A better statistic might be the prediction interval but in my view the community needs to develop more experience with it before we can use it to confidentially judge relevance of heterogeneity (if ever). And unfortunately it does not work very well in a meta-analysis with only a few studies

True. But what else can we do? A paper in the new Core GRADE series (currently under revision in the BMJ) offers a practical solution based on visual inspection of both overlap of confidence intervals (similar to I2) but also where point estimates lie in relation to a relevance threshold.

Great! Will include it in @lightsdatabase.bsky.social