R help - Mar 2004 - contrast lme and glmmPQL and getting additional results...

I have a longitudinal data analysis project.  There are 10 observations 
on each of 15 units, and I'm estimating this with randomly varying 
intercepts along with an AR1 correction for the error terms within 
units.  There is no correlation across units.  Blundering around in R 
for a long time, I found that for linear/gaussian models, I can use 
either the MASS method glmmPQL (thanks to Venables and Ripley) or the 
lme from nlme (thanks to Pinheiro and Bates).  (I also find that the 
package lme4 has GLMM, but I can't get the correlation structure to work 
with that, so I gave up on that one.)

The glmmPQL and lme results are quite similar, but not identical.

Here are my questions.

1. I believe that both of these offer consistent estimates. Does one 
have preferrable small sample properties?  Is the lme the preferred 
method in this case because it is more narrowly designed to this 
gaussian family model with an identity link?  If there's an argument in 
favor of PQL, I'd like to know it, because a couple of the Hypothesis 
tests based on t-statistics are affected.

2. Is there a pre-made method for calculation of the robust standard errors?

I notice that model.matrix() command does not work for either lme or 
glmmPQL results, and so I start to wonder how people calculate sandwich 
estimators of the standard errors.

3. Are the AIC (or BIC) statistics comparable across models?  Can one 
argue in favor of the glmmPQL results (with, say, a log link) if the AIC 
is more favorable than the AIC from an lme fit?  In JK Lindsey's Models 
for Repeated Measurements, the AIC is sometimes used to make model 
selections, but I don't know where the limits of this application might 
lie.


-- 
Paul E. Johnson                       email: pauljohn at ku.edu
Dept. of Political Science            http://lark.cc.ku.edu/~pauljohn
1541 Lilac Lane, Rm 504
University of Kansas                  Office: (785) 864-9086
Lawrence, Kansas 66044-3177           FAX: (785) 864-5700

On 20 Mar 2004 at 11:06, Paul Johnson wrote:
>From what you say, it seems like you have a linear normal (mixed) 
model. This is what lme is made for, and there is no reason to use 
glmmPQL (which calls lme iteratively).

Kjetil Halvorsen
> I have a longitudinal data analysis project.  There are 10
> observations on each of 15 units, and I'm estimating this with
> randomly varying intercepts along with an AR1 correction for the error
> terms within units.  There is no correlation across units.  Blundering
> around in R for a long time, I found that for linear/gaussian models,
> I can use either the MASS method glmmPQL (thanks to Venables and
> Ripley) or the lme from nlme (thanks to Pinheiro and Bates).  (I also
> find that the package lme4 has GLMM, but I can't get the correlation
> structure to work with that, so I gave up on that one.)
> 
> The glmmPQL and lme results are quite similar, but not identical.
> 
> Here are my questions.
> 
> 1. I believe that both of these offer consistent estimates. Does one
> have preferrable small sample properties?  Is the lme the preferred
> method in this case because it is more narrowly designed to this
> gaussian family model with an identity link?  If there's an argument
> in favor of PQL, I'd like to know it, because a couple of the
> Hypothesis tests based on t-statistics are affected.
> 
> 2. Is there a pre-made method for calculation of the robust standard
> errors?
> 
> I notice that model.matrix() command does not work for either lme or
> glmmPQL results, and so I start to wonder how people calculate
> sandwich estimators of the standard errors.
> 
> 3. Are the AIC (or BIC) statistics comparable across models?  Can one
> argue in favor of the glmmPQL results (with, say, a log link) if the
> AIC is more favorable than the AIC from an lme fit?  In JK Lindsey's
> Models for Repeated Measurements, the AIC is sometimes used to make
> model selections, but I don't know where the limits of this
> application might lie.
> 
> 
> -- 
> Paul E. Johnson                       email: pauljohn at ku.edu
> Dept. of Political Science            http://lark.cc.ku.edu/~pauljohn
> 1541 Lilac Lane, Rm 504 University of Kansas                  Office:
> (785) 864-9086 Lawrence, Kansas 66044-3177           FAX: (785)
> 864-5700
> 
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://www.stat.math.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide!
> http://www.R-project.org/posting-guide.html