Page 5 - Volume 4, Issue 1, January 2011

© Benaki Phytopathological Institute
Uncertainty in pest risk analysis 3
cases, the traditional approach is to take a
model selection process to find the best
model from which one makes practical ap-
plications. Several criteria have been pro-
posed for selecting models using a test
dataset (e.g. Smith et al., 1999; Townsend
Peterson et al., 2008). However, potential
problems have been recognized by statisti-
cians. An important concern is that the un-
certainty in model selection is basically ig-
nored once a final model is found (Chatfield,
1995; Draper, 1995). Final estimation, inter-
pretation of the parameter values, and mod-
el predictions are generally based on the
selected model only. In some cases, the in-
stability of the result of a selection process
is high; Yuan and Yang (2005) showed that,
when the model errors are large, a selection
process is likely to lead to a completely dif-
ferent selected model when a slightly differ-
ent dataset is used. The selected model may
also depend on the criterion used for mod-
el selection and, as shown by Townsend Pe-
terson et al . (2008) and by Lobo et al. (2008),
there is no consensus in the scientific com-
munity on the best criterion for selecting
models for predicting biological invasion.
For all these reasons, it is never sure that the
selected model is the most appropriate one
for practical applications.
2. Objectives of uncertainty and
sensitivity analysis
Uncertainty analysis consists in evalu-
ating quantitatively uncertainty in model
components (input variables, parameters,
equations) for a given situation, and deduc-
ing an uncertainty distribution for each out-
put variable rather than a single value (Vose,
2000; Monod et al., 2006). It can be used,
for instance, to compute the probability of
an output variable of interest (e.g. number
of spores entering in a given area) to ex-
ceed some threshold (e.g. Peterson et al.,
2009). Uncertainty analysis is a key compo-
nent of model-based risk analysis because
it provides risk assessors and decision mak-
ers with information about the accuracy of
model outputs. In pest risk analysis, uncer-
tainty analysis was used by several authors
to estimate probability of entry and estab-
lishment (Stansbury et al., 2002; Peterson et
al., 2009; Yen et al., 2010), spread of invasive
species (Koch et al., 2009), and to assess ef-
ficiency of management options (Yen et al.,
2010).
The aim of sensitivity analysis (SA) is to
determine how sensitive the output of a
model is with respect to elements of the
model which are subject to uncertainty.
For dynamic models, sensitivity analysis is
closely related to the study of error prop-
agation. As in SA, input variables and pa-
rameters have the same role, uncertain in-
put variables and parameters will be further
denoted as uncertain factors. Two types of
sensitivity analysis are usually distinguished,
local sensitivity analysis and global sensitiv-
ity analysis (Saltelli et al., 2000). Local SA fo-
cuses on the local impact of uncertain fac-
tors on model outputs and is carried out by
computing partial derivatives of the output
variables with respect to the input factors.
With this kind of methods, the uncertain
factors are allowed to vary within small in-
tervals around nominal values, but these in-
tervals are not related to the uncertainty in
the factor values. Contrary to local SA, glo-
bal SA considers the full domain of uncer-
tainty of the uncertain model factors. In glo-
bal SA, the uncertain factors are allowed to
vary within their whole ranges of variation.
Sensitivity analysis may have various ob-
jectives, such as:
to study relationships between model y
outputs and model inputs;
to identify which input factors have a y
small or a large influence on the output;
to identify which input factors need to be y
estimated or measured more accurately;
to detect and quantify interaction effects y
between input factors;
to determine possible simplification of y
the model;
In pest risk analysis, sensitivity analy-
sis techniques were used to study the sen-
sitivity of spatial model predictions to input
factor values (Koch et al., 2009) and to data