Food Safety and Toxicology
CD Carrington, College Park, MD, USA
PM Bolger, Annapolis, MD, USA
Ó 2014 Elsevier Inc. All rights reserved.
Background
The Federal Food Safety Program resides in part under the
authority of the US Food and Drug Administration. The specific
statutory authority is the Federal Food, Drug, and Cosmetic Act
(FFDCA). The areas of responsibility include the consideration
of the safety or risk of food and color additives, both direct
and indirect, and food-borne contaminants, both natural and
anthropogenic. The FFDCA prescribes somewhat different
standards of safety/risk for intentional food additives that
undergo premarket assessment of safety, specifically, versus
those dietary constituents that are found in food as contami-
nants because they occur naturally or because they arise from
anthropogenic sources.
The prototypical safety assessment for food-borne
compounds is the acceptable daily intake (ADI) methodology,
which was first documented in 1954, and has come to be
employed throughout the world. This paradigm has also been
codified in the consideration of food (e.g., aspartame) and
color additives (e.g., Red Dye No. 2) and pesticides (e.g.,
atrazine). It is also routinely used in the consideration of
incidental food-borne chemical contaminants (e.g., lead),
particularly as a tool for screening out trivial incidents of
exposure. This procedure specifies that an acceptable dose of
a chemical may be calculated with the following equation:
ADI ¼ NOEL=SF
where ADI is the acceptable daily intake, NOEL is the no-
observed-effect level, and SF is a safety factor. In subsequent
considerations, an identical methodology has been imple-
mented where new terms have been substituted for those
originally described for the ADI safety assessment method-
ology. These terms are the reference dose (RfD), minimal risk
level (MRL), and tolerable daily intake (TDI) for the ADI. In
these methodologies, NOEL has been further defined to be the
no-observed-adverse-effect level (NOAEL) and the SF is called
an uncertainty factor (UF). These alternative terms have sup-
planted the ADI in considerations of food-borne contaminants
because unlike food additives, ‘acceptance’ is not an applicable
term for chemical contaminants that do not have a premarket
approval evaluation, and therefore, these alternative terms were
devised. Since contaminants are not deliberately added, and for
some are fairly widespread in the environment and the food
chain, it is often much more difficult to achieve population
exposures that are below an RfD/MRL/TDI.
As a general rule, the NOEL/NOAEL is a dose from
a controlled animal experiment where no adverse effect
(i.e., an effect not considered harmful) is noted. The experi-
ment does not establish that no effect can possibly occur at
that dose under any conditions – it only denotes that none of
the effects looked for in the experiment was observed. Since
a statistical significance test is typically used to establish
whether or not an effect occurred, the NOEL/NOAEL will tend
Encyclopedia of Toxicology, Volume 2 http://dx.doi.org/10.1016/B978-0-12-386454-3.00022-1
to diminish as the sensitivity of the measurement or the
number of observations increases. Since the burden of proof is
on science to show that an effect has occurred, greater
uncertainty tends to raise the level of exposure that is deemed
acceptable/tolerable. The selection of the effect that is
considered adverse is a matter of societal values (i.e., local-
ized, reversible, mild discomfort versus frank, and irreversible
systemic toxicity). That is, establishing that an effect has
occurred is a separate consideration from how much one cares
if it will occur or not.
In the original safety assessment paradigm, a single SF of
100 was used to derive an ADI from an NOEL. The justification
of the SF (also called a UF) was based on scientific consider-
ations as well as a judgment about how to manage the inherent
uncertainty associated with the assessment. Scientific issues
raised included the notion that humans may be more sensitive
to chemicals than rodents used in a laboratory test and that
there may be substantial variability among individuals in
a population. The necessity of managing uncertainty was
