[R-sig-ME] [Lme4-authors] AIC calculation

[Ben Bolker]

   [forwarding to r-sig-mixed-models]

On 13-10-15 04:44 AM, Miguel Boubeta Martínez wrote:
> Hello,
> The lastest versions were 0.999999-2 and 1.0-4. I fit a GLMM, as
>   glmm=glmer(y~ x + (1|dd),family=poisson())
> where 'y' is the response, 'x' is the covariate and 'dd' is the
> area. With these options, the AIC results were different. Could you give
> me the reference of the calculation of AIC used in lme4?
> Sincerely,z``
>   Miguel.


  From the NEWS file for lme4, e.g.

> news(grepl("consistent",Text) & Version=="1.0-0",package="lme4")


: As another side effect of matching glm behaviour, reported
log-likelihoods from glmer models are no longer consistent with those
from pre-1.0 lme4, but _are_ consistent with glm; see glmer examples.

  The AIC calculation (given the log-likelihood) is completely standard:
it's the log-likelihood calculation you have to be careful about -- see
lme4:::logLik.merMod (which also specifies the computation of the number
of parameters).

  Ben Bolker




z``>
> ------------------------------------------------------------------------
> *De:* lme4 maintainer <bbolker at gmail.com>
> *Para:* Miguel Boubeta Martínez <miguel.boubeta at yahoo.es>;
> "lme4-authors at lists.r-forge.r-project.org"
> <lme4-authors at lists.r-forge.r-project.org>
> *Enviado:* Martes 15 de octubre de 2013 5:03
> *Asunto:* Re: [Lme4-authors] AIC calculation
> 
>   Certainly not without a reproducible example, or at least a lot more
> details (what were the "latest versions", i.e. are you referring to
> versions 0.999x and 1.0x , or 1.0-4 and 1.1-0 ?)  Was there anything
> else different about the fits (parameter values, log-likelihood)? Was
> this a GLMM or a LMM? etc etc etc.
> 
>   Our recommended advice is to send general questions first to
> r-sig-mixed-models at r-project.org
> <mailto:r-sig-mixed-models at r-project.org> , where someone else may be
> able to
> answer.  Please read http://tinyurl.com/reproducible-000 on how to
> create good reproducible examples, and the posting guide for the R
> lists, first (google "R mailing list posting guide")
> 
>   sincerely
>     Ben Bolker
> 
> 
> 
>