Date:         Sat, 12 Jan 2008 06:14:38 -0800
Reply-To:     Oliver.Kuss@MEDIZIN.UNI-HALLE.DE
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Oliver.Kuss@MEDIZIN.UNI-HALLE.DE
Organization: http://groups.google.com
Subject:      Re: Pseudo-AIC & Glimmix
Comments: To: sas-l@uga.edu
Content-Type: text/plain; charset=ISO-8859-1

Hello Tobias,
Robin and Dale are right, you should not use the pseudo-AIC in proc
glimmix to compare models, and they explained very well why this is
the case.
However, if your models are not too complicated and you can code them
in proc nlmixed, then you could use the AIC values from proc nlmixed,
even for non-nested models. Moreover, if your models are nested, you
wouldn't even need the AIC and you can do simple LR tests for model
comparison.

Hope that helps,
Oliver

On 11 Jan., 16:37, tobias.jepps...@EKOL.SLU.SE (Tobias Jeppsson)
wrote:
> Hi list
>
> This is my first post here.
>
> I'm using proc glimmix and is having problems with model selection, and the
> reported Fit statistics. My initial idea was to use AIC to choose between
> competing models, but after I read the (short) info on pseudo-AIC I feel
> more hesitiant.
>
> The SAS printout states:
> "Fit statistics based on pseudo-likelihoods are not useful for comparing
> models that differ in their pseudo-data."
>
> And in the glimmix help file you can read:
> "Note that the (residual) log pseudo-likelihood in a GLMM is the (residual)
> log likelihood of a linearized model. You should not compare these values
> across different statistical models, even if the models are nested with
> respect to fixed and/or G-side random effects. It is possible that between
> two nested models the larger model has a smaller pseudo-likelihood."
>
> My interpretation is that you cannot use the pseudo-AIC to compare e.g. two
> nested models:
> model resp = k range;
> model resp = k|range;
>
> I've tried to find more information on the pseudo-AIC, but without much
> results. Is my interpretation correct? If so, can the pseudo-AIC still be
> used to evaluate and choose between different correlation structures in the
> residuals (e.g. sp(exp) vs AR(1))?
>
> I would be very thankful for all comments.