[R-sig-ME] Fwd: RE: [R] errors with hurdle negative binomial mixed effect models

[Ben Bolker]

  [forwarded to r-sig-mixed-models]


-------- Original Message --------
Subject: 	RE: [R] errors with hurdle negative binomial mixed effect models
Date: 	Fri, 9 Aug 2013 19:39:33 +0200
From: 	Marta Lomas <lomasvega at hotmail.com>
To: 	Ben Bolker <bbolker at gmail.com>

> Thanks Ben! Just 2 questions before getting into the site you adviced me:

> So if some combinations of the variable categories are missing, meaning
> that not all the combinations are present in the data set, will not be
> possible to run these models?

   Please see the FAQ I sent you to for information about dealing with
interactions of categorical variables with incomplete coverage.

> If understand, perfectly multicollinear variables means that they are
> correlated. But the Pearson coefficients show that they are not.
> Is it possible that it is because in many observations I have the same
> category level for those variables? for example, GVG= 1 and sward = 1;

  It is possible for a set of more than two variables to be multicollinear
(e.g. A+B+C=constant) even when no pair is perfectly correlated,
although I don't know if that's the case here ...

Ah! the summary is:


> summary(SW_GVG)
   Year     Week    Count            Sward     GVG09          Cluster
 2013:510   1:102   Min.   : 0.000  0:  5     20,525:325      Min.:1.000
            2:102   1st Qu.: 0.000  1:169     28,125:100     1st Qu.:1.000
            3:102   Median : 0.000  2:160   34,775: 20         Median:4.000
            4:102   Mean   : 1.316  3:158   51,375: 5         Mean   :3.951
            5:102   3rd Qu.: 0.000  4: 18    74,2  : 55       3rd Qu.:6.000
                    Max.   :95.000           78    :5         Max.   :8.000

      GVG        SCluster
 Min.   :1.000     1:135
 1st Qu.:1.000    2: 85
 Median :1.000   3: 25
 Mean   :1.784    4: 35
 3rd Qu.:2.000    6:125
 Max.   :6.000    7: 95
                  8: 10


  I don't know exactly what's going on here, but you should look
at

with(SW_GVG,table(factor(GVG),Sward,Count>0))

I suspect you will find there are empty cells.

  Ben Bolker


> To: r-help at stat.math.ethz.ch
> From: bbolker at gmail.com
> Date: Fri, 9 Aug 2013 17:07:31 +0000
> Subject: Re: [R] errors with hurdle negative binomial mixed effect models
>
> Marta Lomas <lomasvega <at> hotmail.com> writes:
>
> >
> > Hello!
>
> > I am new in the mailing list for R help and I hope to be able to
> > formulate a good question easy to understand.
>
> We hope so too :-)
>
> [snip]
>
> I will take a first crack at this here, but follow-ups should
> probably be redirected to the r-sig-mixed-models at r-project.org
> mailing list, which is more appropriate for questions dealing
> with (G)LMMs.
>
> > I am modeling my data set with hurdle negative binomial mixed
> > effects, to find the correlation of some bird counts with
> > environmental (categorical and continuous) variables.
>
> > When I run different models I have always an error. For instance:
> >
> > -For the truncated modeling of the non-zero counts:
> >
> > > HURgvgsw <- glmmadmb(count~ GVG*sward + (1|week),
> > data=subset(SW_GVG,count>0), + family=
> > "truncnbinom")
> >
> > Error en glmmadmb(count ~ GVG * sward + (1 | week), data =
subset(SW_GVG, :
> > rank of X = 6 < ncol(X) = 10
> >
> > -Or the binomial part where the zeros are modeled with the non-zeros:
> >
> > > HURgvgsw <- glmmadmb(count~ sward*GVG + (1|week) + (1|cluster),
> > data=SW_GVG, family= "binomial")
> >
> > Error en glmmadmb(count ~ sward * GVG + (1 | week) +
> > (1 | cluster), data = SW_GVG, :
> > rank of X = 13 < ncol(X) = 15
> >
> > Would you have the solution to this?
>
> This error message is telling you that some of your fixed-effect
> variables (which are, internally, combined into the fixed-effect
> design matrix X) are perfectly multicollinear. This is most likely
> happening because sward and GVG are categorical variables (or at
> least are being treated as categorical variables) and some
> combinations are missing from the data set (for future reference:
> the output of summary(SW_GVG) is useful for diagnosis).
>
> For more information, search http://glmm.wikidot.com/faq for the
> word 'rank'
>
> Good luck
> Ben Bolker
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.