Journal of Exposure Science & Environmental Epidemiology (2020) 30:469–477

https://doi.org/10.1038/s41370-019-0185-7

ARTICLE

A regression-based model to predict chemical migration from

packaging to food
Mélanie Douziech1 Ana Benítez-López2 Alexi Ernstoff3 Cecilia Askham4 A. Jan Hendriks1 Henry King5
● ● ● ● ● ●

Mark A. J. Huijbregts1

Received: 15 May 2019 / Revised: 22 July 2019 / Accepted: 28 August 2019 / Published online: 22 October 2019
© The Author(s), under exclusive licence to Springer Nature America, Inc. 2019

Abstract
Packaging materials can be a source of chemical contaminants in food. Process-based migration models (PMM) predict the
chemical fraction transferred from packaging materials to food (FC) for application in prioritisation tools for human exposure.
These models, however, have a relatively limited applicability domain and their predictive performance is typically low. To
overcome these limitations, we developed a linear mixed-effects model (LMM) to statistically relate measured FC to properties
of chemicals, food, packaging, and experimental conditions. We found a negative relationship between the molecular weight
1234567890();,:
1234567890();,:

(MW) and FC, and a positive relationship with the fat content of the food depending on the octanol-water partitioning
coefficient of the migrant. We also showed that large chemicals (MW > 400 g/mol) have a higher migration potential in
packaging with low crystallinity compared with high crystallinity. The predictive performance of the LMM for chemicals not
included in the database in contact with untested food items but known packaging material was higher (Coefficient of
Efficiency (CoE) = 0.21) compared with a recently developed PMM (CoE = −5.24). We conclude that our empirical model is
useful to predict chemical migration from packaging to food and prioritise chemicals in the absence of measurements.
Keywords Mixed-effects model Deterministic model Chemical risk assessment Life cycle assessment Food contact
● ● ● ●

material

Introduction chemical migrants in plastic FCMs can transfer into the

food and therefore the use of FCMs is regulated to ensure
Plastics are widely used as a food contact material (FCM) to consumer safety [2, 3]. Regulations on chemicals in FCM
ensure the safety and protection of foodstuffs [1]. However, change as new evidence becomes available [4, 5]. For
instance, in Europe, bisphenol A has been banned in baby
bottles in 2011 [6] and the European Food Safety Authority
is currently looking into revising the tolerable intake rates of
Supplementary information The online version of this article (https://
perfluoroalkylated substances found, amongst other, in food
doi.org/10.1038/s41370-019-0185-7) contains supplementary
material, which is available to authorised users. packaging applications [7]. Safety regulations tend to focus
on limiting exposure to single chemicals. However, given
* Mélanie Douziech the increased number of chemicals in use, risk-based che-
[email protected]
mical priority setting is of increasingly relevance for
1
Department of Environmental Science, Institute for Water and example using aggregate human exposure predictions to
Wetland Research, Radboud University Nijmegen, P.O. Box 9010, different chemicals found across different products [8].
6500 GL Nijmegen, The Netherlands One way to predict human exposure to chemical
2
Estación Biológica de Doñana, Integrative Ecology, Avd. Americo migrants is by quantifying the amount of chemical migra-
Vespucio s/n, 41001 Sevilla, Spain tion. Migration experiments, which measure the amount of
3
Quantis, EPFL Innovation Park-Bâtiment D, 1015 migrant transferred from the FCM into food, are accurate
Lausanne, Switzerland but costly, time consuming, and difficult to conduct on food
4
Ostfold Research, Stadion 4, 1671 Kråkerøy, Norway products. As a result, regulatory bodies accept the use of
5
Safety & Environmental Assurance Centre, Unilever, Colworth recognised process-based migration models (PMM) as
Science Park, Bedfordshire MK441LQ, UK alternatives to testing. Such chemical migration models are
470 M. Douziech et al.

