I2 in Meta-Analysis

RTO20

First time posting on this forum, please forgive me if the post is in the wrong place.

We have used JASP to carry out a random-effects meta-analysis comparing the Area Under the Receiver Operating Characteristics curve (AUC-ROC) for multiple studies. We are struggling to generate the I2 (I-squared) heterogeneity score because the papers do not report the standad error (SE). What is confusing however, is that the egger plots generated by the software have SE on the y-axis, so it is quite confusing why I2 is not generated automatically.

Here is my question:

1. Is there a way to generate I2 from papers which do not report SEs? We discussed generating Cochran's Q and converting to I2 but this would also be difficult given the sparce information.
2. How is JASP generating SEs for these studies in the Egger plots? Surely if JASP can generate SE for these studies then it should also automatically provide I2.

Many thanks for any responses.

Best wishes,

Robert

EJ

Thanks for posting, and apologies for the tardy response. I've forwarded this to our expert.

Cheers,

E.J.

František

We resolved this via email with the help of Thomas Debray. I'm attaching the response if other people have a similar question:

The I2 statistics you report are OK, they usually represent the I2 of the logit transformed c-statistic (unless no transformation is applied to the c-statistic before meta-analysis, which is not recommended).

Briefly, when performing a meta-analysis of the AUC, this is what (usually) happens:

1) the AUC and its CI are transformed to the logit scale, which results in an "effect size" that is approximately normally distributed

2) the "effect sizes" (here logit AUC estimates) are pooled using meta-analysis methods

3) the pooled " effect sizes" are back-transformed to the original scale.

In other words, parameters such as tau squared and I squared are applicable to the scale that is used for implementing the meta-analysis. For the AUC, this is usually the logit scale (default implementation of metamsic). The I2 statistic can be interpreted directly (i.e., no need to "understand" the logit scale). However, a parameter like tau squared requires more effort (as it represents the between-study standard deviation of the logit AUC).

does this help? Briefly, no additional work is needed on your side. Robert can use the I square statistic that is already reported in JAGS. If he wants to calculate it himself, he needs to use the estimated logit AUC and their corresponding standard errors (as this is the scale on which the meta-analysis was performed).