Environmental Engineering Reference
In-Depth Information
This approach goes hand in hand with the adaptive management approach dis-
cussed in Chapter 7, and with the approach of confronting models with data pro-
moted by Hilborn and Mangel (1997) and others. It is an extremely powerful and
elegant way of dealing with uncertainty, allowing its effects on predictions to be
fully incorporated into the modelling process and clearly expressed for managers.
However, developing a full Bayesian model is a task for a specialist.
Bayesian modelling is still a very young field in conservation—this is reflected in
the fact that at the moment there are many more articles telling us what a useful
tool it is than actually using it! Even if you decide not to invest the substantial time
and effort required to become a Bayesian modeller, the underlying philosophy can
still inform your modelling—you can still explore the effects of uncertainty on
your model predictions using the range of techniques discussed in Section 5.3, and
can still give management advice that explicitly takes account of uncertainty.
5.4.4.1 Bayesian networks
Bayesian network (or Bayesian Belief Network, BN) models are currently
increasing in popularity in conservation and natural resource management (e.g.
Marcot
et al
. 2001; Wisdom
et al
. 2002). Their only link to Bayesian statistical
models is that they use Bayes' theorem—they are otherwise a completely different
type of model. Bayesian networks have a set of nodes connected by directional
links, and are usually used to show causative relationships between parameters
(Box 5.4).
Bayesian networks are attractive because they are relatively easy to programme
and to understand, and because they have accessible and well-documented soft-
ware available (see below for links). The modelling software available allows sensi-
tivity analyses, exploration of the effects of the addition of new information to the
model, and also allows dynamic models that show how the system evolves over
time. The graphical presentation can be a very attractive way of presenting results
to end-users. However, there are some caveats as well:
●
Cycles
are not allowed, i.e. a node can't influence itself, even indirectly. This
causes the model to blow up, but can generally be easily solved by using a
dynamic model. For example, animal population size in time t
1 is a
function of animal population size in time t, through the action of density
dependence.
●
The
causative relationships
that are represented in the BN are a major factor
in determining the outcome of the model. Just as for any model, the under-
lying model structure needs to be robust, and if there is doubt about the
causative links, alternative versions of the model need to be tested.
●
Just as for any off-the-shelf package, the range of
in-built functions
is limited,
and may not include what you want to do, particularly as regards model
exploration and data entry. The models may also run quite slowly if they are
large and complex. However, the packages do allow you to programme your
own add-ins, which make them substantially more powerful if you have the
technical expertise required.
