Dear R-helpers
I have generated a suite of GLMs. To select the best model for each set, I am
using the
meta-analysis approach of de Luna and Skouras (Scand J Statist 30:113-128).
Simply
put, I am calculating AIC, AICc, BIC, etc., and then using whichever criterion
minimizes APE (Accumulated Prediction Error from cross-validations on all model
sets)
to select models.
My problem arises where I have noticed my rankings from BIC and AICc are exactly
inverse. I fear this behaviour is a result of my coding as follows:
I calculate BIC from sample size:
stepAIC (mymodel.glm, k=log(n))
I then calculate AICc by:
stepAIC (mymodel.glm,
k=2*sum(mymodel.glm$prior.weights)/(sum(mymodel$prior.weights) -
length(coef(mymodel.glm))-1)).
I base these calculations for:
BIC on Venables and Ripley's MASS ("...Only k=2 gives the genuine AIC:
k = log(n) is
sometimes referred to as BIC or SBC."...)
; and for AICc from that AICc = AIC + ((2K*(K+1))/(n-K-1))
Is this behaviour expected? Or is the coding off? I could find no reference to
this
problem in the archives here, nor at S-news.
Cheers,
Joe
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Joseph J. Nocera
Ph.D. Candidate
Biology Department - Univ. New Brunswick
Fredericton, NB
Canada E3B 6E1
tel: (902) 679-5733
