[R] within group variance of the coeficients in LME

[Andrej Kveder]

Firstly let me thank all for your answers and suggestions. I would like to
followup on your comments.
Currently I'm implementing multilevel models within a simulation run. So I
was looking for an estimate weather a covariate varies across second level
units or not. I was using HLM before so I knew about the test implemented in
the package and tried to reporoduce it in R. However I think I can follow
your logic about not using that.

I would appreciate your reflection on the following. I need a quantitative
figure to evaluate weather the covariate varies across second level units in
the process of simulation. Of course I will be running thousands of them and
would need to program the condition in code. In one of the previous
questions to the group dr. Bates suggested to use the CI estmates, however
he warned me about their very conservative nature (I got the same tip from
the book). I thought about using the lower bound of the CI as an estimate
with the rule "if above 0 then the covariate varies". Would that be a sound
think to do? Do you have any other suggestions? I would really appreciate
the feedback.

thanks

Andrej



-----Original Message-----
From: Douglas Bates [mailto:bates at bates4.stat.wisc.edu]On Behalf Of Douglas
Bates
Sent: Monday, June 30, 2003 6:09 PM
To: J.R. Lockwood
Cc: Harold Doran; R-Help; Andrej Kveder
Subject: Re: [R] within group variance of the coeficients in LME


"J.R. Lockwood" <lockwood at rand.org> writes:

> >
> > 	Dear listers,
> >
> > 	I can't find the variance or se of the coefficients in a multilevel
model
> > 	using lme.
> >
>
> The component of an lme() object called "apVar" provides the estimated
> asymptotic covariance matrix of a particular transformation of the
> variance components. Dr. Bates can correct me if I'm wrong but I
> believe it is the matrix logarithm of Cholesky decomposition of the
> covariance matrix of the random effects.  I believe the details are in
> the book by Pinheiro and Bates.  Once you know the transformation you
> can use the "apVar" elements to get estimated asympotic standard
> errors for your variance components estimates using the delta method.
>
> J.R. Lockwood
> 412-683-2300 x4941
> lockwood at rand.org
> http://www.rand.org/methodology/stat/members/lockwood/

First, thanks to those who answered the question.  I have been away
from my email for about a week and am just now catching up on the
r-help list.

As I understand the original question from Andrej he wants to obtain
the standard errors for coefficients in the fixed effects part of the
model.  Those are calculated in the summary method for lme objects and
returned as the component called 'tTable'.  Try

library(nlme)
example(lme)
summary(fm2)$tTable

to see the raw values.

Other software for fitting mixed-effects models, such as SAS PROC
MIXED and HLM, return standard errors along with the estimates of the
variances and covariances of the random effects.  We don't return
standard errors of estimated variances because we don't think they are
useful.  A standard error for a parameter estimate is most useful when
the distribution of the estimator is approximately symmetric, and
these are not.

Instead we feel that the variances and covariances should be converted
to an unconstrained scale, and preferably a scale for which the
log-likelihood is approximately quadratic.  The apVar component that
you mention is an approximate variance-covariance matrix of the
variance components on an unbounded parameterization that uses the
logarithm of any standard deviation and Fisher's z transformation of
any correlations.  If all variance-covariance matrices being estimated
are 1x1 or 2x2 then this parameterization is both unbounded and
unconstrained.  If any are 3x3 or larger then this parameterization
must be further constrained to ensure positive definiteness.
Nevertheless, once we have finished the optimization we convert to
this 'natural' parameterization to assess the variability of the
estimates because these parameters are easily interpreted.

The actual optimization of the profiled log-likelihood is done using
the log-Cholesky parameterization that you mentioned because it is
always unbounded and unconstrained.  Interpreting elements of this
parameter vector is complicated.

I hope this isn't too confusing.