[R-sig-ME] Can you solve a debate between colleagues?

Ben Bolker bbolker at gmail.com
Tue Jul 11 18:14:32 CEST 2017

  Hi Joanne (please call me Ben),

  I'm going to take the liberty of cc'ing this to the
r-sig-mixed-model at r-project.org list, since it's of general interest
(that's usually a good venue for this kind of question -- you do need to
subscribe to post easily, at
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models ).

The key clause below is "that are fitted by REML".  Since lme4 doesn't
use REML for *generalized* linear mixed models as fitted by glmer
(indeed, there is not a universally accepted definition of REML for
GLMMs: see
the statement you give doesn't apply to GLMMs.

   Ben Bolker

On 17-07-10 10:33 AM, Joanne Lello wrote:
> Dear Prof Bolker, 
> I am a lecturer in Biosciences at Cardiff University where (along with
> my colleague mentioned below) we do a fair amount of statistics
> teaching, but we ourselves are not statisticians and have learned ‘on
> the job’ as it were. I would like to think we have a reasonable depth of
> understanding but it is gleaned from many different places and of course
> sometimes the sources contradict one another. I have recently moved to
> using glmer in R (previously I used ASREML for my mixed modelling). As a
> result of reading around the use of this package I came across a number
> of sources, including your own quote below, which state that using AIC
> to compare the fixed model is useless. 
>  ‘...using likelihood-based methods (including AIC) to compare two
> models with different fixed effects that are fitted by REML will
> generally lead to nonsense.’
> I have been debating this point with my colleague; she is convinced that
> this does not apply if the fixed models being compared are nested. 
> e.g. glmer(y ~ a + b + (1|d), data = dframe1) 
> may be compared via AIC with 
> glmer(y ~ a  + (1|d), data = dframe1)  
> but could not be compared with 
> glmer(y ~ a + f + (1|d), data = dframe1) 
> I had read your comments (and others) to mean that AIC could not be used
> to assess the fixed terms in either scenario. 
> We would be very grateful if you could advise on which interpretation is
> correct
> Sincerely
> Joanne Lello

More information about the R-sig-mixed-models mailing list