[R-sig-ME] lme with heteroscedacity- how to plot within-person variability

[Emmanuel Curis]

Hello,

On Fri, May 04, 2012 at 11:31:18AM +0000, Rapsomaniki, Eleni wrote:
« Finally, am I right to think that heteroscedastic variances are
« still not possible with lme4?

I'm still discovering mixed models formalism, and its application in
nlme and lme4, so I would appreciate to be corrected if I am wrong,
but I was guessing that heteroscedasticty between groups can also be
seen as an interaction between the random effect and the fixed effect
defining the group --- but I am not sure if it always apply for more
complex models, or models with other structures than the one I use
currently?

I assume random effect is for variable « subject » and the group is
defined as a fixed effect factor in the variable « group » with only a
few levels.

If so, it may be used in lme4 as (1 + group|subject) in lme4. To get
rid of the correlation however seems more difficult, if one only wants
heteroscedaticity: I tried to use (1|subject) + (0+group|subject), but
it seems to remove only the correlation between the base level and
others, not between other levels. I guess one has to manually define
the indicator variables for each level and do something like

(1|subject) + (is_level_2|subject) + (is_level_3|subject) + ... 

to have full removal or correlations.

PS: the model I have in mind, in case my ideas are not generalizable,
  to check it is a least correct in that case, is a model with three
  factors :

 - genotype (fixed, 3 levels)
 - visit (fixed, 4 levels)
 - patient (random)

with patient nested in genotype and only one measure for each
(patient, visit, genotype) set; my idea was that assuming different
variances between genotypes was formally equivalent to add a
genotype:patient interaction, and so on...

Any comment really appreciated,

-- 
                                Emmanuel CURIS
                                emmanuel.curis at univ-paris5.fr

Page WWW: http://emmanuel.curis.online.fr/index.html