not only used in a regulatory context, but also in high- PVC-PA-PVC-PU since it was not clear what type of
throughput aggregate exposure predictions developed to material they described (51 data points). Further, 369 data
prioritise human exposure to chemicals [9–11]. Pocas et al. points did not report the polymer’s surface area or the
[12], and more recently Gavriil, Kanavouras [13], reviewing polymer’s thickness so that the FC could not be calculated.
migration models, concluded that PMMs are computation- We also discarded data that were considered to be clear
ally complex, require measured chemical diffusion coeffi- outliers in the database. For example, experimental results
cients, and are applicable to a limited set of FCMs only. with exposure times equal to 0 (19 data points), initial or
Here, we developed a linear mixed-effects model (LMM) final chemical masses equal to 0 (120 data points), and final
of chemical migration from packaging into food as a function masses higher than initial ones (79 data points). Our final
of the properties of the food, chemical migrant, and FCM. database included 11,563 migration experiment results for
LMMs can link measured chemical transfer to a number of 23 different food simulants and 94 different foods. All the
relevant predictor variables, including chemical, food and experimental results included were from time-dependent
packaging properties. LMMs are particularly useful to control migration experiments conducted on mono-layer polymeric
for potential bias and correlation in the data available, e.g., packaging. Further details about the data cleaning procedure
from overrepresented packaging or food material [14]. This is are provided in SI, S2, Tables S1, S2.
a major advantage in the case of repeated measurements or
hierarchical data, as it allows for accurate predictions despite Predictor variables
the presence of correlated errors [15]. In various scientific
domains, including ecology and atmospheric pollution We selected a number of predictor variables influencing
research, LMM have shown high predictive performance and migration and classified them in the following groups:
large applicability domains [16–18]. To our knowledge, (a) migrant properties, (b) food properties, (c) packaging
LMMs have never been applied to predict the transfer of material properties, and (d) differences in experimental
chemicals from food packaging materials into food. We design. When the study did not provide values for the prop-
compare our results with a recently developed PMM [19]. erties of interest, we used literature values and assumptions to
fill the data gap (N = 3,277). For example, we used food
proxies or the indications of the EU Annex III [21] to assign
Methods specific fat contents to some of the data (e.g. heptane, a food
simulant, was assigned the same value as olive oil) (SI, S3).
Chemical transfer fraction data A final set of predictor variables was then selected based
on their explanatory power, their correlation to the other
The specific fraction of chemical, also called migrant, variables, and the number of missing data (see S1, S4,
transferred from the FCM, hereafter packaging, into food Tables S3–S5, Figs. S1, S2 and Table 1). A total of 10,235
(FC) was defined as shown in Eq. (1) [19]. migration experiment results had information for all of the
mC; food predictor variables chosen (see SI, S4 Tables S3–S5,
FC ¼ ð1Þ Figs. S1, S2 for more details). In the final predictor set, the
mC; packaging
migrant’s ability to transfer from packaging material into
Where mC, food represents the mass of the chemical in the food is described by the molecular weight and the octanol-
food at the end of the migration experiment and mC, packaging water partitioning coefficient as these are recurring drivers
the mass of the chemical present in the packaging at the of migration in literature [22–28]. According to the Eur-
beginning of the migration experiment. More information opean legislation, the food is best described by its pH and
about how this measure of migration was derived is fat content so both were kept as predictor variables [29].
provided in the Supporting Information (SI, S1). The data The crystallinity of the packaging as well as the ability of
used to derive FC was accessed by Ernstoff et al. [19] and the migrant to dissolve in the packaging, the latter sum-
consisted of experimental outcomes reported to the US marised as the total Hansen solubility, also influence
Environmental Protection Agency (US EPA). The database migration (Table 1) [23, 24, 30, 31]. Finally, the experi-
can be accessed upon request to Ernstoff et al. [19]; and mental conditions, described by the temperature and contact
more information on the database can be found at [20]. duration time, can influence the migration modelling
Overall, the database summarised 12,773 experimental results: higher temperature and contact duration times are
results stemming from 127 different literature sources. We known to increase the faction migrated [22, 23, 31–34].
focused our analysis on food and food simulants, so that Given the data availability and the current mechanistic
572 data points representing non-food experiments or knowledge on migration modelling, the predictor variables
unclear food simulants were discarded. We also removed chosen are expected to describe most of the variability
migration values reported for wax-coated paper and found in the database.
A regression-based model to predict chemical migration from packaging to food 471

Table 1 Predictor variables considered for the modelling of the fraction of chemical transferred from packaging into food
Predictor variables Importance in migration modelling

M Molecular weight The lower the molecular weight, the smaller the molecule, the larger its diffusion into the food [1–4].
Octanol-Water partitioning coefficient (KOW) Represents the lipophilicity of the migrant: the higher, the more lipophilic and therefore the more migration
can be expected, given that the food has a high fat content [3, 5–7].
F Ethanol (EtOH) equivalency The ethanol equivalency is a measure of the food’s fat content [8]. Many migrants are lipophilic, thus the
higher the fat content of the food, the higher the migration into the food [2, 5, 6, 9–11].
pH Fang, Wang [12] found higher migration rates in acidic food.
P Crystallinity The higher the degree of crystallinity, the lower the migration [2, 3, 9, 13].
Total Hansen solubility The migration is likely to be lower when the solubility of the chemical in the polymer is high which can be
quantified with the Hansen solubility [2, 3].
E Temperature An increase in temperature increases the mobility of the migrant. Further it can lead to a swelling of the
polymer and a decrease of its crystallinity thus creating more free space for the migrant [1, 2, 10, 13–15].
Duration of contact The migration increases with prolonged contact between the packaging and the food [9, 10, 15, 16].

