re: st: RE: meta analysis in stata

[Red Owl <rh.redowl@liu.edu>]

I would like to add to Tim Mak's suggestion about weighting in
meta-analysis when the variance or standard errors of the effect size
estimates of the constituent studies are not available.  Tim suggested
that "the next reasonable step would be to assume equal weights for all
your studies."

In typical meta-analysis, the pooled mean effect size is calculated
using inverse-variance weighting under the premise that studies with
lower variance tend to provide more precise estimates of the true effects.

If the original poster (Azreen) does not have information on the
variance or standard errors of the effects in the studies included in
the meta-analysis, Azreen can use -metan- (by Michael Bradburn, Jonathan
Deeks, and Douglas G Altman and available from SSC) with the weighting
option -wgt(weightvar)- and set the weight equal to a variable
containing the sample sizes of the constituent studies.

I suggest that weighting by sample size is a generally a better strategy
than providing equal weights when inverse-variance weighting is not
possible in meta-analysis, because relatively larger samples tend to
(but not always) provide greater degrees of confidence in the point
estimates of the effect sizes, other things (e.g., quality of studies
and research designs) being equal.

If sample sizes are not available for the constituent studies being
meta-analyzed, or if the constituent studies in the meta-analysis tend
to have approximately equivalent sample sizes, or if Azreen does not
accept the premise that larger samples generally tend to produce more
precise estimates, then I would concur with Tim Mak's strategy to use
equal weighting.

Red Owl
redowl@liu.edu
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/