Le vendredi 27 mars 2009 ? 15:30 -0700, mrint a ?crit :
Hi, I'm having a problem using LME and hopefully I can get some help.
Here is what I'm doing:
I have a model like this:
yi = alpha + theta * xi + beta * zi + error, errors are normally
distributed mean 0, var sigma^2
xi and zi are generated from normal distributions within a specified
range.
True values of alpha, theta, beta and sigma^2 are chosen with a
specific mean and variane tau^2.
I have a function which generates the yi from the other variables,
then a function that does a linear regression using lm to create the
estimates.
I then want to use this data to do a meta-anaylsis with the above
repeated between 10-20 times. Within this, I want to use lme to
create estimates for the average true value, sample mean and average
standard error for alpha, theta, beta and the respective tau^2 values
for each of these. For the lme part, I'm using this
a<-summary(lme(alp~1,random=~1|alp, weights=varFixed(~staalp^2)))
This is the one for alpha. This isn't producing the type of results I
would expect, can anyone see where I'm going wrong?
(I suppose that your simulation aims to assess a specific model)
This, and closely related subjects, have already been discussed on this
very list. To make a long story short : lme doesn't (currently) accepts
means and variances of groups as an input, it needs individual data.
Someone (that should be Wolfgang Vischbauer, but I'm almost surely
mutilating his name's spelling ; apologies, Wolfgang !) has written,
specifically for meta-regression purposes, a "mima" function that does
what you're requesting. Wolfgang has stated his intentions to turn this
function into a full-fledged R package (with calling conventions closer
to what other regression functions use), but the "mima" function
available on his site still his 2 years old 0.4 version. For further
details, look for "mima" or for "meta-regression" in the list archives.
RSiteSearch() is your friend...
However, if what you're interested with is strictly speaking a
meta-analysis of 2-samples comparisons (i. e. your theta is scalar and
your x_i are logicals), (at least) two R packages available on CRAN are
built for this purpose : rmeta and meta. Both offer separate analyses for
boolean or continuous dependent variables (i. e. y_i logical or
continuous).
If your theta is scalar but your x_i is continuous (i. e. you're
meta-analysing a single regression coefficient), both package offer a
variant for meta-analysis of effects, that might be relevant for you.
A more general solution would be to enhance the forthcoming lme4 package
to accept an alternative specification of random effects
variances-covariances, which would allow "general" meta-regression. But I
understand that Douglas Bates has already way too much work and not too
much time on his hands, and I doubt he might be coaxed to work in this
direction right now...
A suggestion : you might forward your question to the "r-mixed-models"
SIG mailing list with some profit...
Emmanuel Charpentier