Developing predictive models - Conservation and Sustainable Use - page 172

Although the multiple regression approach is powerful, there are some caveats.
Firstly, it is no substitute for getting a full understanding of the model through
the more simple model exploration described above. It is tempting when using this
technique to just let your model produce 'data' which are then amenable to
statistical analysis—but this turns your model into a black box , which is a highly
dangerous thing to do, because you can't then know when your model is not
working as you think it is, through a programming bug or an unrealistic assump-
tion that you have made in developing the conceptual model. Models are extreme
simplifications of reality, which are there to help you explore the consequences of
your assumptions about how the world works. If the model is so complex that you
don't have an intuitive feel for the mechanisms that are operating, then it stops
being a useful tool and becomes a liability. Even complex models can be broken
down into manageable sub-models which can then be explored and understood.
Second, it is important to follow normal statistical good practice when
analysing model outputs—ensuring that the relationships between the model
inputs and outputs are linear before testing them using a linear model, for example.

5.3.5 Incorporating uncertainty
So far we have varied parameter values in order to carry out sensitivity analyses. But
models can also be built explicitly to incorporate environmental variation that
changes demographic rates from year to year, or observation uncertainty in how
managers perceive the system that they are trying to manage. Model uncertainty
(our confidence in the model itself ) can be addressed using similar methods
to those discussed in Section 4.4.1 with respect to statistical models, and is also
ideally suited to Bayesian modelling approaches (Section 5.4.4). We discuss
observation uncertainty in Section 7.5.2. Here we discuss ways to incorporate
process uncertainty into models.
There is a trade-off to be made in deciding how much variability to incorporate
into a model. The real world is variable. There are large-scale events which can
knock a system out of equilibrium, such as a major episode of poaching or a severe
drought or disease epidemic. There is also the normal variation in weather and
chance events such as accidents. The modeller needs to decide how much of
this variability to incorporate into a model, in order to ensure that the results are
realistic, but that the model doesn't become so messy and complex that its heuris-
tic value is lost.
Demographic variation (for example, in the sex of individual offspring or the fate
of particular individuals) is not usually important at population sizes of about 200
or above, but for small populations it is important to incorporate it. For example, in
our deer model we used an adult survival rate of 0.8. With 201 adults in year t , this
would produce 160.8 adults in year t 1, which we can comfortably round to 161.

With 21 adults, it may really matter whether there are 16 or 17 individuals next year.
The easiest (though not the most elegant) way to incorporate demographic
variation into a simulation model is by simulated coin-tossing. For each of the 21