Its not optimal to think about this
as analogous to a fixed-effects ANOVA
with 2000 levels of the explanatory.
Rather, a random-effects model is the
better approach.
The described application boils down
to a nominal logistic regression variance
components model. Workplaces are random.
The task is to model the nominal outcome
with a random intercept term. The estimated
variance of the intercept, u_0, is the
estimate of between-workplace variability.
The withinn workplace variabilty, e_0, is
likely to be derived under the assumption
of equi-dispersion (binomially distributed
errors) as pi**2/3. This approach is
possible using MIXNO a free software
program by Don Hedeker.
The ICC is then u_0/(u_0 + e_0)
Testing the assumption of equi-dispersion
for a random effects nominal logistic
regression model may be possible with MLwiN
(commercial sotware), but I am not sure.
HTH
Steve Gregorich
>The workplaces are randomly sampled. Selection of the individuals was based
>on a random sample design, but there are known biases.
=================================================================
Instructions for joining and leaving this list and remarks about
the problem of INAPPROPRIATE MESSAGES are available at
http://jse.stat.ncsu.edu/
=================================================================