Preventive Veterinary Medicine 110 (2013) 411

Contents lists available at SciVerse ScienceDirect

Preventive Veterinary Medicine

journal homepage: www.elsevier.com/locate/prevetmed

Graphical models and Bayesian domains in risk modelling:

Application in microbiological risk assessment
Matthias Greiner a,b, , Joost Smid c,d , Arie H. Havelaar c,d ,
Christine Mller-Graf a
a
b
c
d

Federal Institute for Risk Assessment (BfR), Berlin, Germany

Federal Institute for Risk Assessment (BfR) and University of Veterinary Medicine Hannover (TiHo), Germany
Institute for Risk Assessment Sciences, Faculty of Veterinary Medicine, Utrecht University, Utrecht, The Netherlands
National Institute for Public Health and the Environment (RIVM), Bilthoven, The Netherlands

a r t i c l e

i n f o

MSC:
62F15
92C42
92D30
92C42
Keywords:
Bayesian graphical models
Quantitative microbiological risk
assessment

a b s t r a c t
Quantitative microbiological risk assessment (QMRA) models are used to reect knowledge
about complex real-world scenarios for the propagation of microbiological hazards along
the feed and food chain. The aim is to provide insight into interdependencies among model
parameters, typically with an interest to characterise the effect of risk mitigation measures.
A particular requirement is to achieve clarity about the reliability of conclusions from the
model in the presence of uncertainty. To this end, Monte Carlo (MC) simulation modelling
has become a standard in so-called probabilistic risk assessment.
In this paper, we elaborate on the application of Bayesian computational statistics in the
context of QMRA. It is useful to explore the analogy between MC modelling and Bayesian
inference (BI). This pertains in particular to the procedures for deriving prior distributions
for model parameters. We illustrate using a simple example that the inability to cope with
feedback among model parameters is a major limitation of MC modelling. However, BI
models can be easily integrated into MC modelling to overcome this limitation. We refer
a BI submodel integrated into a MC model to as a Bayes domain. We also demonstrate
that an entire QMRA model can be formulated as Bayesian graphical model (BGM) and
discuss the advantages of this approach. Finally, we show example graphs of MC, BI and
BGM models, highlighting the similarities among the three approaches.
2013 Elsevier B.V. All rights reserved.

1. Introduction and objectives

Risk assessors in the area of food safety are challenged
with the increasing complexity of global feed and food supply chains and the dynamics of the agro-food business and
associated food safety systems. The formal framework for
conducting quantitative microbiological risk assessment

Corresponding author at: Federal Institute for Risk Assessment (BfR),

Berlin, Germany.
E-mail addresses: matthias.greiner@bfr.bund.de (M. Greiner),
joost.smid@rivm.nl (J. Smid), arie.havelaar@rivm.nl (A.H. Havelaar),
christine.mueller-graf@bfr.bund.de (C. Mller-Graf).
0167-5877/$ see front matter 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.prevetmed.2013.02.008

(QMRA) has been laid out by the Codex Alimentarius

Commission (1999). Safety risks for food of animal origin
have been studied within this framework using so-called
modular process risk models (Nauta et al., 2007). Such
models structure the farm-to-fork continuum into subsets of scenarios and often aim to identify and quantify
those factors that are amenable to intervention in the
presence of other inuencing factors that cannot be mitigated. Depending on the scope of the risk question, the
nal outcome of such models may be the probability of
illness associated with the consumption of a serving of
contaminated food, the incidence of foodborne disease in
the population or some metric of the associated public
health impact. The primary purpose of QMRA models is to

