## Random effects redux

### Random effects redux

Earlier today I thought I posted this note, but have not seen any response.
At the risk of being a pesk, I've reposted the message.

thanks
bill

I am trying to run a one way anova with a single random effect.
The independent variable (Group) has three levels.
Nested within each Group are 4 to 5 sites with about 50 - 60 subjects in
each site.  Thus I have about 15 sites and an total N of about 1000

I am considering site as a random variable because we have randomly selected
(and assigned to conditions) these sites from a
larger sampling frame.

I am using the following syntax.

UNIANOVA
wgt_tot   BY group site
/RANDOM = site
/METHOD = SSTYPE(3)
/INTERCEPT = INCLUDE
/EMMEANS = TABLES(group)
/EMMEANS = TABLES(site)
/EMMEANS = TABLES(group*site)
/PRINT = DESCRIPTIVE
/CRITERIA = ALPHA(.05)
/DESIGN = group site group*site .

The output shows zero degrees of freedom for my "group" variable.  I wonder
if anyone has suggestions about what I am doing wrong.
(I have run the same syntax without the interaction and have the same problem).

TIA

Bill

William N Dudley, PhD
Asst Professor
Dept Behavi*Science and Health Education
Rollins School of Public Health
Emory University
Atlanta 30322
Phone 404 727 2447
FAX   404 727 1369

### Random effects redux

I'm not sure but from looking at your syntax it appears that you are asking for
a
fully crossed factorial design with site crossed with group, rather than nested
within group.

Is it possible to specify a random effect model without estimating a
residual variance?

With GPA measured on six regularly-spaced occasions, I want to treat
occasion as a random effect with an unstructured G matrix and a null R
matrix.  The log looks like this:

118  proc mixed covtest;
119  class occas;
120  model gpa=occas /noint solution;
121  random occas /subject=student type=un g gcorr;
122  run;

NOTE: Convergence criteria met.
NOTE: Estimated G matrix is not positive definite.

This makes sense because the estimated residual variance and one of
the elements of G are redundant. I would like to fix the residual
variance at zero.  Is this possible?  If so, how?

I realize that this model can be easily estimated using the REPEATED
subcommand with an unstructured R matrix and a null G matrix, but I
want to it the other way around.

Any help would be greatly appreciated.

Thanks

--Joe