(G)LMM status

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
7 messages Options
Reply | Threaded
Open this post in threaded view
|

(G)LMM status

Ross Boylan
Is julia's fitting of LMM's currently good enough for "production" use,
e.g. in a published paper?

Any idea when GLMM's will be in a similar state?

I'm a little confused; there was a thread about a year ago
(https://groups.google.com/d/msg/julia-stats/Txid3wkvNOE/bQZ1mUP4SA8J )
in which Doug Bates said he had merged LMM's into the GLM package, but I
don't see a sign of it in the docs https://github.com/JuliaStats/GLM.jl.

The same thread also said GLM's were on the todo list, but I don't see a
sign of them either.

Thanks.
Ross Boylan

P.S. cc appreciated; I thought I was subscribed but I don't seem to be
getting emails from the list.

--
You received this message because you are subscribed to the Google Groups "julia-stats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.
Reply | Threaded
Open this post in threaded view
|

Re: (G)LMM status

Milan Bouchet-Valat
Le mercredi 20 août 2014 à 14:31 -0700, Ross Boylan a écrit :
> Is julia's fitting of LMM's currently good enough for "production" use,
> e.g. in a published paper?
>
> Any idea when GLMM's will be in a similar state?
>
> I'm a little confused; there was a thread about a year ago
> (https://groups.google.com/d/msg/julia-stats/Txid3wkvNOE/bQZ1mUP4SA8J )
> in which Doug Bates said he had merged LMM's into the GLM package, but I
> don't see a sign of it in the docs https://github.com/JuliaStats/GLM.jl.
In his last message, he said he had undone the consolidation. So have
you tried using
https://github.com/dmbates/MixedModels.jl

> The same thread also said GLM's were on the todo list, but I don't see a
> sign of them either.
No idea about that, but Douglas will probably comment directly.


Regards


> Thanks.
> Ross Boylan
>
> P.S. cc appreciated; I thought I was subscribed but I don't seem to be
> getting emails from the list.
>

--
You received this message because you are subscribed to the Google Groups "julia-stats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.
Reply | Threaded
Open this post in threaded view
|

Re: (G)LMM status

Ross Boylan


On Wednesday, August 20, 2014 2:37:06 PM UTC-7, Milan Bouchet-Valat wrote:
Le mercredi 20 août 2014 à 14:31 -0700, Ross Boylan a écrit :
> Is julia's fitting of LMM's currently good enough for "production" use,
> e.g. in a published paper?
>
> Any idea when GLMM's will be in a similar state?
>
> I'm a little confused; there was a thread about a year ago
> (<a onmousedown="this.href='https://groups.google.com/d/msg/julia-stats/Txid3wkvNOE/bQZ1mUP4SA8J';return true;" onclick="this.href='https://groups.google.com/d/msg/julia-stats/Txid3wkvNOE/bQZ1mUP4SA8J';return true;" href="https://groups.google.com/d/msg/julia-stats/Txid3wkvNOE/bQZ1mUP4SA8J" target="_blank">https://groups.google.com/d/msg/julia-stats/Txid3wkvNOE/bQZ1mUP4SA8J )
> in which Doug Bates said he had merged LMM's into the GLM package, but I
> don't see a sign of it in the docs <a onmousedown="this.href='https://www.google.com/url?q\75https%3A%2F%2Fgithub.com%2FJuliaStats%2FGLM.jl\46sa\75D\46sntz\0751\46usg\75AFQjCNF4SL3UVcmHfdsLzBAXIP4st4P3fQ';return true;" onclick="this.href='https://www.google.com/url?q\75https%3A%2F%2Fgithub.com%2FJuliaStats%2FGLM.jl\46sa\75D\46sntz\0751\46usg\75AFQjCNF4SL3UVcmHfdsLzBAXIP4st4P3fQ';return true;" href="https://github.com/JuliaStats/GLM.jl" target="_blank">https://github.com/JuliaStats/GLM.jl.
In his last message, he said he had undone the consolidation. So have
you tried using
<a onmousedown="this.href='https://www.google.com/url?q\75https%3A%2F%2Fgithub.com%2Fdmbates%2FMixedModels.jl\46sa\75D\46sntz\0751\46usg\75AFQjCNEmnLtuj-ShufRmuea0JYz9BwwLCA';return true;" onclick="this.href='https://www.google.com/url?q\75https%3A%2F%2Fgithub.com%2Fdmbates%2FMixedModels.jl\46sa\75D\46sntz\0751\46usg\75AFQjCNEmnLtuj-ShufRmuea0JYz9BwwLCA';return true;" href="https://github.com/dmbates/MixedModels.jl" target="_blank">https://github.com/dmbates/MixedModels.jl
Thank you.  I haven't tried it yet; my concern is more about the quality of the implementation, which might not be evident from simple experiments.

BTW, it's a little odd that MixedModels does not show up under https://github.com/JuliaStats


> The same thread also said GLM's were on the todo list, but I don't see a
> sign of them either.
No idea about that, but Douglas will probably comment directly.
They name "MixedModels" seems compatible with GLMM's, though the current comment only mentions linear mixed models.

Ross Boylan 


Regards


> Thanks.
> Ross Boylan
>
> P.S. cc appreciated; I thought I was subscribed but I don't seem to be
> getting emails from the list.
>

--
You received this message because you are subscribed to the Google Groups "julia-stats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.
Reply | Threaded
Open this post in threaded view
|

Re: (G)LMM status

Douglas Bates
Hi Ross,

On Wednesday, August 20, 2014 5:54:26 PM UTC-5, [hidden email] wrote:


On Wednesday, August 20, 2014 2:37:06 PM UTC-7, Milan Bouchet-Valat wrote:
Le mercredi 20 août 2014 à 14:31 -0700, Ross Boylan a écrit :
> Is julia's fitting of LMM's currently good enough for "production" use,
> e.g. in a published paper?
>
> Any idea when GLMM's will be in a similar state?
>
> I'm a little confused; there was a thread about a year ago
> (<a href="https://groups.google.com/d/msg/julia-stats/Txid3wkvNOE/bQZ1mUP4SA8J" target="_blank" onmousedown="this.href='https://groups.google.com/d/msg/julia-stats/Txid3wkvNOE/bQZ1mUP4SA8J';return true;" onclick="this.href='https://groups.google.com/d/msg/julia-stats/Txid3wkvNOE/bQZ1mUP4SA8J';return true;">https://groups.google.com/d/msg/julia-stats/Txid3wkvNOE/bQZ1mUP4SA8J )
> in which Doug Bates said he had merged LMM's into the GLM package, but I
> don't see a sign of it in the docs <a href="https://github.com/JuliaStats/GLM.jl" target="_blank" onmousedown="this.href='https://www.google.com/url?q\75https%3A%2F%2Fgithub.com%2FJuliaStats%2FGLM.jl\46sa\75D\46sntz\0751\46usg\75AFQjCNF4SL3UVcmHfdsLzBAXIP4st4P3fQ';return true;" onclick="this.href='https://www.google.com/url?q\75https%3A%2F%2Fgithub.com%2FJuliaStats%2FGLM.jl\46sa\75D\46sntz\0751\46usg\75AFQjCNF4SL3UVcmHfdsLzBAXIP4st4P3fQ';return true;">https://github.com/JuliaStats/GLM.jl.
In his last message, he said he had undone the consolidation. So have
you tried using
<a href="https://github.com/dmbates/MixedModels.jl" target="_blank" onmousedown="this.href='https://www.google.com/url?q\75https%3A%2F%2Fgithub.com%2Fdmbates%2FMixedModels.jl\46sa\75D\46sntz\0751\46usg\75AFQjCNEmnLtuj-ShufRmuea0JYz9BwwLCA';return true;" onclick="this.href='https://www.google.com/url?q\75https%3A%2F%2Fgithub.com%2Fdmbates%2FMixedModels.jl\46sa\75D\46sntz\0751\46usg\75AFQjCNEmnLtuj-ShufRmuea0JYz9BwwLCA';return true;">https://github.com/dmbates/MixedModels.jl
Thank you.  I haven't tried it yet; my concern is more about the quality of the implementation, which might not be evident from simple experiments.

The MixedModels package is indeed where I have been developing my implementation of Linear Mixed-effects Models.  The lmm() function in that package uses the same formula/data specification as lmer() in the R lme4 package, modulo replacement of the ':' operator to indicate crossing of factors by "&".  (Julia's parser treats ':' specially because it has so many different contexts in which it is used.)

The lmm function only creates the model object.  You must call fit(lmm(...)) to cause the model to be fit.  Many of the extractor functions (ranef, fixed, etc.) have the same names as in R.  Like several other parts of Julia, the documentation is incomplete - very incomplete.  Often I resort to reading the source code to clarify issues.  The good news is that all the source code in the MixedModels package is in Julia.

My focus in the Julia implementation has been on performance.  Because some of the early users are experimental psychologists and psycholinguists, I have recently been working on special methods for models with a single grouping factor for the random effects and models with two crossed or nearly crossed grouping factors, say "Subject" and "Item".  The performance gains in those cases have been remarkable.  Yesterday I fit a model with two crossed grouping factors and badly overspecified vector-valued random effects in 32 seconds in Julia using MixedModels.  The lmer fit takes over an hour in R.  This is a special case for a special purpose (to show that such overspecification is problematic) but factors of 5 or 10 in performance are not uncommon.  This code is currently in a branch.  I hope to merge into the master branch today.  I should update some of the IJulia notebooks before merging.

