Page 4 - Volume 4, Issue 1, January 2011

© Benaki Phytopathological Institute
Makowski 2
egorical rating (e.g. sum, multiplication) as
shown by Makowski and Mittinty (2010).
The use of quantitative models has sev-
eral advantages compared to qualitative
methods for pest risk analysis. Quantitative
models do not require the definition of cate-
gorical ratings and can be used to compute
numerical probabilities of entry and estab-
lishment, and to quantify spread and impact
(EFSA, 2008a). Quantitative models can also
be used to assess and select qualitative scor-
ing systems for PRA (Makowski and Mittinty,
2010).
Quantitative models are generally not
used to perform full PRA, but rather to es-
timate some elements of PRA like probabil-
ity of entry, probability of establishment,
spread, or impact. Several probabilisticmod-
els have been developed to predict prob-
ability of entry of pests through import-
ed commodities (e.g. Roberts et al., 1998). A
great diversity of models has been used to
assess the risk of establishment of pest from
bioclimatic variables: statistical models (e.g.
Roura-Pascual et al., 2009), models based
on machine learning techniques (e.g. Phil-
lips et al., 2006), models taking into account
the ecological processes involved in biolog-
ical invasion like Climex (Young et al., 1999)
and NAPPFAST (Magarey et al ., 2007). Epide-
miological models have been developed to
assess risk of spread and impact (Stansbury
et al ., 2002).
These models are powerful tools, but
they include several sources of uncertainty
that need to be taken into account by risk
assessors and communicated to decision
makers. In this paper, we review the main
sources of uncertainty in models used for
PRA, and discuss the practical interest of
uncertainty and sensitivity analysis for pest
risk assessors. The paper is organized as fol-
lows: Sources of uncertainty in models used
for PRA are presented in section 1. The ob-
jectives of uncertainty and sensitivity anal-
ysis are presented in section 2 and the main
steps of these two types of analysis are de-
scribed in section 3. Finally, a case study is
presented in section 4.
1. Origins of uncertainty in models
used for pest risk analysis
Models used for PRA can include up to
four sources of uncertainty, namely input
variables, parameter values estimated from
expert knowledge, parameter values esti-
mated from data, and equations. Input vari-
ables correspond to variables whose values
vary between sites and/or year and can be
measured. Climatic variables, such as tem-
perature annual range or annual precipita-
tion, are typical examples of input variables.
Climatic variables can be measured from
weather stations, but their values are often
imperfectly known due to error of measure-
ment or due to the absence of weather sta-
tion in the sites of interest. Climate change
can also increase uncertainty (Araujo and
New, 2006; EFSA, 2008a).
Parameters correspond to model com-
ponents whose values cannot be directly
measured but need to be estimated from
expert knowledge, from data, and from
both expert knowledge and data. When pa-
rameters are estimated from expert knowl-
edge, the accuracy of the estimates depends
on expert bias and on the method used for
expert knowledge elicitation (O’Leary et al.,
2008). When parameters are estimated from
data, the accuracy of the parameter esti-
mates depends on the estimation technique
and on the quality of the dataset. Consider,
for example, models used for mapping in-
vasive species distribution from bioclimatic
variables. These models include parameters
that need to be estimated from a set of spe-
cies presence records and, if possible, from
a set of species absence data (Vaclavik and
Meentemeyer, 2009). It was shown that the
performances of these models were related
to the size of the datasets and to the reliabil-
ity of presence and absence data (Wisz et al.,
2008; Vaclavik and Meentemeyer, 2009; Gio-
vanelli et al., 2010).
Model equation is another source of un-
certainty. Several alternative models may be
available for a given practical problem, es-
pecially for predicting invasive species dis-
tribution (Roura-Pascual et al ., 2009). In such