U.K. data (the U.K. results showed no adverse effect from methylmercury but we know that there were methylmercury in the fish) but presumably it also double counts fish benefits slightly through the inclusion of Seychelles data in the methylmercury modeling. In any event, we doubt that the beneficial fish contribution is overstated in this model since the size of the beneficial association seen in the Daniels et al. (2004) study (the source of the fish contribution data for this modeling) is not as large as beneficial effects reported in other studies.
Fatal Coronary Heart Disease
A meta-analysis by He et al. (2004a) that examined the association between fish consumption and fatal CHD also included quantitative dose-response modeling. Consequently, we performed risk and benefit assessment modeling using both the data from the studies that He analyzed and the published He et al. (2004a) dose-response function. We reviewed each study that passed the He et al. (2004a) inclusion criteria, which we adopted as our own for purposes of the risk and benefit assessment. Based on these criteria, we added some studies to our own modeling that were published after He et al. (2004a) published their meta-analysis.
Because CHD death is a binary endpoint there is less information lost by using the population level statistics than would be for a continuous variable. This type of endpoint prompted us to develop a population model for CHD death rather than an individual-severity model, i.e., a model based on degrees of severity, as we did for neurodevelopment.
The studies included in He et al. (2004a) meta-analysis are listed as studies 1-13 in Table AA-14 in Appendix A. We used results from these studies in our assessment. Additional studies (14-16 in Table AA-14) are those that we identified through a literature search as having met the inclusion criteria but that were published after the He et al. (2004a) cut-off date. We incorporated these studies into a second model (the "CHD pooled analysis model") that we performed in addition to the "CHD meta-analysis model," as described below.
Of the 13 studies that He et al. (2004a) analyzed and that we modeled, six involve U.S. study populations. A substantial U.S. contribution to the data can be important because risk factors for CHD, including the potential risk factor of methylmercury in fish, may be affected by population characteristics. Different populations appear to experience different overall risks based on such things as diet (including the types of fish they eat), lifestyle, and genetics.
One of the studies analyzed by He et al. (2004a) included participants from Finland, the location of studies that initially reported an association between relatively high levels of methylmercury in fish and increased risk of CHD(12) (Salonen et al., 1995). The Finland study that was incorporated in He et al. (2004a) is not from the identical population that was studied by Salonen et al. and others. However, data from the Salonen et al. (1995) study population were included in another meta-analysis along with data from various other countries, including the United States (Whelton et al., 2004), that produced results similar to those produced by He et al. (2004a). Whelton et al. (2004) found an association between fish consumption and an approximately 20 percent reduction in the risk of fatal CHD.(13)
As stated previously, we divided the population by age and gender into the following categories: females aged 16-45, males aged 16-45, females aged 46 and above, and males aged 46 and above. The primary question for our assessment was whether fish consumption reduces the risk of fatal coronary heart disease, has no effect, or increases the risk in these population categories. Death rates from coronary heart disease vary by age and gender. For purposes of this modeling, ages 16-45 represent childbearing age for females.
For use in conjunction with our modeling, we estimated baseline rates for fatal coronary heart disease in the United States for these subpopulations by dividing the number of deaths from CHD per year for each subpopulation (NCHS, 2006) by the number of people in each subpopulation per the U.S. Census Bureau. Because the data from NCHS and the Census Bureau are in five year increments, the closest increment to "women of childbearing age" as we are defining it (16-45 years of age) is 15-44 years of age. Consequently, we calculated death rates for the age range of 15-44 and we assume that it is essentially the same as the death rate for the 16-45 age group. We then adjusted these rates for sex differences using data from Ho et al. (2005). Because Ho et al. (2005) did not contain rate information for persons under the age of 45, we used the relative rates for men and women in the youngest age group covered by Ho et al. (2005) (45-50) to correct for sex differences in the 15-44 subpopulations of both sexes. The resulting baseline rate estimates are presented in Table IV-4.
|Sex||Age 15-44||Age 45 and above|
|Female||0.14 per 10,000||38 per 10,000|
|Male||1.3 per 10,000||51 per 10,000|
"CHD Meta-Analysis Model:" He et al. (2004a) used a pooled meta-regression of relative risk to combine the results from all 13 studies into one estimate of effect. Details on the methodology are available in He et al. (2004a). We characterize results as being from the "CHD meta-analysis model" in order to differentiate them from the results from our "CHD pooled analysis model," as explained later. The "CHD pooled analysis model" used a different approach in developing its dose-response function in order to reflect various uncertainties in the data.
