[R-meta] Multilevel / Specification of levels
Michael is absolutely right. You should include the row/estimate-level random effect in the model. See also: http://www.metafor-project.org/doku.php/analyses:konstantopoulos2011#a_common_mistake_in_the_three-level_model Best, Wolfgang -----Original Message----- From: R-sig-meta-analysis [mailto:r-sig-meta-analysis-bounces at r-project.org] On Behalf Of Michael Dewey Sent: Wednesday, 18 September, 2019 9:47 To: Dudenh?ffer, Tina; r-sig-meta-analysis at r-project.org Subject: Re: [R-meta] Multilevel / Specification of levels Dear Tina My understanding of this is that you do need to specify a random term for each effect as well to get the equivalent of a random effects meta-analysis. Do doubt someone more expert than I will chip in here if that is not correct. Michael
On 17/09/2019 10:55, Dudenh?ffer, Tina wrote:
Dear all,
I am running a multilevel model using the rma.mv function.
res4 <- rma.mv(Hedges_g, Hedges_v, method=?REML", random = list( ~ 1 | Study), data=Data)
res4
The question I have is regarding the random argument. If I understand correctly, this is where the levels are specified. My effect sizes (Hedges_g) are nested in studies (Study). The variable Study refers to the unique study ID in my data set. When reading the documentation, it says that the ~ 1 | id is a grouping id and outcomes with the same id receive the same random effect. Hence, in my example, all outcomes from the same study would receive the same random effect.
However, I have also seen other models that seem to include the unique effect size ID and if I translate this to my example it would look like this:
res4 <- rma.mv(Hedges_g, Hedges_v, method=?REML", random = list(~ 1 | EScumulative, ~ 1 | Study), data=Data)
res4
As you can see the difference is the ~ 1 | EScumulative. EScumulative in my data set refers to each unique effect size. Based on the documentation and the notion that the id in the random argument refers to grouping variables, I don?t believe that I should include the EScumulative in the model, because none of them share the same value (unlike the id for Study, where some effect sizes share the same study ID. However, there are some examples (and opinions) that suggest otherwise.
That said: My question is, which of the two models is correct?
I am now working using the second model (including the EScumulative):
The variance components look like this: so to me this almost seems like there is no variation on the study level and I wonder if I should even run multilevel?
estim sqrt nlvls fixed factor
sigma^2.1 0.2283 0.4778 62 no EffectSizeID
sigma^2.2 0.0000 0.0000 28 no Study
I am new to Meta Analysis, new to Multilevel and new to R - excuses in case some of this is not clear.
Thanks for your help and kind regards,
Tina