[R-sig-ME] lme4 error: "Number of levels of a grouping factor . . ."

[Ben Bolker]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 07/27/2011 01:50 PM, Brent Pedersen wrote:
> On Wed, Jul 27, 2011 at 7:34 AM, Ben Bolker <bbolker at gmail.com> wrote:
> On 07/27/2011 03:29 AM, Martin Maechler wrote:
>>>>>>>>> Gabrielle Miller-Messner <gmessner at ucdavis.edu>
>>>>>>>>>     on Tue, 26 Jul 2011 22:26:38 -0700 writes:
>>>>
>>>>     > Hello, I have count/proportion data and I am trying to
>>>>     > account for overdispersion in a logistic regression. With
>>>>     > the following code from lme4 I receive the following error
>>>>     > message:
>>>>
>>>>     >> glmm = glmer(cbind(fertilized, unfert) ~ density.class +
>>>>     >                (1|individual), family = binomial)
>>>>     > Number of levels of a grouping factor for the random effects
>>>>     > is *equal* to n, the number of > observations
>>>>
>>>>     > This problem was mentioned in Warton and Hui 2011, so I
>>>>     > imagine it may have been fixed recently.  I downloaded the
>>>>     > most recent version of lme4 available through the command
>>>>     > install.packages("lme4").  Is there an updated version of
>>>>     > the package available in a different form?
>>>>
>>>> {yes, lme4a on R-forge, but that's not the issue now, and not
>>>>  something I'd recommend to get into just right now ...}
>>>>
>>>> Well, I'm puzzled that you say you get an *error* with the above
>>>> message.
>>>> You should get a message, but no error in the case of glmer().
> 
>  Probably this is indeed just a warning message and not an error?
>   Gabrielle, does the function return an answer (i.e. a fitted model)?
> 
> 
>> Hi, I've hit this as well. with:
> 
> packageVersion("lme4a")
>> [1] 0.999375.67
> 
>> (should I be using another version?)
>> and code:
> 
>> """
>> library(lme4a)
> 
>> d = data.frame(a=runif(200),
>>                b=sample(c(strsplit("ABCDEF", "")[[1]]), 200, replace=T),
>>                c=as.factor(1:200))
>> # d$c[2] = d$c[1]
> 
>> fm = glmer(a ~ b + (1|c), d)
>> """
>> it shows:
>> """
>> Error in lmer(formula = a ~ b + (1 | c), data = d) :
>>   number of levels of each grouping factor must be less than number of obs
>> """
> 
>> and if i uncomment that line to make the number of grouping factors (c) less
>> than the number of obs, it runs successfully.
>> not sure if this is different for binomial...



  This is different because (despite using the function "glmer" and not
"lmer"), you are actually fitting a linear mixed model (LMM) and not a
generalized linear mixed model (GLMM) here.  (If you don't specify a
family, or you specify a gaussian family with an identity link, glmer
automatically reverts to calling lmer -- this is a design decision
(infelicity?) that I hope we will get around to reversing ...)  In the
case of a LMM, it doesn't make sense to fit a separate individual-level
random factor, because it is completely confounded with the residual
error.  You only need to worry about overdispersion, and hence it only
makes sense to fit individual-level random effects, in cases where the
scale parameter is fixed (e.g. binomial, Poisson).

   Ben Bolker
> 
> 
>>>>
>>>>
>>>> Just to be sure:  What does
>>>>      packageVersion("lme4")
>>>> say for you?
>>>>
>>>> Martin
>>>>
>>>>
>>>>     > I am using R version 2.13.1 on OS 10.6.4.
>>>>
>>>>     > thank you!  Gabrielle
>>>>
>>>>     > --
>>>>     > Gabrielle Miller-Messner Graduate Student Center for
>>>>     > Population Biology Section of Evolution and Ecology
>>>>     > University of California, Davis 1 Shields Avenue Davis,
>>>>     > California 95616 gmessner at ucdavis.edu
>>>>
>>>>     >         [[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
> 
>>
_______________________________________________
R-sig-mixed-models at r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.10 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iEYEARECAAYFAk4wVJwACgkQc5UpGjwzenOe5gCfRy8EeieSVwXSxbqnVKTS1E79
VeIAn3z4QqAvAD1lxYCFHrzFWlHX6qkH
=Ckg5
-----END PGP SIGNATURE-----