The difference between AIC and BIC came up in stats beerz this week. Here’s a brief overview from Sean Anderson from several months ago:

I’m pretty sure you could write a novel on this topic. In times like that, I like to ask WWBBD or ‘what would Ben Bolker do?’

See his book, starting on page 277

Also, see this short blog post he wrote a while back:

AIC is by far the most common information criteria. Certainly in ecology. This paper by Aho et al. (2014) found that 84% of papers that used information criteria used AIC and only 14% used BIC (also a good read and more than you probably want to know).

One important thing here, besides the slightly different formulas for AIC and BIC, is that AIC and BIC try to answer different questions. AIC seeks to find the most parsimonious model. In other words, it tries to maximize predictive accuracy. BIC tries to find the ‘true’ model. In other words, if the ‘true’ model was among the models considered and as sample size increases, BIC should correctly rank the ‘true’ model as the best model. A bit hard to wrap your mind around, yes. I think ecologists may tend to focus more on what AIC aims to answer. That’s essentially Ben’s conclusion.