Model evaluation and adaptive
management
15
In ecology, a modelis a hypothesis that is usually expressed mathematically. A math-
ematical description allows a more precise definition of the hypothesis than does
a verbal description, and this precision is particularly important for complex, non-
linear processes. We can often construct more than one model to describe a process,
and these alternatives are equivalent to alternative hypotheses. In this chapter, we
explore the methods for choosing between such alternative models or hypotheses.
In Chapter 16, we will introduce the concept of statistical inference, which uses
standardized criteria for decision-making to help ensure that decisions are not
swayed by personal opinion or pressure brought to bear by politicians or the public.
Despite its widespread use, however, statistical inference is not the only, nor even
necessarily the best, way to choose among a wide variety of alternative hypotheses,
whether these arise in the quest for “pure” or more “applied” knowledge ( Johnson
1999; Anderson et al. 2000; Guthery et al. 2001; Johnson and Omland 2004).
Statistical tests are effective at ruling out null hypotheses. The trouble is, the null
hypothesis is sometimes an explanation that we need not seriously entertain, so
rejecting it is not helpful for increasing our understanding of observations. For
example, the null hypothesis in many wildlife habitat studies is that animals have no
habitat preferences. We would be astounded if this ever proved true, so what
progress do we make in rejecting it?
There are far fewer examples of hypothesis testing that are directed at evaluating
a suite of alternative models or hypotheses that vary subtly from one another. It is
hard enough to gather enough data to discriminate between random versus “signi-
ficant” patterns of association, let alone tease apart subtle variants. More importantly,
however, classic statistical methods are often impossible to use when alternative
models are not special cases of more general models. This situation is particularly
common in the kind of non-linear models that we find in ecology. Such “non-nested”
models, in the jargon of professional statisticians, present special problems for
finding a suitable statistical test.
Statistical inference is also plagued by “statistical” versus “biological” significance.
You will recall that in statistical hypothesis testing, a Pvalue of less than 0.05 is
taken to mean that there is a remote probability (1 in 20) that an observed pattern
could have been produced by chance alone. This probability is quite sensitive,
however, to the amount of data that go into the assessment. Endangered species are
often plagued by a crucial lack of data. This can preclude sufficient sample sizes
and replicated treatments needed for significance testing, leaving us with no reliable
option for making management decisions or informed scientific judgments, if we rely
on standard statistical approaches. Even when we have a large sample of data that
253
15.1 Introduction
