Skip to content

[R-meta] CHE- Model Moderator/ Subgroup Analysis

2 messages · Wilma Charlott Theilig, James Pustejovsky

Wilma Charlott Theilig
# 
Good evening!

I wanted to ask whether it is possible to conduct moderator analyses or subgroup analyses for the CHE model (according to Pustejovsky & Tipton, 2022).

Furthermore, I wanted to ask how this could be implemented in R.

If there are already instructions or explanations somewhere that I have overlooked, I would be very grateful for a hint.

My code for the CHE model currently looks like this:


V <- vcalc(df$vi, cluster=df$StudyID, obs=df$EffectsizeID, data=df, rho=0.6, time1=time,  phi = 0.9)


overall <- rma.mv( yi, V = V,

                  data = df,

                  level = 95,

                  method = "REML",

                  slab = Study..author..year.,

                  random = list(~ 1 | StudyID, ~ 1 | EffectsizeID)) |> robust(cluster = StudyID, clubSandwich = TRUE)



I want to do a moderator analysis with a continuous predictor and a subgroup analysis with a categorical predictor.

Would it be necessary to dummycode the predictor in advance?

I thank you in advance for the great help you always get here in the forum!


Wilma

BSc, TU Dresden, Germany
James Pustejovsky
# 
Hi Wilma,

Short answer: Yes, you can conduct subgroup analysis just as in a regular
random effects meta-analysis, by specifying predictors in the mods argument
of rma.mv().

Here is an example using the continuous predictor X:
rma.mv(
  yi = yi, V = V,
  mods = ~ X,
  random = list(~ 1 | StudyID, ~ 1 | EffectsizeID),
  data = df,
  level = 95,
  method = "REML"
) |>
  robust(cluster = StudyID, clubSandwich = TRUE)

Here is an example using the categorical predictor Cat, with the model
specified to estimate average effect sizes for each level of Cat:
rma.mv(
  yi = yi, V = V,
  mods = ~ 0 + Cat,
  random = list(~ 1 | StudyID, ~ 1 | EffectsizeID),
  data = df,
  level = 95,
  method = "REML"
) |>
  robust(cluster = StudyID, clubSandwich = TRUE)

With CHE or other working models for dependent effect sizes, there is the
further, somewhat nuanced question of distinguishing within-study and
between-study effects of the predictor(s). Tanner-Smith, Tipton, and
Polanin (2016; https://doi.org/10.1007/s40865-016-0026-5) recommend
centering the predictors by study so that the between-study effect and
within-study effect can be separately estimated. So if you have a
continuous X, you would end up using two predictors (the within-study
centered X and the study-level averaged X). With a categorical predictor,
implementing this strategy would entail creating dummy variables for each
category and then centering the dummy variables.

James

On Mon, Jul 17, 2023 at 2:31?PM Wilma Charlott Theilig via
R-sig-meta-analysis <r-sig-meta-analysis at r-project.org> wrote: