INVITED PAPER: INFORMATION THEORY IN WILDLIFE SCIENCE: CRITIQUE AND VIEWPOINT

We question whether the growing popularity of model selection based on information theory (IT) and using the Akaike's Information Criterion (AIC) represent a useful paradigm shift in data analysis or a substitution of 1 statistical ritual for another, which leaves in place long-standing problems in wildlife science. We discuss the relevance of model selection in science, problems in the IT-AIC algorithm, errors of commission and omission in IT-AIC-based studies, and the role of IT-AIC in knowledge accrual. Model selection is just another minor tool in the grand panorama of science. The human mind, not statistical methods, produces scientific breakthroughs. Although IT-AIC might include elements of hypothetico-deductive science, it is arguably a form of sensitivity analysis, magnitude of effects estimation, or simple description as currently applied. Accordingly, it is largely an inductive approach to knowledge accrual and, therefore, subject to the pitfalls of induction. The algorithm tends to over fit data (i.e, use too many variables), resulting in models that contain useless variables and that generalize poorly. Errors of commission in IT-AIC-based papers include hopelessly uninformative lists of encrypted models and imposition of the model-selection approach on studies better executed in a simple, descriptive format. The major error of omission is an almost universal failure to test selected models on independent data. From our perspective, IT-AIC is a harmless human construct that is being ritualistically applied and therefore cannot be expected to correct long-standing problems in the conduct of wildlife science, such as failure to apply the hypothetico-deductive method. We view the growing application of IT-AIC as problematic because that growth might discourage use of the full panoply of available methods of inquiry. Accordingly, we urge colleagues to avail themselves of the rich pageant of available analytical techniques that can be applied in wildlife research under the hypothetico-deductive method and to keep ecology, rather than statistics, in the forefront of wildlife science.