[R-sig-ME] Julia vs nlme vs lme4 implementation of fitting linear mixed models

Phillip Alday phillip.alday at staff.uni-marburg.de
Thu Oct 16 11:23:10 CEST 2014

Hash: SHA1

Then this might be more what you're looking for:


although the mathematical discussion is largely restricted to the

Doug will probably comment when it's a decent hour in Wisconsin, but
my understanding was that the Julia code was based on roughly the same
computational approach as lme4, but plays to Julia's strengths. The
standard reference for nlme is Pinherio and Bates, I'm still waiting
for my copy to be delivered, but the portions viewable on Google books
seem to suggest that the computational approach behind nlme is also
discussed in there. The approach in lme4 is based on sparse matrix
methods, so lme4 tends to be faster and more memory efficient (this
was recently hinted at on this list:
https://stat.ethz.ch/pipermail/r-sig-mixed-models/2014q4/022752.html )

Some aspects of the computational approach behind lme4 were
tangentially discussed recently
(https://stat.ethz.ch/pipermail/r-sig-mixed-models/2014q4/022773.html), but
I suspect that discussion was part of the motivation for your question!


On 16.10.2014 10:55, W Robert Long wrote:
> Thanks Phillip
> I've seen the glmm wikidot pages. I've also seen some of the code 
> comparisons between lme4 and julia on github.  I've also seen the
> book by Pinheiro and Bates about nlme and the recent paper on lme4
> at http://arxiv.org/abs/1406.5823 [this paper gives some hints
> about what is done differently in julia, but I would really like
> more detail if possible]
> I'm interested in the differences in internal implementation,
> rather than the differences that a typical user of the packages
> would be concerned with.
> Thanks Rob
> On 16/10/2014 09:40, Phillip Alday wrote: There's a bit on the FAQ
> under "Which R packages (functions) fit GLMMs?":
> http://glmm.wikidot.com/faq
> whihc is fleshed out on this page:
> http://glmm.wikidot.com/pkg-comparison
> And check out this question on StackOverflow:
> http://stats.stackexchange.com/questions/5344/how-to-choose-nlme-or-lme4-r-library-for-mixed-effects-models
> Those pages only discuss R packages, for the Julia package(s), you 
> should check out the Julia package from Doug Bates, which has
> examples worked as parallels to the lme4 examples:
> https://github.com/dmbates/MixedModels.jl
> If I recall correctly, he also has an iPython Notebook with a more 
> involved technical examination of mixed models in Julia, but I
> can't find the link at the moment.
> Best, Phillip
> On 16.10.2014 10:31, W Robert Long wrote:
>>>> What are the resources that compare how linear and
>>>> generalised linear mixed models are fitted in julia, lme4 and
>>>> nlme in terms of the how they differ in their implementation,
>>>> and what advantages/disadvantages each has. I'm asking about
>>>> the theoretical and computational issues rather than
>>>> comparing speeds for any particular dataset/model.
>>>> _______________________________________________ 
>>>> R-sig-mixed-models at r-project.org mailing list 
>>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>> _______________________________________________ 
>> R-sig-mixed-models at r-project.org mailing list 
>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> _______________________________________________ 
> R-sig-mixed-models at r-project.org mailing list 
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
Version: GnuPG v2.0.22 (GNU/Linux)


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