M. Greiner et al. / Preventive Veterinary Medicine 110 (2013) 411

generate insight into the interdependence of model input

and output variables and in particular to quantify the effect
of optional mitigation measures.
The parameters of QMRA models may be derived from
empirical data or from expert opinion. The use of point
estimates for model input parameters is referred to as
deterministic modelling. However, we often need to
quantify the nal outcome against the background of statistical uncertainty and known or assumed heterogeneity
of model parameters. Statistical parameter uncertainty is
mainly a reection of the sample size in relation to the size
of the referent population and is linked to the concept of
precision in statistical inference. Furthermore, structured
or unstructured parameter heterogeneity may be an issue
if, for example, model parameters are known to depend
on other observed and unobserved factors or if empirical
estimates of the parameter vary among aggregated units
of observations, respectively. Moreover, a situation often
encountered in QMRA is the complete lack of empirical
knowledge about certain parameter values. In this case,
the distributional assumptions about a parameter reect
expert opinion (belief) about the unknown quantity. The
use of probability density distributions accommodates all
cases outlined above and is referred to as probabilistic modelling. Probabilistic risk assessment models are
implemented using Monte Carlo (MC) simulation, that
is random sampling from input probability distributions
to produce probability distributions of relevant output
parameters.
Another notion of randomness is relevant in QMRA
models. Each iteration of the simulation model results in
one discrete realisation of the nal outcome which could
occur in the real world, while statistical summary of the
entire simulation reveals the range and probability distribution of possible outcomes (Vose, 2008). Therefore,
probability density distributions or functions thereof are
also used to generate random events, such as for example
the number of contaminated dishes consumed per year.
The random outcome itself represents irreducible variability while the parameters of the associated distribution
models may be parameterised to account for uncertainty as
described above. It has been emphasised that the impact of
variability and uncertainty should be addressed in QMRA
models (WHO, 2009) and handled separately (Nauta, 2000).
Uncertainty distributions of input parameters may be
estimated from data using classical frequentist statistics
but more commonly are based on Bayesian inference (BI,
Vose, 2008). In MC modelling, parameter uncertainty is
only propagated in one direction (that is forward from
input to output parameters and not backwards). An alternative approach involves Bayesian inference and directed
acyclic graphs (DAGs) for propagating uncertainty and variability throughout the full model. We will refer to such
models as Bayesian graphical models (BGMs). In this paper,
we introduce MC modelling using a simple QMRA model
and compare MC and BI approaches (Section 2). Then, we
demonstrate the use of BI in the context of MC simulation
models (Section 3). We explain that QMRA models can be
formulated as Bayesian graphical models and illustrate this
using an example (Section 4). Finally, we summarise and
provide some conclusions (Section 5).

Table 1
Parameters of a simple QMRA model for E. coli O157:H7 in ground beef
differentiating variability (v) and uncertainty (u).
c0 N(10, 2): mean number of bacteria in ground beef according to
expert opinion (u).
k = 0: growth rate for model described for item ct under assumption
of storage under freezing.
t betapert(6, 24, 12): storage time. Minimum, maximum and most
likely values are based on expert opinion (v).
ct = c0 exp(kt): exponential growth model for number of bacteria at
time t.
i discrete(X(1, 1/5, 1/50), p(0.027, 0.373, 0.600)): surviving fraction
X of bacteria after cooking rare, medium and well-done,
respectively, and corresponding probability p (v).
s gamma(3.93, 0.0806): serving size (g) of a dish for children <3
years of age prepared from ground beef. Gamma model was
selected and tted to (hypothetical) empirical consumption data
(sdata) (v).
n Poisson(ct is): number of viable bacteria in one ready-to-eat
serving (u and v).
r Uniform(0.0005, 0.0015): probability of illness per bacterial cell
(single hit theory) given by experts (u).
P = 1 (1 r)n : probability of illness in children <3-year old
associated with bacterial exposure as per one serving of a dish
made of ground beef (outcome).
Adapted from Pouillot et al. (2009).

2. Monte Carlo (MC) simulation models and

Bayesian inference (BI)
For illustration of Monte Carlo (MC) modelling, we
begin with a simplistic QMRA model, adapted from Pouillot
et al. (2009), who used this example for introducing the
R package mc2d for two-dimensional Monte Carlo modelling. Escherichia coli O157:H7 is a microbiological hazard
in food such as ground beef. The risk question pertains to
the probability of illness in children less than three years
of age due to consumption of a single serving prepared
from ground beef contaminated with E. coli O157:H7. The
model is described in Table 1. The model accounts for variability and uncertainty of model input parameters, both
expressed as probability distributions. Functional dependencies among model items can be visualised as a graph
(Fig. 1).
The MC approach encompasses the classical simulation paradigm, whereby a stochastic system is constructed
to mimic individual parametric and process uncertainties
by combining together many distinct probabilistic models
(for example mass or density functions). In contrast, in a
Bayesian inference (BI) model the joint probability distribution is formed across all parameters and processes and
data (if available). The key advantage of the BI approach is
that knowledge from one part of the model migrates to
other parts of the model.
MC simulation models and BI models share the Bayesian
paradigm of interpreting population parameters as random
variables to capture (parameter) uncertainty. The information about model parameters can be specied in terms of
an appropriately parameterised probability density function. A statistical summary consists of a measure of central
tendency (mean or median) and percentiles. The analogy
between BI and MC allows that the output from MC models can be interpreted in a Bayesian way. The interpretation

M. Greiner et al. / Preventive Veterinary Medicine 110 (2013) 411

describe variability (for example lognormal concentration)

with uncertain parameters by dening these as hyperparameters.
3. Bayesian inference in the context of Monte Carlo
modelling

Fig. 1. Dependencies among Monte Carlo model items visualised by a

graph (model adapted from Pouillot et al. (2009)). The shape of items
corresponds to the item type (square: data; triangle: MC variate; circle:
numerical value; and hexahedron: function of MC variates). The colours
of the items were adjusted between green and red, reecting the lowest (most uncritical) and highest (most critical) level of uncertainty and
knowledge base, respectively (qualitative assessment details not shown
here), while the lightblue colour is used for the outcome function.

of MC model outputs in a frequentistic framework appears

(unnecessarily) more complicated.
Another analogy between BI and MC is that priors (BI)
and model input parameters (MC) can be specied empirically or subjectively. We emphasise that, although the
latter option exists, great care should be exercised to justify
subjective input parameters. Formal schemes for assessing
expert assumptions (see for example Cooke and Goossens,
2000; Van der Sluijs et al., 2005) could be considered to
characterise the choices made by experts. Furthermore, it
is good practice to investigate the impact of alternative
parameterisations.
A common approach in elicitation of expert opinion
makes use of questions such as Between which values
do you think that this parameter is with what condence,
and which value do you consider most likely? The R
package rriskDistributions has the functionality to t a
choice of distributions to quantiles that satisfy the experts
assumptions (Belgorodski et al., 2012b).
An important difference between MC and BI models
is how uncertainty and variability are handled. In MC,
a two-dimensional model must be formulated with the
variable parameters in the inner loop and the uncertain
parameters in the outer loop. This greatly increases computational intensity and often results in a limited number
of uncertainty iterations. For the example model shown
in Fig. 1, the inner loop consists of sampling from Monte
Carlo model items representing variability only (i, s, n)
and the outer loop consists of sampling from other items
reecting uncertainty only or an unresolved combination
of variability and uncertainty (c0, t, r). In a BI model
this is not necessary, the model structure denes this
naturally. It is also quite natural to introduce factors that

As pointed out by Smid et al. (2010), Bayesian models

have the advantage of allowing backwards reasoning and
account for correlations among model parameters as long
as they are jointly estimated. For illustration we extend
the simple E. coli model introduced in the previous section.
Assume that there is some evidence that not all batches of
fresh beef are contaminated with E. coli O157:H7 and that
the empirical apparent prevalence estimate of contaminated batches is given by AP = X/n = 12/319 (pooled results
from German monitoring, communication by Dr. Matthias
Hartung, BfR, Germany). Lets further assume that a pilot
validation study for the concerned laboratory method has
 = 104/125 and speciyielded estimates of sensitivity Se
 = 94/100 (Beutin et al., 2010). MC modelling of
city Sp
these data consists of a sequential realisation of MC samples using standard parameterisation (see for example
Vose, 2008) and a function thereof, that is
1. Generate se as MC sample from a beta(105, 22) to reect
uncertainty about the diagnostic sensitivity estimated in
the validation study.
2. Generate sp as MC sample from a beta(95, 7) to reect
uncertainty about the diagnostic specicity estimated in
the validation study.
3. Generate ap as MC sample from a beta(13, 308) to reect
uncertainty about the apparent prevalence estimated in
the application study.
4. Establish pi=(ap+sp-1)/(se+sp-1) as prevalence estimate adjusted for misclassication (Rogan and Gladen,
1978).
The distribution obtained in step 4 includes negative
values, suggesting that the results of the validation study
are not consistent with the results of the application study.
A dilemma for the MC modeller: Which study results are
more trustworthy, from the validation study or from the
application study? How do we account for the sample
sizes in estimating the apparent prevalence, sensitivity and
specicity? BI is a simple and elegant solution because it
provides the joint posterior distribution for the misclassication parameters and the adjusted (true) prevalence.
Negative values for the prevalence cannot occur and the
empirical information from the validation study and the
application study is used in combination. The sample sizes
are implicitly used in the parameterisation of the prior beta
distributions. The model is shown as graph in Fig. 2 generated using OpenBUGS (Lunn et al., 2009). A corresponding
model is implemented in the R package rriskBayes as
described previously (Belgorodski et al., 2012a).
In our example, the (BI) joint posterior distribution
allows that evidence from the application study informs
also the posterior misclassication estimates (backward
reasoning). The BI approach yields the empirical joint

M. Greiner et al. / Preventive Veterinary Medicine 110 (2013) 411

parameter can be used to account for non-contaminated

batches of ground beef (details not shown). We refer those
parameters of an MC model which are linked by one
BI model (pi, se, sp in our example) to as a Bayes
domain. Resampling from the joint posterior distribution
in R is an option which is facilitated by the rriskBayes
package for R.
Bayesian models have been used in the context of QMRA
to some extent (see for example Pouillot et al., 2003; Duffy
et al., 2006; Pollino et al., 2007; Delignette-Muller and
Cornu, 2008; Teunis et al., 2008; Donald et al., 2009; Barker
and Gomez-Tome, 2011; Rigaux et al., 2012).
Fig. 2. Graph of a simple Bayesian inference (BI) model for the count variable X, modelled as binomial with parameters ap=pi*se+(1-pi)*(1-sp)
and n generated using the doodle function of OpenBugs (Lunn et al., 2009).
See text for further details.

posterior distribution of model parameters including

empirical correlations (see Fig. 3), which can be used
for further MC modelling. The correlation pattern among
model parameter estimates could be introduced in an MC
model on the basis of frequentistic estimates as well. However, this would require distributional assumptions (for
example multivariate normal or copula models) and a further modelling step. In our example, the true prevalence

4. Comparison of MC and BGM

An overview of the advantages and disadvantages of
Monte Carlo (MC) simulation models and Bayesian graphical models was provided by Smid et al. (2010), who
referred to these models as Bayesian network models. A
clear strength of MC models is their wide use and that
they are intuitive, relatively easy to implement and exible. Complex systems with correlated parameters and all
kinds of probability distributions can be implemented to
capture both variability and uncertainty.
An important limitation of MC models is the difculty
validating the model parameters. It is often an ad hoc

Fig. 3. Paired scatterplot of joint posterior estimates for the model shown in Fig. 2.

M. Greiner et al. / Preventive Veterinary Medicine 110 (2013) 411

100
0
0
0
0
0
0
0
0

0
0
100

6E-2
14
86
0

0
0
0
0
100
0
0
0
0

Fig. 4. Bayesian network model for Salmonella in pork.

Reprinted with permission from Smid et al., 2011.

approach which includes consecutively running the model

with different parameter values and investigating the effect
of these values on outcome variables of the model, for
which data are available. The purpose of this analysis is to
establish a set of parameter values for which the outcome
variables are close to the data. This approach of parameter
calibration is in many cases a time-consuming process. The
results may equally well be explained by several combinations of parameter values, which is similar to the problem
of identiability in inferential models. MC models concentrate on marginal distributions of the separate variables
instead of the joint distribution of all variables, which
complicates an analysis of the joint effects of multiple
variables.
However, BGM models include the joint probability distribution of all variables. Storage of a joint distribution
becomes computationally challenging when performing
exact probabilistic inference across a graph, in which
case secondary structures, called junction trees, are used
(Lauritzen and Spiegelhalter, 1988). Still, this can be a
very memory-intensive task if the same complexity of the
domain is included in a BGM that would be used in MC models and it may not always be feasible, even with modern
(desk-top) computers. Therefore, the joint distribution may
have to be simplied, for example by appropriate sensitivity analyses such as proposed by Smid et al. (2011). BGMs
are acyclic and thus do not support feedback loops (Jensen,
2001). Recursions may be implemented as dynamic BGMs.
Yet, in practice, such networks may become very large.
Smid et al. (2011) described a way to approximate recursive
equations in a BGM.
However, the focus on a joint distribution has many
advantages. Validation of the model parameters is much
more straightforward than in an MC model. Learning from

