[R-sig-ME] What does "number of groups < 50"

[Maria Paola Bissiri]

Dear Prof. Bolker,
thank you very much for your answer. Yes, in my model the random  
effects are the experiment participants:
(1 + predictor | participant).
There are 86 participants.

My question is if likelihood ratio tests are reliable for calculating  
p-values for the parameters of my glmer model (family=binomial).
I am trying parametric bootstrapping with bootMer and confint, but  
those scripts have been running since more than 1700 minutes (is it  
normal that it takes so long?). So I would prefer to use the LRT  
method, provided that it gives reliable results.

How can I find out if LRT is suitable for my model? Is a number of  
groups > 50 sufficient?
What is meant with "finite-size cases" in this Faq?  
http://glmm.wikidot.com/faq#toc5
For using LRT, are there requirements also regarding the total number  
of samples and of parameters?

Below I copy information about my model and the fitting.
I would be thankful for any suggestion you could give me.
Kind regards,
Maria Paola

> fallmid.glmer6
Generalized linear mixed model fit by maximum likelihood ['glmerMod']
  Family: binomial ( logit )
Formula: resp_X ~ lang * ini_pch + lang * manner + lang * fin_B + (1 +  
      fin_B | subj_ID)
    Data: fallmid
       AIC       BIC    logLik  deviance
1731.5874 1875.8948 -838.7937 1677.5874
Random effects:
  Groups  Name        Std.Dev. Corr
  subj_ID (Intercept) 1.242
          fin_Bm      1.956    -0.72
Number of obs: 1548, groups: subj_ID, 86
Fixed Effects:
     (Intercept)           langde           langen           langsw     
      ini_pchm         ini_pchr         mannerla         mannerna       
      fin_Bm
        -1.76590          0.71972          0.14369          0.49608     
       0.25780         -0.02228          0.02569         -0.24087       
     2.13501
langde:ini_pchm  langen:ini_pchm  langsw:ini_pchm  langde:ini_pchr   
langen:ini_pchr  langsw:ini_pchr  langde:mannerla  langen:mannerla   
langsw:mannerla
        -0.52407          0.11792          0.26505         -0.73270     
      -1.01174         -0.03017         -0.19740         -0.48532       
     0.49438
langde:mannerna  langen:mannerna  langsw:mannerna    langde:fin_Bm     
langen:fin_Bm    langsw:fin_Bm
         0.41269          0.65082          0.61956         -1.57816     
      -1.17330         -2.32019

> probs.fallmid.glmer6 = 1/(1+exp(-fitted(fallmid.glmer6)))
> probs.fallmid.glmer6 = binomial()$linkinv(fitted(fallmid.glmer6))
> library(Hmisc)
> fit.probs.fallmid.glmer6 = somers2(probs.fallmid.glmer6,  
> as.numeric(fallmid$resp_X)-1)
> fit.probs.fallmid.glmer6
            C          Dxy            n      Missing
    0.8606880    0.7213759 1548.0000000    0.0000000



Zitat von Ben Bolker <bbolker at gmail.com>:

> [forwarding to r-sig-mixed models]
>
> -------- Original Message --------
> Subject: What does "number of groups < 50"
> Resent-Date: Tue, 22 Oct 2013 07:42:08 -0400
> Date: Tue, 22 Oct 2013 11:41:04 +0000
> From: Maria Paola
> To: bolker at mcmaster.ca
>
> Dear Prof. Bolker,
> I am carrying out an analysis of my data fitting a GLMM with glmer()
> (from lme4), family binomial.
>
> In the chapter "pvalues" in the manual (page 59)
> http://cran.r-project.org/web/packages/lme4/lme4.pdf
> you recommend the starred (*) methods "when the number of groups is < 50".
>
> What is meant exactly with "number of groups"?
>
> I have in total 1548 observations and four groups of perception
> experiment participants: German, Chinese, Swedish and English natives
> (language is a predictor in the model), with a maximum of 30
> participants per language.
>
> Using the starred (*) methods for GLMMs means bootstrapping, but I am
> experiencing problems with that (e.g. too long calculation time), so I
> would prefer to use Likelihood Ratio Tests.
> However, I am not sure if this method is suitable. I am not sure if my
> data fulfill the criterium of a number of group > 50, since I do not
> know what it means.
> =============
>
>       It's not 100% clear since I don't know the structure of your
> model exactly (presumably it's something like response ~ [fixed
> effects] + (1|participant), or response ~ [fixed effects] +
> (?|participant), i.e. you are using participant as a random effect),
> but the 'number of groups' is the number of levels of the
> random-effects grouping variable, i.e. probably the total number of
> participants in your experiment (which sounds like it is probably >
> 50). These numbers are reported in the output of glmer, in your case
> it would look something like "Number of obs: 1548, groups: participant, ??

-- 
Dr. Maria Paola Bissiri

TU Dresden
Fakultät Elektrotechnik und Informationstechnik
Institut für Akustik und Sprachkommunikation
01062 Dresden

Barkhausen-Bau, Raum S54
Helmholtzstraße 18

Tel: +49 (0)351 463-34283
Fax: +49 (0)351 463-37781
E-Mail: Maria_Paola.Bissiri at tu-dresden.de
http://wwwpub.zih.tu-dresden.de/~bissiri/index.htm