Estimating exposure to migrants from food contact articles

6.7.1 Introduction

The percentile of the population to consider and any at risk or vulnerable groups in an exposure assessment was discussed in section 6.3. Today there is no universally recognised single approach to estimating exposure, particularly to migrants from food packaging. ILSI (2002) undertook a workshop focused on determining exposure to migrants from food contact materials, but a standardised approach or methodology was never agreed; however, the need for more data was identified. Exposure assessments can be determined in a number of ways which can be summarised as:

• Simplistic (or simplified/straightforward), normally using a worst-case assumption, as is the case in the EU today with the 6 dm2/kg food/person/ day. This approach is a subset of a deterministic approach.

• Deterministic, where a fixed value of consumption (normally of a given foodstuff or family of foodstuffs) normally at the high end is combined with a high or most likely the highest level of migration found.

• Probabilistic, where statistical modelling is used to predict those values related to the unknown inputs required in order that a more refined estimate of exposure may be obtained.

Today for risk assessment in the EU it is assumed that each, and every, person eats 1 kg of food, packaged in the same material, every day of their life. Furthermore, it is assumed that the surface area to volume ratio of this package is 6 dm2/kg. In the allocation of migration limits for a substance, it is assumed that the substance always migrates at the highest level, corresponding to its TDI, for all packaging and all foodstuffs, therefore migration at 1 mg/kg equals exposure at 1 mg/person/day. This has the advantage that exposure to a substance for any migrant is independent of its packaging, with a few exceptions. This is unlike the US FDA, where the use of a substance can be restricted to a particular application, including type of package or foodstuff. The major disadvantage is that in most cases, different types of packaging contain different substances and their migration behaviour frequently depends upon the foodstuff and its processing in the package. Consequently, the current EU approach arguably over-estimates exposure to migrants and therefore applies stricter restrictions, which do not necessarily improve consumer safety but could restrict consumer choice. There are some counter-arguments that this approach under-estimates the exposure for infants, but this is currently the subject of debate (Castle 2004) and in reality there are a number of projects being undertaken or in the process of being reported which may clarify this situation.

Whilst most people recognise the shortcomings of the current EU approach for food contact materials, there is no simple solution, as the data required with the necessary detailed information do not exist. EU Member states have nutritional surveys of varying quality. Their prime purpose is to enable authorities, etc., to determine the 'nutritional health' of their population. For example, the amount of salmon, vegetables, fruit, salt or sugar consumed is, quite rightly, considered more important for the surveys than the packaging of the foodstuffs consumed. In some cases, such as exposure to heavy metals in tuna or dioxins in salmon, then this is an invaluable first step in assessing the exposure. Should the resulting value give cause for concern then there is a need to refine the assessment by obtaining more details about the foodstuff items containing the substance of concern consumed.

The deterministic approach is better suited to food additives or contaminants and not migrants from food packaging, provided there are adequate data on the foodstuffs consumed. In these cases the types of food which could contain the additive are known from the food item description, as well as the amount consumed. The concentration of a food additive in the foodstuff is either predetermined or can be measured. Taking a high consumption (97.5 percentile) of the foodstuffs containing relatively high levels of the additive and using a relatively high concentration value will enable an exposure assessment to be made. If the exposure is below the TDI, then there is no cause for concern and no further efforts should be devoted to improving the estimate of exposure. However, it is necessary to consider the relevance of taking high concentration values with high consumption and in some cases only one of the two input parameters would be at the high levels with the other at a lower level (possibly the average level). The deterministic approach is also of value for groups of concern where model diets or actual consumption data exist. The simplistic approach described above could be considered as a simplified deterministic one.

Using data from food frequency questionnaires and portion sizes in combination with a concentration value would give a deterministic estimate of exposure (Parmar et al. 1997), particularly if the contaminant could occur in more than one type of foodstuff. If average or maximum concentration data are used, then an estimate of exposure can be obtained. This approach would also cover rarely consumed foodstuffs where the migrant was specific to these. If the contamination is only in one foodstuff or one source of foodstuff is known to be the major contributor to the intake, then it is possible to use a single point estimate using a single average or maximum concentration data value. A more realistic approach would be to use a distribution of concentration data. Probabilistic modelling is ideally suited where there are data gaps, allowing confidence limits to be put on any exposure estimate. This approach is considered later.

Estimates of exposure need to be in units which are meaningful. Many toxicologists use mg/kg body weight/day to assess risk. The body weight can be that recorded in surveys or the EU assumption of 60 kg body weight per person. The use of mg/person/day enables one to determine if a threshold has been exceeded. Another expression for exposure is mg/kgdiet. All assumptions made should be clearly stated and all sources of data used indicated. An estimate of the likely error bounds for any estimate is important, as is a statement as to who is being 'protected' by the estimate.

6.7.2 Approaches to determining exposure to migrants from packaging of foodstuffs

There are a number of different approaches (Rees and Tennant 1993, Parmar et al. 1997, Kroes et al. 2002) which can be used in order to obtain estimates of consumption of different foodstuffs and any exposure to chemical contaminants in the foodstuffs. In essence they fall into three categories (Rees and Tennant 1994) which are:

• model diets or worst case scenarios

• surveillance methods or duplicate diets.

Per capita estimates of food chemical intake can be made for virtually every European country. They permit comparison between different European countries. There are in essence two basic approaches for undertaking this estimate: multiply the average food consumption of the whole population by anticipated or actual levels of the migrant; divide the total available food chemical by the number of individuals in the population. Obviously this approach is unsuitable for migrants, but ideally suited for food additives.

An advantage of the per capita approach is that it is cost effective and relatively straightforward. If the estimated levels of exposure are significantly lower than those which could cause concern, then arguably further refinement of the estimate of exposure is unjustified. Wherever possible per capita data should not be averaged, but a range retained in any subsequent calculations. The better the manufacturing data and demographic data the 'better' and more reliable the per-capita estimate.

The total diet method, which can be used for food intake studies, utilises data on food purchases based on household budget surveys. They can give the average consumption of different foodstuffs, normally grouped. It is possible for these foodstuffs to be purchased and analysed. This enables the food groups which contribute the major sources of exposure for a given migrant to be identified. If the migrant is restricted to one type of packaging, then it is necessary to select those foods packaged in that packaging. This does not facilitate estimating exposure for individual consumers as the household budget surveys are frequently for families which may or may not be consumers of all items.

Another approach is the model diet where, based on consumer statistics, a diet is proposed that models that of the average or possibly (with adequate data) the non-average consumer. This is of value when limited consumption data exist or when the major source of contamination is one food group. This approach can be considered to be cost effective for a deterministic approach but is subject to errors when many foodstuffs are involved. If the groups for the model diets are based upon 'good' and 'appropriate' data then it may be possible to use model diets for different age groups or 'at risk' groups.

It is also possible to model different scenarios as shown in Table 6.2 (Rees and Tennant 1994) where extreme and reasonable scenarios are used. If the worst-case scenario does not give cause for concern then there is no need to refine any exposure estimate. For example, assume that for a given migrant, migration is at the SML and that all the foodstuffs which could be packaged in the material are packaged in that material. This can be considered as a useful screening technique because if the exposure estimate does not give

Table 6.2 Matrix of intake estimates using the 'scenario' approach (Rees and Tennant 1994) and Rees, N., private communication (1998)

Food consumption

Typical (per capita or average)

Worst case (above average or high level)

Occurrence of migrant in foodstuffs

Typical (mean or most common value)

Worst case (maximum value or high percentile)

Likely. The exposure calculated is likely in the majority of consumers, particularly if averaged over a year Possible. Fewer consumers are likely to have exposure at this level but may require consideration if there is a good chance of selecting high migrant levels on a regular basis

Possible. Fewer consumers are likely to have exposure of this level, but may require consideration if the dietary pattern is habitual

Unlikely. Whilst this may be possible, there should be an assessment of whether it is probable on a regular basis and what the toxicological implications may be cause for concern, then there is no need for further refinement. This approach is useful if there are limited data or a sudden 'unexpected' issue occurs where no data exists. These estimates by definition are imprecise.

An approach to provide data with reduced uncertainty limits is the purpose of surveillance surveys. Simulants are used for relating concentrations in real foodstuffs, whereas a surveillance survey actually measures the concentration, thus there is no doubt about the relevance of simulant derived data to what it would be in the foodstuff the simulant is simulating. However, it is still necessary to allocate migration concentration values to those foodstuffs which were not part of the survey. This can be achieved by either assuming certain foodstuffs are similar in migration characteristics to those which have been tested in the survey or by using stimulant migration data. Surveillance surveys enable those foodstuffs which are most at risk from contamination to be identified, consequently these surveys are frequently 'targeted' and if due consideration for this is not made then any data could be skewed to a higher concentration level and hence an unrealistically high exposure assessment. However, surveillance surveys need to be run in conjunction with consumption surveys to maximise the accuracy of the estimates.

One of the most sophisticated approaches is the duplicate diet method where for every item consumed an identical amount of the item is put aside for analysis at a later date. This approach enables all sources of the substance of interest to be included and gives a realistic and arguably the most accurate way of obtaining an exposure assessment. This is a good way of examining at-risk groups, but it is expensive. The interpretation of the data beyond the survey period is difficult (Kroes et al. 2002). In the case of migrants from packaging, not only is it necessary to ensure that the foodstuffs consumed were representative, but it is necessary to know if the packaging of the foodstuffs consumed during the survey period was typical for each consumer.

6.7.3 The USFDA approach to estimating consumer exposure to migrants from food contact materials

The USFDA approach to assessing exposure to migrants from FCMs is explained in CFSAN/Office of Food Additive Safety, April 2002 and is available on their web site ( It describes the use of exposure estimates for use in food contact notifications (FCNs) which would normally be based upon simulant rather than food migration data, as is the case for new materials. The USFDA approach is described in more detail in Chapter 2. In the USFDA approach a consumption factor is combined with a food distribution factor and concentration data to derive an estimate of exposure from all food types and all FCMs containing the substance of interest.

The consumption factor (CF) describes the fraction of the daily diet expected to contact specific packaging materials and represents the ratio of the weight of all food contacting a specific packaging material to the weight of all food packaged. To account for the variable nature of food contacting each food-contact article, the FDA has calculated food-type distribution factors (fT) for each packaging material to reflect the fraction of all food contacting each material that is aqueous, acidic, alcoholic and fatty. Tables for both factors are supplied by the USFDA. This is then combined with concentration data to obtain an exposure estimate, assuming a daily consumption of food and drink of 3 kg per person per day. This gives an estimated daily intake (EDI) for a substance per source of packaging. If there is more than one source the EDIs are combined to give a cumulative estimated daily intake (CEDI).

The concentration of the substance in the food contacting the food-contact article, <M>, is derived by multiplying the appropriate fT values by the migration values, Mi, for simulants representing the four food types. This, in effect, scales the migration value from each simulant according to the actual fraction of food of each type that will contact the food-contact article.

<M> _ faqueous and acidic(M10% ethanol) + falcohol (M50% ethanol) + ffatty(Mfatty)

where Mfatty refers to migration into a food oil or other appropriate fatty-food simulant. The concentration of the substance in the diet is then obtained by multiplying <M> by CF. The EDI is then determined by multiplying the dietary concentration by the total weight of food consumed by an individual per day, assuming that an individual consumes 3 kg of food (solid and liquid) per day.

A concentration in the daily diet of 1 ppm corresponds to an EDI of 3 mg substance/person/day. This approach is designed to deal with single use (e.g. food packaging) rather than repeated use (e.g. non-stick frying pan) FCMs.

6.7.4 Probabilistic (stochastic) modelling

This section explains the background to the use of probabilistic, also known as stochastic, modelling for estimating exposure to migrants from the packaging of foodstuffs. Not all exposure assessments need the refined approach of probabilistic modelling. However, it is a tool gaining greater acceptance for assessing exposure where there are data gaps. Probabilistic modelling has been used by Lambe et al. (2002) to assess the intakes of flavours. Petersen (2000) compared theoretical and practical aspects of probabilistic modelling.

