Date: Fri, 13 Apr 2007 09:02:30 +0200
Reply-To: Johanna Lepeule <jlepeule@GMAIL.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Johanna Lepeule <jlepeule@GMAIL.COM>
Subject: Re: Fwd: error message Nlmixed
In-Reply-To: <828381.65959.qm@web32215.mail.mud.yahoo.com>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Hi Dale,
Thank you for your answer !
Johanna
2007/4/12, Dale McLerran <stringplayer_2@yahoo.com>:
>
> Johanna,
>
> Yes, missing data is not gracefully handled for you by the NLMIXED
> procedure, unlike most other SAS procedures. You need to exclude
> data with missing values from entering into the NLMIXED procedure.
>
> Dale
>
>
> --- Johanna Lepeule <jlepeule@GMAIL.COM> wrote:
>
> > ---------- Forwarded message ----------
> > From: Johanna Lepeule <jlepeule@gmail.com>
> > Date: 11 avr. 2007 12:31
> > Subject: Re: error message Nlmixed
> > To: SAS-L@listserv.uga.edu
> >
> > Hi,
> >
> > I have posted this message in november of last year.
> >
> > I tested what was suggested by Dale.
> >
> > Finally, I realized that my problem was missing data in explicative
> > variables. In fact, I deleted all missing data and my model converged
> > without any problem.
> > I did not know that missing data could be a problem for Nlmixed...?!
> >
> > Is somebody already saw that ?
> >
> > Thank You for your answers
> >
> > Johanna
> >
> >
> > --- Johanna LEPEULE <lepeule@VET-NANTES.FR> wrote:
> >
> > > Hi,
> > >
> > > I have an error message with this proc NLMIXED :
> > > Does anyone have suggestions ?
> > >
> > > proc nlmixed data=work.sr6brutbin01 tech=quanew itdetails cov;
> > > parms beta0=-1.183 beta1=0.8159 beta2=0.7847 beta3=1.115
> > beta4=0.4867
> > > beta5=0.2137 beta6=0.7075 beta7=0.6242 beta8=1.5296 beta9=0.7343
> > > beta10=0.1479 beta11=0.6167 beta12=1.1004 beta13=0.8128
> > beta14=0.2717
> > > beta15=1.3167 beta16=-1.134 beta17=0.6624 beta18=0.48 beta19=0.2944
> > > sigma=1;
> > >
> > > y=beta0 + beta1*HG30j_corr_quarti2 + beta2*HG30j_corr_quarti3+
> > > beta3*HG30j_corr_quarti4 +
> > > beta4*racei1+ beta5*racei2 + beta6*pentehg_corr_2cla +
> > > beta7*moyconcGL_quarti2 +
> > > beta8*moyconcGL_quarti3 + beta9*moyconcGL_quarti4 +
> > > beta10*prof_zncui1+
> > > beta11*prof_zncui3+
> > > beta12*prof_zncui4+ beta13*primipare + beta14*agemerei0
> > > +beta15*agemerei2 +
> > > beta16*duree1+ beta17*prof_madccai1 +beta18*prof_madccai2
> > > +beta19*prof_madccai3 +u;
> > >
> > > expy=exp(y);
> > > p=expy/(1+expy);
> > >
> > > model sr6net_bin01 ~ binary(p);
> > >
> > > random u ~ normal(0,sigma)
> > > subject=haras1;
> > > run;
> > >
> > > WARNING: Gaussian integration failed for subject 1 during iteration
> > > -1.
> > > WARNING: Gaussian integration failed for subject 1 during iteration
> > > -1.
> > > WARNING: Gaussian integration failed for subject 1 during iteration
> > > -1.
> > > WARNING: Gaussian integration failed for subject 1 during iteration
> > > -1.
> > > WARNING: Gaussian integration failed for subject 1 during iteration
> > > -1.
> > > WARNING: Gaussian integration failed for subject 1 during iteration
> > > -1.
> > > WARNING: Gaussian integration failed for subject 1 during iteration
> > > -1.
> > > WARNING: Gaussian integration failed for subject 1 during iteration
> > > -1.
> > > ERROR: Quadrature accuracy of 0.000100 could not be achieved with
> > 31
> > > points. The achieved accuracy was 1.000000.
> > >
> > > Thanks,
> > > Johanna Lepeule
> > >
> >
> > Johanna,
> >
> > I cannot tell you right offhand why you are getting the warning/
> > error messages for your model. However, I'll tell you the secret
> > to discovering for yourself what is giving rise to your estimation
> > difficulties. But before I tell you this secret, you have to swear
> > by your firstborn not to share this secret with anyone else.
> >
> > OK, now that you have forsworn your firstborn, here is the secret.
> > Simplify, simplify, simplify!
> >
> > So, "What can I simplify?" you ask? Well, your model specifies
> > 21 parameters (20 fixed effect parameters plus a variance). Start
> > by fitting just a mean model without any random effects. So, try
> >
> > proc nlmixed data=work.sr6brutbin01 tech=quanew itdetails cov;
> > y = b0;
> > expy = exp(y);
> > p = expy / (1 + expy);
> > model sr6net_bin01 ~ binary(p);
> > run;
> >
> >
> > Does this run? Does it provide results which are consistent with
> > standard procedures for fitting a fixed effect logistic regression
> > model?
> >
> > If the answer is "No" to either of these questions, then you need
> > to back up even further. What might be wrong with the model
> > presented above? Well, what if your response (SR6NET_BIN01) is
> > not coded 0/1 but rather coded 1/2? The procedures LOGISTIC and
> > GENMOD easily handle a binary response variable coded as 1/2.
> > When the NLMIXED procedure is presented with such data, it will
> > produce an error message stating that "Quadrature accuracy of xxxx
> > could not be achieved with 31 points. The achieved accuracy was
> > 1.0000."
> >
> > There may be another area to examine if the simplified code I
> > present above does not work. I would note that you have specified
> > that the linear combination beta0 + beta1*HG30j_corr_quarti2...
> > is assigned to a variable named Y. That is probably not a good
> > choice. Suppose that Y is used elsewhere (by you, perhaps?).
> > Who knows what problems that might present? The statistical
> > literature usually uses the term ETA to hold the linear combination
> > of the parameters in your model. You might try rewriting your
> > code using
> >
> > proc nlmixed data=work.sr6brutbin01 tech=quanew itdetails cov;
> > eta = b0;
> > expeta = exp(eta);
> > p = expeta / (1 + expeta);
> > model sr6net_bin01 ~ binary(p);
> > run;
> >
> >
> > OK, so now you have the simple fixed effect model functioning.
> > At this point, you can start adding terms into your model. You
> > might try adding your random effect to the mean model. Any problems?
> > Is it possible that the response in a number of clusters is uniformly
> > 0 or uniformly 1? If that is the case, you may not be able to
> > properly estimate the between cluster variance.
> >
> > You passed that test? OK, then add in fixed effects one at a
> > time. At some point, the problems you have previously encountered
> > will arise. Try to isolate which variable (or combination of
> > variables) gives rise to your estimation difficulties. Also, you
> > might examine whether you can estimate the full fixed effect model
> > without the random effects.
> >
> > If your problems arise because of the addition of some variable(s)
> > to your fixed effect model, then the error and warning messages
> > are probably telling you that you don't have good data for
> > estimating all the parameters of your model or you are starting
> > your estimation process from some parameters which are really
> > badly chosen. Often, these are more or less the same issue. If
> > you have minimal data for estimating a particular model, then
> > you may have great difficulty finding initial parameter values
> > which allow the model to converge.
> >
> > Good luck. Let us know if you are able to resolve the problem.
> >
> > Dale
> >
> >
> > ---------------------------------------
> > Dale McLerran
> > Fred Hutchinson Cancer Research Center
> > mailto: dmclerra@NO_SPAMfhcrc.org
> > Ph: (206) 667-2926
> > Fax: (206) 667-5977
> > ---------------------------------------
> >
> >
> >
> >
> >
>
> ____________________________________________________________________________________
> > Sponsored Link
> >
> > Try Netflix today! With plans starting at only $5.99 a month what are
> > you
> > waiting for?
> > http://www.netflix.com/Signup?mqso=80010030
> >
>
>
> ---------------------------------------
> Dale McLerran
> Fred Hutchinson Cancer Research Center
> mailto: dmclerra@NO_SPAMfhcrc.org
> Ph: (206) 667-2926
> Fax: (206) 667-5977
> ---------------------------------------
>
>
>
>
> ____________________________________________________________________________________
> No need to miss a message. Get email on-the-go
> with Yahoo! Mail for Mobile. Get started.
> http://mobile.yahoo.com/mail
>
| 
