Discussion:
[R-sig-ME] GLMMs with Adaptive Gaussian Quadrature
D. Rizopoulos
2018-06-15 14:26:36 UTC
Permalink
Dear R mixed-model users,

I’d like to announce the release of my new package GLMMadaptive for fitting generalized linear mixed models using adaptive Gaussian quadrature. You may read more about it here: https://goo.gl/7pi8Sh

Any comments or suggestions are more than welcome.

Best,
Dimitris


Professor of Biostatistics
Erasmus University Medical Center
The Netherlands

[[alternative HTML version deleted]]
Ben Bolker
2018-06-15 15:06:56 UTC
Permalink
It looks interesting (at an admittedly *very* quick initial glance).
Can you clarify how it differs from using lme4::glmer with nAGQ>1 ?

On Fri, Jun 15, 2018 at 10:26 AM, D. Rizopoulos
Post by D. Rizopoulos
Dear R mixed-model users,
I’d like to announce the release of my new package GLMMadaptive for fitting generalized linear mixed models using adaptive Gaussian quadrature. You may read more about it here: https://goo.gl/7pi8Sh
Any comments or suggestions are more than welcome.
Best,
Dimitris
Professor of Biostatistics
Erasmus University Medical Center
The Netherlands
[[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
D. Rizopoulos
2018-06-15 16:34:32 UTC
Permalink
AFAIK, lme4::glmer with nAGQ>1 *only* works for scalar random effects. At least, when I try setting nAGQ > 1 for a random intercepts and random slopes model in lme4::glmer (lme4_1.1-17) I get the error message:

Error in updateGlmerDevfun(devfun, glmod$reTrms, nAGQ = nAGQ) :
nAGQ > 1 is only available for models with a single, scalar random-effects term

GLMMadaptive::mixed_model implements the AGQ in such settings.

My main motivation to create this package is the longitudinal data analysis setting in which including something more than random intercepts is very typical. At least the students in my Repeated Measurements course (https://github.com/drizopoulos/Repeated_Measurements) have had some difficult times getting lme4::glmer() with a Laplace approximation to work in such cases.


-----Original Message-----
From: Ben Bolker <***@gmail.com>
Sent: Friday, June 15, 2018 5:07 PM
To: D. Rizopoulos <***@erasmusmc.nl>
Cc: r-sig-mixed-***@r-project.org
Subject: Re: [R-sig-ME] GLMMs with Adaptive Gaussian Quadrature

It looks interesting (at an admittedly *very* quick initial glance).
Can you clarify how it differs from using lme4::glmer with nAGQ>1 ?
Post by D. Rizopoulos
Dear R mixed-model users,
I’d like to announce the release of my new package GLMMadaptive for
fitting generalized linear mixed models using adaptive Gaussian
quadrature. You may read more about it here: https://goo.gl/7pi8Sh
Any comments or suggestions are more than welcome.
Best,
Dimitris
Professor of Biostatistics
Erasmus University Medical Center
The Netherlands
[[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
Ben Bolker
2018-06-15 17:56:44 UTC
Permalink
Good point. Extending AGQ to more complex models in lme4 is something
that's been on my list for a long time, but it's great to see someone
meeting the need. Even if I or someone does eventually get it working
in lme4, two implementations are always better than one ...

For those interested in this topic, there are a few other approaches
to improved frequentist estimates (i.e. without going full-Bayesian)
that are implemented in R: Helen Ogden's glmmsr package implements
sequential reduction and importance sampling methods, The glmm and
bernor packages use other flavors of importance sampling/MC likelihood
approximations. glmmADMB has importance sampling; TMB (the engine
underlying glmmTMB) has an importance-sampling method, but it hasn't
(yet) been integrated in glmmTMB ...

