Nicky
AGQUAD can get as close as you want to the likelihood by specifying the
number of quadrature points. This is the method used in NLMIXED, which you
had referred to in one of your emails. The LAPLACE approximation is
equivalent to AGQUAD with a single quadrature point.
AIC can be used for model comparisons quite comfortably for AGQUAD for
example. For nested models you might want to use LRTs though.
AIC is a bit of an approximation in the case of PQL, especially in certain
binary data situations. Wolfinger and I did quite extensive simulations
comparing PQL to quadrature and behavior was consistently quite good in the
case of Poisson with log link. Anyway, you can use LRTs and AGQUAD where you
can if you are concerned...
Some of the literature eg following Lin and Breslow (96), focusses a lot on
some negatives of PQL in certain binary data situations. However, PQL is nice
functionality in most other situations eg Poisson log-link for example. Im
glad we have it in glme and I think SAS has now followed suit in PROC GLIMMIX
(they didn't have it in PROC NLMIXED).
This topic is an interesting one for a longer conversation and Id be happy to
take it off line with you.
Best regards
Michael
-----Original Message-----
From: Nicola Koper [mailto:nkoper@yahoo.com]
Sent: Friday, November 02, 2007 1:35 PM
To: Michael O'Connell; A&B Penner; s-news@lists.biostat.wustl.edu
Cc: Meg Krawchuk
Subject: RE: [S] glme for count data...
Hi,
Someone correct me if I am wrong, but I am sure that AIC CANNOT be used with
those methods, because none of them are based on a true log-likelihood; all
the likelihoods are approximate (e.g. penalized QUASI-likelihood, because
it's not a true likelihood, Laplacian APPROXIMATION to the likelihood, etc).
I see Michael is from insightful so maybe he knows more than I do! But, I was
definately explicitly taught that you cannot use AIC with any of these.
If that is incorrect, could you direct me to a publication (NOT from Splus or
insightful; instead, an independent text or journal publication) that
demonstrates that you can use AIC with these approximate methods?
Thanks,
Nicky
--- Michael O'Connell <moconnell@insightful.com>
wrote:
> For Poisson model with log link (which is probably ok for these data)
> there are several 'likelihood' methods available for
> glme() ie
> 1) PQL - penalized quasi-likelihood.
> 2) MQL - (restricted) marginal quasi-likelihood.
> 3) AGQUAD - adaptive gaussian quadrature
> 4) LAPLACE - Laplacian approximation to the likelihood.
> The number of quadrature points can be specified through the control
> formal argument.
>
> For Poisson, we've found PQL to work quite well. For binary data, you
> should use AGQUAD or LAPLACE in most situations. You can use AIC with
> the above methods.
>
> See the glme() help file and user guide for more information.
>
> For references, you can see Breslow and Clayton (93) or Wolfinger and
> O'Connell (93) for some early examples. The book by Geert Verbeke and
> Geert Molenberghs is quite good, but a little dated now
> (2001)
>
http://www.amazon.com/Linear-Mixed-Models-Longitudinal-Data/dp/0387950273
>
> Michael
>
> -----Original Message-----
> From: s-news-owner@lists.biostat.wustl.edu
> [mailto:s-news-owner@lists.biostat.wustl.edu] On Behalf Of Nicola
> Koper
> Sent: Thursday, November 01, 2007 4:43 PM
> To: A&B Penner; s-news@lists.biostat.wustl.edu
> Subject: Re: [S] glme for count data...
>
> Direct comparison among models can be problematic, so you may not like
> this reply at all. You can really only use AIC if you use maximum
> likelihood estimation, which is theoretically possible with GLME but I
> don't think you can do it in Splus at all. I have turned to R
> (glmmML) or NLMIXED in SAS to
> do so, in the past.
>
> I'll be interested in whether other people know of other methods to do
> this with glme.
>
> I really like Applied Longitudinal Analysis 2004 by Fitzmaurice et al.
