R help - Oct 2012 - LMMs with some variance terms forced constant

Hello,

I have been asked to help perform a meta-analysis with individual- and
aggregate-level data. I found a nice article on this, and the idea is easy to
understand: use a mixed effects models, but for the studies where there is only
aggregate level data, force the variance to be that which was observed.
Unfortunately, I am struggling to see how to implement this in R. I am familiar
with "standard" mixed model formulations using nlme and lmer, but not
able to see how to force this specific variance structure on my model so that
some variance terms are not estimated.

Thanks for your time and any advice.

Best wishes,
Erica



Erica Moodie
Associate Professor, Biostatistics
Biostatistics Program Director
Department of Epidemiology, Biostatistics, & Occupational Health
McGill University
http://www.biostat.mcgill.ca/moodie


-------------- next part --------------
A non-text attachment was scrubbed...
Name: Meta-analysis of continuous outcomes combining individual patient data and
aggregate data-StatMed.pdf
Type: application/pdf
Size: 202242 bytes
Desc: Meta-analysis of continuous outcomes combining individual patient data and
aggregate data-StatMed.pdf
URL:
<https://stat.ethz.ch/pipermail/r-help/attachments/20121005/90f048a6/attachment-0002.pdf>

Erica E.M. Moodie, Dr. <erica.moodie <at> mcgill.ca> writes:
> I have been asked to help perform a meta-analysis with individual-
> and aggregate-level data. I found a nice article on this, and the
> idea is easy to understand: use a mixed effects models, but for the
> studies where there is only aggregate level data, force the variance
> to be that which was observed. Unfortunately, I am struggling to see
> how to implement this in R. I am familiar with "standard" mixed
> model formulations using nlme and lmer, but not able to see how to
> force this specific variance structure on my model so that some
> variance terms are not estimated.
  No solutions, but a few ideas:

 * I assume this is beyond the capacities/flexibility of the metafor package?

 * It's possible that the regress package
 http://cran.r-project.org/web/packages/regress/index.html
could be made to do this

 * There *may* be a way to do this with the pdStruct structures in nlme
(read the relevant bits of Pinheiro and Bates 2000, then stare at
your computer really hard for a while and see if anything comes to you),
although I suspect not

 * you can 'roll your own' mixed models in a flexible way with
WinBUGS/JAGS, maybe now Stan (if you want to be an early adopter!),
all of which have good R interfaces; it's conceivable that AS-REML
can do this too
 
 * follow-ups to r-sig-mixed-models please