1
$\begingroup$

I have a question regarding the way BIC, AIC and ICL are computed in the packages mclust, HDclassif and fisherEM. Both of these packages use the negation of AIC, BIC, ICL (bigger is better). This is alright for in-package model comparison, but if I want to compare models between those packages (regarding the same data), I don't know if the values are comparable.

Thanks in advance!

$\endgroup$
2

1 Answer 1

1
$\begingroup$

By going through it manually I found a partial solution to my question. Apparently mclust and the hddc() function of HDclassif both use the known bigger is better variant of BIC: 2 * loglik - nparams * log(n). Not so with sfem() in FisherEM. Due to it being a penalized approach to subspace clustering, there seems to be a penalty term in the computation of BIC. I don't know what kind. I still don't know if this penalized BIC is comparable to the common kind used by mclust and HDclassif. I would think so, but be astounded, because my sfem() model should be much sparser than the other ones.

EDIT: ...and I seem to have found the rest of the solution, here: stats.stackexchange.com/questions/25817/… . sfem() uses the lasso penalty and the LARS algorithm, so the yielded corrected BIC seems to be comparable to the normal one.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.