acknowledged through the recognition that it is impossible to
demonstrate that no adverse affect could not occur under any
circumstances. As the ADI approach evolved, the single SF was
replaced by two or more SF/UFs of 10, with each factor applied
as a response to a particular scientific issue. For example, an
NOEL/NOAEL from an animal experiment is normally divided
by a factor of 10 as a matter of policy in response to the
supposition that humans may be more sensitive than labora-
tory animals and by another factor of 10 to account for vari-
ability of response in the population of concern. SF/UFs dictate
the impact of uncertainty of some quantity on the decision.
Although the magnitude of the uncertainty is not precisely
described, the general idea is that the greater the uncertainty,
the larger the SF/UF needs to be. Thus, uncertainty in SF/UFs
and uncertainty in the derivation of the NOAEL push in
opposite directions. As a matter of practice, the uncertainty
underlying an SF/UF application is not quantified, so that
a factor of 10 is almost always employed. Even if the uncer-
tainty were quantified, the magnitude of the UF would still
depend on some judgment about what degree of risk adversity
is appropriate.
One outcome of the dependence of the NOEL/NOAEL on
the statistical significance test is that it tends to penalize
chemicals for which there is more or better data. To remedy this
problem, the benchmark dose (BMD) concept was introduced
as an alternative approach. The BMD depends on the specifi-
cation of a low-level effect that would typically be unobserv-
able. The end point may be the specified percentage (5 or 10%)
above background of a population for an end point deemed to
be adverse. Since the end point is defined, determinations for
different chemicals and different data sets tend to be more
comparable.
Other efforts have endeavored to give SF/UFs a scientific
basis by assembling a range of observations that are analogous
639
640 Food Safety and Toxicology
to the occasions for SF/UF application. As indicated previously,
while the RfD/MRL/TDI methodology is essentially the same as
for the ADI derivation, the end point has been recast as
a threshold estimate. However, the ADI/MRL/RfD/TDI inter-
pretation of the product of an NOAEL as a threshold presents
some difficulties. First, since the number generated is partly
a product of the management of the uncertainty, it does not
make sense to say that threshold of safety itself is uncertain.
Second, there is always some uncertainty about whether or not
there is a threshold, especially when considering subpopula-
tions of individuals who may already be ill. Perhaps more
importantly, even if the NOAEL/UF procedure is considered as
a rough estimate of a threshold dose, it only provides infor-
mation about what may happen at that particular dose; there is
no indication about the likelihood or frequency of an adverse
effect at higher doses.
An important part of the safety assessment process lies in
establishing the impact of scientific uncertainties on the
eventual decision. This is typically done by making conser-
vative estimates that deliberately err on the side of safety.
Since scientists are more familiar with the scientific issues, the
responsibility for this is often delegated to them. There are
two potential difficulties with this. The first problem is that
the technical person is entrusted with a decision that is partly
a social/political one. This is particularly true if there is an
interaction between degree of uncertainty and degree of harm.
For example, a larger degree of uncertainty may be tolerated
when the adverse effect is relatively trivial than when it
is severe. However, if the technical person is not aware of all
the competing dietary and nondietary risks that impinge on
the decision, then they may not be in a position to judge the
relative importance of a particular risk. The second problem is
that a technical person may dictate the impact of the uncer-
tainty for only part of the problem, without being aware of the
other uncertainties involved. There is therefore no way to
judge the ‘appropriate’ degree of conservativeness for the
particular problem. Because the safety assessment process
may have many steps, the decision process may be fragmented
