[R-sig-ME] alternative suggestions to glmmTMB family=beta
Ben Bolker
bbolker at gmail.com
Fri Apr 13 01:08:16 CEST 2018
The alternative/simpler way to fit a compound symmetry model is as a
nested model, (1|id/bmt). However, this only allows for *positive*
compound symmetry.
In the example you have, I think this model is actually overspecified,
because the (1|id:bmt) term (the nested specification expands to (1|id)
+ (1|id:bmt)) has one random effect for every observation. An
observation-level random effect underlying a Beta distribution is
equivalent to fitting a logistic-Normal-Beta model, and this will be
unidentifiable (or nearly so) because the Beta distribution has its own
dispersion parameter ...)
On 2018-04-12 01:58 PM, Nat Holland wrote:
> Yes, 10 levels is large... the data set has 277 studyid for total of
> 2770 rows. So, large, but not so giant. I am working on the cs and dia
> specifications now.... thanks for that suggestion. brms package is also
> another way I would not have seen... I will consider. I may simply
> randomly select one bmt level randomly from each studyid and then work
> with reduced data set in betareg package that lacks the
> pseudoreplication of (bmt|studyid). Thanks again...
>
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> J. Nathaniel Holland, Ph.D.
> Research and Data Scientist
> e-mail: jnhollandiii at gmail.com
> <mailto:jnhollandiii at gmail.com>LinkedIn:
> https://www.linkedin.com/in/jnhollandiii/
> Google Scholar:
> https://scholar.google.com/citations?user=VbHqPXEAAAAJ&hl=en
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>
>
> On Thu, Apr 12, 2018 at 12:44 PM, Ben Bolker <bbolker at gmail.com
> <mailto:bbolker at gmail.com>> wrote:
>
>
> A couple of things here:
>
> (1) fitting (bmt | studyid) is very ambitious if bmt is a factor
> with 10 levels - that means fitting a 10x10 variance-covariance
> matrix (55 variance-covariance parameters). OK if you have a giant
> data set, but otherwise (a) consider structured (compound symmetric
> or diagonal) var-cov matrix; (b) use brms package (which does have
> beta distribution, and allows/requires a prior on the
> variance-covariance matrix which will make things better).
>
> (2) the marginal distribution not looking right doesn't
> necessarily mean the residual distribution isn't OK.
>
> cheers
> Ben Bolker
>
>
> On Thu, Apr 12, 2018 at 12:54 PM, Nat Holland
> <jnhollandiii at gmail.com <mailto:jnhollandiii at gmail.com>> wrote:
>
> Thanks...
> bmt is a factor with 10 levels.
>
> Here is freq distribution untransformed percentages...
>
>
>
> I am ok with good ole fashioned transformation, but when I logit
> transform proportions using (car) logit function, then i still
> get the spike at upper end:
>
>
> This distribution is not sufficient for linear model...
>
> Nat
>
>
>
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> J. Nathaniel Holland, Ph.D.
> Research and Data Scientist
> e-mail: jnhollandiii at gmail.com
> <mailto:jnhollandiii at gmail.com>LinkedIn:
> https://www.linkedin.com/in/jnhollandiii/
> <https://www.linkedin.com/in/jnhollandiii/>Google Scholar:
> https://scholar.google.com/citations?user=VbHqPXEAAAAJ&hl=en
> <https://scholar.google.com/citations?user=VbHqPXEAAAAJ&hl=en>~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>
>
> On Thu, Apr 12, 2018 at 11:25 AM, Ben Bolker <bbolker at gmail.com
> <mailto:bbolker at gmail.com>> wrote:
>
> What is bmt? numeric or factor? if factor, how many levels
> does it
> have? If numeric, centering the predictor often helps.
>
> - the mgcv package can fit beta-distributed responses; I'm
> not sure
> if it does "unstructured" (general positive-definite)
> variance-covariance matrices or not. (It doesn't seem
> straightforward:
> https://stat.ethz.ch/R-manual/R-devel/library/mgcv/html/random.effects.html
> <https://stat.ethz.ch/R-manual/R-devel/library/mgcv/html/random.effects.html>)
>
> - you could take the good old-fashioned approach of
> logit-transforming your responses and fitting a linear model
>
> - you could try simplifying the model: i.e. perhaps a diagonal
> (diag(bmt|studyid)) or compound-symmetric (cs(bmt|studyid))
> variance-covariance model would be adequate?
>
> - as a last resort, or if you're really attached to this
> particular
> model, you could try to understand precisely which
> parameters are
> flat/strongly correlated. If you want to do that, respond
> here and I
> (or Mollie Brooks) can try to talk you through extracting
> the Hessian
> of the fit and figuring out which components/directions are
> non-positive ...
>
>
> On Wed, Apr 11, 2018 at 4:26 PM, Nat Holland
> <jnhollandiii at gmail.com <mailto:jnhollandiii at gmail.com>> wrote:
> > I have tried to use the following model to fit beta
> distribution response
> > variable, with high frequency of data at upper end of 0 to
> 1.0 range of
> > histogram.
> >
> > glmmTMB(vas2 ~ bmt + (bmt|studyid),
> family=list(family="beta",link="logit"))
> >
> > I get the following warning messages:
> > Warning messages:
> > 1: In fitTMB(TMBStruc) :
> > Model convergence problem; non-positive-definite Hessian
> matrix. See
> > vignette('troubleshooting')
> > 2: In fitTMB(TMBStruc) :
> > Model convergence problem; false convergence (8). See
> > vignette('troubleshooting')
> >
> > Reading about this on the troubleshooting pages suggests
> "Models with
> > non-positive definite Hessian matricies should be excluded
> from further
> > consideration, in general."
> >
> > Any suggestions on alternative means of analyses to
> evaluate the above
> > model?
> >
> > Thanks in advance,
> > Nat
> >
> > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > J. Nathaniel Holland, Ph.D.
> > e-mail: jnhollandiii at gmail.com <mailto:jnhollandiii at gmail.com>
> > LinkedIn: https://www.linkedin.com/in/jnhollandiii/
> <https://www.linkedin.com/in/jnhollandiii/>
> > Google Scholar:
> >
> https://scholar.google.com/citations?user=VbHqPXEAAAAJ&hl=en
> <https://scholar.google.com/citations?user=VbHqPXEAAAAJ&hl=en>
> > ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> >
> > [[alternative HTML version deleted]]
> >
> > _______________________________________________
> > R-sig-mixed-models at r-project.org
> <mailto:R-sig-mixed-models at r-project.org> mailing list
> > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> <https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models>
>
>
>
>
More information about the R-sig-mixed-models
mailing list