I think BIC is defined incorrectly
helios1014 opened this issue · 1 comments
helios1014 commented
Looking at the code, BIC is defined as n * np.log(sq_error / n) + k * np.log(n).
I have traditionally seen it defined as -2 * loglik + k * log(n)
logLik is only used in your definition of aic and so I am confused as to what the source is for the formulation of BIC you are using.
cokelaer commented
@helios1014 you right, the current version is specific to gaussian distribution but here we should use the more generic version indeed. I have changed the code and it should appear in v1.6.0