Skip to content

[R-meta] Threshold values for Cook's Distances and DFBETAS

4 messages · Thölking, Theresa, Michael Dewey, Wolfgang Viechtbauer

#
Hello everybody,


I am trying to apply the outlier and influence diagnostics as described by Viechtbauer and Cheung (2010) on a meta-analysis I conducted, using a four-level random-effects model with 528 included effect sizes. I have calculated the Cook's Distances and DFBETAS in R, using the functions for model diagnostics for rma.mv objects as described here: https://wviechtb.github.io/metafor/reference/influence.rma.mv.html


However, I don't know how to decide about which studies are influential based on these values. I read the paper by Viechtbauer and Cheung, but I don't quite understand it. Is there some kind of formula to determine the threshold values of CD and DFBETAS for a meta-analytic model with k outcomes? Thanks in advance for your help.


Best,

Theresa
#
Dear Theresa

My feeling is that these diagnostics are better seen as identifying 
values which may benefit from further investigation. If I was concerned 
that they were unduly influential in the model I might try fitting 
leave-one-out models but since those models are essentially data-driven 
they are also only useful in an exploratory sense.

So, even if there is a formula I would be sceptical about its value in 
answering the scientific question.

Michael
On 29/09/2022 21:35, Th?lking, Theresa wrote:

  
    
  
#
And to follow-up on this:

There isn't any formula anyway. It's a bit like asking if there a formula that we can use to determine when a person is 'unusually tall' based on having measured people's height. It's all relative. We can just compare the height of the tallest person with that of the rest. If he/she sticks out (in the literal sense), then we might say that this person is quite tall compared to the rest.

Therefore, even the rules mentioned here are essentially arbitrary:

https://wviechtb.github.io/metafor/reference/influence.rma.uni.html

Best,
Wolfgang
#
Hello Wolfgang, hello Michael,

thanks for taking the time to answer my question. This was very helpful.

Best,
Theresa