downstream data is a central element in BGM modelling

and is relatively easy to perform by consecutively using
Bayes rule in the Bayesian graphical model. Automated
algorithms exist in BGM software to do this, for example
sequential adaptation (Kjaerulff and Madsen, 2008) which
consists of using consecutive datasets describing the combined states of variables to update the prior parameter
distributions in light of the evidence about downstream
variables.
The elicitation of priors is described in Section 2. Inherently, the information provided by experts is subjective and
may depend on the expert asked. In a BGM (downstream)
data are combined with the prior distributions and used to
update them, improving or decreasing the initial beliefs of
the expert. This provides an elegant way of interpreting the
inclusion of expert opinion in the model.
The interactive link between the parameters of a BGM
and downstream data also allows for efcient sensitivity
analyses. It has been shown that there is a surprisingly simple correlation between the probability of a set of evidence
and the probability of a parameter in the model having
a certain value: the probability of the evidence is a linear function of the parameter (Coupe and van der Gaag,
2002; Kjaerulff and Madsen, 2008). In BGM software algorithms exist to rapidly determine the coefcients of these
relations and, thus, to study how sensitive an outcome
variable in the model is to small changes in the prior probability distribution of a parameter. This can be done for
individual parameters (one-way parameter sensitivity
analysis) or for combinations of parameters (n-way parameter sensitivity analysis) (Kjaerulff and Madsen, 2008).
Another type of sensitivity analysis which needs Bayesian
inversion is evidence sensitivity analysis. This type of sensitivity analysis may give answers to questions like Which

M. Greiner et al. / Preventive Veterinary Medicine 110 (2013) 411

Expected transferred number of Salmonella to exterior

1

Expected transferred number of Salmonella to interior

1

0.5

0.5

7
6

5
4

4
3

2
1

7
6

6
5

3
2

2
1

1
0

Variance of transferred number of Salmonella

Transfer of Salmonella from house flora

0.5

0.5

to

to

to

bo

ex

tf

in

no

th

Expected transferred number of Salmonella to exterior

Expected transferred number of Salmonella to interior

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

Variance of transferred number of Salmonella

Transfer of Salmonella from house flora

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

to

to

to

bo
th

t
ex

tf

in

no

Fig. 5. Barplots representing the discrete prior (dashed lines) and posterior (solid lines) distributions of the parameters quantifying the transfer of Salmonella
from house ora. Upper 4 plots: 2009 data, lower 4 plots: 2011 data (numbers are log-transformed; data from Smid et al. (2012)).

evidence supports or refutes a hypothesis? and How

would the conclusions of the model change if another value
had been observed instead of the observed value? Analyses to assess the value of collecting additional information
may be seen as another type of sensitivity analysis. Information measures such as the entropy of a variable (which

measures the spread of the probability mass over the states

of a variable) and the mutual information between two
variables (which provides a measure of the information
shared between two variables in a model) are examples of
two outputs of a value of information analysis. These measures are easy to compute using probabilistic inference and

10

M. Greiner et al. / Preventive Veterinary Medicine 110 (2013) 411

this functionality is included in most BGM software (for

example Hugin).

