[R-sig-ME] Robust SEs in GLMMs

Sharon Poessel sharpoes at gmail.com
Wed Nov 26 04:41:52 CET 2014


Thank you to all of you who have responded.  I will consider your
suggestions.  The information is very much appreciated!

Sharon

On Tue, Nov 25, 2014 at 4:08 PM, Ken Beath <ken.beath at mq.edu.au> wrote:

> It is not too hard to set up a cluster bootstrap (i.e. clusters are
> sampled) for clustered data, although I'm not certain what happens with
> spatial data. It then takes a bit of time to run.
>
> I've used it where I had additional possible random effects and was
> concerned about fitting models with that many random effects. Nicely,
> results for models with 5 random effects fitted with nlme gave the same
> results for the fixed effects as the 3 random effect models with nlme with
> the clustered bootstrap.
>
> On 26 November 2014 at 09:10, Ben Bolker <bbolker at gmail.com> wrote:
>
>> I hate to sound contrary, but ... I actually think that implementing
>> the robust standard errors would be the best way to go here.  I don't
>> have time to work on it myself right now, but someone reasonably
>> experienced in R should be able to look at the `sandwich` package and
>> figure out how to write "bread" and "meat" methods for `merMod`
>> objects ...
>>
>> On Tue, Nov 25, 2014 at 2:11 PM, Tim Meehan <tmeeha at gmail.com> wrote:
>> > Thanks for clarifying the problem with correlation functions and binary
>> > responses, Doug.  Regarding the random effects approach, how would one
>> set
>> > that up?  Would you divide the data into spatial or temporal blocks, and
>> > use the blocks in the random statement, for example?
>> >
>> > On Tue, Nov 25, 2014 at 11:16 AM, Douglas Bates <bates at stat.wisc.edu>
>> wrote:
>> >
>> >> You have to be careful when modeling auto-correlation in a binary
>> >> response.  When using a Gaussian distribution it is possible to model
>> the
>> >> variance and correlation separately from the mean.  No so for a
>> Bernoulli
>> >> distribution (or binomial or Poisson).  In some sense the whole
>> purpose of
>> >> generalized linear models is to take into account that the variance of
>> each
>> >> response is determined by its mean in these distributions.
>> >>
>> >> glmmPQL is a wrapper around the lme function from the nlme package. But
>> >> lme, which provides for modelling correlations, was not intended for
>> this
>> >> purpose.  I personally don't think it would make sense to use a
>> correlation
>> >> function with a binary response.
>> >>
>> >> A preferred approach is to incorporate Gaussian-distributed random
>> effects
>> >> that have the desired auto-correlation pattern.
>> >>
>> >>
>> >> On Tue Nov 25 2014 at 11:58:08 AM Tim Meehan <tmeeha at gmail.com> wrote:
>> >>
>> >>> Hi Sharon,
>> >>>
>> >>> I just looked over a paper by Bolker et al. (2008. GLMMs: a practical
>> >>> guide
>> >>> for ecology and evolution. TREE).  Turns out that while it is
>> possible to
>> >>> model binary data with glmmPQL, it's not really recommended.
>> Nonetheless,
>> >>> you might look for other options that involve modeling autocorrelation
>> >>> rather than correcting for it after the fact.
>> >>>
>> >>> Best,
>> >>> Tim
>> >>>
>> >>>
>> >>> On Tue, Nov 25, 2014 at 10:19 AM, Tim Meehan <tmeeha at gmail.com>
>> wrote:
>> >>>
>> >>> > Hi Sharon,
>> >>> >
>> >>> > Take a look at glmmPQL in the MASS package.  This function allows
>> you to
>> >>> > model a binary response, with random effects, and temporally and
>> >>> spatially
>> >>> > correlated errors.  If you model the correlations, there is less of
>> a
>> >>> need
>> >>> > for adjusting standard errors.
>> >>> >
>> >>> > Best,
>> >>> > Tim
>> >>> >
>> >>> >
>> >>> > On Sun, Nov 23, 2014 at 2:04 PM, Sharon Poessel <sharpoes at gmail.com
>> >
>> >>> > wrote:
>> >>> >
>> >>> >> When computing resource selection functions for animal telemetry
>> data
>> >>> with
>> >>> >> a binary response variable, where the 1s represent animal location
>> >>> data,
>> >>> >> which are spatially and temporally correlated, and the 0s represent
>> >>> random
>> >>> >> locations, which are not correlated, it is recommended to calculate
>> >>> >> robust,
>> >>> >> or empirical, standard errors instead of using the model-based
>> standard
>> >>> >> errors to account for this differing correlation structure.  As
>> far as
>> >>> I
>> >>> >> can tell, none of the glmm packages in R calculate these robust
>> SEs.
>> >>> Does
>> >>> >> anyone know of a way to use glmms that calculate these?  Thanks.
>> >>> >>
>> >>> >> Sharon
>> >>> >>
>> >>> >>         [[alternative HTML version deleted]]
>> >>> >>
>> >>> >> _______________________________________________
>> >>> >> R-sig-mixed-models at r-project.org mailing list
>> >>> >> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>> >>> >>
>> >>> >
>> >>> >
>> >>>
>> >>>         [[alternative HTML version deleted]]
>> >>>
>> >>> _______________________________________________
>> >>> R-sig-mixed-models at r-project.org mailing list
>> >>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>> >>>
>> >>
>> >
>> >         [[alternative HTML version deleted]]
>> >
>> > _______________________________________________
>> > 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
>>
>
>
>
> --
>
> *Ken Beath*
> Lecturer
> Statistics Department
> MACQUARIE UNIVERSITY NSW 2109, Australia
>
> Phone: +61 (0)2 9850 8516
>
> Building E4A, room 526
> http://stat.mq.edu.au/our_staff/staff_-_alphabetical/staff/beath,_ken/
>
> CRICOS Provider No 00002J
> This message is intended for the addressee named and m...{{dropped:11}}



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