[R-sig-ME] (no subject)

[Douglas Bates]

On Thu, Apr 7, 2011 at 5:59 AM, Lim Yonghao <oahgony at hotmail.com> wrote:
>
> Thanks for the advice. I will definitely try out lme4a.
>
> As for the error message, is there anywhere i can find out what it actually means?

The first thing to do is to add verbose=FALSE to your call to glmer.
That provides information about the iterations when optimizing the
deviance.  You will see that you have several parameters over which
the optimization is taking place and, apparently, some of those
parameters are being driven to extreme values so that a matrix that
should be positive definite is becoming close to singular.  That is
what the warnings from the cholmod code are about.

The later warning about false convergence is from the optimizer
itself, "nlminb" in this case.  You may, in fact, have gotten to the
mle's - it's just that nlminb is very cautious about declaring a
(locally) optimal value.

Possible remedies are to simplify the structure of the model.  A term
of the form (time1_18 + time2_18|id) requires estimation of 6
variance-covariance parameters, which may be causing an unstable
estimation situation.



>> Date: Wed, 6 Apr 2011 12:44:52 -0500
>> Subject: Re: [R-sig-ME] (no subject)
>> From: bates at stat.wisc.edu
>> To: Thierry.ONKELINX at inbo.be
>> CC: oahgony at hotmail.com; r-sig-mixed-models at r-project.org
>>
>> On Wed, Apr 6, 2011 at 8:48 AM, ONKELINX, Thierry
>> <Thierry.ONKELINX at inbo.be> wrote:
>> > Dear Yonghao,
>>
>> > The problem is that your model is too complex for the dataset. So you have two options: 1) simplify the model, 2) collect more data.
>>
>> A third option is to try to fit the model in lme4a which should be
>> more robust in fitting GLMMs.  That's not a guarantee that it will
>> work but it is sometimes worth a try.
>>
>> I'll send a message in a moment about a class I am teaching in which
>> students are to do a data analysis project involving mixed-effects
>> models.  Those who would like to have a student consider their data
>> are encouraged to respond to that request.  Data confidentiality can
>> be requested.
>>
>> >> -----Oorspronkelijk bericht-----
>> >> Van: r-sig-mixed-models-bounces at r-project.org
>> >> [mailto:r-sig-mixed-models-bounces at r-project.org] Namens Lim Yonghao
>> >> Verzonden: woensdag 6 april 2011 15:39
>> >> Aan: r-sig-mixed-models at r-project.org
>> >> Onderwerp: [R-sig-ME] (no subject)
>> >>
>> >>
>> >> Dear all,
>> >>
>> >> I wish to conduct a piecewise growth model in a dataset
>> >> (n=40) with 9 timepoints. The dependent variable is
>> >> vocabulary size. Initially, i ran 3 models (using glmer with
>> >> family = poisson) comparing models with a linear growth and
>> >> knots at 2 different timepoints. The knots are suggested both
>> >> by theory and graphical observation. The models ran
>> >> successfully and i chose the model with the lowest AIC.
>> >> However, when i added level 2 predictors to the model, the
>> >> following error message came up
>> >>
>> >> model2 <- glmer(vocab ~ time1_18*seg + time2_18*seg +
>> >> (time1_18 + time2_18|id), data = growth, family = poisson)
>> >>
>> >> Warning messages:
>> >> 1: In mer_finalize(ans) :
>> >>   Cholmod warning 'not positive definite' at
>> >> file:../Cholesky/t_cholmod_rowfac.c, line 432
>> >> 2: In mer_finalize(ans) :
>> >>   Cholmod warning 'not positive definite' at
>> >> file:../Cholesky/t_cholmod_rowfac.c, line 432
>> >> 3: In mer_finalize(ans) : false convergence (8)
>> >>
>> >> I did a brief search of the archives and it seems to suggest
>> >> that having a small sample size might result in this. I am
>> >> unsure what is causing this. Any advice on how to resolve this?
>> >>
>> >> Cheers,
>> >> Yonghao
>> >>
>> >>       [[alternative HTML version deleted]]
>> >>
>> >> _______________________________________________
>> >> R-sig-mixed-models at r-project.org mailing list
>> >> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>> >>
>> > _______________________________________________
>> > R-sig-mixed-models at r-project.org mailing list
>> > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>> >
>
>        [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>