Modelling
The overall aim of the study was to estimate health risks and benefits of important compounds (dioxins, dioxin-like PCBs, methylmercury, omega-3 fatty acids [consisting of eicosapentaenic acid EPA and docosahexaenic acid DHA], and vitamin D) found in Baltic herring and salmon in the current situation. The assessment model was implemented in an open and modular way at Opasnet web-workspace (en.opasnet.org). In practice, this means that all the data and code used for different parts, or modules, of the model are located on different pages at Opasnet. These pages are called knowledge crystals, as their structure and workflow follow certain rules (Tuomisto et al 2019, forthcoming). In this section, we give an overview of the model and describe the input data and assumptions used; the Result section consists of model results. Links to the module pages and all details can be found from the assessment page[8]. The whole model with data and codes is available on the page and also at Open Science Framework research data repository[9].
The benefit-risk assessment was based on a modular Monte Carlo simulation model, which had a hierarchical Bayesian module for estimating dioxin concentrations. The different modules are briefly described below, with references and links to further material.
The input distributions were derived either directly from data or from scientific publications. If no published information was available (as was the case with e.g. disability weights for non-typical endpoints such as tolerable weekly intakes or infertility), we used author judgement and wide uncertainty bounds (these judgements are described later in the text). The model was run with 3000 iterations using R statistical software (version 3.5.3, https://cran.r-project.org).
Consumption survey
The data used were from an internet-based survey that was conducted at the end of 2016. The survey focused on consumers’ eating habits of Baltic herring and salmon in four Baltic Sea countries: Denmark, Estonia, Finland, and Sweden (Figure 1). The questionnaire was designed and the results analysed by the authors, but the survey was administered by a professional market research company Taloustutkimus oy, which has an established internet panel since 1997. The survey company recruited over 500 consumers from each country (total 2117) to respond to the survey questionnaire, which is above the required sample size to generalise the results to each case study country (with a 95% confidence level and 5% margin of error)[10]. The survey was targeted to adult population, i.e. 18 years or older.
The survey questionnaire comprised 32 questions, including sociodemographic questions as well as questions relating to fish consumption frequency and amount in general, and to Baltic herring and salmon in particular. There were also questions about reasons to eat or to not to eat those species, and policies that may affect the amount eaten. The questionnaire was translated into the national language of each case study country (Finnish, Swedish, Estonian and Danish). The country and gender of the respondents were provided directly by the internet panel and were therefore not included in the questionnaire.
Only those respondents who reported fish consumption in general were asked follow-up questions about herring and salmon consumption, and are included in the analysis presented in this paper. As the survey focused specifically on consumption of herring and salmon originating from the Baltic Sea, a distinction had to be made in the questionnaire between Baltic herring and herring originating from elsewhere, e.g. the North Sea or North Atlantic, as well as between the salmonids (Baltic and Norwegian salmon, farmed salmon, rainbow trout). Regarding herring consumption, the respondents that reported eating some type of herring were asked explicitly whether they consume Baltic herring. Concerning Baltic salmon, the respondents were asked to choose from a list of salmonids, which ones they consume. In addition to the analysis presented in this paper, the survey was conducted for the purpose of a consumer perception and consumption study[11] and therefore only part of the survey results are presented in this paper.
Individual long-term fish consumption (in kilograms per year) was estimated from consumption frequency and amount questions. Consumption distributions were produced for subgroups defined by country, gender, and age by random sampling (with replacement) of the individual estimates. People's reactions to changes in policies or fish market (e.g. what if fish consumption is recommended or restricted; what if the availability and usability of these species improves; what if the price of fish changes) were predicted based on their answers. These decision scenarios were used to alter the business-as-usual scenario and compare results between scenarios.
Also a few technical scenarios were developed: what if nobody ate either fish more than ca. 1 kg per year; and what if fish is considered a primary versus a secondary source of nutrients. The latter scenario is important if dose-responses are non-linear, as is the case with vitamin D where the benefits occur only if exposure exceeds the recommended daily value. In such a case, the incremental health benefits of a primary source may differ from those of a secondary source.
The data analysis was conducted using R software (version 3.5.3, http://cran.r-project.org). Because the survey was conducted on an internet panel rather than on a random sample from the general population, the respondents may not be fully representative of the actual population distributions of the countries. Therefore, the respondents were weighted based on actual age, gender, and region distributions of each country to produce population representative results.
To support transparency, the anonymised data and all the results will be available online: http://en.opasnet.org/w/Goherr:_Fish_consumption_study
Concentrations
Fish-size-specific PCDD/F and dioxin-like PCB concentration distributions for each fish species and country were estimated based on EU Fish II study[12]. The results were based on pooled and individual fish samples (98 Baltic herring and 9 salmon samples) and analysed for 17 dioxin and 37 PCB congeners. A hierarchical Bayesian module was developed with the JAGS package of R software. The model assumed ca. 7 per cent annual decrease in dioxin concentrations, based on long time trends measured in Finland. The fish samples were caught between 2009 and 2010.
The concentrations in Baltic herring were found out to be highly sensitive to fish size, while size-dependency was much weaker in salmon. Herring sizes in different scenarios came from a fish growth model developed in BONUS GOHERR project[13].
The fish samples came mostly from the Bothnian Sea, which is an important area for Finnish and Swedish catch. The concentration distributions for the studied countries were derived from the concentration model results by scaling them with the average concentration on a catch area of interest relative to the average from the Bothnian Sea. The Danish and Estonian catch areas were assumed to be Baltic west of Bornholm and the Gulf of Finland, respectively. The Swedish catch areas for herring and salmon were assumed to be the Baltic Main Basin and the Bothnian Sea and Bay, respectively. The area selection was based on landing statistics provided by the International Council for the Exploration of the Sea (ICES)[14][15].
Dioxin and PCB concentrations were weighted and summed up to toxic equivalency quantities (TEQ) by using WHO 2005 toxic equivalency factors (TEF)[16]. Levels of fatty acids and vitamin D in Baltic herring were based on measurement data obtained from the Finnish Food Safety Authority, and those in salmon are based on Fineli food database[17]. Methylmercury concentrations were based on Kerty database[18].
Exposures
Exposures to pollutants and nutrients were simply products of consumption amounts as assessed from the survey and concentrations in the consumed fish, with possibly an uncertain background intake from other sources. An exception to this were the infant's exposures to dioxins and methylmercury during pregnancy and breast-feeding, as they were derived from the mother's exposure using toxicokinetic models.
Infant's exposure during pregnancy and breast-feeding was estimated with this equation:
where Cs,i = serum (s) concentration of dioxins in the infant (i) in pg/g fat; Ia,m = average daily intake of dioxins of the mother (m) in absolute (a) amounts pg/day; t1/2,m = the half-life of dioxins in the mother (2737.5 d = 7.5 a); fm = fraction of ingested dioxins actually absorbing from the gut in the mother (0.80); FE = fraction of mother's dioxin load that is transported to the infant during breast feeding (0.25); BF = body fat amount in the infant (into which the dioxins are evenly distributed) during the period when tooth and testis are sensitive to defects and the exposure at its highest (ca. six months of age) (1 kg)[19]
Exposure-responses
Several benefits and risks were assessed (Table 1). We tried to choose impacts that are arguably large enough that could affect the benefit-risk balance. Effects of omega-3 fatty acids on coronary hear disease mortality and child's intelligence, as well as vitamin D effects on vitamin deficiency belong to this category. In contrast, there are several endpoints that have been linked to fish or omega-3 intake, but they were not included because current evidence is controversial: diabetes[20], prostate cancer[21], asthma[22], and stroke[23].
In addition, many studies have linked health benefits to fish consumption rather than a specific nutrient[24]. We included depression[25], breast cancer[26], and all-cause mortality[27], because the results have been fairly consistent in meta-analyses. In summary statistics, CHD and breast cancer mortalities were subtracted from all-cause mortality to avoid double-counting.
The dioxin effect on sperm concentration[6] and methylmercury effect on child's intelligence[28] are the most sensitive risks of these pollutants, and they were therefore included.
In addition, we included some other dioxin effects. Tolerable weekly intakes from 2001 and 2018 were included for comparing methods of quantitative benefit-risk assessment (based on a single health aggregate, DALY) and more qualitative benefit-risk assessment (based on assessing whether a beneficial or harmful threshold is exceeded). Cancer effect was included because the news media often refers to dioxins as "the super poison causing cancer" even if researchers have thought for years that developmental rather than cancer risks are more relevant; a quantitative assessment could give guidance to news communication. And finally, tooth defects were included because it is a sensitive dioxin endpoint, but no study has compared its magnitude to sperm effects. Interestingly, some of the key papers of this effect had studied Finnish mothers, which had been exposed to dioxins mostly from Baltic herring[29][30][2].
The exposure-response function of methylmercury was a synthesis of EFSA tolerable weekly intake and a linear function from Cohen et al.[31]. This was necessary because although the EFSA estimate is fairly recent, it does not quantify the magnitude of effect if the TWI is exceeded. The function by Cohen was based on concentrations measured from mothers' hair. A conversion from hair concentrations to daily exposures was performed according to U.S.EPA[28].
We derived the exposure-response functions for infertility and tooth defects indirectly from published results, so the rational of those endpoints is described here in more detail.
In humans, sperm concentrations have been shown to decrease permanently if boys are exposed to dioxins before nine years of age. The data come from Seveso[38][39] and a Russian children's study[32].
EFSA recently assessed this risk from the Russian children's study and concluded that significant effect was seen already in the second quartile with median PCDD/F TEQ concentration 10.9 pg/g fat, when measured from the serum of the boys at the age of ca. 9 years. Mean sperm concentration was ca. 65 (95 % CI 50-80) million/ml in the lowest quartile, while in all other quartiles the concentration was ca. 40 (95 % CI 30-55) million/ml. Due to the shape of the effect, we used a non-linear exposure-response curve with half of the maximum effect (effective dose 50, ED50) occurring at TEQ concentration 10 pg/g fat.
However, sperm concentration as such is not an adverse health effect. It only manifests itself if the concentration is low enough to prevent conception in a reasonable time window, say, five years. According to a review, the success rate of couples who try to get pregnant is 65 % in 6 months if the sperm concentration is above 40 million/ml[40]. Below that concentration, the probability is fairly proportional to the sperm concentration.
Based on this, we estimated that (assuming independent probabilities between 6-month periods), the probability of not getting pregnant in five years follows this curve:
P(infertility after 5 a) = (1 - 0.65 (1+ (-0.39c)/(c + 10 pg/g)))^10,
where c is the dioxin concentration in boy's fat tissue. This curve is pretty linear below TEQ concentration 50 pg/g with slope ca. 0.00006 g/pg, meaning that for each 1 pg/g increase in dioxin concentration the boy's fat tissue (or serum fat), there is an incrementally increased probability of 0.00006 that he cannot get a child even after five years of trying.
Exposure-response function for tooth defect was also derived from several studies. Alaluusua and coworkers have studied dioxin exposure in small children and the development of permanent molar teeth. They have found defects in both general population in Finland from the exposures in the 1980's [29] [30] and children exposed during the Seveso accident[41].
Based on these studies, we approximated that the effect is linearly correlated with the logarithm of the dioxin concentration in the child.
Disease burden
Disease burden[42] was estimated in one of two alternative ways (Figure 2): if an exposure agent affects the burden of a particular disease in relation to the background of the disease, the attributable fraction of a particular compound exposure was calculated. If the relation was not relative to background, the attributable number of cases due to the exposure was estimated, and this was multiplied by the years under disease per case and the disability weight of the disease (Table 2.).
BoDi = BoD * PAFi = BoD * f * (RRi - 1) / (f * (RRi - 1) + 1), or
BoDi = Ni * L * Dw = P * f * URi * Ei * L * Dw,
where BoD is the burden caused by the disease under study, i is an exposure agent affecting the risk of the disease, PAF is population attributable fraction, f is the fraction of population that is exposed, RRi is the relative risk that the population faces due to the studied level of exposure to exposure agent i (as compared with a counterfactual scenario with no exposure), Ni is the number of disease cases attributed to exposure agent i, L is the duration of a disease incident, Dw is the disability weight of the disease (0=perfect health, 1=death), UR is the absolute unit risk, and E is the exposure to the agent.
Background disease burdens were needed for all-cause and coronary heart disease mortality, as well as breast cancer and depression; they were obtained from the Institute for Health Metrics and Evaluation (IHME) (Table 3.)[45]. Disease burden of a cancer case was based on IHME data. Also disability weights of diseases were based on their estimates, if available. Duration estimates of diseases were mostly based on the time window considered (one year) or lifetime (in the case of permanent infertility, tooth or IQ effects due to infant exposure). We tried to be realistic with estimates but also not to underestimate the risks of fish consumption, so that potential conclusions about safety of fish would not be unfounded.
With the non-typical health effects, namely exceeding tolerable weekly intakes and deviation from the vitamin D recommendation, we used very wide uncertainty distributions, as it was unclear how much weight should be given to endpoints that are only indications of potential health risk rather than actual adverse effects. A value of information analysis was performed to test the importance of these uncertainties.
Childlessness can be viewed as tragedy of life, so the disability weight could be in the order of 0.1 DALY per year permanently (50 years). However, the disability weight applies to only half of the children (boys). Therefore, we used 0.1*50*0.5 DALY/case = 2.5 DALY/case, with rather high uncertainty (0-5 DALY/case).
Population data for each country for year 2016 was available from Eurostat database. Data was separated for gender and age (18 – 45 years and > 45 years) groups[46].
Value of information analysis
Value of information is a mathematical method that compares the difference of utility (money, DALYs or other measure of the objective) in two scenarios: that some additional information is obtained before a decision is made, or that the decision is made with the current information. This can be formulated as
VOI = E(maxi(U(di))) - maxi(E(U(di))),
where VOI is value of information, E is expected value, U is the utility of decision d, and i is an index of decision options[47]. In this study, we also estimated the value of including or excluding an option to the decision making.