[R-sig-ME] Variance explained by random factor

Douglas Bates bates at stat.wisc.edu
Wed Aug 27 00:02:01 CEST 2008

On Tue, Aug 26, 2008 at 9:54 AM, Ben Bolker <bolker at zoology.ufl.edu> wrote:
> Douglas Bates wrote:
>> Yesterday I was looking at what is done in the glm function.  Several
>> glm families (the non-"quasi" ones) have a member function to evaluate
>> the AIC then use the AIC value to determine the log-likelihood.  I'm
>> not sure how that will extend to the generalized linear mixed model.
>> I think I will be able to reconstruct a log-likelihood with the
>> appropriate scaling factors for the binomial, poisson, gaussian and
>> Gamma families.  I don't know about the inverse Gaussian and my guess
>> is that returning NA for the log-likelihood of the quasi families is a
>> sensible approach.  Ben mentioned the QAIC definition which may be
>> reasonable.
>  Does it seem backwards to anyone else that GLM etc. stores the
> AIC and then reconstructs the number of parameters in order to
> calculate the log-likelihood (rather than, e.g., storing the
> log-likelihood and the number of parameters)?

It does indeed seem a bit backwards.  I'm not sure why the decision to
do it that way was made.  I haven't played around with GLMs that much
until I started working on GLMMs.

More information about the R-sig-mixed-models mailing list