[R-sig-ME] glmer and influence.me - complaining about nAGQ==0

Gabriel Baud-Bovy b@ud-bovy@g@br|e| @end|ng |rom h@r@|t
Fri Jun 4 18:29:21 CEST 2021


I agree that data should not be log transformed twice. In practice,
using the log link with the
Gamma model do solve  errors occutring when negative values occurs
during the fitting process.

However, I think that one should  avoid log-transforming RTs data
ideally because it can affect the significance
of an interaction. I would prefer to assess effect on RTs on its
original scale if possible as suggested by Lo and Andrew
(@Alday: am I wrong in this respect?).

In fact, two years ago, I was able to fit RT data with a Gamma model and
identity link but
I could not understand the results:

https://stats.stackexchange.com/questions/391076/how-to-interpret-significant-factors-in-a-glmm-gamma-model-that-appears-to-be-go

see also my question in this list (3/21/2019) Fitting RT:
underdispersion with gamma and identity link.  At the
time, I caclutated DHARMA  residuals but I was still unable to
understand what was going on with my dataset
(I had highly statistically significant fixed effects that, looking at
the plots, should not have been statistically
significant imho).


Gabriel




On 4/26/2021 1:54 PM, Phillip Alday wrote:
>
> On 24/4/21 11:08 pm, Ben Bolker wrote:
>>     Don't have much to add to John's comments. You can see
>> vignette("lmerperf") for a few suggestions on improving performance.
>>
>>    I'm a little surprised that your response variable is "logRT" *and*
>> you have a log-link; that seems like double-logging?  (I was going to
>> suggest that if you aren't wedded to the Gamma model, a log-Normal model
>> (lmer(log(logRT) ~ ...) would probably be a lot faster ...)
> Are you (Cátia) basing this model off the Lo and Andrews paper? I'm not
> sure I really agree with that paper -- they seem very worried about
> transformations, but then they use alternative error distributions and
> links, which doesn't help interpretation for many users in my experience.
>
>>    It's possible that other platforms (glmmTMB, Julia::MixedModels.jl)
>> would be faster ... but then you might be stuck without influence
>> diagnostics again ...
>>
> We don't have influence currently implemented in MixedModels.jl, but
> that wouldn't actually be hard. The bigger issue is that GLMMs with a
> dispersion parameter, including Gamma, don't currently work in
> MixedModels.jl
>
> _______________________________________________
> R-sig-mixed-models using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> .


--
---------------------------------------------------------------------
Gabriel Baud-Bovy               tel.: (+39) 348 172 4045     (mobile)
UHSR University                       (+39) 02 2643 3429 (laboratory)
via Olgettina, 58                     (+39) 02 9175 1540  (secretary)
20132 Milan, Italy              email:       gabriel.baud-bovy using hsr.it
---------------------------------------------------------------------

[https://www.5xmille.org/images/img_01.png]<https://www.5xmille.org/?utm_source=firmamail&utm_medium=email&utm_campaign=5xmille2021>

CODICE FISCALE 07636600962
Scopri di più su www.5xmille.org<https://www.5xmille.org/?utm_source=firmamail&utm_medium=email&utm_campaign=5xmille2020>


Rispetta l’ambiente: non stampare questa mail se non è necessario.
Respect the environment: print this email only if necessary.



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