[R] Robust regression with groups
secchi at sssup.it
Thu Oct 21 16:57:56 CEST 2004
Bert you are definitely right I've been confuse
and unclear on the nature of my problem (sorry about that).
In my message "robust regression" was referred to techniques able to
deal (when you estimate the variance of your coefficients) with
departures from the set of assumptions in a standard linear regression,
like for example the presence of heteroskedaciticy. In this case the
robust estimator of the variance of \beta (i.e. the coefficients) is
obtained considering a correction that take into account the
contribution from each observation to the score(d(ln L)/d\beta). Now I
would like to consider also the possibility that observations are not
independent as they are but they can be divided into groups that are
independent. In this case to obtain an estimator for the variance
that take into account this departure from the standard assumptions I
need a correction that take into account the contribution of each group
(and not of each observation) to the score(d(ln L)/d\beta). In summary,
I do not need more sophisticated way to estimate my coefficients but
only a routine to obtain a meaningful estimate for the variance of them.
Does this routine already exist in R?
PS Thanks Dimitris but it seems that I cannot use a random effects model
since the Hausmann specification test casts doubt on the assumptions
justifying the use of a GLS estimator.
On Thu, 21 Oct 2004 12:11:22 +0200
"Dimitris Rizopoulos" <dimitris.rizopoulos at med.kuleuven.ac.be> wrote:
> Hi Bert,
> Regarding the sensitivity in the choice of the random-effects
> distribution, I know that usually estimates of the fixed-effects (and
> their std.errors) do not have serious problems, even if you assume
> normality where in fact you have log-normality. However, you do have a
> problem in the EB estimates of the random-effects.
> More info could be found in:
> G. Verbeke and E. Lesaffre (1996). A linear mixed-effects model with
> heterogeneity in the random-effects population, JASA, 91, 217-221.
> W. Ghidey, E. Lesaffre and P. Eilers (2004). Smooth random-effects
> distribution in linear mixed model, Biometrics, 60, 945-953.
> (to appear in December)
> Dimitris Rizopoulos
> Ph.D. Student
> Biostatistical Centre
> School of Public Health
> Catholic University of Leuven
> Address: Kapucijnenvoer 35, Leuven, Belgium
> Tel: +32/16/396887
> Fax: +32/16/337015
> Web: http://www.med.kuleuven.ac.be/biostat/
> ----- Original Message -----
> From: "Berton Gunter" <gunter.berton at gene.com>
> To: "'Dimitris Rizopoulos'" <dimitris.rizopoulos at med.kuleuven.ac.be>;
> "'Angelo Secchi'" <secchi at sssup.it>
> Cc: <r-help at stat.math.ethz.ch>
> Sent: Wednesday, October 20, 2004 5:42 PM
> Subject: RE: [R] Robust regression with groups
> > Angelo and Folks:
> > Beware! It is not at all clear what you mean by "robust" regression.
> > The
> > sandwich estimator is often said to be "robust" to model
> > misspecification in
> > the sense that it converges to the correct covariance matrix whether
> > or not
> > the correlation structure in the GEE has been correctly specified
> > (as
> > Dmitris implied). Is this what you mean? Mixed effect models are
> > often said
> > to be "robust" in the sense that individual group "estimators"
> > (blups) are
> > shrunk toward the overall fixed effect estimates. Is this what you
> > mean?
> > In other applications, "robustness" can mean insensitivity to
> > distributional
> > assumptions. Mixed effects models for continupus responses commonly
> > assume
> > normality (as the estimates solve likelihood equations), as do
> > GLMM's for
> > the random effects. I know of no definitive work that has examined
> > sensitivity of estimates (or inferences, which are, at best,
> > asymptotic
> > anyway) to those assumptions. (in the simple independent errors
> > case, it is
> > usually the case that estimates are not at all sensitive). However,
> > I am a
> > novice here, so others may be able to illuminate the issue more.
> > Finally, "robustness" is often used to mean "outlier resistance."
> > Here the
> > situation is yet murkier. Do you mean resistance to individual
> > "outlying"
> > observations within a subject or resistance to outlying subjects?
> > Shrinkage
> > should help with both, but, again, I know of no definitive work,
> > especially
> > regarding resistance to individual extreme values. Given the
> > sensitivity of
> > covariance estimates to heavy tails and the consequent inferential
> > inefficiency, this presumably could be a problem. Finding methods
> > that could
> > deal with this may be nearly impossible, as you are adding yet
> > another layer
> > of nonlinear estimation (that of determining optimal case
> > weights/parameters
> > for mixture contamination models/or whatever...) to the problem; it
> > is easy
> > to come up with examples where the data are inherently ambiguous and
> > parameter estimates for resistant case weights and the model would
> > trade off
> > with each other depending on starting values. That is, too many
> > nonlinear
> > parameters are being estimated and the model estimates are therefore
> > unstable.
> > Again, I am happy to leave more definitive resolution and correction
> > of any
> > errors in my comments to the experts, but, at the least, I think you
> > need to
> > think more and communicate more clearly about what you mean by
> > "robust."
> > Cheers,
> > -- Bert Gunter
> > Genentech Non-Clinical Statistics
> > South San Francisco, CA
> > "The business of the statistician is to catalyze the scientific
> > learning
> > process." - George E. P. Box
> >> -----Original Message-----
> >> From: r-help-bounces at stat.math.ethz.ch
> >> [mailto:r-help-bounces at stat.math.ethz.ch] On Behalf Of
> >> Dimitris Rizopoulos
> >> Sent: Wednesday, October 20, 2004 7:08 AM
> >> To: Angelo Secchi
> >> Cc: r-help at stat.math.ethz.ch
> >> Subject: Re: [R] Robust regression with groups
> >> Hi Angelo,
> >> There are two possible options (at least to my knowledge):
> >> 1. to use a random-effects model, either using `lme' (packages:
> >> nlme,
> >> lme4) if you have normal data or `glmmPQL' (package: MASS) or
> >> `GLMM'
> >> (package: lme4) or `glmmML' (package:glmmML) if you cannot use the
> >> normal distribution.
> >> 2. to use a gee model with a robust (sandwich) std.error
> >> estimation.
> >> See at `gee' (package: gee) and `geese' (package: geepack).
> >> I hope this helps.
> >> Best,
> >> Dimitris
> >> ----
> >> Dimitris Rizopoulos
> >> Ph.D. Student
> >> Biostatistical Centre
> >> School of Public Health
> >> Catholic University of Leuven
> >> Address: Kapucijnenvoer 35, Leuven, Belgium
> >> Tel: +32/16/396887
> >> Fax: +32/16/337015
> >> Web: http://www.med.kuleuven.ac.be/biostat/
> >> http://www.student.kuleuven.ac.be/~m0390867/dimitris.htm
> >> ----- Original Message -----
> >> From: "Angelo Secchi" <secchi at sssup.it>
> >> To: <r-help at stat.math.ethz.ch>
> >> Sent: Wednesday, October 20, 2004 3:22 PM
> >> Subject: [R] Robust regression with groups
> >> >
> >> >
> >> > Hi,
> >> > I have data on a group of subjects in different years. I should
> >> > assume
> >> > that observations regarding different individuals are independent
> >> > but
> >> > observations for the same individual in different years are not
> >> > and
> >> > I
> >> > would like to have an estimated standard error (and
> >> > variance-covariance
> >> > matrix) taking into account this problem.
> >> >
> >> > More in general is there a way in R to run a (robust)regression
> >> > having
> >> > different groups in the observations and specifying that the
> >> > observation
> >> > are independent across groups but not necessarily independent
> >> > within
> >> > groups?
> >> >
> >> > Thanks
> >> > a.
> >> >
> >> > ______________________________________________
> >> > R-help at stat.math.ethz.ch mailing list
> >> > https://stat.ethz.ch/mailman/listinfo/r-help
> >> > PLEASE do read the posting guide!
> >> > http://www.R-project.org/posting-guide.html
> >> >
> >> ______________________________________________
> >> R-help at stat.math.ethz.ch mailing list
> >> https://stat.ethz.ch/mailman/listinfo/r-help
> >> PLEASE do read the posting guide!
> >> http://www.R-project.org/posting-guide.html
Angelo Secchi PGP Key ID:EA280337
Research Fellow Scuola Superiore S.Anna
Piazza Martiri della Liberta' 33, Pisa, 56127 Italy
ph.: +39 050 883365
email: secchi at sssup.it www.sssup.it/~secchi/
More information about the R-help