M stands for the migrant properties, F for food properties, P for packaging properties, and E for experimental design properties

Model fitting We assessed the best LMM’s fit based on the marginal and
the conditional R2 (R2M and R2C), where R2M describes how
We used LMMs to relate the log-transformed fraction of much of the variability can be explained by the fixed effects
chemical transferred from the packaging into food with the of the best model, and the R2C by both the fixed and random
predictor variables, also referred to as fixed effects. Chemical effects [37]. The relative importance of the random effects
migration might be increasingly influenced under certain was then quantified following Assink and Wibbelink [38],
combinations of predictor variables, so we also included and for the fixed effects by comparing the changes in
interactions in the model namely between: the experimental explained variance after randomising the values of each
temperature and the duration [33], the migrant’s molecular variable [39].
weight and the polymer’s crystallinity, and the migrant’s KOW
and the EtOH-equivalency of food. Further, we included a Predictive performance
quadratic term for the experiment’s duration to account for
potential non-linear responses [34, 35]. The fraction of che- The predictive performance of the best LMM was quantified
mical transferred from the packaging was log-transformed using a fivefold block cross-validation (CV), with folds
(base 10 logarithm) to ensure normality in the model resi- based on the random effect with the largest variation
duals, as were the migrant’s KOW, molecular weight, and the (migrant name) [40]. We used the coefficient of efficiency
experiment’s duration. In addition, the continuous variables (CoE Eq. (2)) and the root mean squared error (RMSE, Eq.
were scaled for the model selection procedure. (3)) as predictive performance metrics.
LMMs allow to control for dependencies between P
experimental data or replicated experimental results using ðyi  ^yi Þ2
CoE ¼ 1  Pi 2
ð2Þ
the so-called random effects [14]. The first step in the model i ð yi  yÞ
selection procedure consisted of choosing the optimal ran- sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
dom effects structure by comparing fitted models with the i ðyi  ^ yi Þ2
RMSE ¼ : ð3Þ
same full fixed effects structure and different random effects. N
The models were fitted with the restricted maximum like-
lihood method (REML). The random effect structure of the With yi being the measured migrated fraction, ^yi the
LMM with the smallest Bayesian Information criteria (BIC) predicted migrated fraction, y the average of the measured
value was kept [14]. It included the citation ID, analytical migrated fractions, and N the number of observations. A
method, migrant, packaging, and food name nested in the CoE of 1 represents a perfect fit, while negative values
EU food simulant categories [29, 36]. The second step was indicate that the observed mean is a better predictor than the
to fit all possible fixed effects structures with the maximum model. For the RMSE, the higher the value, the larger the
likelihood method, using the optimal random effects struc- difference between predicted and observed values, thus the
ture [14], and to choose the best fixed effects structure based poorer the fit.
on the lowest BIC value. Finally, the best LMM, with The predictive performance of regression models can drop
optimal random and fixed effects structure and used for when predicting for random effect levels that were not
inferences and predictions, was fitted using REML. The included during model fitting. Given that we wanted to
validity of the assumptions of normality, homogeneity, and expand the applicability domain of our model to untested
independence was tested. migrants, food types, and packaging materials, we assessed
472 M. Douziech et al.

the variation in the model’s predictive performance for several temperature, the crystallinity of the polymer, the ethanol
scenarios: equivalency of the food, and the migrant’s molecular weight
(1) an untested migrant, and KOW. In addition, the interactions between the migrant’s
(2) an untested migrant in an unknown laboratory using molecular weight and polymer’s crystallinity, the migrant’s
another analytical method, KOW and the food’s EtOH equivalency, and the experiment’s
(3) an untested migrant transferring to an untested food duration and temperature were also included. The formula
type in an unknown laboratory using another analytical describing the model is shown in Eq. (4). The ten models with
method, and lowest BIC are shown in the SI, Table S6.
(4) an untested migrant present in an untested packaging    
logðF C Þ ¼ log MWMigrant þ log KOWMigrant þ EtOH  eqFood
transferring to an untested food type in an unknown
þ CrystallinityPolymer þ Temperature þ logðTimeÞ
laboratory using another analytical method.  
þ ðlogðTimeÞÞ2 þlog MWMigrant CrystallinityPolymer
 
