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: Appendix A, Technical Description of the Risk and Benefit Assessment Methodology
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.
Overview — Data Sources
Estimates of daily fish consumption were developed from several different data sources: 1) The U. S. Department of Agriculture's (USDA) Continuing Survey of Food Intake by Individuals (CSFII) survey conducted between 1989 and 1991 (USDA 1993); 2) the National Health and Nutrition Survey (NHANES) conducted in 2001-02 (CDC 2004); and 3) market share data obtained from the National Marine Fisheries Service (NMFS 2002). The aspects of the consumption estimate addressed with the use data from each of these sources is listed in Table AA-1.
CSFII: As the exposure model was designed to generate estimates for each individual in the CSFII survey, the data from this source figured in just about every aspect of the estimates. Records for all fish consumption events were selected for all individuals for whom a full three-day record was included in the survey (3,525 individuals). The survey data were provided with demographic weights that were used to project the survey to the U.S. population. Although more recent data are available, the 89-91 data were accumulated from surveys which tabulated consumption over a three-day period, rather than more recent data which contained records for only two days (CSFII 94-98) or one day (NHANES) The additional day makes the 1989-91 survey a better instrument for characterizing the chronic behavior of fish consumers. Daily intakes from CSFII 89-91 and CSFII 1994-98 are similar.
NHANES: Data from the 30-day fish consumption survey from NHANES were used for two purposes. First, they were used to adjust the short-term population distribution to generate long term fish consumption frequency population distributions and to estimate. Second, they were used to estimate the extent to which different individuals eat a variety of different fish.
NMFS Market Share: Data describing the extent to which different fish species are marketed in the United States were obtained from the National Marine Fisheries Service. Market share data was used to allocate frequency of consumption under two different circumstances. First, it was used to allocate species consumption for CSFII food categories that were composed of multiple species. Second, it was used to allocate species consumption when the short-term survey was considered inadequate for a particular serving (see Variation in Fish Species Consumed, below). In addition, a correction factor was applied to portion sizes from the CSFII survey so that total intake matched per capita estimates from NMFS.
|Exposure Parameter||CSFII 1989-1991||NHANES 2003||NMFS Market Share as of 2005|
|Frequency of Fish Consumption||X||X||-|
|Variety of Consumption||X||X||-|
Adjustments for Chronic Frequency of Intake
Short-term surveys often do not provide accurate estimates of long-term food consumption (Paustenbach 2000). In particular, short-term surveys tend to misrepresent infrequent consumers since they will either not account for consumers who did not eat a specific food item during the survey period and they will project a higher average intake for an infrequent consumer who did happen to eat the specific food item during the survey period. As a result, a short-term survey will underestimate the number of eaters and overestimate average daily intake for eaters for longer periods of time. Furthermore, a short-term survey may not accurately reflect the pattern of fish consumption, i.e., individuals who consume a particular species during the survey period may consume other species over a longer period of time.
To compensate for the inaccuracy of short-term food intake surveys, several adjustments were made. First, the number of fish consumption events was decreased and the number of eaters increased by a Long Term-to-Short Term Consumer Ratio (LTSTCR) with an uncertain range of 2.3 to 2.5. Adjusting the survey data for LTSTCR results in an estimate that in a given year, 85 to 95 percent of the total U.S. population consumes fish. This range is consistent with the food consumption/frequency information available from the 30-day National Human and Nutrition Estimation Survey (NHANES; CDC 2001). Since equal and opposite LTSTCRs were applied to the frequency of consumption and number of consumers, the long-term per capita mean consumption of fish was held constant to short-term consumption.
Because short term surveys are better at monitoring consumption patterns for frequent consumers than for infrequent consumers, the LTSTCR in serving frequency was reduced for frequent fish consumers using an exponential function that reduced the LTSTCR as the number of servings increased according to the following equation:
AS = Annual Servings
D3S = 3 Day Servings
LTSTCR = Long Term to Short Term Consumer Ratio
a, b = model parameters
The model parameters used to extrapolate long-term frequency of consumption from short- term records were chosen to be consistent with the 30-day fish consumption data collected by NHANES (see Figure AA-1).
The CSFII based projection employed the exponential function described in Carrington and Bolger (2002b), using values of 0.696 and 0.356 for the alpha and beta parameters, respectively. These parameters were obtained by fitting the projected frequency distribution to 30 day survey data obtained from NHANES III (CDC, 2003).
Variation in Fish Species Consumed
Short-term surveys also may also fail to portray variation in the types of fish consumed. For example, an individual who consumes a particular species every day of a three-day survey may consume other species at other times during the year. Since the levels of mercury in fish may vary considerably by species, this may significantly influence the exposure estimate for that individual. Therefore, individual exposure estimates employed both the survey data and per capita market share information to build a consumption pattern for each individual. This distribution was derived from the NHANES survey, by calculating the fraction of total fish consumption in the fish category with the highest number of eating occasions for the 403 adult women who consumed fish on four or more occasions (see Figure AA-2).
A distribution representing the extent to which a single species dominates fish consumption over a 30-day period. This distribution was used to determine the extent to which the short-term survey was used to predict long-term fish consumption behavior. Specifically, the fraction of fish belonging to a single category was used to determine the fraction of the species determined by the CSFII. For example, if all the fish for an individual was an identical species, then the CSFII survey was considered to adequately characterize long term consumptions. If the fraction was low, indicating that the individual ate a wide variety of species, then most of the fish were selected from the market share distribution.
Individual variation in species consumption and overall frequency of consumption were assumed to be independent. The assumption is supported by the observation that the repetition ratio and 30-day frequency of seafood in NHANES was used to determined the extent of variation in species consumption is largely uncorrelated (r = -0.11). Some of the overall variation in species selection by each consumer also comes from CSFII. Since these data are paired, whatever correlation there is between species variation and frequency is represented in the exposure model.
Water Loss during Food Preparation
A concentration factor was applied to serving sizes to reflect water loss during food preparation. These factors were based on water loss of 11 percent for fried fish, 21 percent for poached or steamed fish, and 25 percent for baked or broiled fish (EPA 2002, pages 2-5 and 2-6). Group specific correction factors were calculated based on the frequency of different food preparation procedures (e.g. baking, steaming, or frying) within each fish group. A value of 20 percent was used for fish groups represented in the methylmercury surveillance data but not in the CSFII survey. The resulting concentration factors are listed in Table AA-3. Correction factors were not needed for canned tuna since the methylmercury concentration values in that fish group were obtained after cooking and draining of water or oil from the can.
Portion Size Adjustment
A correction factor of 1.15 was applied to portion sizes from the CSFII survey so that total intake matched per capita estimates from NMFS. This correction was factor was calculated as follows:
Average Intake from CSFII (1989-91): 14.3 g/day
Average Intake from NMFS (2005): 16.2 lbs/year = 20.1 g/day
Weight loss During Cooking: 20 percent
Correction factor = 20.1 * 0.8/ 14.2 = 1.125
Methylmercury Levels in Fish
Mercury Concentrations in Individual Species
Most surveys of mercury in fish, as well as biomarkers in blood and hair measure total mercury, and as a result do not distinguish between inorganic mercury and methylmercury. However, when the forms are speciated it has been shown that most (over 90 percent) of the mercury in fish is methylmercury (WHO 1990; Hight & Cheng 2006).
In order to combine the fish consumption data with the levels of mercury in fish, it was necessary to map the 268 food codes employed in the CSFII survey with the groups used for reporting methylmercury levels (see Table AA-2). The mapping resulted in a total of 51 fish groups. In most cases, the correspondence was either direct or the fish ingredient in the survey food code was a member of a methylmercury contamination group. For several species, an analog (or surrogate) was chosen. If there was no other species that was very similar, several new distributions were created that combined multiple methylmercury contamination groups. Specifically, groups were created for crabs, lobster, shellfish, finfish, and all other fish. Per capita market share was used to assign histogram frequencies for each group.
Distributions of methylmercury levels in fish were constructed for each of the 51 fish groups which represented over 99 percent of the fish consumed in the United States. Three different methods were used to construct the distributions:
- For fish categories (fresh tuna, canned light tuna, canned albacore tuna, shark, and swordfish) for which there were over 100 observations, distributions were generated empirically by directly sampling from FDA surveillance data.
- For other species for which additional raw survey data are available, distributions were developed by fitting the distributions to the portions of the cumulative distribution above the levels of detection. A battery of ten distributions was fit to each data set and the four that provided the best fit were used to construct a probability tree. An example is shown in Figure AA-3. See Carrington (1996) for further description of the methodology.