cheers
Ben Bolker


On Fri, Jun 15, 2018 at 12:34 PM, D. Rizopoulos
Post by D. Rizopoulos
nAGQ > 1 is only available for models with a single, scalar random-effects term
GLMMadaptive::mixed_model implements the AGQ in such settings.
My main motivation to create this package is the longitudinal data analysis setting in which including something more than random intercepts is very typical. At least the students in my Repeated Measurements course (https://github.com/drizopoulos/Repeated_Measurements) have had some difficult times getting lme4::glmer() with a Laplace approximation to work in such cases.
-----Original Message-----
Sent: Friday, June 15, 2018 5:07 PM
Subject: Re: [R-sig-ME] GLMMs with Adaptive Gaussian Quadrature
It looks interesting (at an admittedly *very* quick initial glance).
Can you clarify how it differs from using lme4::glmer with nAGQ>1 ?
Post by D. Rizopoulos
Dear R mixed-model users,
I’d like to announce the release of my new package GLMMadaptive for
fitting generalized linear mixed models using adaptive Gaussian
quadrature. You may read more about it here: https://goo.gl/7pi8Sh
Any comments or suggestions are more than welcome.
Best,
Dimitris
Professor of Biostatistics
Erasmus University Medical Center
The Netherlands
[[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
D. Rizopoulos
2018-06-15 18:31:58 UTC
Permalink
Indeed! GLMMadaptive::mixed_model is also more flexible in allowing users to define their own mixed models by specifying the log-density of the repeated measurements outcome, i.e., something similar to what Proc NLMIXED in doing in SAS. More info in the vignette: https://cran.r-project.org/web/packages/GLMMadaptive/vignettes/Custom_Models.html

Best,
Dimitris


-----Original Message-----
From: Ben Bolker <***@gmail.com>
Sent: Friday, June 15, 2018 7:57 PM
To: D. Rizopoulos <***@erasmusmc.nl>
Cc: r-sig-mixed-***@r-project.org
Subject: Re: [R-sig-ME] GLMMs with Adaptive Gaussian Quadrature

Good point. Extending AGQ to more complex models in lme4 is something that's been on my list for a long time, but it's great to see someone meeting the need. Even if I or someone does eventually get it working in lme4, two implementations are always better than one ...

For those interested in this topic, there are a few other approaches to improved frequentist estimates (i.e. without going full-Bayesian) that are implemented in R: Helen Ogden's glmmsr package implements sequential reduction and importance sampling methods, The glmm and bernor packages use other flavors of importance sampling/MC likelihood approximations. glmmADMB has importance sampling; TMB (the engine underlying glmmTMB) has an importance-sampling method, but it hasn't
(yet) been integrated in glmmTMB ...

cheers
Ben Bolker
Post by D. Rizopoulos
nAGQ > 1 is only available for models with a single, scalar
random-effects term
GLMMadaptive::mixed_model implements the AGQ in such settings.
My main motivation to create this package is the longitudinal data analysis setting in which including something more than random intercepts is very typical. At least the students in my Repeated Measurements course (https://github.com/drizopoulos/Repeated_Measurements) have had some difficult times getting lme4::glmer() with a Laplace approximation to work in such cases.
-----Original Message-----
Sent: Friday, June 15, 2018 5:07 PM
Subject: Re: [R-sig-ME] GLMMs with Adaptive Gaussian Quadrature
It looks interesting (at an admittedly *very* quick initial glance).
Can you clarify how it differs from using lme4::glmer with nAGQ>1 ?
Post by D. Rizopoulos
Dear R mixed-model users,
I’d like to announce the release of my new package GLMMadaptive for
fitting generalized linear mixed models using adaptive Gaussian
quadrature. You may read more about it here: https://goo.gl/7pi8Sh
Any comments or suggestions are more than welcome.
Best,
Dimitris
Professor of Biostatistics
Erasmus University Medical Center
The Netherlands
[[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
Sean MacEachern
2018-06-15 19:37:18 UTC
Permalink
Looks interesting. Would it be possible to fit a Numerator relationship
matrix as a random effect similarly to MCMCglmm or Asreml for binary or
categorical datasets?

Regards,

Sean MacEachern
Post by D. Rizopoulos
Indeed! GLMMadaptive::mixed_model is also more flexible in allowing users
to define their own mixed models by specifying the log-density of the
repeated measurements outcome, i.e., something similar to what Proc NLMIXED
https://cran.r-project.org/web/packages/GLMMadaptive/vignettes/Custom_Models.html
Best,
Dimitris
-----Original Message-----
Sent: Friday, June 15, 2018 7:57 PM
Subject: Re: [R-sig-ME] GLMMs with Adaptive Gaussian Quadrature
Good point. Extending AGQ to more complex models in lme4 is something
that's been on my list for a long time, but it's great to see someone
meeting the need. Even if I or someone does eventually get it working in
lme4, two implementations are always better than one ...
For those interested in this topic, there are a few other approaches to
improved frequentist estimates (i.e. without going full-Bayesian) that are
implemented in R: Helen Ogden's glmmsr package implements sequential
reduction and importance sampling methods, The glmm and bernor packages use
other flavors of importance sampling/MC likelihood approximations. glmmADMB
has importance sampling; TMB (the engine underlying glmmTMB) has an
importance-sampling method, but it hasn't
(yet) been integrated in glmmTMB ...
cheers
Ben Bolker
Post by D. Rizopoulos
AFAIK, lme4::glmer with nAGQ>1 *only* works for scalar random effects.
At least, when I try setting nAGQ > 1 for a random intercepts and random
Post by D. Rizopoulos
nAGQ > 1 is only available for models with a single, scalar random-effects term
GLMMadaptive::mixed_model implements the AGQ in such settings.
My main motivation to create this package is the longitudinal data
analysis setting in which including something more than random intercepts
is very typical. At least the students in my Repeated Measurements course (
https://github.com/drizopoulos/Repeated_Measurements) have had some
difficult times getting lme4::glmer() with a Laplace approximation to work
in such cases.
Post by D. Rizopoulos
-----Original Message-----
Sent: Friday, June 15, 2018 5:07 PM
Subject: Re: [R-sig-ME] GLMMs with Adaptive Gaussian Quadrature
It looks interesting (at an admittedly *very* quick initial glance).
Can you clarify how it differs from using lme4::glmer with nAGQ>1 ?
On Fri, Jun 15, 2018 at 10:26 AM, D. Rizopoulos <
Post by D. Rizopoulos
Dear R mixed-model users,
I’d like to announce the release of my new package GLMMadaptive for
fitting generalized linear mixed models using adaptive Gaussian
quadrature. You may read more about it here: https://goo.gl/7pi8Sh
Any comments or suggestions are more than welcome.
Best,
Dimitris
Professor of Biostatistics
Erasmus University Medical Center
The Netherlands
[[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
[[alternative HTML version deleted]]
D. Rizopoulos
2018-06-15 19:57:21 UTC
Permalink
No, this is not currently possible. I will need to think if it can be “easily” incorporated in the package…

Best,
Dimitris


From: Sean MacEachern <***@gmail.com>
Sent: Friday, June 15, 2018 9:37 PM
To: D. Rizopoulos <***@erasmusmc.nl>
Cc: ***@gmail.com; r-sig-mixed-***@r-project.org
Subject: Re: [R-sig-ME] GLMMs with Adaptive Gaussian Quadrature

Looks interesting. Would it be possible to fit a Numerator relationship matrix as a random effect similarly to MCMCglmm or Asreml for binary or categorical datasets?