"CHD Pooled Analysis Model:" We also estimate the effect of fish consumption on CHD death with an alternative model (the "CHD pooled analysis model"). This model incorporated results from the same studies as were used in the "CHD meta-analysis model" (i.e., studies 1-13 in Table AA-14 in Appendix A), plus three additional studies (i.e., studies 14-16 in that table). However, its methodological approach produces results that can be more reflective of uncertainties in the estimates than those predicted by other models.
The description of the differences between the "CHD pooled analysis model" and the "CHD meta-analysis model" is explained in detail in Appendix A. A brief summary follows.
First, the "CHD pooled analysis model" used separate dose-response functions that were developed from the data in each of the individual studies. These dose-response functions were then integrated into a common dose-response function by weighting according to sample size. By contrast, the "CHD meta-analysis model" involved averaging across studies to yield a single dose-response function, essentially treating all the data as if they were drawn from the same underlying population. This treatment does not allow for the possibility that these populations have significant differences in terms of confounding risk factors for CHD.
Second, the confidence intervals for the "CHD pooled analysis model" were based on sampling error for each individual data point. This was done so that we did not have to assume a common variance across all studies and dose groups, as was done in the "CHD meta-analysis model."
Third, in addition to a linear model, alternative non-linear ("sigmoidal") models were used to describe the data. The linear model included a maximum effective dose parameter, meaning that the benefits from fish consumption peak at some point. All these models permitted greater effects at particular dose ranges than did the simple linear model used by He et al. and incorporated into the "CHD meta-analysis model." A probability tree was used to include model choice as a source of uncertainty.
Finally, rather than using relative risk, the "CHD pooled analysis model" used adjusted group events. This approach allows sampling error from the low dose group to be represented instead of being fixed to a relative risk of one. As a result, the model is not forced through the illness rate reported for the control group.
The practical consequence of this approach is that the "CHD pooled analysis model" has confidence intervals that are wider than those produced by the "CHD meta-analysis model." These wider confidence intervals reflect the uncertainty that arises from using data from different study populations, each with its own risk factors for CHD, and applying those results to the entire U.S. population. The narrower confidence intervals in the "CHD meta-analysis model" derive from the assumption that the study populations are collectively analogous to each other. In addition, because the models employed a maximum effect parameter, the dose-response function was nonlinear, with most of the benefit being conferred at levels of consumption below 25 g. per day.
Stroke is the third leading cause of death in the United States and the leading cause of adult disability according to the National Stroke Association (NSA 2008). It involves interrupted blood flow to an area of the brain due to an obstruction of an artery (ischemic stroke) or a break in a blood vessel (hemorrhagic stroke). Our modeling involved both types of stroke.
A meta-analysis that estimated a quantitative dose-response relationship between fish consumption and stroke was Bouzan et al. (2005). We developed a risk and benefit assessment on the basis of the Bouzan et al. (2005) dose-response relationship and the data that Bouzan et al. (2005) had incorporated into their meta-analysis. We refer to this model as the "stroke meta-analysis model.") We also estimated the effect of fish consumption on stroke death with an alternative model, as described below (the "stroke pooled analysis model").
Because CHD death is a binary endpoint there is less information lost by using the population level statistics than would be for a continuous variable. This type of endpoint (death) prompted us to develop a population model for stroke death rather than an individual-severity model, i.e., a model based on degrees of severity, as we did for neurodevelopment.
Table IV-5 shows the studies that were used in the "stroke meta-analysis" and "stroke pooled analysis" models. The "stroke meta-analysis model" used the six studies that Bouzan et al. (2005) used in their meta- analysis.(14) One of these studies (Caicoya 2002) did not involve multiple exposure groups per the inclusion criteria but we regarded the use of the Bouzan et al. (2005) published dose-response function as sufficiently important to justify using all the data that Bouzan et al. (2005) used. It would not have been possible to extract one study from that dose-response function. We could fully apply the inclusion criteria to the data used for the "stroke pooled analysis model," however, since it involved the development of our own dose-response function. The "stroke pooled analysis model" did not incorporate the results from the Caicoya (2002) study.
Another meta-analysis, by He et al. (2004b), also investigated the relationship between fish consumption and stroke, but did not estimate a dose-response relationship. The "stroke pooled analysis model" used all but one of the studies identified in the He et al. meta-analysis. Consequently, the "stroke pooled analysis model" utilizes a larger database than does the "stroke meta-analysis model." As Table IV-5 shows, the "stroke pooled analysis model" used five of the six studies that were used in the "stroke meta-analysis model" (there was significant overlap in the studies used by Bouzan et al. and He et al.) in addition to three used solely by He et al. and two others that were published after the He meta-analyses but that met the inclusion criteria: Nakamura et al. (2005), and Mozaffarian et al. (2005). In addition to omitting the Caicoya (2002) study, the "stroke pooled analysis model" also omitted the study by Keli et al. (1994) used by He et al. (2004b) because it only contained two exposure groups.(15)
|Study||Population Size||# of Events||Average Age||Avg. Follow up (yrs)||% Male||Site||Bouzan2005||He et al., 2004b||"Stroke Pooled
|Orencia et al. (1996)||1,847||76||47.6||30||100||USA||X||X||X|
|Iso et al. (2001)||79,839||574||34||14||0||USA||X||X||X|
|He et al. (2002)||43,671||608||53.4||12||100||USA||X||X||X|
|Caicoya (2002)||440 cases/ 473 controls||-||-||n/a||Spain||X||-||-|
|Morris et al. (1995)||21,185||281||52||4||100||USA||-||X||X|
|Yuan et al. (2001)||18,244||460||54||12||100||USA||-||X||X|
|Sauveget et al. (2003)||40,349||1,462||56||16||100||Japan||-||X||X|
|Keli et al. (1994)||552||42||15||100||Netherlands||-||X||-|
|Nakamura et al. (2005)||8,879||288||58.3||12||44||Japan||-||-||X|
|Mozaffarian et al. (2005)||4,775||626||58.3||12||42||USA||-||-||X|
To parallel the risk and benefit assessment for coronary heart disease, we divided the population by age and gender into the following categories: females aged 16-45, males aged 16-45, females aged 46 and above, and males aged 46 and above. The primary question for our assessment was whether fish consumption reduces the risk of stroke, has no effect on stroke, or increases the risk in these population categories. We present the results in terms of population-level effects. The results are reported in terms of median (50th percentile) and a range of lower and upper bounds (5th and 95th percentiles, respectively).
For use in conjunction with our modeling, we first estimated baseline rates for stroke death in the United States for females ages 15-45 and ages 46+ and for males ages 15-45 and ages 46+ by dividing the number of stroke deaths per year for each subpopulation (NCHS 2006) by the number of people in each subpopulation per the U.S. Census Bureau. NCHS and the Census Bureau provide data in five year increments. The closest such increment to "women of childbearing age" as we are defining it (16-45 years of age) is 15-44. Consequently, we calculate the death rates for the 15-44 age groups and assume that it is essentially the same as the death rate for the 16-45 age groups.
The baseline rates of death from stroke are shown in Table IV-6.
|Sex||15-44||45 and above|
|Female||0.25 per 10,000||18 per 10,000|
|Male||0.24 per 10,000||13 per 10,000|
"Stroke Meta-Analysis Model:" Bouzan et al. (2005) conducted a regression analysis with data from the five studies listed in Table IV-5 that investigated the relationship between the frequency of fish consumption and stroke. Their regression analysis generated a linear slope that did not go through zero, as shown in Figure IV-4. Bouzan et al. (2005) interpreted the intercept at the y-axis as an indicator of risk reduction associated with any quantity of fish consumption, even a small quantity.
We were not willing to adopt an assumption that a minute amount of fish consumption could have a substantial health impact. Consequently, we modified the Bouzan et al. (2005) dose-response function in order to reflect a more biologically plausible relationship between fish consumption and stroke. Specifically, we assumed that the effect at low doses occurs between zero and 50 grams of fish per week. This amount roughly corresponds to the low end of the range of the data used in the Bouzan et al. (2005) analysis. Thus, the Bouzan et al. (2005) model's 12 percent reduction in risk that it had attributed to a fish consumption of zero was modeled, instead, as a gradual increase up to 50 grams of fish per week. The resulting dose-response function is shown in Figure IV-4.
"Stroke Pooled Analysis Model:" As explained previously, we developed a second model for the effect of fish consumption on risk of fatal strokes that used the data from four of the studies used in the "stroke meta-analysis model" as well as most of the data that had been evaluated in the meta-analysis conducted by He et al. (2004b). In addition to utilizing a larger database, the "stroke pooled analysis model" used a methodological approach in which the uncertainties produced larger confidence intervals than were produced by the "stroke meta-analysis model."
The "stroke pooled analysis model" developed separate dose-response functions from the data in each of the individual studies. These dose-response functions were then integrated into a common dose-response function by weighting according to sample size. By contrast, the "stroke meta-analysis model" involved averaging across studies to yield a single dose-response function, essentially treating all the data as if they were drawn from the same underlying population. This treatment does not allow for the possibility that these populations have significant differences in terms of confounding risk factors for stroke.
The confidence intervals for the "stroke pooled analysis model" were based on sampling error for each individual data point. This was done so that we did not have to assume a common variance across all studies and dose groups, as was done in the "stroke meta-analysis model."
Finally, rather than using relative risk, the "stroke pooled analysis model" used group disease rates. This approach allows sampling error from the low dose group to be represented instead of being fixed to a relative risk of one. As a result, the model is not forced through the illness rate reported for the control group.
