Skip to content

[R-meta] Question about rma(measure = "COR", ..)

3 messages · Jalynskij, Maria, Michael Dewey, Guido Schwarzer

Jalynskij, Maria
# 
Dear all,


while conducting a meta-analysis I asked myself a few questions.


a)When I'm conduction a random effects meta-analysis with correlation-coefficients as effects sizes using


rma(measure = "COR", ri = corr, ni = n, data = data)

, does R conduct the analysis with Fisher's-z-transformed coefficients in the background and compute the main-effect results back into the r metric for the output or not?


b) Or should I use rma(measure="ZCOR", ri = corr, ni = n, data = data), and conduct the main result back by myown at the end?


In the second version the Trim and Fill method doesn't make any sense.


I would be glad about some support.


Sincerely,

Maria
Michael Dewey
# 
Dear Maria

If you want it to analyse on the transformed scale you need your second 
option ZCOR and then back transform the results. You may want to use the 
function transf.ztor() provided to help you.

I am not sure about your query about trim and fill here but it does seem 
to have become a less popular option recently.

Michael
On 21/10/2019 14:22, Jalynskij, Maria wrote:

  
    
  


  


      

    

    

      
      

        


  

        

      Guido Schwarzer
    

    

      
      #
    

  


  

      
Am 21.10.19 um 15:22 schrieb Jalynskij, Maria:

            

      
      
      
    

        
Dear Maria,

While one may question whether Trim and Fill is the best method to 
adjust for small study effects in meta-analysis, I do not understand why 
the method should make sense for untransformed correlations but not for 
Fisher z-transformed correlations.

Trim and Fill removes the most extreme studies (trim step), determines a 
symmetry axis and adds the extreme studies on the opposite side (fill 
step). This works both for untransformed and Fisher transformed 
correlations.

Best wishes, Guido

P.S. I would always prefer to meta-analyse Fisher transformed 
correlations as the meta-analysis of untransformed correlations could 
result in impossible confidence limits outside -1 and 1. Accordingly, 
the Fisher method is the default in /metacor()/ of R package *meta* 
(which automatically does the back-calculation for the user).