DAfshartous at med.miami.edu
Mon Nov 9 14:50:39 CET 2009
For lme4, search the archives for "random effect variance per treatment group in lmer", I wrote a summary e-mail on this issue on 7/13/07. For details on variance modeling per group in nlme, see Pinheiro & Bates (2004; chapter 5); for a discussion of issue with respect to SAS, see Little et al (2006; chapter 9) (SAS for mixed models).
On 11/8/09 11:27 AM, "Emmanuel Charpentier" <emm.charpentier at free.fr> wrote:
Le dimanche 08 novembre 2009 à 00:01 +0000, Thomas Mang a écrit :
> Suppose may data consist of groups (which also define the levels for
> random effects), which show group-wise heteroscedasticity, that is for
> some groups the variance of residuals is larger than for the others.
> Based on specific knowledge of the data and the problem this even makes
> perfect sense and is actually a good sign. Technically however it's not
> good of course, to put it mildly.
> Is there a way in lme4 to handle heteroscedasticity (with known grouping
> for the different variances) ?
> Any suggestions ?
Well, you might try to equalize dependent variable variances by the
"classical" transformations (log, sqrt) and their generalizations
(Box-Cox nd siblings, see "boxcox" and "logtrans" in MASS). Lrger sets
of transformations of both dependents nd independent variables are
proposed in ace and avas, and Harrell's Design (now rms) package has lso
some functions aimed at this kind of problems.
Be aware, however, that most of these functions im t finding optimal
transformations in fixed-effect context, and that using them with
random-effects models is not necessarily a good solution.
R-sig-mixed-models at r-project.org mailing list
More information about the R-sig-mixed-models