As with the CHD analysis, the practical consequences of these methodological approaches are wider confidence intervals, particularly at low doses, and a non-linear dose-response relationship with most of the benefits occurring with fish intake levels of 25 g. per day or less.
CHD and Stroke: Non-Fatal Events
The risk and benefit assessment for coronary heart disease and stroke estimated the effect of commercial fish consumption on fatal events only and did not estimate the effect on non-fatal events. For modeling purposes there were more data available on fatal than non-fatal events so we chose to defer modeling non-fatal events at this time. This would be an important area for future research.
(11) Appendix D contains an estimation of the exposures to methylmercury that were experienced by this cohort and compares them to U.S. exposures.
(12) Because methylmercury exposure has been hypothesized to be a risk factor, it is important for this analysis to include data that can help to investigate the question of effect of methylmercury on CHD.
(13) We did not use the Wheldon et al. (2004) meta-analysis as the basis for our modeling because it did not include a dose-response function. He et al. (2004a) included such a function.
(14) The point of departure for Bouzan et al. was a literature search conducted of Medline by Wang et. al (2004). Wang et al. screened the studies they found based on matters such as size and age of the study group, duration of the study, whether the study reported exposure only in terms of biomarker levels, and similar matters. Bouzan et al. then imposed three more criteria to ensure that the studies were appropriate for the purpose of quantitative dose-response evaluation, as follows:
- The studies had to quantify risk relative to a no-intake or very-low-intake reference group
- Only studies with designs rated by Wang et al. as "A" (least bias, results are valid) or "B" (susceptible to some bias but not sufficient to invalidate the results) are included
- Includes both fatal and non-fatal strokes.
The application of these criteria lead Bouzan et al. to the five studies they utilized in their meta-analysis as listed in Table IV-5.
(15) Unlike the He et al., (2004a) meta-analysis for CHD, the He et al. (2004b) meta-analysis for stroke did not require that studies include more than two exposure groups due to the relatively limited number of studies available on stroke. For that reason He et al. (2004b) does not include a dose-response estimate for stroke.
Descriptions of Figures IV-1 to IV-4
- FIGURE IV-1 Basic Modeling Structure
Figure IV-1 starts with a box at the top that represents the exposure assessment. Results from this assessment are used in the assessments of risk and benefit for fetal neurodevelopment, risk of fatal coronary heart disease, and risk of fatal stroke. Lines go from the exposure assessment box to boxes for these three assessments. The fetal neurodevelopment assessment is represented by three boxes with lines connected them to each other. The three boxes are labeled "methylmercury contribution to net effect," "fish contribution to net effect," and "net effect." The assessment for coronary heart disease is represented by a single box labeled "net effect of eating a variety of commercial fish on risk of fatal CHD." The assessment for stroke is represented by a single box labeled "net effect of eating a variety of commercial fish on risk of fatal stroke."
- Figure IV-2: Flow diagram of the exposure modeling. The numbers at various steps in flow correspond to numbers in Table IV-1, which is presented later.
This is a flow diagram for the exposure modeling. It starts with the consumption surveys, the results of which are used on animal intake simulation. This simulation brings together the amounts of fish consumed (taking into account market share of species and serving sizes per meal) and the amounts of methylmercury in the fish. The results of the simulation are on an average annual methylmercury intake and an average annual fish intake.
- Figure IV-3: Flow Diagram for the Dose-Response Modeling. The numbers correspond to numbers in Table IV-6 that describe knowledge gaps, assumptions and implications. The numbers start with “8”, Figure IV-2. Box “8” here carries the results of the exposure modeling over to this flow diagram.
This is a flow diagram with boxes and arrows that shows the flow of the dose-response modeling. It begins with with the exposure assessment then moves to the cardiovascular dose-response model and then a cardiovascular rate, i.e., the effect of fish consumption on the rate of fatal cardiovascular heart disease. It does the same for stroke. For fetal neurodevelopment it divides between the neurobehavioral fish response model and the neurobehavioral methylmercury-response model and comes back together for and ??? of neurobehavioral net effect.
- Figure IV-4. Dose-response function for Stroke. The intersection of the dotted line and the straight line represents the lowest dose that Bouzan et al. (2005) modeled
Figure IV-4 depicts a pair of straight-line graphs plotted on a vertical scale of Change in Stroke rate from 0% down to -30% and a horizontal scale of Fish Servings (100 g) per week from 0 to 5. The dotted line (labeled Modified-Bouzan) extends from (0, 0%) down to approximately (0.5, -13%), where it meets the solid line (labeled Bouzan). The solid line extends from (0,0%) to (0, -12%), and then to (5, -22%).