[R-sig-ME] anova (lm, lmer ) question

[Ben Bolker]

On 14-10-04 10:13 AM, Ben Pelzer wrote:
> Dear romunov, Ben and Ken,
> 
> Thanks for your replies. From these I conclude that:
> - for linear (lmer vs. lm) models there's no problem in using the
> deviance difference
> - for generalized  linear models (glmer vs. glm) it's ok to use the
> deviance difference as long as nAGQ=1.
> Would you agree with me? Best regards,
> 
> Ben.


  Yes, I believe so, but you might want to check the archives.  I think
I've posted examples to this effect in the past.  (The way to
double-check this would either be to set up an example where the RE
variance was estimated as exactly zero (e.g. a small/noisy data set with
a small number of levels of the grouping variable), or to extract the
deviance function via devFunOnly=TRUE and force the random effects to
zero -- for lmer this is trivial since the fixed effects estimates are
profiled out; for glmer you would have to put the deviance function
inside a wrapper function that set the variance parameters to zero while
filling in specified values for the fixed effects, and optimize over
this function ...)

  Ben Bolker

> 
> On 4-10-2014 2:48, Ben Bolker wrote:
>> Thanks for checking.  The comparison with Stata isn't necessarily
>> relevant
>> though -- or question is whether `lm` and `lmer` (or `glm` and `glmer`)
>> include/exclude the same additive constants, so that their
>> log-likelihoods
>> are directly comparable.
>>
>> On Fri, Oct 3, 2014 at 8:38 PM, Ken Beath <ken.beath at mq.edu.au> wrote:
>>
>>> nAGQ=1 and greater than 1 give different results, and the nAGQ=1 matches
>>> fairly closely the log likelihood from Stata for 3 quadrature points, so
>>> presumably is correct. Stata's Laplace didn't converge with my data.
>>>
>>>
>>> Ken
>>>
>>>
>>>
>>> On 4 October 2014 09:06, Ben Bolker <bbolker at gmail.com> wrote:
>>>
>>>> romunov <romunov at ...> writes:
>>>>
>>>>> FWIW, this is from the glmm faq site <http://glmm.wikidot.com/faq>.
>>>>>
>>>>> How can I test whether a random effect is significant?
>>>>>
>>>>    ...
>>>>
>>>>>     - *do not* compare lmer models with the corresponding lm fits, or
>>>>>     glmer/glm; the log-likelihoods are not commensurate (i.e., they
>>>> include
>>>>>     different additive terms)
>>>>    For what it's worth, I believe this is out of date, _except_ for
>>>> glmer fits with nAGQ>1.  It should be possible to implement
>>>> anova(<merMod>,<lm>/<glm>) -- it's only a nuisance (sadly, if we
>>>> were still using S4 classes at this level it would be easier ...)
>>>>
>>>>    Ben Bolker
>>>>
>>>> _______________________________________________
>>>> R-sig-mixed-models at r-project.org mailing list
>>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>>>
>>>
>>>
>>> -- 
>>>
>>> *Ken Beath*
>>> Lecturer
>>> Statistics Department
>>> MACQUARIE UNIVERSITY NSW 2109, Australia
>>>
>>> Phone: +61 (0)2 9850 8516
>>>
>>> Building E4A, room 526
>>> http://stat.mq.edu.au/our_staff/staff_-_alphabetical/staff/beath,_ken/
>>>
>>> CRICOS Provider No 00002J
>>> This message is intended for the addressee named and m...{{dropped:11}}
>> _______________________________________________
>> 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