Regards,

Sean MacEachern

On Fri, Jun 15, 2018 at 12:09 PM D. Rizopoulos <***@erasmusmc.nl<mailto:***@erasmusmc.nl>> wrote:
Indeed! GLMMadaptive::mixed_model is also more flexible in allowing users to define their own mixed models by specifying the log-density of the repeated measurements outcome, i.e., something similar to what Proc NLMIXED in doing in SAS. More info in the vignette: https://cran.r-project.org/web/packages/GLMMadaptive/vignettes/Custom_Models.html

Best,
Dimitris


-----Original Message-----
From: Ben Bolker <***@gmail.com<mailto:***@gmail.com>>
Sent: Friday, June 15, 2018 7:57 PM
To: D. Rizopoulos <***@erasmusmc.nl<mailto:***@erasmusmc.nl>>
Cc: r-sig-mixed-***@r-project.org<mailto:r-sig-mixed-***@r-project.org>
Subject: Re: [R-sig-ME] GLMMs with Adaptive Gaussian Quadrature

Good point. Extending AGQ to more complex models in lme4 is something that's been on my list for a long time, but it's great to see someone meeting the need. Even if I or someone does eventually get it working in lme4, two implementations are always better than one ...

For those interested in this topic, there are a few other approaches to improved frequentist estimates (i.e. without going full-Bayesian) that are implemented in R: Helen Ogden's glmmsr package implements sequential reduction and importance sampling methods, The glmm and bernor packages use other flavors of importance sampling/MC likelihood approximations. glmmADMB has importance sampling; TMB (the engine underlying glmmTMB) has an importance-sampling method, but it hasn't
(yet) been integrated in glmmTMB ...

cheers
Ben Bolker
Post by D. Rizopoulos
nAGQ > 1 is only available for models with a single, scalar
random-effects term
GLMMadaptive::mixed_model implements the AGQ in such settings.
My main motivation to create this package is the longitudinal data analysis setting in which including something more than random intercepts is very typical. At least the students in my Repeated Measurements course (https://github.com/drizopoulos/Repeated_Measurements) have had some difficult times getting lme4::glmer() with a Laplace approximation to work in such cases.
-----Original Message-----
Sent: Friday, June 15, 2018 5:07 PM
Subject: Re: [R-sig-ME] GLMMs with Adaptive Gaussian Quadrature
It looks interesting (at an admittedly *very* quick initial glance).
Can you clarify how it differs from using lme4::glmer with nAGQ>1 ?
Post by D. Rizopoulos
Dear R mixed-model users,
I’d like to announce the release of my new package GLMMadaptive for
fitting generalized linear mixed models using adaptive Gaussian
quadrature. You may read more about it here: https://goo.gl/7pi8Sh
Any comments or suggestions are more than welcome.
Best,
Dimitris
Professor of Biostatistics
Erasmus University Medical Center
The Netherlands
[[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
_______________________________________________
R-sig-mixed-***@r-project.org<mailto:R-sig-mixed-***@r-project.org> mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models

[[alternative HTML version deleted]]
D. Rizopoulos
2018-06-16 12:54:05 UTC
Permalink
Yes, this is in my plans to include in future versions of the package … For now I’m focusing on finalizing/extending the methods for the standard generics. Most notably, including subject-specific (dynamic) predictions with standard errors in the predict() method. The development version of the package is on my dedicated GitHub repo.

Best,
Dimitris


From: Christopher Stanley <***@yahoo.com>
Sent: Saturday, June 16, 2018 10:09 AM
To: Sean MacEachern <***@gmail.com>; D. Rizopoulos <***@erasmusmc.nl>
Cc: r-sig-mixed-***@r-project.org
Subject: Re: [R-sig-ME] GLMMs with Adaptive Gaussian Quadrature

Dimitris, the flexibility sounds great. Will this package allow users to specify zero-inflated (poisson/negative binomial) count data as well?

Best
Christopher

On Friday, June 15, 2018, 9:57:41 PM GMT+2, D. Rizopoulos <***@erasmusmc.nl<mailto:***@erasmusmc.nl>> wrote:


No, this is not currently possible. I will need to think if it can be “easily” incorporated in the package…

Best,
Dimitris


From: Sean MacEachern <***@gmail.com<mailto:***@gmail.com>>
Sent: Friday, June 15, 2018 9:37 PM
To: D. Rizopoulos <***@erasmusmc.nl<mailto:***@erasmusmc.nl>>
Cc: ***@gmail.com<mailto:***@gmail.com>; r-sig-mixed-***@r-project.org<mailto:r-sig-mixed-***@r-project.org>
Subject: Re: [R-sig-ME] GLMMs with Adaptive Gaussian Quadrature

Looks interesting. Would it be possible to fit a Numerator relationship matrix as a random effect similarly to MCMCglmm or Asreml for binary or categorical datasets?

