[R-meta] moderator and adjusted effects

Dear experts,

In my sample of articles for meta-analysis there are three categories, or
three conditions, that may influence the effect of interest.

I am more interested in estimating different effects from these conditions
than in explaining heterogeneity in effect sizes.
1) I can do a meta-analysis for each of these conditions separately and get
three different mean effect sizes.
2) Or I can do a meta-analysis of the whole sample, then include a
condition as a moderator and calculate adjusted effects as described here :
http://www.metafor-project.org/doku.php/tips:computing_adjusted_effects

Which option is better?

Additional question:  when I include a categorical moderator, is it the
same as including a dummy variable in a regression?
How can I specify that the variable is categorical with 3 levels?

Best,
Valeria

	[[alternative HTML version deleted]]

Dear Valeria

Comments in-line

On 14/12/2020 17:13, Valeria Ivaniushina wrote:

Dear experts,

In my sample of articles for meta-analysis there are three
categories, or
three conditions, that may influence the effect of interest.

I am more interested in estimating different effects from these
conditions
than in explaining heterogeneity in effect sizes.
1) I can do a meta-analysis for each of these conditions separately
and get
three different mean effect sizes.
2) Or I can do a meta-analysis of the whole sample, then include a
condition as a moderator and calculate adjusted effects as
described here :

I would go for option two as it will give you estimates of the 
differences between the levels of your moderator.

Which option is better?

Additional question:  when I include a categorical moderator, is it
the
same as including a dummy variable in a regression?
How can I specify that the variable is categorical with 3 levels?

If you make the moderator a factor then R will take care of this for
you.

Best,
Valeria

	[[alternative HTML version deleted]]

Dear Valeria,

another difference between your options (1) and (2) is that in case of
(1) you get three different, independent estimates of the heterogeneity
(tau), whereas in (2) you assume a common heterogeneity parameter for
all three groups.

In case you have "many" studies in each group (say, 20), it may not
make much of a difference, but if you have "few" studies (say, 5) in
some, and the assumption of a common heterogeneity parameter seems
plausible, then borrowing information on the heterogeneity across the
three groups may help.

Cheers,

Christian


On Mon, 2020-12-14 at 18:18 +0000, Michael Dewey wrote:

Dear Valeria

Comments in-line

On 14/12/2020 17:13, Valeria Ivaniushina wrote:

Dear experts,

In my sample of articles for meta-analysis there are three
categories, or
three conditions, that may influence the effect of interest.

I am more interested in estimating different effects from these
conditions
than in explaining heterogeneity in effect sizes.
1) I can do a meta-analysis for each of these conditions separately
and get
three different mean effect sizes.
2) Or I can do a meta-analysis of the whole sample, then include a
condition as a moderator and calculate adjusted effects as
described here :

http://www.metafor-project.org/doku.php/tips:computing_adjusted_effects

I would go for option two as it will give you estimates of the
differences between the levels of your moderator.

Which option is better?

Additional question:  when I include a categorical moderator, is it
the
same as including a dummy variable in a regression?
How can I specify that the variable is categorical with 3 levels?

If you make the moderator a factor then R will take care of this for
you.

Best,
Valeria

    [[alternative HTML version deleted]]