Skip to content

Multilevel Modeling using R

3 messages · WONG, Ka Yau, Doran, Harold, Stas Kolenikov

Doran, Harold
# 
That's by intent, by the way. The standard errors of the variance
components are only useful if the distribution is symmetric, and this is
not always true. If you were using lmer, and not lme, then you could use
the mcmcsamp function to look at the distribution of the random effects
to see if it is approximately normal. But, there are other ways to check
and see if retaining the variance component is necessary or not, such as
LRT statistics.
Stas Kolenikov
# 
In most biometric applications, those variances are treated as
nuisance parameters. They only need to be controlled for, while the
main purpose is to get the right point estimates and standard errors
for the fixed effects. In social science multilevel modeling (of which
education is probably the heaviest user), the variances usually mean
something, so there is interest in conducting inference on them (as
you probably want to do). As noted by Harold Doran, whatever you do
with these random effects is quite sensitive to their distributions.
Getting the standard errors on those variances usually comes from
assuming a particular model such as the normal one.

What you do looks more like ANOVA to me. So you can use aov() to get
some F-statistics on your within- and between-school variability.
On 3/17/09, WONG, Ka Yau <kayau at ied.edu.hk> wrote: