[R] Meta-regression with lmer() ? If so, how ?

[Emmanuel Charpentier]

Dear List,

I am (again) looking at meta-regression as a way to refine meta-analytic
results. What I want to do is to assess the impact of some fixed factors
on the results of a meta-analysis. Some of them may be crossed with the
main factor of the meta-analysis (e. g. clinical presentation of a
disease, defining subgroups in each of the studies under analysis), some
of them may be a grouping factor amond studies (e. g. study design
characteristics).

Homework : the R packages meta and Rmeta do not allow for multiple
independent factors. Looking hard at R-help archives, I found one of my
previous posts on this subject (which got me an answer suggesting using
lme()), a discussion of lme() use concluding that the R version of lme()
could not be used for this purpose, and a 2002 discussion ending up with
a custom function for (part of) this purpose (lockwood at rand.org, Sat 31
Aug 2002 - 06:31:11 EST).

Now, lmer() introduced a new method of specification of the variance
sources (it has also the advantage of being usable with various models,
such as linear, binomial or Poisson). However, I have been unable to
extract from lmer() documentation a way to specify variances associated
with each individual mean. I might also use some help for the
specification of variance structures.

In short, what I want to be able to do is to write something along the
lines of :

analysis<-foo(
              Outcome~Treatment*(Presentation+Design),
              random=(1|Study %in% Design),
              sd=Sds
              data=RawData, ...)

where RawData is a data frame with a line for each level of
Study*Treatment*Presentation, each study having only one value for
Design ; Outcome is the mean value of the judgement criterion in the
corresponding subgroup of the study, and Sds is its standard deviation.

The results of interests are mainly the effects of
	- Treatment (of course)
	- Treatment:Presentation (of great clinical interest)
	- Treatment:Design (is there evidence of bias introduced by
      	  study design ?)

Of secondary interest are :
	- Presentation (i. e. natural history of the disease)
	- Design (How should future studies be done ?)

Any hint on writing foo() using lmer() (or otherwise) and using the
resulting object would be very much appreciated.

					Emmanuel Charpentier

PS : a CC to charpent at bacbuc.dyndns.org would also be appreciated : I'm
reading the list through the web archives.