[R-sig-ME] quasi-binomial family in lme4

[Jarrod Hadfield]

Hi Florian,

This comes up regularly and the list (nearly) always stays silent. As  
far as I am aware quasi models in lmer do not, and never have, given  
sensible results. To model over-dispersion you can try fitting an  
observation-level random effect. For example:

data$resid<-as.factor(1:dim(data)[1])

and fitting (1|resid) in the model formula.

This year  I have reviewed four papers that have used quasi models in  
lmer, and its no fun to tell the authors that their results may not be  
meaningful. To paraphrase an earlier post, why are they there if they  
do not work - it's irresponsible?

Jarrod





On 8 Nov 2010, at 03:30, T. Florian Jaeger wrote:

> Hi,
> I am analyzing some data that seems to require a quasibinomial model,
> but the model returns incredibly small standard errors (and
> correspondingly inflated t-values) that do not seem to be justified
> given my data.
> I've been reading (I think) all available posts on problems with the
> quasi-binomial family in lme4. But I can't judge from the posts
> whether all issues with the  quasi-binomial models in lme4 are assumed
> to be resolved. So, I am wondering, are there any known issues with
> quasi-binomial models in lme4?
>
> I am using
>
> Package:            lme4
> Version:            0.999375-35
> Date:               2010-08-18
>
> in environment:
>                _
> platform       i386-pc-mingw32
> arch           i386
> os             mingw32
> system         i386, mingw32
> status
> major          2
> minor          11.1
> year           2010
> month          05
> day            31
> svn rev        52157
> language       R
> version.string R version 2.11.1 (2010-05-31)
>
> cheers,
> Florian
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>


-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.