- Since raw data were unavailable for some species, distributions were generated with modeled distributions that reflected reported arithmetic mean values published from a National Marine Fisheries Service survey (NMFS 1978a) for each group and a range analogous to those obtained from tuna, shark, and swordfish. Lognormal and Gamma distributions were used to represent the data, with each model assigned a probability of 0.5 to represent model uncertainty. The magnitude of the shape parameters (the geometric standard deviation of the lognormal distribution and the beta parameter of the gamma distribution) were represented as uniform distributions that encompassed the range of values resulting from fitting the shark, swordfish, and tuna data. The scale parameters (the geometric mean of the lognormal distribution and the alpha parameter of the gamma distribution) were calculated from the arithmetic mean in the NMFS survey and the shape parameter (Carrington and Bolger, 2002).
The type of distribution used for each species is identified in Table AA-3. The one percent of the fish market not included was presumed to follow the same distribution as the rest of the market.
|SPECIES||MERCURY CONCENTRATION (PPM)1||NO. OF SAMPLES||SOURCE OF DATA2|
|ANCHOVIES||0.043||N/A||N/A||ND||0.34||40||NMFS REPORT 1978|
|BASS CHILEAN||0.386||0.303||0.364||0.085||2.18||40||FDA 1990-04|
|BUFFALOFISH||0.19||0.14||‡||0.05||0.43||4||FDA SURVEY 1990-02|
|BUTTERFISH||0.058||N/A||N/A||ND||0.36||89||NMFS REPORT 1978|
|CARP||0.14||0.14||‡||0.01||0.27||2||FDA SURVEY 1990-02|
|CRAB 4||0.06||0.03||0.112||ND||0.61||63||FDA 1990-04|
|CROAKER ATLANTIC (Atlantic)||0.072||0.073||0.036||0.013||0.148||35||FDA 1990-03|
|CROAKER WHITE (Pacific)||0.287||0.28||0.069||0.18||0.41||15||FDA 1990-03|
|FLATFISH 5||0.045||0.035||0.049||ND||0.18||23||FDA 1990-04|
|GROUPER (ALL SPECIES)||0.465||0.41||0.293||0.053||1.205||43||FDA 2002-04|
|HADDOCK (Atlantic)||0.031||0.041||0.021||ND||0.041||4||FDA 1990-02|
|HERRING||0.044||N/A||N/A||ND||0.135||38||NMFS REPORT 1978|
|LOBSTER (NORTHERN/AMERICAN)||0.31||N/A||N/A||0.05||1.31||88||NMFS REPORT 1978|
|LOBSTER (Species Unknown)||0.169||0.182||0.089||ND||0.309||16||FDA 1991-2004|
|LOBSTER (Spiny)||0.09||0.14||‡||ND||0.27||9||FDA SURVEY 1990-02|
|MACKEREL ATLANTIC (N. Atlantic)||0.05||N/A||N/A||0.02||0.16||80||NMFS REPORT 1978|
|MACKEREL CHUB (Pacific)||0.088||N/A||N/A||0.03||0.19||30||NMFS REPORT 1978|
|MACKEREL KING||0.73||N/A||N/A||0.23||1.67||213||GULF OF MEXICO REPORT 2000|
|MACKEREL SPANISH (Gulf of Mexico)||0.454||N/A||N/A||0.07||1.56||66||NMFS REPORT 1978|
|MACKEREL SPANISH (S. Atlantic)||0.182||N/A||N/A||0.05||0.73||43||NMFS REPORT 1978|
|MARLIN *||0.485||0.39||0.237||0.1||0.92||16||FDA 1990-02|
|MONKFISH||0.18||N/A||N/A||0.02||1.02||81||NMFS REPORT 1978|
|MULLET||0.046||N/A||N/A||ND||0.13||191||NMFS REPORT 1978|
|ORANGE ROUGHY||0.554||0.563||0.148||0.296||0.855||49||FDA 1990-04|
|PERCH (Freshwater)||0.14||0.15||‡||ND||0.31||5||FDA SURVEY 1990-02|
|PERCH OCEAN||ND||ND||ND||ND||0.03||6||FDA 1990-02|
|SABLEFISH||0.22||N/A||N/A||ND||0.7||102||NMFS REPORT 1978|
|SALMON (CANNED)||ND||ND||ND||ND||ND||23||FDA 1990-02|
|SALMON (FRESH/FROZEN)||0.014||ND||0.041||ND||0.19||34||FDA 1990-02|
|SCALLOP||0.05||N/A||N/A||ND||0.22||66||NMFS REPORT 1978|
|SCORPIONFISH||0.286||N/A||N/A||0.02||1.345||78||NMFS REPORT 1978|
|SHAD AMERICAN||0.065||N/A||N/A||ND||0.22||59||NMFS REPORT 1978|
|SHEEPSHEAD||0.128||N/A||N/A||0.02||0.625||59||NMFS REPORT 1978|
|SKATE||0.137||N/A||N/A||0.04||0.36||56||NMFS REPORT 1978|
|SQUID||0.07||N/A||N/A||ND||0.4||200||NMFS REPORT 1978|
|TILEFISH (Atlantic)||0.144||0.099||0.122||0.042||0.533||32||FDA 2002-04|
|TILEFISH (Gulf of Mexico)||1.45||N/A||N/A||0.65||3.73||60||NMFS REPORT 1978|
|TROUT (FRESHWATER)||0.072||0.025||0.143||ND||0.678||34||FDA 2002-04|
|TUNA (CANNED, ALBACORE)||0.353||0.339||0.126||ND||0.853||399||FDA 2002-04|
|TUNA (CANNED, LIGHT)||0.118||0.075||0.119||ND||0.852||347||FDA 2002-04|
|TUNA (FRESH/FROZEN, ALBACORE)||0.357||0.355||0.152||ND||0.82||26||FDA 2002-04|
|TUNA (FRESH/FROZEN, BIGEYE)||0.639||0.56||0.184||0.41||1.04||13||FDA 2002-04|
|TUNA (FRESH/FROZEN, SKIPJACK)||0.205||N/A||0.078||0.205||0.26||2||FDA 1993|
|TUNA (FRESH/FROZEN, Species Unknown)||0.414||0.339||0.316||ND||1.3||100||FDA 1991-2004|
|TUNA (FRESH/FROZEN, YELLOWFIN)||0.325||0.27||0.22||ND||1.079||87||FDA 2002-04|
|TUNA(FRESH/FROZEN, ALL)||0.383||0.322||0.269||ND||1.3||228||FDA 2002-04|
|WEAKFISH (SEA TROUT)||0.256||0.168||0.226||ND||0.744||39||FDA 2002-04|
|WHITING||ND||ND||‡||ND||ND||2||FDA SURVEY 1990-02|
1 - Mercury was measured as Total Mercury and/or Methylmercury. ND - mercury concentration below the Level of Detection (LOD=0.01ppm). NA - data not available.
2 - Source of data: FDA Surveys 1990-2003, "National Marine Fisheries Service Survey of Trace Elements in the Fishery Resource" Report 1978 , "The Occurrence of Mercury in the Fishery Resources of the Gulf of Mexico" Report 2000
3 - Includes: Sea bass/ Striped Bass/ Rockfish
4 - Includes: Blue, King, and Snow Crab
5 - Includes: Flounder, Plaice, Sole
An example of a fitted distribution. Ten different distributions were fit to the sample Hg data for Crabs. The four best models were used to create a probability tree that describes the frequency distribution with a representation of model uncertainty. A primary advantage of using distributions to describe the data is that they can be used to extrapolate the concentration in the samples that are below the level of detection - which comprised about 50 percent of the crab samples.
|Tuna, Albacore Canned||3.81%||0.353||Empirical||1|
|Lingcod and Scorpion fish||0.02%||0.286||Surrogate||0.802|
|Carp and Buffalo fish||0.04%||0.203||Modeled||0.871|
|Haddock, Hake, and Monkfish||4.86%||0.17||Modeled||0.802|
|Snapper, Porgy, and Sheepshead||0.86%||0.137||Modeled||0.812|
|Tuna, Light Canned||11.41%||0.118||Empirical||1|
|Anchovies, Herring, and Shad||3.06%||0.05||Surrogate||0.737|
|Perch, Ocean and Mullet||0.47%||0.04||Surrogate||0.809|
|Oysters and Mussels||2.22%||0.023||Modeled||0.782|
1 - Market share calculation based on 2005 National Marine Fisheries Service published landings, imports and exports data (NMFS 2005; NMFS 2008).
As a result of species not included in the list, the sum of the market share values is about 99 percent.
2 - Empirical - Direct sampling of data set, used for large data sets with very few values below the limit of detection. Fitted - Modeled distribution with uncertainty about model form (see text for additional explanation). Used for data sets with a limited number of observations, often with many values below the level of detection. Surrogate - Two generic distributional forms (lognormal or gamma) were employed, with a mean value from 1978 National Marine Fisheries Survey, and a shape parameter shape derived from distributions for other species. This technique was used when only mean values are available.
3 - These values reflect weight after food preparation as a percentage of initial weight. Mercury concentrations for fish as eaten were calculated by dividing initial concentration by the correction factor. No correction factor was applied for canned tuna, since the mercury measurements were made after cooking.
Biomarker Calculations: Mercury in Blood and Hair
While many studies have attempted to relate dietary methylmercury exposure to blood mercury levels, in most cases the correlation is very poor, with r values of 0.3 or less (reviewed in WHO 1990). This lack of correlation may be attributed in large part to the failure of short-term measurements of mercury exposure to gauge long-term dietary exposure (Sherlock & Quinn 1988). The study by Sherlock et al. (1984), in which 20 male volunteers consumed controlled fish diets with known methylmercury concentrations over a 100-day exposure period, was selected for use in this assessment. Mercury blood values monitored for the duration of the study were used to project equilibrium values for a chronic diet-blood relationship. The mean body weight for the subjects was 71 kg, with a range of 52 to 102 kg. The relationship between dietary exposure and mercury blood level appeared to be linear with respect to dose. Although the ratio of mercury blood level to dietary exposure was inversely related to body weight, it was not directly proportional to body weight. Therefore, Sherlock et al., (1984) suggested using a body weight (BW) dose conversion factor of BW.1/3 We have determined that a conversion factor of BW0.44 will result in corrected values that have no correlation with body weight (i.e. r=0; see Figure AH-4) .
Sherlock et al. (1984) extrapolated steady-state blood levels from two other parameters (a and b). The extrapolated steady-state levels reported in the paper were not corrected for body weight. Therefore, the values for each of the 20 subjects were recalculated using BW0.44 to normalize all values to a BW of 70 kg. In order to characterize the measurement error for each subject, 40 bootstrap data sets were generated from the standard deviations reported for each parameter estimate. Each bootstrap set was then fit by 10 different frequency distributions using least squares regression. Three weighted models were retained per bootstrap, which were assigned probabilities on the basis of goodness-of fit and number of parameters (Carrington 1996). The resulting 120 models were then employed as a probability tree to characterize uncertainty from measurement error and model selection. When used in a simulation, the contribution of body weight was calculated by applying BW0.44 to the weight of each subject in the food consumption survey.
Exposure to Other Sources of Methylmercury
Since the present model is intended to represent methylmercury exposure from fish, background mercury blood levels were added to the model to acknowledge the possibility of minor exposures from sources other than fish. This range reflected the levels at the low end of the NHANES 30-day fish survey (CDC 2003). Virtually everyone in the NHANES survey had a blood mercury level above zero, yet 10-20 percent of the NHANES survey population reported no fish consumption, suggesting that there are contributions to blood mercury levels from other sources (e.g., dental amalgams) other than fish. To model the population distribution for background blood methylmercury (i.e., methylmercury from sources other than fish), a normal distribution with an uncertain range of 0.05 to 0.1 ppb for the mean and a standard deviation of 0.02 ppb was used. The distribution was truncated at zero.
For the purposes of predicting hair levels from given blood levels, an empirical distribution was constructed from paired observations from the NHANES 1999-2000 survey (CDC 2003). Since this data is more recent, has a larger sample size, and reflects exposures ranges found in the United States, it was used in favor of other data used in previous versions of the model (Carrington et al., 2002; Carrington et al., 2004). The data were used as follows:
- In order to avoid potential errors arising from analytical imprecision at low concentration levels, only individuals with total mercury levels above one ppb were used.
- The 1999-2000 NHANES survey measured both total mercury and inorganic mercury in blood. In order to avoiding confounding contributions from inorganic mercury, only individuals with an inorganic contribution of 25 percent or less of the total were used.
- The survey consisted of children aged one-five and adult women. Since the present model is only concerned with exposure of adults to mercury, only the data from adult women were used.
- An individual with an extremely high hair level (849 ppm) was excluded since this hair level almost certainly did not result from the consumption of fish (see McDowell et al., 2004 for further discussion; see also Section II of this report).
- Methylmercury concentrations for the 526 individuals who met the above criteria were calculated by subtracting the inorganic mercury concentration from the total concentration.
- Hair/Blood ratios were calculated for each individual to develop a population distribution of ratios. This approach presumes that the ratio is independent of dose (i.e. the relationship between blood methylmercury and hair methylmercury is linear.) The ratio is also presumed to be independent of body weight.
- Not all of the variation observed may be true variation in the pharmacokinetic relationships between blood and hair (or blood and hair, which hair is used as a marker for). Of particular concern is the fact that blood measurements fluctuate and are dependent on the time since the last fish meal, and as a result, measurements made at a single time point may not accurately reflect long-term exposure. Since inorganic mercury was not measured independently in hair, it is also possible that there is some contamination of hair from inorganic mercury - perhaps from environmental sources. Regardless of the explanation, because actual pharmacokinetic variation is almost certainly narrower that the apparent distribution, the distribution was truncated with uncertainty ranges of 20 percent at both ends.
Paired measurements of hair and blood mercury concentrations are plotted in Figure AA-5. The distribution of ratios is shown in Figure AA-6. The tails of the distribution of ratios, which were partially (i.e. as an uncertainty) excluded are shown in Figure AA-7 and Figure AA-8. The figures illustrate that while the majority of the hair/blood ratios fall in a relatively narrow range of 0.1 to 0.3, there is significant departure from this range at both the upper and lower tails.
Methylmercury and Neurological Endpoints
Milestones at Two Years - FDA (Carrington & Bolger 2000)
The dose-response function used to represent the relationship between maternal exposure to mercury, using hair mercury as a marker for dose, and the age of onset of walking and talking was based on the analysis described in Carrington and Bolger (2000). This analysis is based on pooled data from the Iraqi poisoning episode in early 1970's and data obtained from the prospective epidemiology study in the Seychelles. The models used in Carrington and Bolger (2000) differ in one or more of 5 different aspects: 1) The primary model used to describe the relationships between methylmercury exposure and outcome, 2) the statistical model used to describe variation the use of background terms, 3) background terms used to describe variation independent of dose, 4) study variables that accounted for differences between Iraq and Seychelles, and 5) the order in which the above components are assembled, which in at least some cases, determined how they interacted with one another. Additional details, including a list of the alternative model components, are given in Appendix C.
One modification was made to the prior analysis. Since the onset of walking and talking in Iraq was recorded in sixth month increments, the reported ages were 0-6 months higher than the actual age of onset. Since milestones in the Seychelles study were recorded in one month increments, this accounts for some, but not all, of the differences in the baseline age of onset between the two populations. To correct for this reporting imprecision, three months were subtracted from the reported milestone ages for the data obtained from Iraq. The data were the reanalyzed as described previously. Although this correction reduced the importance of the study parameter used to account for differences in the two populations, it had little impact on model fitness the other parameter estimates.
IQ at Seven Years - Axelrad et al. (2007)
The analysis developed by Axelrad et al (2007) for the U.S. Environmental Protection Agency (EPA) developed separate integrated estimates of IQ for three different prospective epidemiology studies: New Zealand, Seychelles, and the Faroe Islands which are presented in the following table.
|Study||Linear Slope1||Pop. Size2||Notes|
|New Zealand||-0.50 ± .027||237||Reported in Table III of Crump (1998); outlier child omitted; rescaled to study population variance|
|Seychelles||-0.17 ± 0.13||643||Reported in Table 2 of Myers (2003); rescaled to study population variance|
|Faroe Islands||-0.124 ± .057||917||Reported in Axelrad et al (2007), based on structural equation modeling of three IQ subtests by Budtz-Jørgensen et al. (2005).|
1 ± .Standard Error of the Mean
2 - Population size reflects final study group size used to for the dose-response evaluation.
Axelrad et al. (2007) used a Bayesian analysis to generate an estimate of a single slope of -0.153 and confidence intervals based on the standard error of the mean that ranged from -0.047 to -0.259. That estimate is employed in our analysis as a normal distribution, per Axelrad, with a mean of -0.153 and standard deviation of 0.064.
IQ at Seven Years - Harvard Center for Risk Analysis (HCRA): Cohen et al. (2005b)
Cohen et al. (2005b) conducted an analysis that is very similar to that of Axelrad et al. (2007). However, there are two noteworthy differences. First, the uncertainty analysis was based on a limited number of alternative assumptions. Although it was based on the same three epidemiology studies, it did not generate separate IQ estimates for each study. It was noted that much lower estimates were obtained from the Seychelles than from the other two studies. It is also not clear whether the confidence ranges are statistical (i.e. fifth and 95th percentiles) or encompass the entire range of IQ estimates. Second, the analysis developed linear coefficients from the Faroe Islands study using the low end of the log(dose)-linear slope reported in the study. This range was chosen because it most closely matches exposures in the United States. The results using this analysis yielded a central estimate of -0.7 IQ points per ppm in maternal hair. The confidence range for this estimate was 0 to -1.5 IQ points per ppm in hair. However, because the NAS committee deemed the log(dose) transform used in the Faroe study to be "biologically implausible" (NRC, 2000), Cohen et al. (2005b) also calculated from the first and third quartiles, which yielded an estimate of -0.2 IQ points per ppm in hair - which is much closer to the Axelrad et al. (2007) estimate derived from the entire range.
As pointed out in the National Academy of Sciences methylmercury report (NRC, 2000), the log(dose)-linear and linear models provide a similar description of at least some of the data from the Faroe Islands study. However, the models diverge greatly at doses both below and above the ranges encountered in the study. In addition to the more esoteric and theoretical criticisms of the log(dose) regression, there are empirical grounds for discounting the log(dose) transform from other data in the literature and from common experience. First, the log(dose) transformation implausibly predicts that the size of the effect increases as the dose decreases. In fact, the predicted increase in IQ approaches infinity as the dose approaches zero. If this were even approximately true there would be huge differences in the IQ of populations who do not consume fish. Second, the log(dose) transformation predicts that there is relatively little additional effect on IQ at doses higher that those encountered in the Faroe Islands study. This prediction is inconsistent with the results from Iraq and Minamata where clinical effects that were much more severe than those modeled in the Faroe Islands study occurred at higher levels of exposure. The log(dose) scale yields an estimated decrease of about seven IQ points per 10-fold increase in mercury level. Since the levels in various tissues were 10-100 times higher in Iraq and Minamata than in the Faroe Islands, the log linear model predicts relatively modest further effects corresponding to seven-14 IQ points at the higher dose levels (i.e. less than one standard deviation). However, two children in Iraq were unable to walk or talk at five years of age. Since the standard deviation for these milestones is roughly two months, this represents a developmental delay of about 18 standard deviations, or about 270 points on an IQ scale. Many other children in the Iraq study also displayed overt neurological symptoms that are not predicted by the log linear model (see Figure AA-11). On the other hand, the observations from Iraq are consistent with a linear model (Carrington & Bolger, 2000, and see above) that is close to the slope noted in the secondary analysis from Cohen et al. (2005b). Given its implausibility, the log(dose) transformed doses does not merit serious consideration as a mathematical tool for drawing conclusions from the Faroe Islands study. The secondary analysis based on the range of the Faroe Islands data is more reliable.
Finally, the Cohen et al. (2005b) analysis did not yield a formal quantitative representation of uncertainty. They did present a plausible range of values that was based on various combinations of scores that were grouped by the similarity of the individual measures (e.g. cognitive, motor, language) or the study in the test was conducted (i.e. New Zealand Seychelles, or Faroe Islands). The uncertainty ranges in our simulation models employ a probability tree comprised of the individual measures using the weights assigned in the original paper. Also this approach yields a wider confidence interval than the plausible range in the original report, the average is identical.
Bayley Scales at 12 months - Poland Study (Jedrychowski et al., 2005)
A small study of 233 infants in Krakow, Poland conducted by Jedrychowski et al. (2005) recorded the concentration of mercury in maternal and cord blood and subsequently examined the infants at 12 months of age. The range of mercury exposures was relatively narrow and lower than either the Seychelles or the Faroe Islands. Two test domains were reported - the Bayley Cognitive and the Bayley Psychomotor. After obtaining the raw data from the authors, the relationship between both biomarkers and both test scores was examined by linear regression. It may be observed from Figure AA-9 and Figure AA-10 that the regression slopes are relatively flat relative to the overall variation in the scores. Nonetheless, there is a small negative slope for three of the four regression results which are similar in magnitude (see Table AA-5).
|Hg Biomarker||Bayley Test Score||Regression Slope||Test SD||Z slope||Z Slope Hair Equivalent1|
1 - Conversion factors of 5 and 3.3 were used to convert maternal and chord blood concentrations to hair equivalents, respectively.
Comparison of Neurodevelopmental Dose-Response Functions
Quantitatively, the dose-response functions developed from Iraq, New Zealand, Seychelles, and the Faroe Islands can be grouped into three categories (see Figure AA-11 and Figure AA-12):
- The dose-response functions developed from developmental milestone data from the Iraqi poisoning episode, the IQ functions derived from the Faroe Islands using a linear dose-response model, and the Myers et al. (2003) slope for IQ in the Seychelles are all similar at both high and low doses. Since the Iraqi estimates are anchored on high dose data, only this group of dose-response functions is consistent to what was observed with the Iraqi study where there are data to anchor the high dose estimates.
- The two estimates derived from New Zealand and the linear slope derived from the lower quartile of the log-dose transformation of the Faroe Islands study by Cohen et al. (2005b) yield slopes that are much higher than those in the first group. These slopes are not consistent with Iraq. The Faroe Islands slope is demonstrably different because it is derived from a supralinear dose-response function. However, the New Zealand study does seem to indicate a higher neurodevelopmental effect relative to normal variation than would be expected from the Iraqi epidemic. Although the slope derived from the Poland study might also be placed in the "high" category, it is not so much higher than Iraq that it is necessarily unrealistic.
- The slope for IQ derived by Cohen et al. (2005b) from the Seychelles study is slightly positive (i.e. the net decrease is negative). When extrapolated to high doses, this is also clearly inconsistent with the epidemics in Iraq and Japan. This discrepancy can be explained without resorting to uncontrolled variables. First, a sublinear dose-response model is plausible. Second, the fact that the methylmercury exposure in the Seychelles is almost entirely from fish may provide a beneficial effect of fish consumption that equals or exceeds the negative effect of methylmercury.
The bases of the confidence intervals for the three dose-response functions considered in this analysis are different. These differences reflect the scientific rationale behind the derivation of the dose-response functions from their associated data. Although these differences have relatively little effect on the central estimates, they do affect the width and shape of the confidence intervals. In particular, some dose-response functions reflect statistical notions of probability (i.e., the uncertainty is related to an underlying frequency), while some do not. Uncertainties based on notions of frequency can be represented by continuous statistical distributions. The other sources of uncertainty may be represented with probability trees, where the sum of the probabilities of each model, study, or measure is one (Hacking, 1976; Rescher, 1993.). The sources of uncertainty for each dose-response function considered are summarized in Table AA-6. In spite of the differences in approach, the confidence intervals for the neurobehavioral dose-response functions have a breadth that is quite comparable (see Figure AA-12).
The values plotted are the median estimates of the uncertainty distributions. The dose-response functions are listed in the legend in the order in which they appear on the graph, from left to right at the high-dose end of each function.
The values plotted are the median estimates of the uncertainty distributions. The dose-response functions are listed in the legend in the order in which they appear on the graph, from top to bottom at the high-dose end of each function. The functions labeled from Iraq also include data from the Seychelles.
|Dose-Response Analysis||Sampling Error1||Model Uncertainty2||Study Uncertainty3||Measure Uncertainty4|
|Carrington & Bolger (2000)||No||Yes||Yes||No|
|Axelrad et al.(2007)||Yes||No||Yes||No|
|Cohen et al. (2005b)||No||No||Yes||Yes|
1. Sampling Error. This statistical notion of probability arises when generalizations about a large population are drawn from a smaller population. The confidence intervals reflect the notion that the small sample is randomly drawn from the entire population and that the subset may not be entirely representative of the whole population.
2. Model Uncertainties. Different mathematical equations can often be used to draw a generalization from data. As long as the models are in roughly the same range as the data, then it may make little difference which mathematical form is used since all will be constrained by the data. On the other hand, when extrapolating from high to low doses, the models are often not sufficiently constrained by data at low doses to make model selection an irrelevant issue. Since it is generally not possible to establish that one and only one dose-response model is correct, potential model bias may be eliminated by including model uncertainty in the analysis by using more than one model.
3. Study Uncertainties. It is not uncommon for different studies that are concerned with causal relationships between the same variables to yield different results. This can generally be attributed to the presence of one or more uncontrolled variables in at least one of the studies. Not surprisingly, variations in apparent causal relationships are especially common in epidemiology studies where there are many uncontrolled variables. While epidemiologists try to address this issue by modeling variables that are known to influence an outcome, this introduces additional model uncertainties (i.e. the relationships of the other variables may not be modeled correctly), and there always may be additional factors that are unaccounted for. The Axelrad estimate presumes that an underlying mean value common to all the studies is the true value, and therefore the confidence interval does not reflect differences between studies. Although Cohen et al. (2005b) produced an analysis that averaged the results from all three studies into a single estimate, the confidence intervals reflect the differences in the studies.