4.1. Example Salmonella in pork

Recently, two QMRA models have been developed to
describe the dynamics of Salmonella in the pork production chain by means of mathematical models. One is an MC
model (EFSA, 2010), the other a BGM (Smid et al., 2011,
2012). The BGM uses the same mathematical equations as
the MC model but it was implemented as a network which
includes the joint distribution of all variables in the model.
In this section we compare these models to provide a practical example of the advantages and disadvantages of MC
models and BGM. The graph of the BGM example is shown
in Fig. 4.
The MC model was developed as a generic model valid
for different EU countries which were clustered into a small
number of regions having similar delivery chain characteristics. The model accounts for differences between the
chain characteristics of slaughterhouses within one cluster
and between different product units by including values of
the model parameters as probability distributions. Often,
these were uniform distributions between the smallest and
largest values that were reported for different slaughterhouses in published literature. The variability expressed
by these distributions is propagated through the model by
means of MC simulations. Uncertainty about the parameter
values was assumed to be absent in this model because of
practical and computational convenience.
The BGM aims to describe one specic slaughterhouse in
The Netherlands and does, therefore, not include variability
between slaughterhouses. Yet, uncertainty about the values of certain parameters in this specic slaughterhouse
is considerable. Therefore, the parameter distributions
describing variability between different slaughterhouses,
as used in the MC model, were in this model considered as
uncertainty distributions quantifying the initial absence of
knowledge about the specic parameter values governing
the bacterial dynamics in this slaughterhouse.
By including Salmonella prevalence and concentration
data from carcasses sampled upstream and downstream in
the chain in our specic slaughterhouse, the parameters
of the BGM could be inferred with an increased precision.
The sampling experiment was performed in the spring of
2009 and the data have been fully described previously
by van Hoek et al. (2012). An important nding from the
2009 data was that house ora was likely to have acted
as an independent source of contamination in 2009 and
was likely to have contaminated many carcasses at the end
of the slaughter line (van Hoek et al., 2012). Inclusion of
these data in the BGM provided additional evidence for
this hypothesis. Fig. 5 (upper four plots) shows that the
prior distributions of the parameters of the BGM quantifying the transfer of house ora were conservative estimates
of their posteriors. The transfer of house ora was not considered in the baseline MC model and, in an adapted version
hereof, this transfer was considered to be of only limited
importance. So, if the MC model would have been used to
describe the Dutch slaughterhouse it would have largely

Fig. 6. Marginal prior probabilities of the states of parameter HF, representing the probability of transfer of house ora to the carcass exterior,
interior or both on each date (Day/Month) in 2009 for which the datasets
were obtained.
Reprinted with permission from Smid et al., 2012.

underestimated the Salmonella prevalence and load on carcasses after slaughter.

The experiment was repeated (data not published).
Inclusion of the 2011 data in the BGM showed that house
ora was not an important contributor to contaminated
carcasses after slaughtering in 2011 (Fig. 5, lower four
plots). Fig. 6 also shows that the values of a parameter
(in this case the parameter indicating the probability of
transfer of house ora to a carcass) may change over time.
Because of the unidirectional model structure, the MC
model can only be used to provide expected numbers and
prevalences of Salmonella on a product unit at different
points in the chain and to study the effects of interventions
to reduce Salmonella on the end product. The multidirectionality of the information ow in BGMs allows for
a relatively new purpose of QMRA models, that is biotracing (Barker et al., 2009). In these non-typical cases
the purpose of a farm-to-fork model is to identify the
origin of food-borne hazards, for example the sources of
Salmonella in pork. The BGM provides a realistic example of how this works. The trigger signal for a biotrace in
this model is the detection of a Salmonella contaminated
carcass after slaughtering, for which quantitative data are
available. Given these data and the model structure, posterior estimates of the model parameters are generated and
can be used for identifying a likely source. Posterior probabilities alone are not always reliable as an evaluation of
evidence (Barker and Gomez-Tome, 2011). However, these
may be evaluated by additional variables such as for example report variables (Smid et al., 2011), which return the
probabilities of an effect E for every potential source of
having caused the largest proportion of the bacterial population on the end product; or statistical tests as for example
likelihood ratios P(E|S)/P(E|S) for a source S given evidence
relative to an alternative, S.
The example of Salmonella in pork provided here shows
that dynamical changes of the values of parameters in a
model may be inferred from data, and a Bayesian model

M. Greiner et al. / Preventive Veterinary Medicine 110 (2013) 411

seems the most convenient model structure for this task.

