Has there been any follow up to this question? I have found myself wondering the same thing: How then does SAS fit a beta distributed GLMM? It also fits the negative binomial distribution. Both of these would be useful in glmer/lmer if they aren't 'illegal' as Brian suggested. Especially as SAS indicates a favorable delta BIC of over 1000 when I fit the beta to my data (could be the beginning of a great song..) versus my original binomial fit. Jeff Evans Michigan State University
generalized linear mixed models with a beta distribution
5 messages · Jeff Evans, Douglas Bates, dave fournier +1 more
On Thu, Feb 26, 2009 at 12:04 PM, Jeff Evans <evansj18 at msu.edu> wrote:
Has there been any follow up to this question? I have found myself wondering the same thing: How then does SAS fit a beta distributed GLMM? It also fits the negative binomial distribution.
When SAS decides to open-source their code we'll be able to find out.
Both of these would be useful in glmer/lmer if they aren't 'illegal' as Brian suggested. Especially as SAS indicates a favorable delta BIC of over 1000 when I fit the beta to my data (could be the beginning of a great song..) versus my original binomial fit.
Definitions of generalized linear mixed models are not entirely straightforward, at least for me. I'm making some progress but, as always, it is slower than one would like it to be.
Thanks for responding Doug. I'm sure SAS just hasn't gotten around to releasing their code yet. lme4 does have a leg up on GLIMMIX in other areas, though. The latest SAS release (9.2) is now able to compute the Laplace approximation of the likelihood, but it will only fit an overdispersion parameter using pseudo-likelihoods which can't be used for model selection. I'm not sure what lme4 is doing differently through the quasi-distributions that allows this, but it's enormously useful. Jeff -----Original Message----- From: dmbates at gmail.com [mailto:dmbates at gmail.com] On Behalf Of Douglas Bates Sent: Thursday, February 26, 2009 3:50 PM To: Jeff Evans Cc: r-help at r-project.org Subject: Re: [R] generalized linear mixed models with a beta distribution
On Thu, Feb 26, 2009 at 12:04 PM, Jeff Evans <evansj18 at msu.edu> wrote:
Has there been any follow up to this question? I have found myself
wondering
the same thing: How then does SAS fit a beta distributed GLMM? It also
fits
the negative binomial distribution.
When SAS decides to open-source their code we'll be able to find out.
Both of these would be useful in glmer/lmer if they aren't 'illegal' as Brian suggested. Especially as SAS indicates a favorable delta BIC of over 1000 when I fit the beta to my data (could be the beginning of a great song..) versus my original binomial fit.
Definitions of generalized linear mixed models are not entirely straightforward, at least for me. I'm making some progress but, as always, it is slower than one would like it to be.
You can fit this kind of model (and negative binomial) and more
difficult mixed models with AD Model Builder's random effects module
which is now freely available at
http://admb-project.org/
David A. Fournier P.O. Box 2040, Sidney, B.C. V8l 3S3 Canada Phone/FAX 250-655-3364 http://otter-rsch.com
Jeff Evans-5 wrote:
lme4 does have a leg up on GLIMMIX in other areas, though. The latest SAS release (9.2) is now able to compute the Laplace approximation of the likelihood, but it will only fit an overdispersion parameter using pseudo-likelihoods which can't be used for model selection. I'm not sure what lme4 is doing differently through the quasi-distributions that allows this, but it's enormously useful. Jeff
Sorry, but I wouldn't necessarily take comfort from this. I must confess that I can't keep the distinctions between marginal pseudo/quasi-likelihoods in my head, but on what grounds are you confident that the number that lme4 produces can be used for model selection? (I would guess that) Some people would be happy using QAIC based on pseudo-likelihoods, some people wouldn't be happy with anything other than a true likelihood (or approximation thereof). This discussion is probably better for r-sig-mixed-models ... Ben Bolker
View this message in context: http://www.nabble.com/generalized-linear-mixed-models-with-a-beta-distribution-tp22230099p22238830.html Sent from the R help mailing list archive at Nabble.com.