to a degree that no one can judge whether or not a decision is
reasonable.
From an administrative standpoint, a great benefit of the
safety assessment process is its relative simplicity. It deals with
a broad spectrum of dietary public health issues that may be
quite complex both socially and scientifically with a few short
rules. For small problems, the benefit of either more carefully
considering the scientific issues or encouraging widespread
participation in the decision process may not justify the effort
that must be expended to do so. The efforts to improve
the scientific basis of safety assessment often result in the
complexity that may result from a risk assessment without the
benefit of clarity of purpose. There are occasions where
the safety assessment process may prove to be too simple.
The bigger the problem is economically and politically, the
less comfortable the general public may be with delegating
important social/political decisions to scientists. As a result,
the safety assessment process is more useful for identifying
small problems, than it is in dealing with large problems. For
food additives, the safety assessment process ensures that any
risk from a compound that is intentionally added to food is
trivial.
Food Risk management
safety/ Prediction
Risk
question ? Risk assessment
Model ? Data
Research
Figure 1 A dietary public health decision paradigm.
As with food additives, the safety assessment process for
contaminants is helpful in identifying chemicals that pose
a trivial risk. However, the safety assessment process may not
be useful for those contaminants where the level of exposure in
a population already exceeds what has been identified as the
threshold of safety. The level specified by the safety assessment
process may be unattainable in all circumstances, and control
may best be directed at reducing exposure in general rather
than attaining a particular level. Considering the best options
available may require better information than the safety
assessment process can deliver. In this case, a formal risk
assessment is required, where the goal is to provide an estimate
of the probability of harm for a public health threat. A key
difference between safety and quantitative risk assessment
(QRA) is that while the former is a decision process in and of
itself, a QRA is intended to provide information as part of
a larger decision process – this view of risk assessment is
illustrated in Figure 1.
Since QRA is an applied analysis, the goal is to describe
what is known in response to a particular question. Where
matters of degree of exposure and risk are involved, numbers
provide more precise descriptions than words, particularly for
exposures that exceed the threshold of safety. In QRA, numbers
may be used to describe both empirical quantities and, for the
purpose of describing uncertainty, degrees of knowledge. As
these are two fundamentally different purposes, the benefits of
using quantitative tools are somewhat different. For describing
empirical quantities (e.g., body weight and heart rate),
numbers are useful because they are more accurate in that they
provide a more precise statement of what is known. On the
other hand, when describing uncertainties, the primary
advantage of quantitative methods is that they are more
transparent, thereby allowing more people to participate in the
decision process.
QRA models constructed for dietary risks typically have
two main components: an exposure assessment and a dose–
response assessment. The exposure assessment determines the
amount of a chemical contaminant that consumers will be
exposed to from the diet, and sometimes may include expo-
sure from other sources as well. The dose–response assess-
ment characterizes the causal relationship between dietary
exposure and one or more adverse health outcomes. The
interactions with risk management occur at the beginning
(formulating the question) and end (making the decision) of
the risk assessment process (Figure 1). In particular, risk
managers often have an interest in identifying the toxic effect
that a QRA model is designed to address. The end result is the
risk assessment model that explicitly states how a particular
adverse effect is causally related to a particular dietary expo-
sure to a contaminant.

