Skip to content

[R-meta] To include or not include time in rma.mv

3 messages · Yuhang Hu, James Pustejovsky

#
Hello All,

I'm considering two candidate fixed-effects structures (cat_mod
= categorical moderator) to meta-analyze a group of longitudinal
studies (pre-test and follow-ups):

m1 = rma.mv(yi ~ 0 + cat_mod + covariates, random = ~ 1 | study/effect)
m2 = rma.mv(yi ~ 0 + cat_mod*time + covariates, random = ~ 1 | study/effect)

In several authoritative meta-analyses of longitudinal studies such as
***https://doi.org/10.1080/19345747.2021.2009072*** and
***https://doi.org/10.1007/s11121-021-01246-3***, I noticed the authors
used "m1" over "m2".

But I was wondering, wouldn't, ignoring "time" (as in "m1") in
the fixed-effects structure mix up the pre-test and follow-up effect
sizes together which can blur our understanding of the treatment effect?

Your guidance is highly appreciated,

Yuhang
#
Hi Yuhang,

As one of the authors of the second study you mentioned, I think you may be
misinterpreting our analysis. My understanding is that the data that we
used to develop our examples did not include standardized mean differences
prior to treatment. All of the included effect sizes were for
post-treatment effects (i.e., `postwks >= 0`).

I can't say for certain, but I suspect that the same thing holds for the
Williams et al. meta-analysis (the first study you mentioned). If you look
at their Table 2, the "Outcome timing" variable ranges from "midstream
during intervention" to "follow-up post-test," so there don't appear to be
any pre-treatment effect sizes included in the analysis.

If the data did include pre-test effect sizes, then I fully agree that m1
would blur the interpretation of the treatment effects.

James
On Mon, Sep 5, 2022 at 7:36 PM Yuhang Hu <yh342 at nau.edu> wrote:

            

  
  
#
Dear James,

Thank you very much for the clarification.

I'm assuming that you guys excluded the pretreatment effects mainly because
your study pool consisted only of true experiments (right?).

I might be wrong, but random assignment in single experiments could still
suffer from bad luck and can equalize the groups only on average given
innumerable replications.

So, do you advise including a "time" interaction overall?

Thank you so much for your time.

Yuhang
On Tue, Sep 6, 2022 at 8:04 AM James Pustejovsky <jepusto at gmail.com> wrote: