Prev: create a file using filename FTP method
Next: SAS EG Graph problem??
From: Brian Sauer on 8 Feb 2010 16:42
On Feb 8, 10:26 am, stringplaye...(a)YAHOO.COM (Dale McLerran) wrote:
> --- On Sun, 2/7/10, Brian Sauer <brian.sa...(a)GMAIL.COM> wrote:
>
>
>
> > From: Brian Sauer <brian.sa...(a)GMAIL.COM>
> > Subject: Re: proc logistic: 'out of memory'
> > To: SA...(a)LISTSERV.UGA.EDU
> > Date: Sunday, February 7, 2010, 4:47 PM
> > I am in the process of testing the NLMIXED technique that Dale
> > mentioned for my problem, but I wanted to share the answer I received
> > from SAS support.
> > Brian,
>
> > The problem is caused by the default check for dependencies between
> > strata and the predictors in SAS 9.1.  This can require a large amount
> > of memory when there are many strata.  You can turn off this check by
> > specifying the NOLINDEP option in the STRATA statement.  For example:
>
> >    strata strata / nonlindep;
>
> > This check is off by default in all the current SAS 9.2 releases.
>
> > ----
> > NOTE: If you have any follow-up questions on this matter, please
> >            reply to this email by February 10, 2010.
>
> > - - - - - - - - -
> > David Schlotzhauer
> >   Phone: (919) 677-8008
> > Senior Statistical Consultant   Web:
> > support.sas.com/ts
> > SAS Institute Inc.
>
> OK, that is useful to know.  However, the syntax specified above
> does not match syntax in documentation of version 9.1.3 or 9.2.
> The 9.2 documentation indicates the following syntax which I
> would think would be the official syntax:
>
> strata ... / CHECKDEPENDENCY=NONE;
>
> Neither the CHECKDEPENDENCY option nor the NONLINEDEP option is
> indicated in version 9.1.3 documentation.  Have you tried the
> syntax with NONLINDEP specified as an option in one (or both)
> of these versions?
>
> Just as an aside, I am sure that it is reasonable to perform
> this sort of linear dependency check when the number of strata
> is small.  However, it would seem that SI might have implemented
> the linear dependency checking with an initial determination of
> the number of strata in the data set.  If the number of strata
> is above a certain level (say, 1000), then the linear dependency
> checking would not be enforced (under the assumption that given
> a large number of strata, it is not likely that one would
> encounter complete linear dependency between covariates and
> strata.  But from the standpoint of a computer programmer, I
> can see where it is always good to evaluate whether the data
> conform to the requirements of the estimation procedure.
>
> Dale
>
> ---------------------------------------
> Dale McLerran
> Fred Hutchinson Cancer Research Center
> mailto: dmclerra(a)NO_SPAMfhcrc.org
> Ph:  (206) 667-2926
> Fax: (206) 667-5977
> ---------------------------------------

Dale,
You are correct, the NONLINDEP causes an error. I am testing default
and CHECKDEPENDENCY = NON now

Thanks guys,
Brian
From: oloolo on 8 Feb 2010 18:32
Very nice and elaborated interpretation!!!

Verbeke once (2001, The American Statistican) emphasised this conditional
regression approach in linear mixed model and draw its connection to
conditional logistic regression.


On Mon, 8 Feb 2010 10:29:46 -0800, Dale McLerran <stringplayer_2(a)YAHOO.COM>
wrote:
out of the model. Thus, we cannot
>estimate the probability of the outcome for the j-th
>observation from the i-th stratum.
>
>I have noted before - and Oliver discusses this as well with
>a little different twist from how I typically present the case -
>that the random effects model and the GEE model do estimate
>different effects. The GEE model estimates marginal or population
>average estimates of effects. If you want to know for policy
>purposes what the average effect estimate will be in a population,
>then you would prefer to present estimates obtained from a model
>estimated employing GEE (or similar) effect estimation. However,
>if you want to know what is the benefit to the i-th subject,
>you want a conditional or subject-specific effect estimate.
>
>As for the example of the stratified conditional analysis which
>included gender in the CLASS statement but nowhere else, the
>inclusion of gender in the class statement was no doubt residual
>from a full model in which observation-specific effects (possibly
>including age as a continuous effect) were incorporated into the
>model. As long as every person has a gender specification, then
>naming gender on the CLASS statement would not affect the
>estimates obtained for the MODEL statement which was specified.
>
>HTH,
>
>Dale
>
>---------------------------------------
>Dale McLerran
>Fred Hutchinson Cancer Research Center
>mailto: dmclerra(a)NO_SPAMfhcrc.org
>Ph: (206) 667-2926
>Fax: (206) 667-5977
>---------------------------------------
From: Ryan on 9 Feb 2010 09:35
On Feb 8, 1:29 pm, stringplaye...(a)YAHOO.COM (Dale McLerran) wrote:
> --- On Sat, 2/6/10, Ryan <ryan.andrew.bl...(a)GMAIL.COM> wrote:
>
>
>
>
>
> > From: Ryan <ryan.andrew.bl...(a)GMAIL.COM>
> > Subject: Re: proc logistic: 'out of memory'
> > To: SA...(a)LISTSERV.UGA.EDU
> > Date: Saturday, February 6, 2010, 7:14 PM
> > On Jan 7, 2:31 pm, stringplaye...(a)YAHOO.COM
>
> > Dale,
>
> > I'm curious if and when it is preferable to run a conditional logistic
> > regression model in a 1:1 matching design rather than a GEE (pop.
> > average) or perhaps even a generalized linear mixed model (subj.
> > specific). I read a conditional logistic regression example online
> > (http://www.biostat.umn.edu/~will/6470stuff/Lect21/lecture21H.pdf),
> > where 2 patients from each of the 79 participating clinics were
> > enrolled in the study. Within each clinic, one of the patients was
> > assigned to the treatment condition while the other was assigned to
> > the control condition. I suppose this would be considered a 1:1
> > matching design. The binary outcome was improve/did not improve. The
> > suggested code to run a conditional logistic regression model was:
>
> > Proc Logistic descending ;
> > class center treatment(ref="P") gender(ref="F");
> > model improve = baseline_score treatment;
> > strata center;
>
> > What would be the driving force as to whether you would run a
> > conditional logistic regression, GEE or perhaps generalized linear
> > mixed model for this example? Can you think of a scenario where you
> > would prefer to use a conditional logistic regression in a 1:1
> > matching design rather than one of the other two options? It's also
> > interesting in the example code that gender is placed on the class
> > statement, yet appears no where else in the code.
>
> > Thanks,
>
> > Ryan
>
> Ryan,
>
> Let me point you to a SUGI presentation from 2002 (SUGI 27) by
> Oliver Kuss as a good read on this topic.  See
>
> http://www2.sas.com/proceedings/sugi27/p261-27.pdf
>
> Oliver demonstrates that the conditional logistic regression
> parameter estimates are similar to the parameter estimates that
> one would obtain for the mixed model fitted using NLMIXED or
> GLIMMIX.  It is interesting to note that both a stratified
> analysis and an analysis in which random effects are estimated
> in the model are both referred to as conditional logistic
> regression models.
>
> The stratified model has some advantages in that it is
> semi-nonparametric.  It is not necessary to assume that the
> subject/stratum random effects are normally distributed.
> Also, it is worth noting that even if the subject random
> effects are normally distributed, it is not necessary to
> estimate the subject random effects using the stratified
> analysis.  To the extent that Gaussian quadrature does not
> produce exact integration over the random effects, the results
> from a method which approximates the integration has some loss.
> For shared parameters, the stratified analysis may actually be
> preferred to results from a procedure which estimates the
> random effects.
>
> But the stratified analysis also has some loss in that the
> intercept is conditioned out of the model.  Thus, we cannot
> estimate the probability of the outcome for the j-th
> observation from the i-th stratum.
>
> I have noted before - and Oliver discusses this as well with
> a little different twist from how I typically present the case -
> that the random effects model and the GEE model do estimate
> different effects.  The GEE model estimates marginal or population
> average estimates of effects.  If you want to know for policy
> purposes what the average effect estimate will be in a population,
> then you would prefer to present estimates obtained from a model
> estimated employing GEE (or similar) effect estimation.  However,
> if you want to know what is the benefit to the i-th subject,
> you want a conditional or subject-specific effect estimate.
>
> As for the example of the stratified conditional analysis which
> included gender in the CLASS statement but nowhere else, the
> inclusion of gender in the class statement was no doubt residual
> from a full model in which observation-specific effects (possibly
> including age as a continuous effect) were incorporated into the
> model.  As long as every person has a gender specification, then
> naming gender on the CLASS statement would not affect the
> estimates obtained for the MODEL statement which was specified.
>
> HTH,
>
> Dale
>
> ---------------------------------------
> Dale McLerran
> Fred Hutchinson Cancer Research Center
> mailto: dmclerra(a)NO_SPAMfhcrc.org
> Ph:  (206) 667-2926
> Fax: (206) 667-5977
> ---------------------------------------- Hide quoted text -
>
> - Show quoted text -

Dale,

What you wrote is clear and very helpful. Thank you. I must say that
I'm intrigued by the concept of shared parameter models. I did a quick
google search and found an abstract on using nonlinear mixed models to
account for shared parameters in what appeared to be a typical
clinical trial parallel design. The shared parameters were: (1) event
data (i.e. drop out rate) and (2) repeated measures. I might start a
new thread at some point asking questions/sharing ideas around
developing a NLMIXED code to account for this type of situation. Quite
interesting! Anyway, I've derailed this thread enough.

Thanks again,

Ryan
First  |  Prev  | 
Pages: 1 2 3
Prev: create a file using filename FTP method
Next: SAS EG Graph problem??