Uncertainty
Public health risk assessments used in considerations of food
safety are intended to make predictions, particularly quan-
titative, of plausible negative impacts. Obviously depending
on the underlying scientific information, some predictions
are more reliable than others, and the extent to which
predictions are reliable should be taken into account in the
public health decision process. Since risks of chemical
contaminants in food are generally small and are often
difficult to observe and measure, the uncertainties associated
with estimating the occurrence of potential adverse effects
can be large relative to the estimated effects. Therefore,
consideration of uncertainty is very often an inherent part of
the risk assessment process.
In discussing uncertainty, it is important to distinguish
efforts to quantify uncertainty from the somewhat related
activity of characterizing the nature and extent of frequencies in
a population or series. Although uncertainties may often be
reasonably calculated or based on frequencies (i.e., statistical
uncertainty), frequencies may be of interest on their own
(e.g., population variability), while some uncertainties are not
frequency based (e.g., model uncertainty). The difference
between probability and frequency is often a matter of context.
The same distribution might be used to describe the frequency
of an event when the question is about a population or serve as
the basis for a quantitative probability statement when the
question is about an individual. Consider two examples. First,
a distribution describing the variability of a scientific instru-
ment over time may be used to describe the uncertainty of
individual measurement made by the instrument. Second,
a distribution describing body weight could be used to predict
the frequency of values expected in a population or the
uncertainty of the value in a particular individual for whom
a measurement is unavailable.
When mathematical models are used to draw inferences,
the values of the model parameters may reflect uncertainties
arising from other sources. In particular, traditional statistical
methods often place confidence intervals on parameter
estimates that reflect the potential influence of sampling
error (i.e., the data may not be representative of the pop-
ulation or series that an estimate is needed for). Bootstrap
methods can be used to include other sources of uncertainty
in model parameter estimates, which involve repeating the
modeling process many times using different assumptions
about either the data or the mathematical form of the
models or both.
Model uncertainty arises from the availability of alterna-
tive mathematical expressions with which an inference or
prediction might be made. As a practical matter, model
uncertainty is often not an important issue. In particular, so
long as there are a number of observations to guide the
model that come from similar circumstances as the problem
that the analysis is concerned with, then any plausible model
should yield very similar results. However, when models are
used to extrapolate values, model uncertainty can be an
important source of uncertainty. Prime examples are the use
of models to predict outcomes at low doses (see Figure 2)
and the use of distributions to estimate the frequency of
occurrence of rare events.
Food Safety and Toxicology 641
One hit Weibull Probit
Multistage Logit Data
0
Log population frequency
−1
−2
−3
−4
−5
−6
−7
−6 −5 −4 −3 −2 −1 0 1
Log dose (mg kg−1)
Figure 2 Model uncertainty in cancer risk assessment.
Dose–Response Models
Dose–response models describe a cause–effect relationship.
There are a wide range of mathematical models that have been
used for this purpose. The complexity of a dose–response
model can range from a simple one-parameter equation
to complex multicompartment toxicokinetic/toxicodynamic
models. Many dose–response models, including most cancer
risk assessment models, are population models that predict the
frequency of a disease in a population. Such dose–response
models typically employ one or more frequency distributions
as part of the equation. Dose–response may also operate at an
individual level and predict the severity of a health outcome as
a function of dose. Particularly, complex dose–response
models may model both severity of outcome and population
variability and perhaps even recognize the influence of
multiple causal factors.
Since the estimation of an adverse effect at doses consistent
with exposures normally seen in human populations generally
involves high to low extrapolation, uncertainty will arise from
the fact that different plausible models may result in very
different estimates (see Figure 2). Model uncertainties may be
represented by constructs that are commonly referred to as
probability trees. The basic idea behind a probability tree is that
the sum of the probabilities of all the models under consid-
eration is one. Selecting and assigning probabilities to the
nodes of a probability tree are accomplished by an evaluation
of the weight-of-the-evidence. To some extent, this generally
involves some reliance on expert opinion. If a more formal
process is necessary or desired – possibly to enable the evalu-
ation to be carried out by a computer – an algorithm may be
used. Since presumably the same considerations are involved,
the same criteria that are used to select a ‘best’ model may also
be used to weight them.
Statistics and Public Health
Food safety of chemical contaminants is a public health
problem where the goal is to evaluate the toxicological
642 Food Safety and Toxicology

