Constraining the variance to zero.

[RasulGro]

Hello,

I'm currently modelling a three level growth curve model with a linear and quadratic effect of time.
I followed the instructions of chapter 13 in the MLWin guide, however my is more complexe due to the inclusion of three instead of two levels.
Now, the linear effect of time is random on all three levels, the the quadratic effect is random on level 3 and 2. I want to test, whether I should include random effects for the quadratic term both levels or if I can exclude them.
For this purpose, is it sufficient if I seperately constrain the variance of the quadratic term to zero and compare the -2loglikelihood with the unconstrained model? Or do I have to constrain all associated covariances as well?

Kind regards,
Rasul

[Harvey Goldstein]

My view would be that all associated covariances should be constrained to zero also.
harvey

[ghudson621]

You could constrain the co-variance to zero.