[R-sig-ME] glmer output

[Krause, R.W. (Robert)]

Hello Moshi,

I want to give you an advice for your analysis and tried answer your questions. But have in mind: I am not a mixed models expert and just a student.
I suggest you read the paper by Barr et al. (2014) and change your model.
http://idiom.ucsd.edu/~rlevy/papers/barr-etal-2013-jml.pdf

You did not include your fixed effect as a random effect, which I think is highly recommended after reading Barr. Probably this makes no sense because you only have one obersvation per Family or there is no "random slope" in the population, however, Barr et al. state: "In terms of power, maximal models perform surprisingly well even in a ‘‘worst case’’ scenario where they assume random slope variation that is actually not present in the population."

Somehing like:
y ~ treat + (1 + treat| Family) + (1 | obs)

I do not know your data and experiment well enough and I hope that some of the experts will help with that.

I did not really get your question. When you use dummi coding,as it seems, you see in your first table (treatLH      -1.1329     0.3004  -3.772 0.000162 ***) that "treat" is a significant predictor. As you said there was no other level "(treat which has two levels - High and Low)", if you had a null treatment (no diet?), you could have had contrasts showing the effect of high and low treat vs null-treat.
As you can see in your second table, the over all effect of treat is 0.566 in one or the other direction, depending on your group.
If you just want the significance of including or excluding treat, you should use ML testing. Run the model:
y ~ 1 + (1 + treat| Family) + (1 | obs)
and compare it with
y ~ treat + (1 + treat| Family) + (1 | obs)



For the reporting, I do not know the biology guidelines, but unfortunately psychology is so slow in adaptation that we have none for mixed models, however you can say something like this:

We used a Generalized Linear Mixed Models for binomially distributed outcomes with glmer() function of the lme4 (cite lme4) package in R (cite R). The model included a fixed intercept and a fixed slope for treat, (here again, check if you do not want to use sum-to-zero contrast, I think your intercept will disappear than... options(contrasts=c("contr.sum"))).(If you change your model like Barr recommends, you can go on as follows - I saw that in a published paper:) We followed Barr, Levy, Scheepers, and Tily's(2013) advice to use a maximal random-effects structure: - And now you list the random effects, do not forget the random correlation terms amog the random effects, by now you do not have any, because you only use random intercept models but after adding a slope, the correlation between slope and intercept is also part of the model. If you run into no-convergence now and changing optimizers and increasing iterations do not help, I suggest you split your random effect (1|group) + (0+slope|group) and the correlation is excluded, but you should report that! If you use optimzers, then state which you used and you could also add the number of iterations, everything someone needs to replicate your findings.
You should write the model into your paper so everybody can see directly what you did: y ~ treat + (1 + treat| Family) + (1 | obs)

Lastly, I recommend you use bootstrapping to determine p-values with the package pbkrtest and the bootmer function. It might take some time but is, as far as I know, the "best" (reliable and valid) way to get p-values. And it does not have the theoretical issue of how many parameters and dfs your model really has...


The rest could be written as usual.

Again note: I am just a student, so if one of the masters out there sees something wrong in my explanations, please commend.

Kind regards,

Robert

--

Robert Krause
Student, Master of Bahvioural Science
Radboud University Nijmegen
robert.krause at student.ru.nl

----- Oorspronkelijk bericht -----
> Van: "Moshiur Rahman" <mrahmankufmrt at gmail.com>
> Aan: r-sig-mixed-models at r-project.org, r-sig-mixed-models-request at r-project.org,
> r-sig-mixed-models-owner at r-project.org
> Verzonden: Maandag 7 april 2014 07:54:30
> Onderwerp: [R-sig-ME] glmer output
> Dear mixed model experts,
> 
> 
> 
> I used the generalized linear models (GLM) to explore the dietary
> effects
> entering the number of offspring sired by both groups as the response
> variable (y), dietary treatment as a fixed effect (treat which has two
> levels - High and Low), family as a random effect with a binomial
> error
> distribution and a logit link function. The model was weighted by
> total
> number of offspring, and observation-level was included in the full
> model
> as a random effect to approximate overdispersion which reduced the
> dispersion parameter from 3.39 to 1.02. The output is below:
> 
> 
> 
> y=cbind(Data$Success,Data$Failure)
> 
> 
> Generalized linear mixed model fit by the Laplace approximation
> 
> Formula: y ~ treat + (1 | Family) + (1 | obs)
> 
> Data: Data
> 
> AIC BIC logLik deviance
> 
> 198.6 206.4 -95.28 190.6
> 
> Random effects:
> 
> Groups Name Variance Std.Dev.
> 
> obs (Intercept) 1.0383 1.019
> 
> Family (Intercept) 0.0000 0.000
> 
> Number of obs: 52, groups: obs, 52; Family, 26
> 
> 
> 
> Fixed effects:
> 
> Estimate Std. Error z value Pr(>|z|)
> 
> (Intercept) 0.5664 0.2124 2.667 0.007658 **
> 
> treatLH -1.1329 0.3004 -3.772 0.000162 ***
> 
> ---
> 
> Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
> 
> 
> 
> Correlation of Fixed Effects:
> 
> (Intr)
> 
> treatLH -0.707
> 
> 
> 
> The result gives only the effect of Low dietary group (treatLH). I
> know
> also how to see the effect of High group excluding intercept with the
> following formula:
> 
> Fixed effects:
> 
> Estimate Std. Error z value Pr(>|z|)
> 
> treatHH 0.5664 0.2124 2.667 0.00766 **
> 
> treatLH -0.5665 0.2124 -2.667 0.00765 **
> 
> 
> 
> I also know some experts suggest for likelihood-ratio test. But first
> of
> all, I'd like to see the effects of treat on y (not individual effect
> of
> High or Low group) like ANOVA.
> 
> 
> 
> Can anyone suggest how to get it? If it's not possible, how can I
> report
> the existing results for my thesis?
> 
> 
> 
> Any advice will be highly appreciated.
> 
> 
> 
> Cheers,
> 
> 
> 
> Moshi
> 
> --
> MD. MOSHIUR RAHMAN
> PhD Candidate
> School of Animal Biology/Zoology (M092)
> University of Western Australia
> 35 Stirling Hwy, Crawley, WA, 6009
> Australia.
> Mob.: 061-425205507
> 
> [[alternative HTML version deleted]]
> 
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models