Have you tried the "Centre for Multilevel Modelling" at the
University of Bristol (www.mlwin.com)? This web site includes a
benchmark set of data sets and separate reviews of capabilities of
different "MLM" software with answers and timing. I found no unified
comparison table, but the information required to construct your own
seems to be there.
I have not used SAS in many years, but from what I've heard, I
believe the output will be quite similar in most but not all cases.
If you have access to SAS, could you please try the following silly
little test case for me:
tstDF <- data.frame(group=letters[1:5], y=1:5)
fitOops <- lme(y~1, data=tstDF, random=~1|group)
VarCorr(fitOops)
var(tstDF$y)
In S-Plus 6.2 and R 2.3.1 (with nlme 3.1-75), I get the following:
> VarCorr(fitOops)
group = pdLogChol(1)
Variance StdDev
(Intercept) 2.1917808 1.4804664
Residual 0.3082192 0.5551749
> var(tstDF$y)
[1] 2.5
Using lmer (with lme4 0.995-2 and Matrix 0.995-12), I get the
following:
fitOops4 <- lmer(y~1+(1|group), data=tstDF)
Groups Name Variance Std.Dev.
group (Intercept) 1.81818 1.34840
Residual 0.68182 0.82572
This is, of course, silly, because there are zero degrees of freedom
to distinguish "group" from Residual. It is comforting that the sum of
the variances sum to the variance of "y", though the split between
"group" and Residual is different. However, I would prefer to have the
multilevel software catch this case and optionally return an error or
drop the redundant group with a warning.
Hope this helps.
Spencer Graves
Neung-Hwan Oh wrote:
Hi.
Is there someone who has used both S-Plus lme and SAS repeated measures
anova? Is there a good reference which compared the both methods? I am
wondering whether the output would be exactly the same for an analysis
with repeated measures.
Thanks in advance.
NH
|