Regards,

Sean MacEachern

On Fri, Jun 15, 2018 at 12:09 PM D. Rizopoulos <***@erasmusmc.nl<mailto:***@erasmusmc.nl><mailto:***@erasmusmc.nl<mailto:***@erasmusmc.nl>>> wrote:
Indeed! GLMMadaptive::mixed_model is also more flexible in allowing users to define their own mixed models by specifying the log-density of the repeated measurements outcome, i.e., something similar to what Proc NLMIXED in doing in SAS. More info in the vignette: https://cran.r-project.org/web/packages/GLMMadaptive/vignettes/Custom_Models.html

Best,
Dimitris


-----Original Message-----
From: Ben Bolker <***@gmail.com<mailto:***@gmail.com><mailto:***@gmail.com<mailto:***@gmail.com>>>
Sent: Friday, June 15, 2018 7:57 PM
To: D. Rizopoulos <***@erasmusmc.nl<mailto:***@erasmusmc.nl><mailto:***@erasmusmc.nl<mailto:***@erasmusmc.nl>>>
Cc: r-sig-mixed-***@r-project.org<mailto:r-sig-mixed-***@r-project.org><mailto:r-sig-mixed-***@r-project.org<mailto:r-sig-mixed-***@r-project.org>>
Subject: Re: [R-sig-ME] GLMMs with Adaptive Gaussian Quadrature

Good point. Extending AGQ to more complex models in lme4 is something that's been on my list for a long time, but it's great to see someone meeting the need. Even if I or someone does eventually get it working in lme4, two implementations are always better than one ...

For those interested in this topic, there are a few other approaches to improved frequentist estimates (i.e. without going full-Bayesian) that are implemented in R: Helen Ogden's glmmsr package implements sequential reduction and importance sampling methods, The glmm and bernor packages use other flavors of importance sampling/MC likelihood approximations. glmmADMB has importance sampling; TMB (the engine underlying glmmTMB) has an importance-sampling method, but it hasn't
(yet) been integrated in glmmTMB ...

cheers
Ben Bolker
Post by D. Rizopoulos
nAGQ > 1 is only available for models with a single, scalar
random-effects term
GLMMadaptive::mixed_model implements the AGQ in such settings.
My main motivation to create this package is the longitudinal data analysis setting in which including something more than random intercepts is very typical. At least the students in my Repeated Measurements course (https://github.com/drizopoulos/Repeated_Measurements) have had some difficult times getting lme4::glmer() with a Laplace approximation to work in such cases.
-----Original Message-----
Sent: Friday, June 15, 2018 5:07 PM
Subject: Re: [R-sig-ME] GLMMs with Adaptive Gaussian Quadrature
It looks interesting (at an admittedly *very* quick initial glance).
Can you clarify how it differs from using lme4::glmer with nAGQ>1 ?
Post by D. Rizopoulos
Dear R mixed-model users,
I’d like to announce the release of my new package GLMMadaptive for
fitting generalized linear mixed models using adaptive Gaussian
quadrature. You may read more about it here: https://goo.gl/7pi8Sh
Any comments or suggestions are more than welcome.
Best,
Dimitris
Professor of Biostatistics
Erasmus University Medical Center
The Netherlands
[[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
_______________________________________________
R-sig-mixed-***@r-project.org<mailto:R-sig-mixed-***@r-project.org><mailto:R-sig-mixed-***@r-project.org<mailto:R-sig-mixed-***@r-project.org>> mailing list

https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models

[[alternative HTML version deleted]]

_______________________________________________
R-sig-mixed-***@r-project.org<mailto:R-sig-mixed-***@r-project.org> mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models

[[alternative HTML version deleted]]

Loading...