impact in a population rather than an individual. This adds
a statistical dimension to the assessment. Population vari-
ability is especially important for exposure assessment to
chemical contaminant since the amount of a particular food
consumed and the amount of chemical contaminant in
a food will vary, sometimes to a fairly extensive degree.
Variability in dose–response relationships may also be
important in some cases.
Frequency distributions may be used to represent or draw
inferences about the variation in a particular value among
a population or series (i.e., a single set of values such as body
weight). There are a number of ways of fitting or estimating
the parameters for a frequency distribution. The popularity of
the normal distribution may be at least partly attributed to the
fact that estimates for the parameters (with a particular set of
assumptions about the relationship between the model and the
data) can be calculated directly. There are a large number of
frequency models to choose from. Although there is some
theoretical basis for many of them, the theory is at best only
loosely applicable to describing variability in biological pop-
ulations – the biology is almost certainly more complicated than
the model. It is often a good idea to use several models.
It is also often possible to use empirical distributions to
represent variation in exposure. In particular, simulations are
often constructed with food consumption surveys as the start-
ing point. The simulation model can then add other sources of
variation to produce a risk estimate for a larger population.
Empirical distributions may also be used for other components
of an exposure assessment (e.g., the occurrence of a chemical
contaminant in a food). On the other hand, since dose–
response relationships are always theoretical (i.e., they are not
directly observed), variability in population susceptibility must
be modeled as well.
Risk Characterization and Simulations
Once the overall risk assessment model is constructed, it may
be used to make predictions. Running a large risk assessment
model that includes exposure and a dose–response function
collecting the results is often referred to as a simulation. If there
are statistical components to the model, the model may be run
repeatedly using different random numbers to select values
from the statistical distributions each time. This process is
known as Monte Carlo simulation. In public health models,
distributions can be used to describe variability in populations
or the uncertainty in a value, parameter, or model. Since
Table 1 A methylmercury risk mitigation scenario via the use of a fish
advisory. Estimated IQ increases, relative to baseline, in a population after
a proposed fish consumption advisory
Percentile IQ point change
Average 0.023 (0.009, 0.041)
25th 0.000 (0.000, 0.002)
Median 0.007 (0.003, 0.014)
75th 0.025 (0.009, 0.043)
95th 0.057 (0.023, 0.100)
99th 0.091 (0.036, 0.167)
99.5th 0.218 (0.086, 0.401)
99.9th 0.294 (0.116, 0.573)
will be used in making a decision. It is important to inform the
decision maker on the summary process so that they under-
stand that a distribution lies behind the single number. The
simplest way to do this is to simply display a list of percentiles
along with the summary statistics. The results of a 2D simula-
tion will necessarily be more complicated.
Simulations may also be constructed to estimate the impact
of potential risk mitigation procedures. If simulations are being
run for a number of different scenarios (e.g., expected values
with and without public health intervention), it is preferable to
generate a table of summary statistics as percentiles such as
shown in Table 1. At each percentile, an IQ change is given
along with the corresponding confidence limits.
Although they may allow for quick comparisons, tables
inherently compare one value at a time. Graphing or visuali-
zation can be a better way of presenting the entire distribution.
A 1D simulation will produce a frequency (when simulating
variability) distribution or a likelihood distribution (when
representing uncertainty). There are two ways of plotting
frequency or likelihood curves: the first is to plot density versus
value (Figure 3), which emphasizes the values which are the
most common or likely, and the second is to plot cumulative
percentiles versus value (Figure 4), which allows the percentile
corresponding to a particular value to be read from the plot.
Either plot can be used to represent either a frequency distri-
bution or an uncertainty that has a statistical origin.
0.025
0.02
Probability density

