Skip to content

[R-meta] Effect sizes calculation in Pretest Posttest Control design

4 messages · Michael Dewey, Célia Sofia Moreira

Original
Célia Sofia Moreira
# 
Hi!


I am studying a Pretest Posttest Control group design. I saw the
recommended method (Morris) to compute the effect sizes, presented in one
of the examples from Prof. Wolfgang? s webpage:

http://www.metafor-project.org/doku.php/analyses:morris2008



However, I don?t have pretest-posttest correlations. Prof. Wolfgang
suggests that in this case ?one can substitute approximate values (...) and
conduct a sensitivity analysis to ensure that the conclusions from the
meta-analysis are unchanged when those correlations are varied?. However,
since I have many different outcomes, sensitive analysis will be a very
complex task. So, I was wondering if, instead of measure = "SMCR", I could
use measure ="SMD". More specifically:



datT <- escalc(measure="SMD", m1i=m_post, m2i=m_pre, sd1i=sd_post, sd2i=
sd_pre, n1i=N1, n2i=N2, vtype="UB" , data=datT)

datC <- escalc(measure="SMD", m1i=m_post, m2i=m_pre, sd1i=sd_post, sd2i=
sd_pre, n1i=N1, n2i=N2, vtype="UB" , data=datC)

dat <- data.frame(yi = datT$yi - datC$yi, vi = datT$vi + datC$vi)



If not, can you please explain the problem of this approach and inform
about the existence of any other simpler alternative?



Kind regards
Michael Dewey
# 
Dear C?lia

I do not think the sensitivity analysis needs to be quite so complex as 
you suggest. You can use the same imputed correlation for all your 
primary studies. Then do it for (say) 0.2, 0.5, 0.8 and see what 
happens. If the results are very different then use some intermediate 
values as well to see where it all breaks down.

Michael
On 26/01/2018 22:50, C?lia Sofia Moreira wrote:

  
    
  


  


      

    

    

      
      

        


  

        

      Célia Sofia Moreira
    

    

      
      #
    

  


  

      
Dear Prof. Michael Dewey,

Thank you very much for your encouraging comments. Indeed, I considered
different values for the correlation and the results on the differences
(between the two standardised mean changes values - "SMCR"), for each
outcome, were the same. Only the variances of these differences varied a
bit, according to the following rule: higher correlation --> lower
variance. Thus, following your advice, maybe r=.5 is a reasonable choice.
Do you agree?

Kind regards,
 celia

2018-01-27 13:54 GMT+00:00 Michael Dewey <lists at dewey.myzen.co.uk>:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

  
  


  


      

    

    

      
      

        


  

        

      Michael Dewey
    

    

      
      #
    

  


  

      
Comment in line

        
On 27/01/2018 23:29, C?lia Sofia Moreira wrote:

            

      
      
      
    

        
That may depend on your field of research. For what one might loosely 
call psychological variables (attitude, belief, ...) test-retest 
correlations over any reasonable time period would not be expected to be 
much above 0.5. If you were measuring something harder (systolic blood 
pressure, serum creatinine, .. ) over a shorter period then I might 
expect the correlation to be a bit higher.