Comparison with process-based model þ log KOWmigrant EtOH  eqFood þ Temperature
   
þ logðTimeÞ þ 1jNameMigrant þ 1jNamePolymer
The performance of the LMM was further assessed by com- þ ð1jFood Simulant=NameFood Þ þ ð1jReferenceÞ
paring its predictive power to the PMM, recently developed þ ð1jAnalytical MethodÞ;
by Ernstoff, Fantke [19]. This PMM incorporates mechanistic
ð4Þ
descriptions of the Fick’s law of diffusion and chemical
partitioning into a single equation describing the migrated where the superscript “2” stands for quadratic term, “1|”
fraction of chemicals. We used the PMM as published by denote the random effects, and “/ ” shows the nested
Ernstoff et al. [19], including the implemented equations for structure of the random effects.
the diffusion and packaging-food partitioning coefficients. For the final model fitted to the entire dataset, the fixed
The comparison of the LMM and PMM was based on two effects accounted for 10.4% of the observed variability
subsets of the US EPA database. Subset 1 consisted of 4,949 (R2m) while both random and fixed effects explained 93.5%
data points where measured diffusion coefficients were (R2C). The model met the homogeneity and independence
available, excluding the data points of the study named assumptions, with a slight deviation from normality for the
“Models for the Migration of Low Molecular Weight Addi- extreme residuals, which had a small influence on the
tives in Polyolefins. National Bureau of Standards. Report overall model fit (SI, S7, Tables S7, S8 Figs. S5–S11).
NBSIR 81 e 2264, April 1981.” because the reported diffu- Amongst the random effects, the migrant’s name con-
sion coefficients represented clear outliers compared with the tributed the most to the total variance (31.1%), followed by
other reported values (SI, S5 Figs. S3, S4). Subset 2 included the citation ID (23.0%), the polymer name (22.7%), the
all data points for which the diffusion coefficients could be food name and simulant category (11.8%), and the analy-
estimated (N = 10,235). The performance of the PMM was tical method (4.1%).
compared with the predictive performance of the LMM for an
untested migrant in an unknown laboratory using another Predictive power
analytical method by comparing the coefficient of efficiencies
and root mean squared errors of both models. Higher CoE and The CV shows that the median CoE is 0.30 (standard
smaller RMSE mean a better predictive performance of the deviation SD = 0.18) for the model predictions for chemi-
model considered. Further, we chose to consider all data cals not included in the database. Further excluding the
points for which both models could be applied, disregarding citation ID and analytical method as prior information
the applicability domain of the PMM (meaning the require- reduced the median CoE to 0.22 (SD = 0.12). Excluding
ment for KOW and EtOH-equivalency values to be inside the also the food name in the prediction of the transfer fractions
range of empirical data used to derive the partitioning resulted in a median CoE of 0.09 (SD = 0.30). Finally
coefficient). excluding all random effects results led to a CoE of −0.11
(SD = 0.51). In other words, for predictions of unknown
chemicals in known food products and/or packaging
Results materials, the LMM has added value. However, model
predictions for combinations of a new chemical, food item,
Mixed-effects model and packaging material are likely to be worse than the
average measured transfer fraction from the full dataset.
The final LMM included, as fixed effects, the experiment’s Comparing the root mean squared error leads to the same
duration (both linear and quadratic terms) and the conclusions (SI, S7, Fig. S11).
A regression-based model to predict chemical migration from packaging to food 473

Transfer fraction relationships

From all fixed effects included in the model, the migrant’s

molecular weight and the experiment duration influenced
the fraction migrated the most followed by the temperature
and the EtOH equivalency of the food (SI, S7, Table S8).
Overall, the fraction of chemical migrated increases with
increasing experimental time and temperature (Fig. 1). For
higher temperatures (above 75 °C), however, this relation-
ship seems to plateau after experiment durations of 4 days.
Figure 1 also shows that the fraction of chemical migrating
from packaging into food increases with increasing EtOH-
equivalency value, representing the fat content, of the food.
The overall influence of the migrant’s KOW is rather small.
Still, it can be seen that for food with higher EtOH equiv-
alencies, a higher KOW of the migrant implies an increase in
migrated values (Fig. 1). On the contrary, food with EtOH
equivalencies around 25% show a decrease in migration
with increasing KOW. A similar but less strong relationship
applies for food with EtOH equivalencies around 50%.
Finally, Fig. 1 illustrates that larger chemicals (molecular
weight above approximately 400 g/mol) migrate more from
packaging materials with low crystallinity values (45%)
than packaging with high crystallinity values (65%). The
opposite is true for chemicals with molecular weights below
400 g/mol. Overall, larger chemicals migrated less than
smaller ones.

Comparison to process-based model

The coefficient of efficiency of the LMM derived here is

always larger than the one for the PMM developed by
Ernstoff et al. [19] (Fig. 2). The negative CoE of the PMM
shows that its performance is worse than using the mea-
sured average of the data selected. This, therefore, shows Fig. 1 Relationship between the common logarithm of the chemical
fraction migrated and the common logarithm of the experiment’s time
that the LMM has a larger predictive power than the PMM. for three different temperatures (a), the common logarithm of the KOW
In addition, the applicability domain of the LMM is larger value for three different food ethanol equivalency values (b), and the
than that of the PMM. In fact, Fig. 2 highlights in blue the common logarithm of the chemical’s molecular weight for three dif-
points outside the applicability domain of the PMM but ferent packaging crystallinities (c)
within the one of the LMM.
The 90% confidence interval (CI) of the ten logarithm of
the predicted migrated fraction from the deterministic model descriptors of experimental design by using a data-driven
is ± 4.6 for measured diffusion coefficients, and ±4.8 for LMM approach. The LMM clearly has a better predictive
modelled diffusion coefficients. For the LMM the 90% CI performance compared with the PMM given the higher CoE
of the prediction is ±1.4 and ±1.5 for measured and mod- and lower RMSE. Below, we compare our results with other
elled diffusion coefficients respectively. findings in the literature.

Parameters influencing migration

Discussion
From all the parameters kept in the LMM, the migrant’s
We showed that it is possible to predict the transfer of molecular weight influenced the migrated fraction the most.
chemicals using a limited set of properties for the chemical The KOW, another predictor variable of the migrant’s
migrant, food, and packaging material, combined with properties, was the least influential parameter. While the
474 M. Douziech et al.

