> On May 28, 2017, at 11:53 PM, Brigitte Mangin <brigitte.mangin at
inra.fr> wrote:
>
> Thanks Ron,
>
> In fact, I want to make a model choice using different fixed structures and
using the results of:
> Gurka MJ (2006) Selecting the best linear mixed model under reml. The
American Statistician 60(1):19{26,
> the best criterium uses the reml likelihood.
>
> I asked the ASREML-r developpers and they answered that their results were
checked against GENSTAT.
>
> I think it is not really a good think for the R community to compute a REML
likelihood that is probably not the REML likelihood.
Is it your understanding that REML values should be different somehow than other
likelihoods with respect to the fact that you should only be comparing
_differences_ in model likelihoods calculated on the same data? The value of a
likelihood is only specified up to a constant (as Thierry Onkelinx already
pointed out.)
I can get different deviances (-2*log(likelihood) in glm poisson models by just
grouping data elements and modeling counts. But varying models will have the
same differences in deviance regardless of grouping or not.
Looking at this copy of that citation It appears to me that differences
(comparing full to reduced) in various criteria for models is what is under
discussion:
http://users.jyu.fi/~hemipu/itms/Gurka%202006,%20TAS,%20REML.pdf
You should show some results rather than letting this discussion remain so
vague.
--
David.
>
> Brigitte
>
>
>
> Brigitte Mangin, INRA, LIPM, CS 52627, 31326 CASTANET-TOLOSAN
> tel: 33 + (0)5 61 28 54 58
>
> ________________________________________
> De : Crump, Ron <R.E.Crump at warwick.ac.uk>
> Envoy? : mardi 23 mai 2017 10:29
> ? : r-help at r-project.org; Brigitte Mangin
> Objet : Re: R-help Digest, Vol 171, Issue 20
>
> Hi Brigitte,
>
>> Did somebody know why asreml does not provide the same REML loglikehood
>> as coxme, lme4 or lmne.
>
> I don't know the answer to this, but I'd guess it is either to do
with the
> use of the average information REML algorithm or asreml-r is for some
> reason ending up with a different subset of the data.
>
>> If it was just a constant value between the two models (with or without
>> the fixed effect) it would not be important. But it is not.
>> I checked that the variance component estimators were equal.
>
> I'm still not clear that it is important (if the data subset analysed
is
> the same). You would only use the REML likelihoods to compare models with
> different random effects and the same fixed effect structure (is there
> another use for the REML likelihood other than that?), so then it is
> really a question of whether for a given pair of random effect models and
> the same data the likelihood ratio test statistic changes across analysis
> methods. Unless for some reason you are comparing two random effect models
> fitted with different routines (one of which is asreml-r).
>
> Ron.
>
>
> ______________________________________________
> R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
David Winsemius
Alameda, CA, USA