Methodological review demonstrates widespread misinterpretation of uninformative parameters in ecological AIC model selection, highlighting strategies to prevent false conclusions.
Key Points
To evaluate the prevalence and consequences of misinterpreting uninformative parameters in AIC-based model selection and assess strategies to resolve this problem.
Examined the mathematical behavior of Akaike's Information Criterion (AIC) penalties relative to changes in model deviance when adding extra variables.
Audited AIC-based articles published in the Journal of Wildlife Management during 2008 to quantify the frequency of misinterpreting models with ΔAIC ≤ 2.
Evaluated five statistical solutions for handling uninformative parameters across both small a priori model sets and large exploratory model sets.
Models containing uninformative parameters were incorrectly presented as competitive candidates in 72% of AIC-based papers in the Journal of Wildlife Management in 2008.
An uninformative parameter adds a +2 AIC penalty without reducing deviance sufficiently, meaning models within 2 ΔAIC units often lack meaningful ecological effects.
Reporting all models while dismissing those with uninformative parameters works best for small a priori candidate sets, while all-possible subsets or model-culling approaches suit larger sets.