of their AIC scores. A wise researcher or wildlife manager would consult all of the
models whose ∆AIC values were within a difference of 2 of the leading model, because
all such models are consistent with the evidence. The robustness of future popula-
tion projections can then be re-evaluated as further evidence is accrued. This learn-
ing procedure is especially useful if comparisons can be made under conditions in
which plausible models make different predictions.Nowhere is model evaluation more important than in the emerging concept of
adaptive management (Walters 1986; Walters and Holling 1990). No ecosystem is
completely understood. As a consequence, we can never predict with certainty how
any ecosystem will respond to human intervention, such as harvesting (Chapter 19),
or conservation programs (Chapter 18). What this means, of course, is that any man-
agement policy that we choose to adopt should be viewed as an experiment whose
outcome is uncertain. Good wildlife managers have always recognized this, at least
subconsciously, and accordingly gather data to monitor the status of species with
which they are charged. Where “adaptive” management departs from simply “good”
management is in formalizing a mechanism by which management policies can be
improved over time, by reducing at least some of the uncertainties.
One way to go about this is to make the best use possible of historical data to
judge which model out of many possible models is most useful. We have already
demonstrated such an analysis to explain the demographic patterns of Serengeti
wildebeest. In sifting through the possible candidate explanatory models, informa-
tion theory (likelihood, AIC, and the like) provides a powerful set of tools for data
analysis. This process is known as passive adaptive management, because it relies on
natural variability to expose the underlying relationships. Without wide variation in
wildebeest abundance, for example, we would have had great difficulty discriminat-
ing among the three subtle, but dynamically different, classes of density-dependent
model that might apply.
The variation in numbers of wildebeest occurred through a fortuitous accident, the
unintentional elimination of rinderpest. This event immediately suggests a radically
different approach: manipulative experimentation with the express intention of
clarifying our understanding of the system. Such intentional experimentation at the
ecosystem level is known as active adaptive management. In principle, all of the
attributes of good experimental design (Chapter 16) should be incorporated: use
of controls versus experimental treatments, replication, and a factorial design to
identify possible interactions among processes. In practice, however, it is difficult
to implement an ideal experimental design at the enormous spatial scale at which
wildlife management policy is typically conducted. Consequently, these difficulties
argue against using conventional inferential statistics to evaluate the outcome of active
adaptive management, such as the ANOVAdesigns described in Chapter 16. Rather,
information theory and Bayesian methods of analysis offer a more practical toolkit,
based on the objective of finding the best explanatory model among many possible
models, rather than definitively rejecting a null model. Active adaptive management
policies trade off the short-term goal to maximize some output, such as a harvest,
against the long-term goal of gaining greater understanding of important ecological,
physical, or social processes. In that sense, active adaptive management is rather
like industrial research and development: reinvesting current revenue to enhance
future profits.264 Chapter 15