Fig. 2 Scatterplots displaying the fraction migrated estimated with the diffusion coefficients (a, b) and data points with estimated diffusion
deterministic model of Ernstoff et al. [19] compared with the linear coefficients (c, d). The points shown in blue display the data points
mixed effects model without citation ID, analytical method and that are outside the applicability domain of the deterministic model of
migrant name in the random effects for data points with measured Ernstoff et al. [19]

importance of the molecular weight is supported by pre- the KOW might influence partitioning differently when food
vious findings (e.g. Fang and Vitrac [22]), the small influ- simulants are considered instead of food [23]. Further, the
ence of the KOW is more surprising given research stressing importance of the EtOH equivalency of the food also
its importance in migration modelling [8, 49]. The primary indicates that diffusion as well as partitioning influence the
explanation probably lies in the fact that the database migration of chemicals from the FCM into the food. In fact,
includes results of migration in early stages, thus likely not the food’s fat content was shown to be one of the main
capturing partitioning effects. It is, therefore, not possible to drivers of the partitioning behaviour of migrants [23, 50].
clearly state that migration is less influenced by the Some of the relationships found with the LMM have
hydrophilic character of the migrant than the fat content of also been described in previous studies. For example, in
the food material it is in contact with. This is supported by accordance with previous findings [22, 23, 28], the LMM
the importance of the EtOH-equivalency of the food in the shows an overall increase in the migrated fraction of
LMM. Another explanation could relate to the difference in molecules with lower molecular weights. Smaller mole-
partitioning behaviour between food and food simulants: cules tend to have a larger ability to migrate than larger
A regression-based model to predict chemical migration from packaging to food 475

ones [22–25]. While this relationship might not apply to all of the total variance of the reported results (23.0%) as did
molecules, the LMM identified this general tendency across the name of the migrant (31.1%) and packaging type
the 10,235 migration experiment results. The model also (22.7%). These results indicate that even though great care
demonstrated that crystalline packaging materials were was given to choose all of the potentially relevant para-
associated with lower migration values, which is in line meters explaining migration, additional variables related to
with observations by, amongst others, Canellas, Aznar packaging, migrant characteristics, and experimental con-
[24]. This is explained by the fact that crystalline materials ditions might be necessary to model the migration from
have more ordered polymer structures, which will typically plastics into food more accurately. While the importance of
allow less movement of the migrants [51]. More interest- citation ID probably simply reflects expected variabilities
ingly, the LMM showed an interaction between the crys- between different laboratories, or research groups, the
tallinity of the packaging and the migrant’s molecular importance of the migrant and packaging name will require
weight. The higher the crystallinity, the larger the differ- more investigation. As a preliminary test, we assessed the
ence in migrated fraction between low and high molecular performance of a model including additional molecular
weight migrants. A potential explanation might be related variables from chemo-informatics [56], which have been
to the swelling of the packaging which occurs when rather related to physico-chemical properties [57, 58], metabolic
hydrophilic FCM is put into contact with food with a high characteristics [59–61], and ecotoxicity values [62].
water content [34]. One might argue that the swelling of Including these predictor variables in the model selection
FCM with higher crystallinity will impact on the migration procedure, only slightly improved the performance of our
of small migrants more than for an FCM with lower model (SI, S8, Tables S9, S10, Fig. S12). Considering the
crystallinity, where parameters other than the available increased level of complexity implied when these predictors
space are likely to influence migration more. So, while the are added, we did not present this model in this work. We
swelling of the packaging was not directly represented by only considered 2D predictor variables while 3D variables,
the predictors included in the LMM, the identified inter- such as the Van der Waal’s volume, can also be relevant
action between the crystallinity of the packaging and the (Fang and Vitrac [22], Vitrac et al. [63]).
migrant’s molecular weight likely accounts for this beha- The small share of variance explained by the food and
viour. The LMM also showed that for fatty food (high food simulant category (11.8%) indicates that the identified
EtOH equivalency), increased KOW of the migrant, thus relationships most likely apply invariably for both.
increased lipophilicity, meant increased migrated fractions. Unlike citation ID and migrant and packaging name, the
The positive relationship between KOW and migrated frac- analytical method explained a relatively low fraction of the
tion was also reported by Sanches-Silva et al. [28], Cai total explained variance. (4.1%), suggesting that migration
et al. [52]. Less attention is given in literature to the pos- experiments conducted with any of the recognised analy-
sible dependency of this relationship on the EtOH equiv- tical methods will lead to comparable results.
alency of the food. In fact, models considering both The model developed here is, to our knowledge, the first
partitioning and diffusion processes usually include a mixed-effects migration model estimating the fraction of
directly positive relationship between migrated fraction and migrants transferred from packaging into food. Our model
the food’s fat content. The LMM accounts for this rela- outperforms the predictive accuracy of previous PMMs [19]
tionship, as well as taking into account that lipophilic independently of whether estimated or experimental diffusion
migrants (higher KOW values) will remain in the packaging coefficients are used. We also showed, that our LMM could
(known to be rather lipophilic) rather than migrate when in be applied to a larger set of chemicals than the PMM devel-
contact with food with lower EtOH equivalencies (e.g. oped by Ernstoff et al. [19], thus showing its larger applic-
Kawamura et al. [53]). Finally, similarly to the PMM of ability domain. Despite the careful choice of the predictor
migration, the LMM predicts an increase in the migrated variables, the LMM is still a simplification of reality and
fraction for higher temperatures and longer contact times. might not capture all complex interactions. In addition, the
Higher temperatures imply higher diffusion rates for the performance of this LMM is also highly dependent on the a
migrant from the packaging into the food [22, 54], while priori knowledge of the food and/or packaging tested. In fact,
longer contact times give chemicals a longer time-window the model performed best when predicting the migrated
to migrate [35, 55]. fraction for packaging and food combinations for which
experimental results were available in the database used for
Explained variance model fitting. As a result, the predictions of migrated fractions
for new migrants of our LMM are most accurate whenever the
Most of the explained variance was captured by the random packaging and food materials are known. While the prediction
effects. Differences in how the studies were conducted, accuracy of the LMM is not high enough for application in
represented by the citation ID, explained an important share packaging compliance or consumer protection purposes, this
476 M. Douziech et al.