4. Measure Uncertainties. The relative public health significance of different measures can also be a source of significant uncertainty. This is particularly apt to occur in an economic or cost-benefit analysis where an abstract concept of value is used. In addition, there may be uncertainties in how different measures are related. This is especially true for "IQ" measures which are generally a collection of different measures that are partially related. This issue can be treated as a statistical problem by modeling the extent to which two measures are correlated. However, there can still be additional uncertainty over whether or not two scales are measuring the same attribute, even if they are highly correlated.
Fish Consumption and Neurodevelopmental Endpoints
Daniels et al. (2004)
Daniels et al. (2004) studied the relationship between maternal fish intake during pregnancy and cognitive development from questionnaires posed to mothers of 7,421 English children born in 1991-1992. In a subset of that population, they also studied the relationship between pre-natal exposure to methylmercury and cognitive development. Finally, they studied the relationship between postnatal fish intake by the children themselves and their cognitive development. Each individual child's cognitive development was evaluated using adaptations of the MacArthur Communicative Development Inventory at 15 months of age and the Denver Developmental Screening Test at 18 months.
Their study measured and categorized the maternal fish intake (mum) of oily and white fish intake as follows: rarely or never, once per two weeks, one to three times per week, and four or more times per week. The estimated average fish intake per meal was 4.5 ounces or 127.6 grams. The child's fish intake was monitored at ages six months (child6) and 12 (child12) months by simply noting whether or not at least one fish meal was consumed per week. The study also recorded the age of the child in weeks (age) at the completion of the MacArthur Communicative Development Inventory (MCDI) and the Denver Developmental Screening Test (DDST).
We analyzed the data from this study using multivariate linear regression analysis. Each of the six outcomes (i.e., three different MCDI scores and three different DDST scores) was analyzed four different ways:
Each of the six outcomes was analyzed four different ways:
- Maternal fish intake; age of child at testing; children's fish intake at six months, mercury concentration in cord tissue;
- Maternal fish intake; age of child at testing; children's fish intake at 12 months; mercury concentration in cord tissue;
- Maternal fish intake; age of child at testing; children's fish intake at six months;
- Maternal fish intake; age of child at testing; children's fish intake at 12 months;
Results are shown in Tables AA-7-10. Children's fish intake was a discrete (zero or one) variable and was used as such in the regression analysis. However, for the purpose of deriving a slope for the relationship between the child's intake and test outcome, we employed an estimated average fish intake per meal of two ounces, which yields an average daily intake of eight grams/day.
|Behavioural Test||Mum (g/d)||Child6 (g/d)||Cord Mercury (ppm)||Age (weeks)||Subjects (n)|
|MCDI Social activity at 15m||0.0058||0.16||0.71||0.32||1053|
|Denver total development score (18m)||0.0055||0.18||0.23||0.39||1009|
|Denver communication score (18m)||0.0009||0.06||0.61||0.19||1013|
|Denver social achievement score (18m)||-0.0002||0.05||-0.90||0.13||1013|
|Behavioural Test||Mum (g/d)||Child6 (g/d)||Cord Mercury (ppm)||Age (weeks)||Subjects (n)|
|MCDI Social activity at 15m||0.007||0.06||0.44||0.33||1053|
|Denver total development score (18m)||0.0085||-0.0008||-0.08||0.39||1009|
|Denver communication score (18m)||0.0021||-0.012||0.50||0.20||1013|
|Denver social achievement score (18m)||0.0005||0.006||-0.99||0.14||1013|
|Behavioural Test||Mum (g/d)||Child6 (g/d)||Age (weeks)||Subjects (n)|
|MCDI Social activity at 15m||0.0113||0.099||0.34||7466|
|Denver total development score (18m)||0.0042||0.093||0.38||7204|
|Denver communication score (18m)||0.0024||0.032||0.15||7223|
|Denver social achievement score (18m)||0.0004||0.028||0.10||7215|
|Behavioural Test||Mum (score per g/d)||Child12 (score per g/d)||Age (weeks)||Number of Subjects|
|MCDI Social activity at 15m||0.011||0.10||0.34||7466|
|Denver total development score (18m)||0.0042||0.084||0.38||7204|
|Denver communication score (18m)||0.0026||0.020||0.146||7223|
|Denver social achievement score (18m)||0.00022||0.031||0.10||7215|
In order to evaluate the relative importance of each variable, the relative contribution of each variable on each outcome measure was gauged calculating a "maximum Z-score," which was calculated as follows:
Maximum Z = Variable Range * Slope / Test Score SD
The extent to which these values (Table AA-10 and Table AA-12) deviate from zero (i.e. positive or negative) indicate the relative strength of the relationship between the independent and dependent variables. The following general conclusions may be drawn from Tables AA-7 through AA-12:
- Since it has a much bigger contribution to the variation in outcome, age of testing is clearly an important variable for all outcomes (see Table AA-7 through AA-10. The slopes are uniformly positive and the magnitudes of the slopes are not greatly affected by which of the other variables are included.
- With the smaller data set that included mercury (Table AA-7 and Table AA-8), there are no clear trends for cord mercury, maternal fish intake, or children's fish intake. Not only are both positive and negative slopes attained from the regression analyses, somewhat discrepant results are obtained when child's fish consumption at 12 months is used instead of a 6 months.
- With the full data set (without mercury; Table AA-9 and Table AA-11), there a consistent, albeit small, positive relationship between fish intake by both mother and child and test outcomes. On a per gram basis (i.e. if one meal is assumed to correspond to eight grams per day), the slopes are considerably higher for direct consumption by the children. More of the total variation is accounted for by children's intake as well (Table AA-10 and Table AA-12).
|Behavioural Test||Maximum Z-score||Standard Deviation|
|MCDI Verbal Comprehension||0.00||0.00||0.00||0.26||30.7|
|MCDI Social activity at 15m||0.10||0.09||0.05||1.20||5.4|
|Denver total development score (18m)||0.11||0.00||-0.01||0.56||5.6|
|Denver communication score (18m)||0.06||-0.04||0.12||0.64||2.4|
|Denver social achievement score (18m)||0.02||0.02||-0.28||0.49||2.2|
|Behavioural Test||Maximum Z Score||Standard Deviation|
|MCDI Verbal Comprehension||0.08||0.18||0.57||31.5|
|MCDI Social activity at 15m||0.15||0.15||1.61||5.5|
|Denver total development score (18m)||0.06||0.12||0.62||5.6|
|Denver communication score (18m)||0.08||0.07||0.55||2.4|
|Denver social achievement score (18m)||0.01||0.12||0.42||2.2|
In order to evaluate potential net benefits to infants from mothers eating fish, Z-score slope from MCDI Verbal Comprehension and DDST Communication Scores were used. The results from the various regression analyses are given in Table AA-12. It may be observed that the slopes derived from the full data set all fall in a range of 0.0010 to 0.0012. The slopes derived from the partial data set that included cord mercury as a variable are less consistent. In particular, in the analyses where the cord slope mercury was positive (i.e. better scores were obtained from mothers with higher mercury levels) the slope for maternal fish consumption was diminished. This result may be explained by the fact that blood and fish consumption are highly correlated.
As a summary of these results, a triangular distribution with a minimum of 0, and most likely value of 0.001 and a maximum value of 0.0012 was used in our simulation model intended to illustrate combined effects. An illustration of the combined dose-response with methylmercury concentrations fixed at the market average is shown in Figure AA-14.
|Analysis||MCDI Comprehension||Denver Communication|
|Partial Data Set, with Cord Hg, Children at 6 months||0.0010||0.0003|
|Full Data Set, without Cord Hg, Children at 6 months||0.0012||0.0010|
|Partial Data Set, with Cord Hg, Children at 12 months||0.0000||0.0009|
|Full Data Set, without Cord Hg, Children at 12 months||0.0011||0.0011|
All units are for ΔZ per g of fish consumed per day.
A combined dose-response function for fish and mercury on neurobehavioral development that combines the methylmercury-delayed talking dose-response function and the Daniels et al. (2004) fish-verbal test score dose-response function. For this graph, it was assumed that all fish contained a market average concentration of 0.086 ppm. Only the median response is displayed. Although there is no variability represented in the fish dose-response, the dose-response function for mercury includes population variability from the biomarker relationships and the hair-response model. Although the results displayed in the graph are entirely positive (i.e. increased fish consumption results in increased neurobehavioral performance), at the low end of the population distribution there is a small negative component.
Fish Consumption and Cardiovascular Disease
Cardiovascular Death - He et al. (2004a)
The main result of the meta-analysis produced by He et al. (2004a) was an estimate of the relative risk associated with each of four different rates of fish consumption. This analysis was based on studies 1-13 in Table AA-14. However, they also reported the results of a regression analysis of the pooled data which yielded a slope of seven percent per 20 grams of fish per day with a confidence interval (CI) of one to 13 percent. The confidence intervals were used to calculate the standard error of the mean (SEM) of 3.65 percent. This translates to a linear slope with a normal distribution of 0.35 percent per g-day and an SEM of 0.18 percent.
