Ecological Archives E096-274-A2
P. S. Petraitis and S. R. Dudgeon. 2015. Variation in recruitment and the establishment of alternative community states. Ecology 96:3186–3196. http://dx.doi.org/10.1890/14-2107.1
Appendix B. Description of Burnham and Anderson’s approach for multimodel inference.
Burnham and Anderson’s (2002) approach relies on AIC differences, Akaike weights and evidence ratios to evaluate a group of competing models. AIC difference or Δi is the Akaike Information Criteria (AIC) for the ith model minus smallest AIC; this is a relative measure and models with Δi < 2 have substantial support. Akaike weights, wi, are the relative likelihood that the ith model is the best supported model given the data and the specified group of competing models. By definition, Δi = 0 for the best supported model and Σwi =1. Evidence ratio is the comparison of two competing models and is wi / wj. This gives the relative odds of one model against another. For example, a ratio of six would suggest model i is six times more likely to be the best model than model j. Ideally, one would like to find a single model for which Δi < 2 and all evidence ratios are much larger than 1. In practice, usually a subset of models are equally plausible and when this occurred, we used model averaging in which calculation of the average parameter estimate is weighted by wi.
Literature cited
Burnham, K. P., and D. R. Anderson. 2002. Model Selection and Multi-Model Inference: A Practical Information-Theoretic Approach. Second edition. Springer.