model can be very useful when combined with prioritisation 6. European Commission. Bisphenol A: EU ban on baby bottles to
techniques for human exposure to chemicals. In fact, while for enter into force tomorrow: European Commission; 2011. http://
europa.eu/rapid/press-release_IP-11-664_en.htm.
packaging compliance it is important to quantify in absolute
7. European Food Safety Authority. Contaminants update: first of
terms whether a migration level is too high or not, such two opinions on PFAS in food 2018 [03.2019]. https://www.efsa.
prioritisation relies on a comparative assessment of the europa.eu/en/press/news/181213.
migration in different packaging-food applications. In risk 8. Biryol D, Nicolas CI, Wambaugh J, Phillips K, Isaacs K. High-
throughput dietary exposure predictions for chemical migrants
assessment, such a model is useful when screening packaging
from food contact substances for use in chemical prioritization.
materials for potential health impacts as it can be applied Environ Int. 2017;108:185–94.
quickly to a large set of packaging, migrants, and food types 9. Wambaugh JF, Wang A, Dionisio KL, Frame A, Egeghy P,
and could help refine dietary exposure models [64]. Judson R, et al. High throughput heuristics for prioritizing human
exposure to environmental chemicals. Environ Sci Technol.
2014;48:12760–7.
Code availability 10. Isaacs KK, Glen WG, Egeghy P, Goldsmith MR, Smith L, Vallero
D, et al. SHEDS-HT: an integrated probabilistic exposure model
All the analyses were carried out in R version 3.5.1 [41] for prioritizing exposures to chemicals with near-field and dietary
sources. Environ Sci Technol. 2014;48:12750–9.
using the following packages: dplyr and reshape2 to rear-
11. Ernstoff A, Niero M, Muncke J, Trier X, Rosenbaum RK,
range the data [42, 43], ggplot2 and ggeffects to generate the Hauschild M, et al. Challenges of including human exposure
figures presented [44, 45], lme4 to fit the models [46], to chemicals in food packaging as a new exposure pathway in
MuMIn and hydroGOF to compute the goodness of fit life cycle impact assessment. Int J Life Cycle Assess.
2019;24:543–52.
measures [47, 48]. The code necessary to generate the
12. Pocas MF, Oliveira JC, Oliveira FA, Hogg T. A critical survey of
models can be accessed upon request to the corresponding predictive mathematical models for migration from packaging.
author. Crit Rev Food Sci Nutr. 2008;48:913–28.
13. Gavriil G, Kanavouras A, Coutelieris FA. Food-packaging
Acknowledgements We would like to thank Tanja Radusin (Nofima, migration models: a critical discussion. Crit Rev Food Sci Nutr.
Institute for food technology), Jorunn Nilsen, and Siw Bodil Fre- 2018;58:2262–72.
driksen (Norner) for their help in better understanding the properties 14. Zuur A, Ieno E, Walker N, Saveliev A, Smith G. Mixed effects
driving migration from plastic packaging. We also thank two anon- models and extensions in ecology with R. New York: Springer;
ymous reviewers for their comments to improve the paper’s structure 2009.
and readability. This project was funded by the European Union’s 15. Garson GD. Hierarchical linear modeling: guide and applications.
Horizon 2020 research and innovation programme under the Marie California, London, New Dehli, Singapore, Washington DC:
Sklodowska-Curie grant agreement No 641459 (RELIEF). Sage; 2012.
16. Sun W, Palazoglu A, Singh A, Zhang H, Wang Q, Zhao Z, et al.
Prediction of surface ozone episodes using clusters based general-
Compliance with ethical standards ized linear mixed effects models in Houston–Galveston–Brazoria
area, Texas. Atmos Pollut Res. 2015;6:245–53.
Conflict of interest The authors declare that they have no conflict of 17. Bohora SB, Cao QV. Prediction of tree diameter growth using
interest. quantile regression and mixed-effects models. For Ecol Manag.
2014;319:62–6.
Publisher’s note Springer Nature remains neutral with regard to 18. Uzoh FCC, Oliver WW. Individual tree diameter increment model
jurisdictional claims in published maps and institutional affiliations. for managed even-aged stands of ponderosa pine throughout the
western United States using a multilevel linear mixed effects
model. For Ecol Manag. 2008;256:438–45.
19. Ernstoff A, Fantke P, Huang L, Jolliet O. High-throughput
References migration modelling for estimating exposure to chemicals in food
packaging in screning and prioritization tools. Food Chem Tox-
1. Muncke J. Food packaging materials. 2012. https://www.foodpacka icol. 2017;109:428–38.
gingforum.org/food-packaging-health/food-packaging-materials. 20. US. Food and Drug administration. Guidance for industry: pre-
2. FDA. Guidance for industry: preparation of premarket submissions paration of premarket submissions for food contact substances
for food contact substances: chemistry recommendations. 2007. (Chemistry recommendations). 2007. https://www.fda.gov/Food/
https://www.fda.gov/Food/GuidanceRegulation/GuidanceDocuments GuidanceRegulation/GuidanceDocumentsRegulatoryInformation/
RegulatoryInformation/ucm081818.htm. ucm081818.htm.
3. EFSA. Food contact material applications: regulation and gui- 21. Commission Regulation (EU) No 10/2011 of 14 January 2011 on
dance. 2017. http://www.efsa.europa.eu/en/applications/foodconta plastic materials and articles intended to come into contact with
ctmaterials/regulationsandguidance. food, 2011R0010. 2011.
4. Healy BF, English KR, Jagals P, Sly PD. Bisphenol A exposure 22. Fang X, Vitrac O. Predicting diffusion coefficients of chemicals in
pathways in early childhood: Reviewing the need for improved and through packaging materials. Crit Rev Food Sci Nutr.
risk assessment models. J Exposure Sci Environ Epidemiol. 2017;57:275–312.
2015;25:544. 23. Tehrany EA, Desobry S. Partition coefficients in food/packaging
5. Cohen Hubal EA. PFAS: insights from past actions to inform systems: a review. Food Addit Contam. 2004;21:1186–202.
today’s decisions. J Exposure Sci Environ Epidemiol. 24. Canellas E, Aznar M, Nerín C, Mercea P. Partition and diffusion
2019;29:129–30. of volatile compounds from acrylic adhesives used for food
A regression-based model to predict chemical migration from packaging to food 477