Probabilistic modelling overcomes the lack of data by estimating the most likely exposure to a given migrant(s), using input data with uncertainties but also deriving confidence limits for any assessment. The treatment of data with uncertainties is one of the strengths of probabilistic modelling. Probabilistic models can deal with data rich and data poor inputs. Another factor to consider is variability and whether it should be separated from uncertainty. This gives rise to one- or two-dimensional probabilistic models, with the one-dimensional model combining uncertainty and variability and the two-dimensional model propagating them separately (Hart A., private communication). The theory of probabilistic modelling is outside the scope of this chapter.

Probabilistic modelling uses the principles of statistics and in essence is based upon the Monte Carlo approach. It repeatedly (typically > 1000 iterations) calculates the exposure to obtain an estimate of the mean and the uncertainty for any given percentile using different input parameters, some of which are randomly generated. Where there is uncertainty about the actual value of a parameter, lower and upper limits can be set, with a most likely value. The statistical model randomly generates values between the lower and upper limits for each iteration, with the majority of the values being distributed about the most likely value, using whatever distribution between minimum and maximum centred around the most likely is considered appropriate. This means that values near to the most likely one are used more times in the total number of iterations than those values towards the lower and higher limits. A more detailed description of such a model is given in Holmes et al. (2005) It is recognised that a number of groups are working in this area including, for example CSL, Crème and Rikilt. It should be borne in mind that there are only a few mathematical/statistical models. The differences in use are in how the data are input and the results obtained.

In order to estimate exposure to migrants from food packaging materials, the data required are similar to those for any other method of exposure assessment, namely:

1. consumption of foodstuffs

2. packaging of foodstuffs consumed

3. concentration data of migrant(s) in foodstuffs consumed.

Clearly not all of these data are readily available and, in most cases, very limited data exist. Thus it is necessary to make assumptions, but in doing so the impact of the assumptions on the estimate needs to be evaluated. Probabilistic modelling facilitates this requirement.

Inputs can be considered as being either fixed or variable being varied between upper and lower limits, depending upon the accuracy, amount and availability of the required input data. Where some input data have considerably greater uncertainty or variability than others, then it is questionable if the amount of effort and treatment required for the less uncertain or less variable parameters is justified. The case of migrants originating from packaging is a good example. Today, for most cases, significantly less is known about the packaging of all the foodstuffs consumed than the types and amounts of foodstuff consumed, thus uncertainty arising from the accuracy of the amount of foodstuff consumed could be considered insignificant compared to what it was packaged in. Possible treatments for a few of the more important variable input parameters will be considered here as examples of how the uncertainty in input data was treated using as an example the CSL model (Holmes et al. 2005).

In the UK NDNS surveys, in some instances, the food item description described the packaging. Examples are 'canned' or 'not canned' or 'bottled'. In these cases it is possible to associate the likely concentration of the migrant(s) with the foodstuffs consumed, being either definitely present or absent. In other cases, where the packaging of the foodstuff was unclear, it is necessary to use 'expert judgement' or market share to allocate the most likely packaging to the food item description and the most likely concentration of migrant in that food item. Being undefined, there is some uncertainty about the value to be used. Thus a lower and upper limit are allocated, and for each iteration of the Monte Carlo model a different value for the packaging, between lower and upper limits, is selected, with the majority being around the most likely value using a triangular distribution.

Obtaining concentration data was considered earlier (section 6.5). In the case of the UK NDNS food surveys, the description of some food items consumed are a home-made meal rather than a pre-packaged one, without a breakdown of the amount of each ingredient. Having identified those items which could be in the packaging(s) of interest, it is then necessary to apply a correction factor because if only part of the food item contains the migrant, then the concentration data need to be adjusted accordingly. For example, if a food item description consists of three food components and 100 g of the food item consumed and the migrant is present in only one of the three food components, then the 100 g of food item consumed does not contain 100 g of the food containing that migrant. Therefore it is necessary to reduce the concentration of the migrant in that weight of foodstuff consumed by the use of a factor. Values can be varied between lowest estimate of its content in the meal, most likely and highest with a different value being used for each iteration, with the majority of the values being around the most likely.