uncertainty is ever present, the presence of distributions in the

model that are intended to describe variability usually results in 0.015
a two-dimensional (2D) distributional model, where one
dimension represents population variability and another 0.01
represents uncertainty in the outcome.
Whether it is 1D or 2D, Monte Carlo simulation produces
many individual estimates – one per iteration. While these 0.005
numbers are the end result of the risk assessment, no one can
look at all these numbers and make sense of them. Therefore, 0
they need to be sorted and tabulated and summarized. 0 20 40 60 80 100
However, deciding how to do this involves considering how Distribution value
the results are going to be used and by whom. For example,
reporting a distribution as a mean entails that the mean is what Figure 3 A probability density curve.
 Food Safety and Toxicology 643

100% 100
90%
80%
Cumulative frequency

70%

Percent
60%
50%
40%
30%
20% 0
0.50 1.50
10% Mercury blood to diet ratio
0%
0 20 40 60 80 100 Figure 6 Variability in mercury blood to diet ratios.
Distribution value

Figure 4 A cumulative frequency curve. Research proposals are justified on the basis of an expectation
that they will reduce uncertainty. There are three general
purposes that additional research may be designed to address:
1. Measurement of values used directly in risk assessment (e.g.,
90 in empirical distribution functions).
2. Collection of data that will allow more accurate estimates of
80 model parameters.
3. Collection of data that will allow discrimination among
70
models or influence weight-of-the-evidence evaluations.
60
Distribution value

50 Summary
40
Although contaminants are not intentionally added, the
30 assessment of contaminants typically begins with the applica-
Cumulative likelihood

0.99
0.9
0.5
0.1
0.01
0.01 0.05 0.1 0.25 0.5 0.75 0.9 0.95
Cumulative frequency
Figure 5 Three-dimensional representation of risk: severity (distribution
value), frequency, and uncertainty (cumulative likelihood).
Two-dimensional results are more difficult to display. Two
strategies for adding an extra dimension are illustrated in
Figures 5 and 6. The first uses 3D perspective to portray the
third dimension (Figure 5). The second uses shading, where
darker hues are used to represent either higher density or more
central values (Figure 6). This is particularly useful for dis-
playing uncertainty, as the more probable parts of a curve are
more defined.
tion of the same safety assessment process that is used for the
20
evaluation of food additives. In most instances, such an
10 assessment is sufficient in providing the assurance of safety of
the potential exposure to many dietary contaminants. There
0
0.99 are, however, other instances where the environmental
contaminant is so ubiquitous, and therefore difficult to avoid,
that a more informative analysis needs to be considered. The
output of such an assessment is intended to describe the degree
of harm expected in the population and the degree of uncer-
tainty associated with the estimates. Specific areas of uncer-
tainty can also be identified that will inform and suggest
meaningful areas of research.
See also: Cumulative (Combined Exposures) Risk Assessment;
Monte Carlo Analysis; Safety Pharmacology; Uncertainty
Analysis; Uncertainty Factors.

Identifying Data Gaps and Planning Research

Putting a formal QRA together will inevitably raise areas of
uncertainty that can be addressed through further research.
Since it has identified the issues that are important for the
decision, a risk assessment can be useful in planning the
research that will have the most impact on future decisions.
Further Reading
Carrington, C.D., Bolger, P.M., 2010. The limits of regulatory toxicology. Toxicol. Appl.
Pharmacol. 243, 191–197.
Carrington, C.D., Montwill, B., Bolger, P.M., 2004. An intervention analysis for the
reduction of exposure to methylmercury from the consumption of seafood by
women of child-bearing age. Regul. Toxicol. Pharmacol. 40, 272–280.
644 Food Safety and Toxicology

Gaylor, D., Axelrad, J., Brown, R., et al., 1997. Human risk assessment prac-
tices in the US Food and Drug Administration. Regul. Toxicol. Pharmacol. 26,
307–321.
National Research Council, 1983. Risk Assessment in the Federal Government:
Managing the Process. National Academy of Science Press, Washington, DC.
http://foodsafe.msu.edul – National Food Safety and Toxicology Center (at Michigan
State University).
http://toxnet.nlm.nih.gov – TOXNET, Specialized Information Services, National Library
of Medicine. Search for FoodSafety and Toxicology.
http://www.cfsan.fda.gov – US Food and Drug Administration, Toxicological Principles
for the Safety Assessment of Food Ingredients.

Relevant Websites

http://www.atsdr.cdc.gov – Agency for Toxic Substances and Disease Registry.

Toxicological Profile for FoodSafety and Toxicology.
http://www.who.int – Joint Expert Committee on Food Additives of the World Health
Organization and Food and Agriculture Organization, Environmental Health Criteria
70 - Principles for the safety assessment of food additives and contaminants in
food.
http://www.who.int/foodsafety/chem/principles/en/index1.html – Joint Expert
Committee on Food Additives of the World Health Organization and Food and
Agriculture Organization, Environmental Health Criteria 240 Principles and methods
for the risk assessment of chemicals in food.