lme correlation grouped by sub when random effects are not sub-specific

[<Jeri.Forster@UCHSC.edu>]

Dear S-World,

 

Thanks in advance for any help.

 

I am using Splus 7.0 for Windows and depending on the computer, either

Windows 2000 or Windows XP.  My short question is, using the correlation

argument in lme, how do you get a block diagonal (multiple of an identity) error

covariance structure, blocked by subject, when the random effects are not

subject specific?

 

The long version is as follows.  I have a mixed model that is run with this

code in SAS v9.1:

 

proc mixed data="">

class z;

model y=x1 x2 x1*x2;

random w1-w3/type=toep(1);

random ww1-ww3/type=toep(1);

repeated / subject=z;

run;

 

The 'w' variables in the random statements are not subject-specific

(hence no 'subject=' argument) which is why the repeated statement is

added - to tell SAS how to distinguish the subjects and this gives

a block diagonal covariance structure for the errors.  I can't seem to

replicate this exactly in Splus.  I have tried the following:

 

fit_lme(fixed=y~x1*x2,

           random=pdBlocked(list(pdIdent(~w-1),pdIdent(~ww-1))),
           correlation=corBand(~1 | z,ord=1),

           data="">

 

I am comfortable that the random effects are correct and they are

not grouped by subject.  The corBand is a banded structure, but I

believe that with ord=1, it is the multiple of an identity. 

 

Although the above code is what the Splus help suggests that I use

(correlation=corBand(~1 | z,ord=1)),  this doesn't seem to group it by 

subject.  The reason that I don't think it is grouping by subject is that

in summary(fit) it shows:

 

Correlation Structure: Independent

Formula:  ~ 1 | 1

 

and I would think that it should say Formula: ~1 | z. 

 

Any tricks or ideas??

 

Thanks to anyone who can help!

Jeri Forster

University of Colorado Health Sciences Center

4200 E 9^th Ave

Denver, CO 80262