Yet another use of this factor could be to correct for subsequent dilution of a packaged item to one that is consumed. As an example 10 g of powdered soup may give 100 g of soup reported as being consumed. Another example is that 2 g of milk may be in 100 g of a cup of coffee reported as being consumed. Expressing the concentration data in a weight per unit area enables a surface volume to weight ratio to be applied rather than the conventional factor of 6 dm2/kg. In most cases, if not all, the actual surface to volume ratio is unknown. Therefore it is necessary to allocate appropriate ranges, with again a different value between minimum, most likely and maximum being used per iteration.

It should be possible to derive a cumulative exposure curve for the whole of the population and, ideally, for consumers only. It is desirable to obtain estimates of exposure for any selected sub-set of the population, to ensure adequate protection of vulnerable groups. Depending upon the quality of the food consumption survey and associated data this may or may not be easy to achieve. If non-consumers are included in an exposure estimate, as is the case in typical per-capita estimates, then the estimate of exposure will be below the actual exposure and the amount of under-estimating will depend upon the ratio of non-consumers to consumers.

Having obtained a value it is necessary to determine the confidence limits (error bounds) for this estimate. In this way, the impact of any uncertainty surrounding the initial assumptions used can be evaluated. A representation of a mean value for the population and 95% confidence limits for any given percentile is shown in Fig. 6.2. As can be seen, the difference between the upper and lower values of an exposure estimate for a given percentile increases the nearer the selected percentile is to 100%.

From a probabilistic approach it should be possible to derive the exposure for any age group or any percentile, as shown for example in Fig. 6.3. In Fig. 6.3, upper and lower limits are represented by the line either side of the mean estimate represented by the diamond. In addition the effect of gender, socioeconomic class, ethnicity, etc., can be evaluated, provided they have been identified in the food intake survey, although as the group studied becomes smaller greater uncertainties inevitably arise. Exposure can be expressed as mg/kg actual body weight or mg/person/day or mg/kg diet.

It is also of value to identify the main contributors to any estimate of exposure in order that any parameters driving the estimate can be investigated further, as illustrated in Fig. 6.4 (Holmes et al. 2005). In some cases major contributors to exposure may be due to high concentrations of the migrants in the foodstuffs, but in others it may be that even though the concentration

90 80

I 60

1 40

Exposure mg/day

Fig. 6.2 Mean estimate for any given percentile, with upper and lower confidence limits.

Exposure mg/day

Fig. 6.2 Mean estimate for any given percentile, with upper and lower confidence limits.

Youth male Youth female Adult male Adult female Seniors male Seniors female

Fig. 6.3 97.5th percentile exposure estimate (with a 95% confidence range) for the three age groups for BADGE using the Food Standards Agency, Food Surveillance Information Sheet 9, November 2000. Median dietary exposure estimate 0.04 mg/kg bw/day (0.03-0.053 mg/kg bw/day) (Holmes et al. 2005).

data are low, the foodstuffs are consumed in much greater quantities than others. In some cases identifying the foodstuffs which contribute the major part of the exposure may help mitigate the exposure by, for example that particular packaging being substituted. In other cases, revisiting the model may enable the estimate of exposure to be reduced by refining some of the assumptions used. For example, reducing the LOD for BADGE in beverages substantially reduced the estimated exposure, because the actual levels were still below the reduced LOD. On the other hand, canned meat which had measurable levels of BADGE had only a minor contribution to the exposure due to the relatively low consumption of canned meat. Consult Holmes et al. (2005) and Oldring et al. (2006) for further information.

Knowledge of the impact of the major sources of uncertainty in input data on the estimate of exposure is essential. Correlation coefficients and scatter











n io






o o


Ca ev b


Canned vegetables

Canned pasta

Canned soup

Canned fruit

Canned meat

Canned desserts

Fig. 6.4 The % contribution of major food groups (with 95% confidence range) toward the overall exposure for a high level basket set up for BADGE using the Food Standards Agency, Food Surveillance Information Sheet 9, November 2000. The data shown is for all age categories and indicates the effect of the uncertainties in the model run (Holmes et al. 2005, Oldring et al. 2006).

