[R-sig-ME] sandwich variance estimation using glmer?

Thu Nov 4 15:18:45 CET 2010

Let me push on this just a bit to spark further discussion. The OP was interested in robust standard errors given misspecification in the likelihood. So, one possible avenue was to explore Huber-White standard errors, or the sandwich estimator, to account for this misspecification and obtain "better" standard errors, but still use the point estimates of the fixed effects as given.

Some discussion on this has noted that the misspecification occurs in many ways, sometimes given that distributional assumptions were not met. Let's assume someone was willing and skilled to code up the HW as a possible solution within lmer to account for not meeting certain distributional assumptions.

My question is now why not directly code up models that permit for different distributional assumptions, such as t-distributions of residuals (random effects) or whatever the case might be? In other words, why not write code that addresses the problems directly (misspecification of the likelihood) rather than focusing on HW estimates.

Isn't it a better use of time and energy to focus on properly specifying the likelihood and estimating parameters from that model rather than HW?

> -----Original Message-----
> From: r-sig-mixed-models-bounces at r-project.org [mailto:r-sig-mixed-models-
> bounces at r-project.org] On Behalf Of Tyler Dean Rudolph
> Sent: Wednesday, November 03, 2010 6:07 PM
> To: Andrew Robinson
> Cc: r-sig-mixed-models at r-project.org
> Subject: Re: [R-sig-ME] sandwich variance estimation using glmer?
> 
> Hi Andrew,
> 
> Unfortunately I do not have access to SAS, so that is simply not an option
> for me, though I do welcome your clarification.  If this is a sensitive
> topic perhaps I will abstain from mentioning it in future, but to me it was
> a simple observation and not a value statement requiring qualification.
> 
> Perhaps I should put this another way: can anyone confirm that this
> functionality does NOT exist or is NOT presently being worked on somewhere
> within the R sphere?
> 
> Tyler
> 
> 
> 
> On Wed, Nov 3, 2010 at 5:49 PM, Andrew Robinson <
> A.Robinson at ms.unimelb.edu.au> wrote:
> 
> > Hi Tyler,
> >
> > I think that there's something that you're missing.
> >
> > R is not motivated by comparisons with SAS or any package.  So, your
> > impression that R was ahead of SAS or behind SAS is mistaken, or at
> > least, it's your impression, so you are responsible for it.  R
> > responds exactly to the community's needs because the community
> > supports it.  If the functionality that you want isn't there, it's
> > because noone else has wanted it badly enough to
> >
> > a) code it up, or
> >
> > b) pay someone else to code it up.
> >
> > If you want that function, and you know that SAS has it, then use SAS.
> > If you want to use that function in R, then see the above two points.
> >
> > Good luck,
> >
> > Andrew
> >
> >
> > On Wed, Nov 03, 2010 at 04:04:23PM -0400, Tyler Dean Rudolph wrote:
> > > Indeed, in this case the correlation structure of the random effects is
> > not
> > > fully appreciated or known, in which case the standard errors are likely
> > > underestimated.  The use of sandwich estimators should render variance
> > > estimates, and therefore inference, somewhat more realistic.  While this
> > is
> > > currently possible with GEEs, that approach does not ask the same
> > question
> > > as a GLMM (i.e. marginal or "population" estimates vs. conditional or
> > > "subject-specific" estimates).
> > >
> > > I used to think R updates were ahead of SAS upgrades in terms of new
> > > approaches but apparently that is often not the case.  Does anyone have
> > the
> > > know-how required to implement this in R, or is there something I'm still
> > > missing?
> > >
> > > Best,
> > > Tyler
> > >
> > >
> > > On Wed, Nov 3, 2010 at 9:12 AM, Dimitris Rizopoulos <
> > > d.rizopoulos at erasmusmc.nl> wrote:
> > >
> > > > On 11/3/2010 1:57 PM, Doran, Harold wrote:
> > > >
> > > >> Out of curiosity, why would you want a sandwich estimator from lmer?
> > That
> > > >> estimator is typically used when the likelihood is misspecified, but
> > you
> > > >> still want standard errors that account for correlations among units
> > within
> > > >> a cluster.
> > > >>
> > > >> Since this is what lmer standard errors already account for, is there
> > a
> > > >> need for the sandwich?
> > > >>
> > > >
> > > > well, it is possible that the random-effects structure that you have
> > > > specified is not the correct one (i.e., in order to fully account for
> > the
> > > > correlations), and in this case it makes sense to use the sandwich
> > estimator
> > > > (of course, the sandwich estimator has its own problems, but this is
> > > > probably another discussion...)
> > > >
> > > > Best,
> > > > Dimitris
> > > >
> > > >
> > > >
> > > >  -----Original Message-----
> > > >>> From: r-sig-mixed-models-bounces at r-project.org [mailto:
> > > >>> r-sig-mixed-models-
> > > >>> bounces at r-project.org] On Behalf Of Tyler Dean Rudolph
> > > >>> Sent: Tuesday, November 02, 2010 4:41 PM
> > > >>> To: r-sig-mixed-models at r-project.org
> > > >>> Subject: [R-sig-ME] sandwich variance estimation using glmer?
> > > >>>
> > > >>> Are there any current functionalities in R that permit estimation of
> > > >>> robust
> > > >>> sandwich variances based on lmer (mixed model) objects??  I'm aware
> > of
> > > >>> the
> > > >>> sandwich package and gee implementations but to my knowledge these
> > are
> > > >>> not
> > > >>> yet compatible with mixed model objects.
> > > >>>
> > > >>> Apparently these are already implemented in SAS....
> > > >>>
> > > >>> Tyler
> > > >>>
> > > >>>        [[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
> > > >>
> > > >>
> > > > --
> > > > Dimitris Rizopoulos
> > > > Assistant Professor
> > > > Department of Biostatistics
> > > > Erasmus University Medical Center
> > > >
> > > > Address: PO Box 2040, 3000 CA Rotterdam, the Netherlands
> > > > Tel: +31/(0)10/7043478
> > > > Fax: +31/(0)10/7043014
> > > > Web: http://www.erasmusmc.nl/biostatistiek/
> > > >
> > >
> > >       [[alternative HTML version deleted]]
> > >
> > > _______________________________________________
> > > R-sig-mixed-models at r-project.org mailing list
> > > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> >
> > --
> > Andrew Robinson
> > Program Manager, ACERA
> > Department of Mathematics and Statistics            Tel: +61-3-8344-6410
> > University of Melbourne, VIC 3010 Australia               (prefer email)
> > http://www.ms.unimelb.edu.au/~andrewpr              Fax: +61-3-8344-4599
> > http://www.acera.unimelb.edu.au/
> >
> 
> 	[[alternative HTML version deleted]]
> 
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models