> I find it more intuitive than some of the other texts that talk about
> GLME.
>
> Nicky
> --- A&B Penner <pennerab@gmail.com> wrote:
>
> > Thank you to everyone who replied to my question:
> I will use the
> > glme() function from the correlatedData library to
> perform a
> > mixed-effects model on my count data.
> >
> > This brings up a new question. Is there a way to
> compare overall glme
> > model fits when the models have different
> fixed-effect terms? For
> > example, lme models estimated by maximum
> likelihood can be compared
> > using a drop-in-deviance test performed by the
> anova function. Is
> > there anything similar for glme models?
> >
> > Additionally, besides Venables and Ripley 2002 and
> the help file for
> > the correlatedData library, are there other basic
> references for
> > performing and interpreting generalized linear
> mixed models?
> >
> > Many thanks,
> > A. Penner
> >
> > On 10/29/07, Michael O'Connell
> > <moconnell@insightful.com> wrote:
> > >
> > > yes, the correlatedData library and associated
> > glme() function postdate
> > > the Pinheiro and Bates text.
> > >
> > > gee() in the correlatedData library will also
> work
> > for these data. I
> > > prefer the glme() implementation in this case.
> > >
> > > Michael
> > >
> > > ------------------------------
> > > *From:* s-news-owner@lists.biostat.wustl.edu
> > [mailto:
> > > s-news-owner@lists.biostat.wustl.edu] *On Behalf
> > Of *Austin, Matt
> > > *Sent:* Thursday, October 25, 2007 8:14 PM
> > > *To:* A&B Penner; s-news@lists.biostat.wustl.edu
> > > *Subject:* Re: [S] lme for count data?
> > >
> > > In the past I would have suggested looking at
> > glmmPQL() in the MASS
> > > library or glme() in the correlatedData library
> if
> > you have grouped count
> > > data, but I can't find the correlatedData
> library
> > in version 8.
> > >
> > >
> > >
> > > Another alternative would be a GEE approach.
> > >
> > >
> > >
> > > The MASS text by Venables and Ripley gives a
> short
> > overview of both
> > > methods (PQL and GEE) and good references for
> > deeper study. It also gives a
> > > brief explanation of the difference in
> > interpretation between the methods.
> > > I believe the correlatedData library and
> > associated glme() function postdate
> > > the Pinheiro and Bates text.
> > >
> > >
> > >
> > >
> > >
> > > --Matt
> > >
> > >
> > > ------------------------------
> > >
> > > *From:* s-news-owner@lists.biostat.wustl.edu
> > [mailto:
> > > s-news-owner@lists.biostat.wustl.edu] *On Behalf
> > Of *A&B Penner
> > > *Sent:* Thursday, October 25, 2007 1:50 PM
> > > *To:* s-news@lists.biostat.wustl.edu
> > > *Subject:* [S] lme for count data?
> > >
> > >
> > >
> > > Hi,
> > >
> > > I'd like to fit a mixed-effects model to my
> data.
> > Pinheiro and Bates
> > > (2000) write that "These models are intended for
> > grouped data in which the
> > > response variable is (at least approximately)
> > continuous." My response
> > > variable is count data with range from 0-4, mean
> > 1.7, median 2.0, stdev
> > > 0.8, and an approximately normal distribution.
> Can
> > I use the lme function
> > > for this data?
> > >
> > > Thank you,
> > > A. Penner
> > >
> >
>
>
>
>
> __________________________________________________
> Do You Yahoo!?
> Tired of spam? Yahoo! Mail has the best spam protection around
> http://mail.yahoo.com
>
--------------------------------------------------------------------
> This message was distributed by
> s-news@lists.biostat.wustl.edu. To
> unsubscribe send e-mail to
> s-news-request@lists.biostat.wustl.edu with the BODY of the message:
> unsubscribe s-news
>
__________________________________________________
Do You Yahoo!?
Tired of spam? Yahoo! Mail has the best spam protection around
http://mail.yahoo.com
|