packaging multilayers manufacturing. J Mater Chem. 46. Bates D, Maechler M, Bolker B, Walker S. Fitting linear mixed-
2010;20:5100–9. effects models using lme4. J Stat Softw. 2015;67:1–48.
25. Widen H, Leufven A, Nielsen T. Migration of model contaminants 47. Barton K. MuMIn: Multi-model inference. R package version
from PET bottles: influence of temperature, food simulant and 1.42.1. (2018). https://CRAN.R-project.org/package=MuMIn.
functional barrier. Food Addit Contam. 2004;21:993–1006. 48. Bigiarini MZ, Bigiarini MMZ Package “hydroGOF”. R-package.
26. Chea V, Angellier-Coussy H, Peyron S, Kemmer D, Gontard N. 2013. wwwr-project.org/. Accessed 7 May 2018.
Poly(3-hydroxybutyrate-co-3-hydroxyvalerate) films for food 49. Ozaki A, Gruner A, Störmer A, Brandsch R, Franz R. Correlation
packaging: physical–chemical and structural stability under food between partition coefficients polymer/food simulant, KP, F, and
contact conditions. J Appl Polym Sci. 2016;133:41850–8. octanol/water, log POW-a new approach in support of migration
27. Goulas A. Overall migration from commercial coextruded modeling and compliance testing. Dtsch Lebensm-Rundsch.
food packaging multilayer films and plastics containers into 2010;106:203–8.
official EU food simulants. Eur Food Res Technol. 2001; 50. Seiler A, Bach A, Driffield M, Paseiro Losada P, Mercea P, Tosa
212:597–602. V, et al. Correlation of foodstuffs with ethanol-water mixtures
28. Sanches-Silva A, Andre C, Castanheira I, Cruz JM, Pastorelli S, with regard to the solubility of migrants from food contact
Simoneau C, et al. Study of the migration of photoinitiators used materials. Food Addit Contam Part A Chem Anal Control Expo
in printed food-packaging materials into food simulants. J Agric Risk Assess. 2014;31:498–511.
Food Chem. 2009;57:9516–23. 51. Marcato B, Guerra S, Vianello M, Scalia S. Migration of anti-
29. Commission regulation (EU) No 10/2011, 10/2011. 2011. oxidant additives from various polyolefinic plastics into oleagi-
30. Lau O-W, Wong S-K. Contamination in food from packaging nous vehicles. Int J Pharm. 2003;257:217–25.
material. J Chromatogr A. 2000;882:255–70. 52. Cai H, Ji S, Zhang J, Tao G, Peng C, Hou R, et al. Migration
31. Limam M, Tighzert L, Fricoteaux F, Bureau G. Sorption of kinetics of four photo-initiators from paper food packaging to
organic solvents by packaging materials: polyethylene ter- solid food simulants. Food Addit Contam Part A Chem Anal
ephthalate and TOPAS®. Polym Test. 2005;24:395–402. Control Expo Risk Assess. 2017;34:1632–42.
32. Mercea P. Physicochemical processes involved in migration 53. Kawamura Y, Ogawa Y, Mutsuga M. Migration of nonylphenol
of bisphenol A from polycarbonate. J Appl Polym Sci. 2009; and plasticizers from polyvinyl chloride stretch film into food
112:579–93. simulants, rapeseed oil, and foods. Food Sci Nutr. 2017;5:390–8.
33. Bhunia K, Sablani SS, Tang J, Rasco B. Migration of chemical 54. Diduch M, Polkowska Z, Namiesnik J. Factors affecting the
compounds from packaging polymers during microwave, con- quality of bottled water. J Exposure Sci Environ Epidemiol.
ventional heat treatment, and storage. Compr Rev Food Sci Food 2013;23:111–9.
Saf. 2013;12:523–45. 55. Pocas MF, Oliveira JC, Brandsch R, Hogg T. Feasibility study on
34. Alin J, Hakkarainen M. Type of polypropylene material sig- the use of probabilistic migration modeling in support of exposure
