Skip to main content
U.S. flag

An official website of the United States government

Truth, models, model sets, AIC, and multimodel inference: a Bayesian perspective

July 8, 2015

Statistical inference begins with viewing data as realizations of stochastic processes. Mathematical models provide partial descriptions of these processes; inference is the process of using the data to obtain a more complete description of the stochastic processes. Wildlife and ecological scientists have become increasingly concerned with the conditional nature of model-based inference: what if the model is wrong? Over the last 2 decades, Akaike's Information Criterion (AIC) has been widely and increasingly used in wildlife statistics for 2 related purposes, first for model choice and second to quantify model uncertainty. We argue that for the second of these purposes, the Bayesian paradigm provides the natural framework for describing uncertainty associated with model choice and provides the most easily communicated basis for model weighting. Moreover, Bayesian arguments provide the sole justification for interpreting model weights (including AIC weights) as coherent (mathematically self consistent) model probabilities. This interpretation requires treating the model as an exact description of the data-generating mechanism. We discuss the implications of this assumption, and conclude that more emphasis is needed on model checking to provide confidence in the quality of inference.

Publication Year 2015
Title Truth, models, model sets, AIC, and multimodel inference: a Bayesian perspective
DOI 10.1002/jwmg.890
Authors Richard J. Barker, William A. Link
Publication Type Article
Publication Subtype Journal Article
Series Title Journal of Wildlife Management
Index ID 70154802
Record Source USGS Publications Warehouse
USGS Organization Patuxent Wildlife Research Center