Stroke - Bouzan et al. (2005)
Bouzan et al. (2005) conducted a regression analysis with data from multiple epidemiological studies that related the frequency of fish consumption to stroke. They included data from studies numbered two, three, five and six in Table AA-16 and an additional case control study (Caicoya 2002). Their regression analysis generated both linear slope and an intercept. Bouzan et al. (2005) interpreted the intercept as an indicator of the benefit or risk associated with any fish consumption. Since the idea that a nanogram of fish would have a substantial health impact is implausible, the Bouzan model was modified to attribute the low dose effect to the first 50 grams of fish per week, which roughly corresponds to the low end of the range of exposures in the analysis. Thus, the reduced risk of 12 percent (CI: - one percent, 25 percent) attributed to the intercept was translated to a slope of 1.68 percent per g/day, an SEM of 1.1 percent, and a maximum response at 7.1 grams/day. For higher doses, and additional reduction was based on the slope estimates provided by Bouzan et al (2006); the reported values of two percent per 100 grams/week (CI: -2.7 percent, 6.6 percent) were translated to a normal distribution with a mean of 0.14 percent per grams/day with an SEM of 0.2 percent. For uncertainty characterization, paired slope and intercept parameters were generated that matched the reported correlation coefficient of 0.77 (Cohen et al., 2005a).
Cardiovascular Death and Stroke - FDA, 2007
Two meta-analyses conducted by He et al. (2004a,b) were used as a starting point for the development of dose-response function for the relationship of average fish consumption (grams/day) and the population frequency of two endpoints -- cardiovascular death and stroke. Although based on the same data, a second analysis (referred to throughout this report as "the stroke pooled analysis model" due to the pooling of models of individual studies, as described below) differs from the work of He in the following respects:
- Dose-response models were fit to data from individual studies rather than pooled data from all of the studies. This allowed for analysis of the uncertainty arising from the idea that the studies may be imperfectly analogous to the U.S. population. The models of the individual studies were pooled into a common dose-response function with a probability tree. Since it is not assumed that all the studies are measuring the same underlying population rate (i.e. there is an underlying "true" mean value for the complete set of cohorts), the confidence intervals in our analysis are greater than those of He. This difference is analogous to using the standard deviation rather than the standard error of the mean to characterize uncertainty.
- Sampling error was represented by binomial sampling from each individual data point instead of assuming a common variance across all studies and dose groups. Specifically, after generating 300 bootstrap data sets from each study, dose-response functions were estimated for each set by using nonlinear regression and weighted least squares as a goodness-of-fit measure. The data points were weighted by the number of person-years associated with each observation.
The data were fit using a linear model, with a maximum effective dose parameter (Hockey Stick) and two sigmoidal models; a three parameter Michaelis-Menten function (KD, minimum, and maximum) and a Hill function (KD, power, minimum, and maximum). The inclusion of both maximum and minimum parameters allowed the model to indicate that a subset of the disease occurrences are reduced or increased by fish consumption. The sigmoidal models were also permitted to have both positive and negative relationships between disease outcome and fish consumption. Since the Hill function has four parameters, it was only used for studies with at least four dose groups. More specifically, the following equations were used:
Dose<MaxDose: Disease Rate = Intercept + Dose * Slope
Dose>=MaxDose: Disease Rate = Intercept + MaxDose * Slope
Disease Rate = Min + (Max-Min) * (KD/(Dose+KD))
Disease Rate = Min + (Max-Min) * (KDk/(Dosek+KDk))
- The relationship between dose and frequency of outcome was modeled for each study instead of pooling the data from all the studies. A probability tree was used to integrate the results of each study into a single dose-response function, instead of averaging the results by either pooling the data or using Bayesian Model Averaging (e.g. the techniques used by Axelrad et al. (2007) for methylmercury and IQ). The probabilities assigned to the studies were weighted by the square root of the sample size. The fundamental difference in this approach is that it is not assumed that there is necessarily a common effect across studies. Instead, the effect may be different in magnitude between studies, and therefore each study is considered independently as a plausible prediction. The net effect of this approach is that the confidence intervals associated with the dose-response function are much wider.
As a means of accounting for known differences among dose groups in each study, adjusted rate estimates were used instead of the relative risks. These were calculated for each non-referent group by calculating the number of events yielding the published relative risk, using the following equation:
Adjusted Group Events = (Adjusted Relative Risk * Referenent Events * Group Person Years) ÷ Referent Person Years
The main difference in this approach compared to that of He is that it allows sampling error from the low dose group to be represented instead of being fixed to a value of one.
- One stroke study (Keli et al. (1994) that was included in the He et al. (2004b) meta-analysis was not included in the present analysis because it only contained two groups, which is insufficient for modeling a dose-response relationship. In addition several studies published after the He meta-analyses that met the original criteria were included: Folson and Demissie (2004), Nakamura et al. (2005), and Iso et al. (2006) were added for CHD, and Nakamura et al. (2005) and Mozaffarian et al. (2005) were added for stroke.
|Study||Pop. Size||Events||Average Age at Baseline||Average Follow Up||Study Median Age||% Male||Nationality|
|1. Kromhout et al. (1985)||852||78||72.5||20||82.5||100||Netherlands|
|2. Fraser et al. (1992)||26,473||260||52||6||55||38||USA|
|3. Ascherio et al. (1995)||44,895||264||55||6||58||100||USA|
|4. Daviglus et al. (1997)||1,822||430||47.6||30||62.6||100||USA|
|5. Mann et al. (1997)||10,802||64||34||13||40.65||38||UK|
|6. Albert et al. (1998)||20,551||308||53.2||11||58.7||100||USA|
|7. Oomen et al. (2000) - Finland||1,088||242||57.8||20||67.8||100||Finland|
|8. Oomen et al. (2000) - Italy||1,097||116||58.3||20||68.3||100||Italy|
|9. Oomen et al. (2000) - Netherlands||553||105||58.4||20||68.4||100||Netherlands|
|10. Yuan et al. (2001)||18,244||187||54||12||60||100||China|
|11. Hu et al. (2002)||84,688||484||42||16||50||0||USA|
|12. Mozaffarian et al. - (2003)||7,389||247||72.5||11||78||39||USA|
|13. Osler et al. (2003)||3,910||247||47||9||51.65||53||Denmark|
|14. Folsom & Demissie (2004)||41,836||922||62||14||69||0||USA|
|15. Nakamura et al. (2005)||8,879||142||51||19||60.5||44||Japan|
|16. Iso et al. (2006)||41,578||62||49.5||12||55.5||48||Japan|
|Cohort||Person-Years||Fish Consumption (g/day)||Adjusted Relative Risk||Unadjusted Events||Adjusted Events|
|Most Likely||Lower Bound||Upper Bound|
The cohort numbers refer to the studies listed in Table AA-1.
The figure numbers refer to the studies listed in Table AA-14. The dose-response functions from each study was normalized to the rates for U.S. males aged 45 and above.
|Study||Pop. Size||Events||Average Age at Baseline||Average Follow Up||Study Median||% Male||Location|
|1. Morris et al. (1995)||21,185||281||52||4||54||100||USA|
|2. Orencia et al. (1006)||1,847||76||47.6||30||62.6||100||USA|
|3. Gillum (1996)||2,059||262||62||12||68||100||USA|
|4. Gillum (1996)||2,351||252||62||12||68||0||USA|
|5. Yuan et al. (2001)||18,244||460||54||12||60||100||USA|
|6. Iso et al. (2001)||79,839||574||34||14||41||0||USA|
|7. He et al. (2002)||43,671||608||53.4||12||59.4||100||USA|
|8. Sauveget et al. (2003)||40,349||1462||56||16||64||100||Japan|
|9. Mozaffarian et al. (2005)||8,879||288||58.3||12||64.3||44||Japan|
|10. Nakamura et al. (2005)||4,775||626||58.3||12||64.3||42||USA|
|Cohort||Person-Years||Fish Consumption (g/day)||Adjusted Relative Risk||Unadjusted Events||Adjusted Events|
|Most Likely||Lower Bound||Upper Bound|
The figure numbers refer to the studies listed in Table AA-16. The dose-response functions from each study was normalized to the rates for U.S. females aged 45 and above.
As an alternative, in addition to cohort size, studies were also weighted by sex and age relative to each subpopulation. This results in dose-response functions that vary by sex and age (results not shown). There are some small, but noticeable, differences. The impact of seafood consumption on CHD mortality appears to be somewhat greater in males than females, and greater in older populations. For stroke, while there is virtually no evidence to suggest that age is an important factor, there appears to be a greater impact in females.
Cardiovascular Disease Rates in the United States
All of the cardiovascular dose-response models used predict relative rates of disease as a function of fish consumption. In order to predict the number of cases, baseline rates of disease were calculated and all the estimates were normalized to U.S. population rate using the rate from each study irrespective of fish consumption. For CHD death rate, data from National Center for Health Statistics (NCHS 2006) and the U.S. Census Bureau (2001) were used to calculate age-specific rates for each population group which were then adjusted for sex differences using data from Ho et al. (2004). Since the latter article did not contain rate information for persons under the age of 45, the relative rates for men and women in the youngest age group (45-50) were used to correct for sex differences in the 15-44 subpopulations of both sexes. For Stroke, Age and sex specific population rates were calculated from data compiled by NCHS (NCHS 2006). The resulting rate estimates are presented in Table AA-18.
|Sex||Age||CHD Death Rate||Stroke Death Rate|
The exposure assessment was constructed using data from the 3,524 selected individuals in the CSFII survey data set. This strategy maintained the information about individual characteristics associated with each estimate of mercury exposure. It also retained the limited information present in the three-day survey about long-term consumption patterns.
The simulation model, constructed in Microsoft Excel, consisted of three iterative loops with the following logical structure:
- Begin Uncertainty Loop
- Randomly Select Distributions for Fish methylmercury Concentration
- Randomly Select percent Consumers (85-95 percent - from NHANES)
- Randomly Select Annual Serving Variability Parameter
- Begin Population Loop (3,525 Individuals in CSFII)
- Calculate Average Serving Size for Individual (from CSFII)
- Calculate three-Day Servings (from CSFII)
- Calculate Annual Servings (using model)
- Randomly Select Fish Consumption Individual Variability
- Begin Annual Exposure Simulation (# of Annual Servings)
- Randomly Select Survey Source (CSFII vs. Market Share)
- If Market Share, Randomly Select Species
- Randomly Select methylmercury Concentration for Identified Species
- Correct for Water Loss During Cooking
- Calculate methylmercury Intake
- Sum Total Fish Intake for Individual
- Sum Total Methylmercury Intake for Individual
- Next Serving
- Calculate and Record Average Daily Methylmercury and Fish Intake
- Record Demographic Characteristics for Individual (from CSFII)
- Next Individual
- Next Plausible Set of Assumptions
The Uncertainty loop consisted of 200 iterations and contained the uncertainty distributions developed for methylmercury concentration in the various fish groups and projection of the short-term consumer survey to long-term fish consumption patterns were re-sampled within this loop. The random numbers used for each iteration were generated prior to running the simulation. This allows post-hoc investigation of individual results and allowed the LTSTCR to be carried forward to the biomarker simulation. Each iteration of the second Variability loop consisted of an individual from the CSFII survey who consumed one or more servings of fish during the three-day survey. The number of servings and average serving size for each individual are calculated at this step.
The annual number of servings was then used to set the number of iterations for the third loop, in which in each iteration simulated a fish consumption event. First, a random number was used to select the information source (CSFII or per capita) to be used for the serving. Specifically, if the random number was less than the percentile ranging from 0.2 to 0.8 selected at the outset of the uncertainty iterations, a randomly selected CSFII record for the individual was used to identify the species and the serving size. Otherwise, a species was randomly selected from a histogram distribution based on per capita disappearance rate, and the average serving size for the individual was used. Second, the mercury concentration for the species consumed by randomly sampling from either an empirical distribution (shark, swordfish, and tuna) or a modeled distribution using a mean value from NMFS data and a distribution selected at the outset of the uncertainty iteration. Methylmercury exposure from the serving was then calculated by multiplying serving size by concentration. After completion of the specified number of servings, total methylmercury exposure for the year was summed from all the servings, and then divided by 366 to yield an average daily methylmercury. This number was recorded along with the age, sex, body weight, and demographic weight for the individual. After completion of the middle and outer loops, a two-dimensional array was produced with dimensions of 200 uncertainty iterations by 3,525 variability iterations. These were stored and used as the basis for the subsequent biomarker simulation.
At the end of each variability loop, per capita population percentiles were calculated. This was accomplished by generating a frequency histogram from the 3,525 estimates where the width is proportional to the demographic weight provided with the survey. Individuals not consuming fish were included in the distribution by introducing a value of zero for the fraction of non-consumers. The percentage of fish consumers was calculated by multiplying the number of consumers in the three-day survey by the LTSTCR for the current uncertainty iteration. Subtraction of the resulting value from one yielded the fraction of non-consumers.
Dose-response simulations were constructed to predict responses for each of the four subpopulations modeled. Individuals belonging to each subpopulation were extracted from the exposure assessment.
Each simulation consisted of a two-dimensional Monte-Carlo routine with an outer uncertainty loop and an inner variability loop with the following logical structure:
- Begin Uncertainty Loop
- Randomly Select Uncertainty Iteration from Exposure Assessment
- Randomly Select Population Model for Diet-Blood Ratio
- Randomly Select Dose-Response Models
- Begin Population Loop
- Randomly Select Individual from Exposure Assessment
- Randomly select Diet-Blood Ratio from Population Model
- Correct for Body Weight
- Add other Mercury Exposures
- Calculate Blood methylmercury
- Randomly select Blood-Hair Ratio from Empirical Distribution
- Calculate Predicted Hair Value
- Calculate methylmercury-Dependent Neurobehavioral Outcomes
- Calculate Fish-Dependent Cardiovascular Outcomes
- Record Output
- Next Individual
- Calculate Population Distributions for Neurobehavioral Outcomes
- Calculate Average Population Rates for Cardiovascular Outcomes
- Next Plausible Set of Assumptions
A simulation for the entire population was run with 5,000 variability iterations and 500 uncertainty iterations.
At the outset of each uncertainty iteration, one of the 200 uncertainty iterations from the exposure assessment and a population model for the diet to blood ratio were randomly selected. The variability loops were then run with random selection of the individual from the exposure assessment, the diet/blood ratio from the population model, and the blood/hair ratio from the empirical distribution. Random numbers for the variability iterations were generated prior to the simulation and the same set of values were used for each uncertainty iteration. These values were then used to calculate blood and hair values for each individual. At the conclusion of each variability loop, per capita population percentiles were calculated in the same manner the percentiles for daily methylmercury exposure.
The logic of the dose-response model is also illustrated in AA-19. Since some of the items displayed pertain to populations and other pertain to individuals, the statistical relationships are not illustrated. In particular, note the following:
- The population value of "% Consumers" is used to randomly determine if the individual is a fish consumer. If not, the methylmercury contribution from fish is ignored.
- The cardiovascular rate calculated as a response in not an individual outcome. For this reason, average or total population rates are reported rather than population distributions.
Also, the arrow indicating a relationship between Fish Intake and Neurobehavioral outcome applies only to the simulation model that includes a Daniels et al. (2004)- derived benefit from fish consumption. The simulation models for men and older women only include cardiovascular outcomes.
Descriptions of Figures AA-1 to AA-18
- Figure AA-1: Long-Term Frequency Extrapolation for Consumption
Figure AA-1 depicts a cumulative distribution of the frequency of fish consumption. One line depicts the data obtained from the 30-day NHANES survey. The other line, labeled CSFII Based Projection is an exponential function that closely matches the data from the survey.
- Figure AA-2: Variation among Individuals of the Variation in Species Consumed
Figure AA-2 depicts a step-wise function representing the cumulative distribution of Major Category Frequency, which corresponds to the ratio of the species consumed most often to the total number of meals consumed over a 30 day period for each individual in the NHANES seafood frequency survey. The values for the distribution of ratios vary from 0.16 to 1, with a median value of 0.5.
- Figure AA-3: Fitted Distributions for Mercury in Crab Meat
Figure AA-3 depicts a cumulative distribution of methylmercury concentrations in Crabs. The plotted data points begin at the 50th percentile at a concentration value of about 0.02 ppm and increase to about 0.18 ppm at about the 98th percentile. There is also an extreme value at about 0.6 ppm. Values below the 50th percentiles, which were below the limit of detection, are not shown. In addition to the data, four different statistical distributions are shown, all of which provide an approximate description of the data. These are the Beta, Gamma, Logistic, and Normal distributions.
