MRE Symonds, A Moussalli (2011) A brief guide to model selection, multimodel inference and model averaging in behavioural ecology using Akaike’s Information Criterion, Behavioral Ecology and Sociobiology 65: 13-21 PDF
What’s it about?
If you want a novice level introduction into how to carry out statistical analyses and model selection (what combination of variables best describes variation in a particular trait) using information theory (and specifically Akaike’s Information Criterion) rather than probability values, this is your paper.
What’s the story behind it?
For a paper which virtually drove Adnan and I insane writing and revising it, this has ultimately been a rewarding experience, generating almost double the number citations (in just a couple of years) of any of my other papers. It’s genesis lies in the same statistical workshop that bred the Garamszegi et al. (2009) review I was involved in. Las Garamszegi asked me to write an introductory paper for a special edition of Behavioral Ecology and Sociobiology that was going to be focussed on new statistical methods in behavioural ecology. I agreed because I felt that the literature on how to use AIC (as its friends call it) was intimidating to floudering novices (as I had been until fairly recently when I converted to floundering sophomore). The main problem we had was that the reviewers a) didn’t see the need for such a paper (“this has been all covered elsewhere”) and b) seemed to have very different ideas about what was the correct way to use AIC – which made it a real bugger to try and tow a revision line through the manuscript – especially when ultimately the paper got sent to SIX reviewers – by some distance the most I have ever had on a paper. One consequence was that from revision to revision we were told to change things that earlier reviewers had insisted on. Thus one early review urged us “the authors must make more of a case for why behavioural ecologists should used AIC”, whilst a later review took us to task for being “too uncritical of AIC, almost breathlessly so”.
Curiously one thing that seemed to be a consistent theme throughout the reviewers comments was something along the lines of ‘You’ve described this wrong…Burnham and Anderson are very clear on this point’ – referring to Burnham and Anderson seminal (2002) text on Model selection and inference (kind of the AIC bible). The funny thing was that this argument would be used to argue exactly opposite points (technical stuff to do with removing or not removing (depending on your viewpoint) poorly weighted parameters). It seemed to confirm to me (something I had felt reading their book, and which had prompted my writing of this paper), that for their many outstanding contributions to the field, clarity of writing was not one of B&A’s stronger points (just my opinion).