I would definitely appreciate users trying out this package and letting me know about problems and/or deficiencies.

Starting in September I will focus on GLMMs.  I don't think it should be extraordinarily difficult to add them but I also have enough experience to realize that it will probably not be smooth sailing entirely.  I have said they will be available by (American) Thanksgiving.

BTW, it's a little odd that MixedModels does not show up under <a href="https://github.com/JuliaStats" target="_blank" onmousedown="this.href='https://www.google.com/url?q\75https%3A%2F%2Fgithub.com%2FJuliaStats\46sa\75D\46sntz\0751\46usg\75AFQjCNGgZd8BTVZz8Rvx6nGfQwGO50DwOQ';return true;" onclick="this.href='https://www.google.com/url?q\75https%3A%2F%2Fgithub.com%2FJuliaStats\46sa\75D\46sntz\0751\46usg\75AFQjCNGgZd8BTVZz8Rvx6nGfQwGO50DwOQ';return true;">https://github.com/JuliaStats. 

Having a package in JuliaStats is more about who can commit than about the need for all statistics-related packages to be listed there.  I have kept it in an archive under my login because I was experimenting with many different formulations and didn't want to subject others to following the twists and turns.  I would be happy to move it to the JuliaStats group if that would improve visibility.
 


> The same thread also said GLM's were on the todo list, but I don't see a
> sign of them either.
No idea about that, but Douglas will probably comment directly.
They name "MixedModels" seems compatible with GLMM's, though the current comment only mentions linear mixed models.

Ross Boylan 


Regards


> Thanks.
> Ross Boylan
>
> P.S. cc appreciated; I thought I was subscribed but I don't seem to be
> getting emails from the list.
>

--
You received this message because you are subscribed to the Google Groups "julia-stats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.
Reply | Threaded
Open this post in threaded view
|

Re: (G)LMM status

Ross Boylan
On Thu, Aug 21, 2014 at 09:01:08AM -0700, Douglas Bates wrote:
...
Thanks for all the info; the speedups are really striking.
One comment on the side issue of the packages being listed under julia-stats on github:

> [Ross wrote]
> BTW, it's a little odd that MixedModels does not show up under
> > https://github.com/JuliaStats.
> >
>
> Having a package in JuliaStats is more about who can commit than about the
> need for all statistics-related packages to be listed there.  I have kept
> it in an archive under my login because I was experimenting with many
> different formulations and didn't want to subject others to following the
> twists and turns.  I would be happy to move it to the JuliaStats group if
> that would improve visibility.
>  
I think it might be easier to discover there, although really an ideal
solution would be for the general package list to have some other ways
of organizing things than alphabetical or search.  Some kind of
tagging or classification system would be good.  Debian's
packagesearch application
(https://packages.debian.org/wheezy/packagesearch) and the
infrastructure behind provide a pretty helpful way of finding stuff in
the huge number of Debian packages.  Note that a lot of the tagging is
done automatically.

One interesting thing about this is that the system defines a number
of dimensions (e.g., substantive field to which the package relates,
formats it supports).  Some of the dimensions only apply to some of
the packages.  I'm sure there's a formal name and a theory
behind this approach, though I don't know either.


Ross

--
You received this message because you are subscribed to the Google Groups "julia-stats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.
Reply | Threaded
Open this post in threaded view
|

Re: (G)LMM status

Iain Dunning
Rest assured Ross that we are well aware of the "discoverability" problem. I maintain pkg.julialang.org, and only thing stopping improvements is a lack of time. For now, organizations are indeed what we rely on to guide the discovery process. There are a few issues on the main Julia github page, e.g. https://github.com/JuliaLang/julia/issues/6807

On Thursday, August 21, 2014 1:38:51 PM UTC-4, Ross Boylan wrote:
On Thu, Aug 21, 2014 at 09:01:08AM -0700, Douglas Bates wrote:
...
Thanks for all the info; the speedups are really striking.
One comment on the side issue of the packages being listed under julia-stats on github:

> [Ross wrote]
> BTW, it's a little odd that MixedModels does not show up under
> > <a href="https://github.com/JuliaStats" target="_blank" onmousedown="this.href='https://www.google.com/url?q\75https%3A%2F%2Fgithub.com%2FJuliaStats\46sa\75D\46sntz\0751\46usg\75AFQjCNGgZd8BTVZz8Rvx6nGfQwGO50DwOQ';return true;" onclick="this.href='https://www.google.com/url?q\75https%3A%2F%2Fgithub.com%2FJuliaStats\46sa\75D\46sntz\0751\46usg\75AFQjCNGgZd8BTVZz8Rvx6nGfQwGO50DwOQ';return true;">https://github.com/JuliaStats.
> >
>
> Having a package in JuliaStats is more about who can commit than about the
> need for all statistics-related packages to be listed there.  I have kept
> it in an archive under my login because I was experimenting with many
> different formulations and didn't want to subject others to following the
> twists and turns.  I would be happy to move it to the JuliaStats group if
> that would improve visibility.
>  
I think it might be easier to discover there, although really an ideal
solution would be for the general package list to have some other ways
of organizing things than alphabetical or search.  Some kind of
tagging or classification system would be good.  Debian's
packagesearch application
(<a href="https://packages.debian.org/wheezy/packagesearch" target="_blank" onmousedown="this.href='https://www.google.com/url?q\75https%3A%2F%2Fpackages.debian.org%2Fwheezy%2Fpackagesearch\46sa\75D\46sntz\0751\46usg\75AFQjCNEpXofUSKn4loagsciyEn4d0mQ0DA';return true;" onclick="this.href='https://www.google.com/url?q\75https%3A%2F%2Fpackages.debian.org%2Fwheezy%2Fpackagesearch\46sa\75D\46sntz\0751\46usg\75AFQjCNEpXofUSKn4loagsciyEn4d0mQ0DA';return true;">https://packages.debian.org/wheezy/packagesearch) and the
infrastructure behind provide a pretty helpful way of finding stuff in
the huge number of Debian packages.  Note that a lot of the tagging is
done automatically.

One interesting thing about this is that the system defines a number
of dimensions (e.g., substantive field to which the package relates,
formats it supports).  Some of the dimensions only apply to some of
the packages.  I'm sure there's a formal name and a theory
behind this approach, though I don't know either.


Ross

--
You received this message because you are subscribed to the Google Groups "julia-stats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.
Reply | Threaded
Open this post in threaded view
|

Re: (G)LMM status

Douglas Bates
In reply to this post by Ross Boylan
I just released version 0.3.10 of the MixedModels package.

On Thursday, August 21, 2014 12:38:51 PM UTC-5, Ross Boylan wrote:
On Thu, Aug 21, 2014 at 09:01:08AM -0700, Douglas Bates wrote:
...
Thanks for all the info; the speedups are really striking.
One comment on the side issue of the packages being listed under julia-stats on github:

> [Ross wrote]
> BTW, it's a little odd that MixedModels does not show up under
> > <a href="https://github.com/JuliaStats" target="_blank" onmousedown="this.href='https://www.google.com/url?q\75https%3A%2F%2Fgithub.com%2FJuliaStats\46sa\75D\46sntz\0751\46usg\75AFQjCNGgZd8BTVZz8Rvx6nGfQwGO50DwOQ';return true;" onclick="this.href='https://www.google.com/url?q\75https%3A%2F%2Fgithub.com%2FJuliaStats\46sa\75D\46sntz\0751\46usg\75AFQjCNGgZd8BTVZz8Rvx6nGfQwGO50DwOQ';return true;">https://github.com/JuliaStats.
> >
>
> Having a package in JuliaStats is more about who can commit than about the
> need for all statistics-related packages to be listed there.  I have kept
> it in an archive under my login because I was experimenting with many
> different formulations and didn't want to subject others to following the
> twists and turns.  I would be happy to move it to the JuliaStats group if
> that would improve visibility.
>  
I think it might be easier to discover there, although really an ideal
solution would be for the general package list to have some other ways
of organizing things than alphabetical or search.  Some kind of
tagging or classification system would be good.  Debian's
packagesearch application
(<a href="https://packages.debian.org/wheezy/packagesearch" target="_blank" onmousedown="this.href='https://www.google.com/url?q\75https%3A%2F%2Fpackages.debian.org%2Fwheezy%2Fpackagesearch\46sa\75D\46sntz\0751\46usg\75AFQjCNEpXofUSKn4loagsciyEn4d0mQ0DA';return true;" onclick="this.href='https://www.google.com/url?q\75https%3A%2F%2Fpackages.debian.org%2Fwheezy%2Fpackagesearch\46sa\75D\46sntz\0751\46usg\75AFQjCNEpXofUSKn4loagsciyEn4d0mQ0DA';return true;">https://packages.debian.org/wheezy/packagesearch) and the
infrastructure behind provide a pretty helpful way of finding stuff in
the huge number of Debian packages.  Note that a lot of the tagging is
done automatically.

One interesting thing about this is that the system defines a number
of dimensions (e.g., substantive field to which the package relates,
formats it supports).  Some of the dimensions only apply to some of
the packages.  I'm sure there's a formal name and a theory
behind this approach, though I don't know either.


Ross

--
You received this message because you are subscribed to the Google Groups "julia-stats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.