nificantly influences the migration of antioxidants from polymer assessment from food contact materials. Risk Anal. 2010;
packaging to food simulants during microwave heating. J Appl 30:1052–61.
Polym Sci. 2010;118:1084–93. 56. Todeschini R, Consonni V. Molecular descriptors for che-
35. Fang H, Wang J, Lynch RA. Migration of di(2-ethylhexyl) moinformatics, Volumes I and II. Mannhold R, Kubinyi H,
phthalate (DEHP) and di- n -butylphthalate (DBP) from Folkers G, editors. Weinheim, German: Wiley-VCH; 2009.
polypropylene food containers. Food Control. 2017;73:1298–302. 57. Gramatica P, Chirico N, Papa E, Cassani S, Kovarich S. QSAR-
36. Schielzeth H, Nakagawa S, Freckleton R. Nested by design: INS: a new software for the development, analysis, and validation
model fitting and interpretation in a mixed model era. Methods of QSAR MLR models. J Computational Chem. 2013;34:
Ecol Evol. 2013;4:14–24. 2121–32.
37. Nakagawa S, Schielzeth H, O’Hara RB. A general and simple 58. Mamy L, Patureau D, Barriuso E, Bedos C, Bessac F, Louchart X,
method for obtaining R2 from generalized linear mixed-effects et al. Prediction of the fate of organic compounds in the envir-
models. Methods Ecol Evol. 2013;4:133–42. onment from their molecular properties: a review. Crit Rev
38. Assink M, Wibbelink CJM. Fitting three-level meta-analytic Environ Sci Technol. 2015;45:1277–377.
models in R: a step-by-step tutorial. Quant Methods Psychol. 59. Pirovano A, Brandmaier S, Huijbregts MA, Ragas AM, Veltman
2016;12:154–74. K, Hendriks AJ. The utilisation of structural descriptors to predict
39. Breiman L, Cutler A. Random forests. 2002. https://www.stat. metabolic constants of xenobiotics in mammals. Environ Toxicol
berkeley.edu/~breiman/RandomForests/cc_home.htm#varimp. Pharmacol. 2015;39:247–58.
40. Roberts DR, Bahn V, Ciuti S, Boyce MS, Elith J, Guillera-Arroita 60. Pirovano A, Huijbregts M, Ragas A, Veltman K, Hendriks AJ.
G, et al. Cross-validation strategies for data with temporal, spatial, Mechanistically-based QSARs to describe metabolic constants in
hierarchical, or phylogenetic structure. Ecography. 2017;40:913–29. mammals. ATLA 2014;42:59–69.
41. R Core Team. R: a language and environment for statistical 61. Papa E, van der Wal L, Arnot JA, Gramatica P. Metabolic bio-
computing. Vienna, Austria: R Foundation for Statistical Com- transformation half-lives in fish: QSAR modeling and consensus
puting; 2016. analysis. Sci Ttal Environ. 2014;470–471:1040–6.
42. Wickham H, Francois R, Henry L, Müller K. dplyr: A grammar of 62. Gramatica P, Cassani S, Sangion A. Aquatic ecotoxicity of per-
data manipulation. R package version 0.8.0.1. (2019). https:// sonal care products: QSAR models and ranking for prioritization
CRAN.R-project.org/package=dplyr. and safer alternatives’ design. Green Chem. 2016;18:4393–406.
43. Hadley W. Reshaping data with the reshape package. J Stat Softw. 63. Vitrac O, Lézervant J, Feigenbaum A. Decision trees as applied to
2007;21:1–20. http://www.jstatsoft.org/v21/i12/. the robust estimation of diffusion coefficients in polyolefins. J
44. Wickham H. ggplot2: elegant graphics for data analysis. J Stat Appl Polym Sci. 2006;101:2167–86.
Softw. 2010;35:65–88. 64. Hartle JC, Fox MA, Lawrence RS. Probabilistic modeling of
45. Lüdecke D. ggeffects: tidy data frames of marginal effects from school meals for potential bisphenol A (BPA) exposure. J Expo-
regression models. J Open Source Softw. 2018;3:772. sure Sci Environ Epidemiol. 2016;26:315–23.