Fig. 6.4 The % contribution of major food groups (with 95% confidence range) toward the overall exposure for a high level basket set up for BADGE using the Food Standards Agency, Food Surveillance Information Sheet 9, November 2000. The data shown is for all age categories and indicates the effect of the uncertainties in the model run (Holmes et al. 2005, Oldring et al. 2006).

o graphs can be used to identify key sources of uncertainty affecting the exposure estimates. The correlation coefficients and graphs can be calculated for each of the uncertain variables against the mean exposures to determine which are the key sources of the uncertainty. Figure 6.5 shows a simplified scatter plot for the mean exposure (mg/kg bw/day) against the mean of the beverage concentration distribution (ND) used to model the level of BADGE in canned beverages. In this case the Pearson correlation coefficient was 0.609. This approach identifies where the risk assessor should concentrate in order to obtain more data to reduce the uncertainty if the estimated exposures are close to levels of concern or to address the data gaps in order to obtain a more realistic assessment of exposure. This is in contrast when the uncertainty in the concentration data for canned fish in oil is considered. As can be seen from Fig. 6.6, there is no correlation between the uncertainty in the concentration values used and the resulting exposure.

Correlation coefficients and scatter graphs of the exposure against consumption of different groups of foodstuffs may also be analysed and these show the correlation between exposure and the variability within the system. Figure 6.7 shows an example of a scatter plot between the consumption in grams of canned beverages and the average exposure for the BADGE model. This plot indicates how the increasing consumption of certain items is correlated with exposure. By combining the approach of the uncertainties in the concentration data and the effect of the amount consumed on exposure (Figs 6.5 and 6.7) it was possible to determine that the ND value for BADGE in beverages was driving any exposure estimate for BADGE, using the input values for ND.

Uncertainty in mean conc. (mg/kg) - beverages

Fig. 6.5 Simplified scatter plot between the average exposure of BADGE with the sampled values used for the mean of the beverage concentration distribution (Holmes et al. 2005).

Uncertainty in mean conc. (mg/kg) - fish in oil

Fig. 6.6 Simplified scatter plot between the average exposure of BADGE with the sampled values used for the mean of the fish in oil concentration distribution.

The advantages and disadvantages of probabilistic approaches can be summarised (Hart A., private communication) as follows.

• Potential advantages of probabilistic approaches include:

- increased realism, representing real-world variation in factors that influence exposure and replacing or refining worst-case assumptions

- makes more use of the available data

- indicates the influence of quantified uncertainties on the assessment outcome

- helps to increase the cost-effectiveness of further data collection by targeting it on major sources of uncertainty

- helps in targeting of risk management actions by identifying key contributors to exposure.

0 500 1000 1500 2000 2500 3000 3500 4000 Weight of canned beverage consumed (grams)

Fig. 6.7 Scatter plot between the total consumed weight (grams) from the canned beverage sub-basket against the average exposure (Holmes et al. 2005). Pearson linear correlation = 0.610.

0 500 1000 1500 2000 2500 3000 3500 4000 Weight of canned beverage consumed (grams)

Fig. 6.7 Scatter plot between the total consumed weight (grams) from the canned beverage sub-basket against the average exposure (Holmes et al. 2005). Pearson linear correlation = 0.610.

• The main disadvantages of probabilistic approaches are:

- they are more complex and require more expertise than deterministic approaches

- the results are more complex than deterministic exposure estimates and can be hard to communicate effectively

- probabilistic methods - especially those for quantifying uncertainty -are still under development and there is not yet an established consensus on which methods are most appropriate for which purposes

- probabilistic assessments are not yet readily accepted by regulatory authorities.

It is frequently stated that probabilistic methods require more data than deterministic methods. This is not literally true; it is possible to perform probabilistic calculations with input distributions based on small datasets or expert judgement. It is true that distributions derived from small datasets or expert judgement are likely to be very uncertain. However, if these uncertainties can be adequately represented within the probabilistic assessment, or dealt with by making conservative assumptions for the affected inputs, then probabilistic methods should still provide a useful refinement. Even in those cases where the uncertainties are too great to provide reliable estimates of exposure, probabilistic analysis may still be useful as a form of sensitivity analysis to identify priorities for data collection.

0 0

Post a comment