- Figure AA-4: Influence of Body Weight on Blood/Diet Ratio
Figure AA-4 shows a scatter plot of body weight vs the mercury blood concentration (ppb) per microgram of mercury consumed per day for 20 individuals. Four different series are shown. The first series of uncorrected values have a clear decreasing trend. In particular, the three individuals with exceeding 0.92 all have body weights below 70 kg, while all four individuals with a body weight greater than 80 kg all have blood-diet ratios under 0.77. The second series, labeled body weight to the power of 1, shows corrected values when it is assumed that blood concentration is proportional to body weight. Rather than decreasing, this series shows an increasing trend, suggesting that this common method of adjusting for body weight overestimates to consequences of increasing body weight on mercury blood levels. The third and fourth series, respectively labeled “body weight to the power of 0.33” and “body weight to the power of 0.44”, show corrected values using a body weight power function. Both of these plots appear to be fairly flat as body weight increases.
- Figure AA-5: Hair and Blood Concentrations in Women of Childbearing Age in the U.S. (from NHANES)
Figure AA-5 shows a scatter plot of mercury hair concentration (ppm) and mercury blood concentration (ppb). The majority of the values have hair levels of less than 2 ppm and blood values of less than 10 ppm. There are also two individuals with very high hair levels greater than 30 ppm (31 and 38), with blood levels of 18 and 8, respectively.
- Figure AA-6: Hair/Blood Ratios in Women of Childbearing Age in the U.S. (from NHANES)
Figure AA-6 shows a cumulative distribution of the ratio of hair (ppm) to blood (ppb) concentrations. 95% of the values are below 0.5. Most of the rest are under one, but there are three values in the range of 1.6 to 5.2.
- Figure AA-7: Hair-Blood Ratios, the Lower Tail of the Distribution (from NHANES)
Figure AA-7 shows the bottom 20% of a cumulative distribution of the ratio of hair (ppm) to blood (ppb) concentrations. While the rest of the distribution looks linear, the bottom 5% show some curvature, with a decline to a hair/blood ratio of 0.01,
- Figure AA-8: Hair-Blood Ratios, the Upper Tail of the Distribution (from NHANES)
Figure AA-8 shows the upper 20% of a cumulative distribution of the ratio of hair (ppm) to blood (ppb) concentrations. 95% of the values are below 0.5. Most of the rest are under one, but there are three values in the range of 1.6 to 5.2.
- Figure AA-9: Regression Analysis for Poland Study - Maternal Blood Hg
Figure AA-8 (2) shows a scatterplot of maternal blood mercury (ppb) vs results on two test scores, with the bloof values ranging from 0. to 5 ppb and test scores that range from 69 to 125. Two series are shown. The first is for the Bayley Cognitive score, while the second is for the Bayley Psychomotor score. The results of two regression analyses are also shown. Although these straight lines look relatively flat with respect to maternal blood mercury, the Bailey Psychomotor regression does show a 3 point decrease over the range of 0 to 5 ppb.
- Figure AA-10: Regression Analysis for Poland Study - Cord Blood Hg
Figure AA-9 shows a scatterplot of cord blood mercury (ppb) vs results on two test scores, with the bloof values ranging from 0 to 5 ppb and test scores that range from 69 to 125. Two series are shown. The first is for the Bayley Cognitive score, while the second is for the Bayley Psychomotor score. The results of two regression analyses are also shown. Both regression lines show a decrease of about 4 points over a range of 0 to 5 ppb.
- Figure AA-11: Developmental Effects of Methylmercury: High Doses on a Log Scale
Figure AA-10 shows graphs of nine different dose-response functions that relate MeHg in Maternal hair to IQ, with dose shown on a log linear scale, with a range of 1 to 1000 ppm. In order of decreasing slope, the functions shown are deceasing order of slope magnitude. From highest to lowest predict effect on IQ, these are “IQ - Faroe Islands - Cohen - Primary”, “IQ - New Zealand -Axelrad”, “IQ - New Zealand - Cohen”, “Delayed Walking - Iraq”, “Delayed Talking - Iraq:, “IQ - Seychelles - Axelrad”, “IQ- Faroe Islands - Cohen - Secondary”, “IQ - Faroe Islands - Axelrad”, and “IQ - Seychelles -Cohen”. Unlike the others, the last function decreases with dose.
- Figure AA-12: Developmental Effects of Methylmercury: Low Doses on a Linear Scale
Figure AA-11 shows graphs of nine different dose-response functions that relate MeHg in Maternal hair to IQ, with dose shown on a linear scale, with a range of 0 to 10 ppm. In order of decreasing slope, the functions shown are deceasing order of slope magnitude. From highest to lowest predict effect on IQ, these are “IQ - Faroe Islands - Cohen - Primary”, “IQ - New Zealand -Axelrad”, “IQ - New Zealand - Cohen”, “Delayed Talking - Iraq:, “IQ - Seychelles - Axelrad”, “IQ- Faroe Islands - Cohen - Secondary”, “IQ - Faroe Islands - Axelrad”, “Delayed Walking - Iraq”, and “IQ - Seychelles -Cohen”. Unlike the others, the last function decreases with dose.
- Figure AA-13: Developmental Effects of Methylmercury: Comparison of Confidence Intervals
Figure AA-12 shows graphs of two different dose-response functions that relate MeHg in Maternal hair to IQ, with dose shown on a linear scale, with a range of 0 to 10 ppm. These two functions, labeled Delayed Talking - Iraq, and Axelrad - IQ - Integrated, appear to be identical, with a linear slope of 0.18 IQ points per ppm in maternal hair. Confidence intervals are also shown. The lower limit for both functions are both very close to zero. The upper confidence limits are quite different, with a slope of about 0.27 IQ points per ppm for Delayed Taling -Iraq and about 0.37 IQ points per ppm for Axelrad - IQ - Integrated.
- Figure AA-14: Combined Dose-Response For Methylmercury and Fish (Daniels et al., 2004)
Figure AA-13 shows a combined dose-response function for fish and mercury on neurobehavioral development that combines the methylmercury-delayed talking dose-response function and the Daniels et al. (2004) fish-verbal test score dose-response function. The function appears to linear with a Z-score increase of about 0.06 per 100 g of consumed per day. Confidence intervals are also shown, these have slopes of about 0.008 Z-score increase per 100 g fish consumed per day, and 0.1 Z-score increase per 100 g fish consumed per day for the lower and upper confidence limits, respectively.
- Figure AA-15: CHD Mortality Dose-Response Models for Individual Studies
Figure AA-14 shows dose-response functions for sixteen different cohorts that relate the amount of fish consumed per day to the mortality from CHD. Each cohort is plotted separately. The data in and number labels for each graph correspond to the values in Table AA-15. Over half of the cohorts, numbered 1, 2, 3, 4, 8, 9, 10, 14, and 15 exhibit decreasing rates of heart disease with increasing fish consumption, with confidence intervals that are both positive. Six of the studies, numbered 5, 6, 9 , 12, 13, and 16, show relatively little relationship between fish consumption and heart disease, with a central estimate that is relatively flat and a both positive and negative confidence intervals. One of the cohorts, number 8, appears to exhibit a negative relationship between fish consumption and CHD mortality; the central estimate and both confidence intervals are negative.
- Figure AA-16: Dose-response Function for Fish Consumption and CHD Death in Men Aged 45+
Figure AA-15 shows an integrated dose-response functions from all sixteen cohorts that relates the amount of fish consumed per day to the mortality from CHD. The central estimate exhibits decrease in estimated CHD mortality. The estimated rate decreases from a rate of just over 0.6% per year with a fish intake to a rate of under 0.5% with in intake of 25 g/day. The dose-response function is relatively flat at intakes greater than 2g g/day. One of the confidence intervals is slightly negative with fish intakes of less than 10 g/day or greater than 50 g/day.
- Figure AA-17: Stroke Mortality Dose-Response Models for Individual Cohorts
Figure AA-16 shows dose-response functions for ten different cohorts that relate the amount of fish consumed per day to mortality from stroke. Each cohort is plotted separately. The data in and number labels for each graph correspond to the values in Table AA-17. Half of the cohorts, numbered 1, 3, 4, 6, 8, and 9 exhibit decreasing rates of heart disease with increasing fish consumption, with confidence intervals that are both positive. Four of the studies, numbered 5, 7, 8, and 10, show relatively little relationship between fish consumption and heart disease, with a central estimate that is relatively flat and a both positive and negative confidence intervals. One of the cohorts, number 2, appears to exhibit a negative relationship between fish consumption and CHD mortality; the central estimate and both confidence intervals are negative.
- Figure AA-18: Dose-response Function for Fish Consumption and Stroke Death in Women Aged 45+
Figure AA-17 shows an integrated dose-response functions from all sixteen cohorts that relates the amount of fish consumed per day to the mortality from stroke. The central estimate exhibits decrease in estimated stroke mortality. The estimated rate decreases from a rate of just over 0.23% per year with a fish intake to a rate of under 0.17% with in intake of 25 g/day. The dose-response function is relatively flat at intakes greater than 2g g/day. One of the confidence intervals is slightly negative at virtually all doses.