Hi All
I have a question regarding the merits of AIC vs. BIC as a criterion for model selection.
Is one more preferable than the other?
I understand that BIC favours models with fewer parameters from the formulas
AIC = -2*loglik + 2*p
BIC = -2*loglik + log(n)*p
where n= #observations in my case 5000 and p = #parameters
I am performing an ordinal logistic regression in R and use the polr() function to fit a model with 30 predictors. I then run the stepAIC() function using AIC (k=2) and my final model has 14 variables. If I run stepAIC with BIC (k=log(n)) my final model has 5 variables.
So given such a difference in the final models which is better to use AIC or BIC? I understand that one should look beyond the AIC and BIC to the actual models selected and if they make sense etc... However I am just curious as to the large difference and wondered if anybody can provide insights, tips, hints etc...
Best Regards,
Mary
---------------------------------
Yahoo! Messenger with Voice. Make PC-to-Phone Calls to the US (and 30+ countries) for 2¢/min or less.
|