We are finishing up a meta-analysis that uses the average magnitude of the effect size (i.e. the absolute value of Hedge's d). Therefore, our range of effect size estimates are all positive. When I generate a funnel plot and run a trim and fill analysis in Metafor, it plots missing studies with negative effect size values. This plot seems spurious and also makes me think that the trim and fill estimate is incorrect as well.
I can't seem to find any examples of funnel plots that compare |d| to SE (reversed). Anyone have any suggestions or feedback?
The terminology here is off. Why would you ever use the "absolute value of Hedge's d" or an absolute value of a Hedges' g? Hedges' g and the like (e.g., standardized mean difference, raw mean difference) depend on the sign of the effect size value when comparing two groups. How would you ever know which group did better? Perhaps you mean to say "raw mean difference" (aka unstandardized mean difference). Regardless, it doesn't matter if all of the study effect sizes are positive. The funnel in a funnel plot is always centered around the overall effect size and study effect size symmetry should exist around that "funnel".
"Average magnitude of the effect size" is odd language. A meta-analysis does not calculate a mathematical average of study effect sizes.
Dear Dr Warren, thank you for the great explanation above.
If it is possible may I ask How would you apply Egger's test and/or Harbord test to interpret funnel plots (publication bias)? As soon as I know, We use Egger's test for continuous data and Harbord test or Rücker test for dichotomous data depending on tau level.
Thanks to all for your feedback. Just for your information, it's actually quite common to use the absolute value of Hedge's d. Feel free to review papers by Fritz et al. 2012; Sullivan and Fein 2012; Morrissey 2016 and the Meta-book by Koricheva et al. 2013 for more information. These resources all discuss the utility, limitations, and biases associated with this approach and adopt the same terminology. It also quite common to estimate the average magnitude of effect sizes (i.e a grand mean of effect sizes from individual studies). In our case, we calculated this effect using either a fixed-effect model (if
- Badges
- Science topic