[R-sig-ME] new version of lme4 - Log-Likelihood of model returned as "-Inf"

[Ben Bolker]

  It looks like the proximal problem here is that you have non-integer
responses in the Poisson distribution.  I haven't dug in to figure out
why this should cause a problem; having non-integer responses in a
nominally Poisson response breaks the nominal statistical model, but
normal GLMs can manage anyway, and there's *probably* no reason that the
internal lme4 code couldn't be made a bit more flexible.
   Where do your non-integer responses come from?

   I would also comment offhand that trying to fit a random effect of
(1|BatID/PlantID) seems rather optimistic; you have a total of 26
observations (after removing NAs), and 121 Bat:Plant combinations, so
very few of the combinations are actually measured, and are measured at
most twice ...  the BatID:PlantID interaction is going to translate
_almost_ to an observation-level random effect.  When I fit without the
single non-integer response, I get a zero variance estimate for the
Bat:Plant interaction ...

  Ben Bolker


On 14-02-13 09:46 PM, juwb08 at hampshire.edu wrote:
> Hi Ben Bolker,
> I tried running the code with lme4 version 1.1-4, and receive a model
> that I can print, but it still lacks the LogLikelihood. When I call
> summary on that model I get a new error message.
> 
> "
> Error in diag(vcov(object, use.hessian = use.hessian)) :
>   error in evaluating the argument 'x' in selecting a method for
> function 'diag': Error in if (nonSymm > stopTol)
> stop(sprintf("calculated variance-covariance matrix is non-symmetric
> (tol=%f)",  :
>   missing value where TRUE/FALSE needed
> "
> 
> I had not received any error message with lme4 version 1.0-5 when
> calling summary() on that model. Still lost about the -Inf as the
> LogLikelihood.
> 
> cheers
> Justin Baldwin
> 
> 
> Quoting Ben Bolker <bbolker at gmail.com>:
> 
>> Still haven't seen the data set ...
>>
>>  cheers
>>    Ben Bolker
>>
>> On 14-02-12 06:14 PM, juwb08 at hampshire.edu wrote:
>>>
>>> ... the dataset failed to attach, so here it is.
>>> Sorry!
>>>
>>> Quoting juwb08 at hampshire.edu:
>>>
>>>> Hi All,
>>>>
>>>> I was working with Poisson glmms in lme4 last year in 2013 using
>>>> Windows XP, lme4 version 0.999375-16 (2008-06-23) and R version 2.15.1
>>>> (2012-06-22), evaluating a set of candidate models with AIC to select
>>>> the best ones.
>>>>
>>>> This worked well until I updated lme4 and R last month.
>>>>
>>>> I now updated to R version 3.0.2 (2013-09-25) -- "Frisbee Sailing" and
>>>> lme4 version 1.0-5 (2013-10-24), still Windows XP, and am running the
>>>> same models again on the same data, and lme4 now returns "-Inf" for
>>>> the log likelihood, so it can't calculate AIC.
>>>>
>>>> Using the previous versions, I got stable values for Log-Likelihoods
>>>> and calculated AICs with them.
>>>>
>>>> Attached is a sample of the dataset and the code I am using.
>>>>
>>>> How can I get a Log-Likelihood estimate with the current versions of
>>>> lme4 and R?
>>>>
>>>> Thanks in advance.
>>>> Justin Baldwin
>>>>
>>>> Hampshire College, MA, USA
>>>
>>>
>>>
>>>
>>> _______________________________________________
>>> R-sig-mixed-models at r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>>
> 
> 
>