Hence, implementing a QMRA model as a BGM and using
targeted data collections can transform a generic model
into a specic one. Such specic models, trained by data,
can be used for additional purposes of QMRA models such
as biotracing.
5. Conclusions
We have shown similarities between Monte Carlo (MC)
modelling for probabilistic risk assessment, Bayesian inference (BI) and Bayesian graphical models (BGMs). Bayesian
methods provide a exible and very powerful approach to
quantitative microbiological risk assessment (QMRA). MC,
BI and BGM have in common the expression of parameter uncertainty using appropriate distribution functions
and can be visualised as network graphs. Conceptually,
the Bayesian approach is preferable for risk assessment
modelling but currently the implementation may be more
challenging in practice than MC.
Conict of interest
None declared.
Acknowledgements
We wish to thank Annett Martin and Dr. Matthias Hartung (both Federal Institute for Risk Assessment (BfR),
Berlin, Germany) for providing information about the validation study and the monitoring results, respectively, for
the E. coli example. We gratefully acknowledge the constructive comments from the editor and two reviewers. In
particular, we are indebted to one of the reviewers who
helped to improve the clarity of some of the central denitions in our manuscript.
References
Barker, G., Gomez, N., Smid, J., 2009. An introduction to biotracing in food
chain systems. Trends Food Sci. Technol. 20, 220226.
Barker, G., Gomez-Tome, N., 2011. A risk-assessment model for enterotoxigenic Staphylococcus aureus in pasteurized milk: a potential route
to source-level inference. Risk Anal. 33, 249269.
Belgorodski, N., Greiner, M., Engelhardt, A., 2012. rriskBayes: Predened Bayes Models Fitted with Markov Chain Monte Carlo
(MCMC) (related to the rrisk project). R Package Version 2.3.
http://www.bfr.bund.de/cd/52158
Belgorodski, N., Greiner, M., Tolksdorf, K., Schueller, K., 2012. rriskDistributions: Fitting Distributions to Given Data or Known Quantiles. R
Package Version 1.8. http://www.bfr.bund.de/cd/52158
Beutin, L., Martin, A., Krause, G., Steege, K., Haby, S., Pries, K., Albrecht,
N., Miko, A., Jahn, S., 2010. Ergebnisse, Schlussfolgerungen und
Empfehlungen aus zwei Ringversuchen zum Nachweis und zur
Isolierung von Shiga (Vero) Toxin bildenden Escherichia coli (STEC)
aus Hackeischproben. J. Consum. Prot. Food Saf. 5, 2134.
Codex Alimentarius Commission (CAC), 1999. Principles and guidelines for
the conduct of microbiological risk assessment, cac/gl-30. Tech. Rep.,
Codex Alimentarius Commission. http://www.codexalimentarius.net
Cooke, R., Goossens, L., 2000. Procedures guide for structural expert judgement in accident consequence modelling. Radiat. Prot. Dosimetry 90,
303309.
Coupe, V., van der Gaag, L., 2002. Properties of sensitivity analysis of
Bayesian belief networks. Ann. Math. Artif. Intel. 36, 323356.

11

Delignette-Muller, M.L., Cornu, M., AFSSA STEC study group, 2008.

Quantitative risk assessment for Escherichia coli O157:H7 in frozen
ground beef patties consumed by young children in French households. Int. J. Food Microbiol. 128, 158164, http://dx.doi.org/10.1016/
j.ijfoodmicro.2008.05.040.
Donald, M., Cook, A., Mengersen, K., 2009. Bayesian network for risk of
diarrhea associated with the use of recycled water. Risk Anal. 29,
16721685.
Duffy, G., Butler, F., Cummins, E., et al., 2006. E. coli O157:H7 in beefburgers produced in the Republic of Ireland: a quantitative microbial
risk assessment. Tech. Rep., Ashtown Food Research Centre, Teagasc,
Ashtown, Dublin 15, Ireland.
EFSA, 2010. Quantitative microbiological risk assessment on Salmonella
in slaughter and breeder pigs: nal report. Report Prepared by VLA in
Consortium with DTU and RIVM.
Jensen, F., 2001. Bayesian Networks and Decision Graphs. Springer-Verlag,
New York.
Kjaerulff, U., Madsen, A., 2008. Bayesian Networks and Inuence Diagrams. Springer-Verlag, New York.
Lauritzen, S., Spiegelhalter, D., 1988. Local computations with probabilities on graphical structures and their application to expert systems. J.
R. Stat. Soc. 50, 157224.
Lunn, D., Spiegelhalter, D., Thomas, A., Best, N., 2009. The BUGS project:
evolution, critique, and future directions. Stat. Med. 28, 30493067.
Nauta, M., 2000. Separation of uncertainty and variability in quantitative
microbial risk assessment models. Int. J. Food Microbiol. 57, 918.
Nauta, M.J., Jacobs-Reitsma, W.F., Havelaar, A.H., 2007. A risk assessment
model for campylobacter in broiler meat. Risk Analysis 27, 845861,
http://dx.doi.org/10.1111/j. 1539-6924.2006.00834.x.
Pollino, C., Woodberry, O., Nicholson, A., Korb, K., Hart, B., 2007. Parameterisation and evaluation of a Bayesian network for use in an ecological
risk assessment. Environ. Model. Softw. 22, 11401152.
Pouillot, R., Albert, I., Cornu, M., Denis, J.B., 2003. Estimation of uncertainty
and variability in bacterial growth using Bayesian inference. application to listeria monocytogenes. Int. J. Food Microbiol. 81 (March (2)),
87104.
Pouillot, R., Delignette-Muller, M., Kelly, D., Dennis, J., 2009. The
mc2d package. Tech. Rep. http://riskassessment.r-forge.r-project.org/
docmcEnglish.pdf
Rigaux, C., Ancelet, S., Carlin, F., Nguyen-Th, C., Albert, I., 2012. Inferring
an augmented Bayesian network to confront a complex quantitative
microbial risk assessment model with durability studies: application
to Bacillus cereus on a Courgette Pure production chain. Risk Anal.,
http://dx.doi.org/10.1111/j. 1539-6924.2012.01888.x.
Rogan, W., Gladen, B., 1978. Estimating prevalence from the results of a
screening test. Am. J. Epidemiol. 107, 7176.
Smid, J.H., Heres, L., Havelaar, A.H., Pielaat, A., 2012. A biotracing model
of Salmonella in the pork production chain. J. Food Prot. 75, 270280,
http://dx.doi.org/10.4315/0362-028X.JFP-11-281.
Smid, J.H., Swart, A.N., Havelaar, A.H., Pielaat, A., 2011. A practical framework for the construction of a biotracing model: application to
Salmonella in the pork slaughter chain. Risk Anal. 31, 14341450,
http://dx.doi.org/10.1111/j. 1539-6924.2011.01591.x.
Smid, J.H., Verloo, D., Barker, G.C., Havelaar, A.H., 2010. Strengths and
weaknesses of Monte Carlo simulation models and Bayesian
belief networks in microbial risk assessment. Int. J. Food
Microbiol. 139 (Suppl. 1), S57S63, http://dx.doi.org/10.1016/
j.ijfoodmicro.2009.12.015.
Teunis, P.F.M., Ogden, I.D., Strachan, N.J.C., 2008. Hierarchical dose
response of E. coli o157:h7 from human outbreaks incorporating heterogeneity in exposure. Epidemiol. Infect. 136, 761770,
http://dx.doi.org/10.1017/S0950268807008771.
Van der Sluijs, J., Craye, M., Funtowicz, S., Kloprogge, P., Ravetz, J., Risbey,
J., Fermi, V., Campus, C., 2005. Combining quantitative and qualitative
measures of uncertainty in model-based environmental assessment:
the NUSAP system. Risk Anal. 25, 481492.
van Hoek, A., de Jonge, R., van Overbeek, W., Bouw, E., Pielaat, A.,
Smid, J., Malorny, B., Junker, E., Lfstrm, C., Pedersen, K., Aarts,
H.M., Heres, L., 2012. A quantitative approach towards a better
understanding of the dynamics of Salmonella spp. in a pork slaughterline. Int. J. Food Microbiol. 153, 4552, http://dx.doi.org/10.1016/
j.ijfoodmicro.2011.10.013.
Vose, D., 2008. Risk Analysis: A Quantitative Guide. Wiley, Chichester.
WHO, 2009. Risk characterization of microbiological hazards in food.
Tech. Rep., World health Organization (WHO). https://www.who.int/
foodsafety/publications/micro/MRA17.pdf