This is a SKY website examining the performances of methods correcting for publication bias.
This website contains:
The main results of the project so far can be summarized in the performance of the various estimators of the true effect compared to pre-registered replications. The analysis behind this data is available here. The following graph shows it all:
Performance of various estimators of an effect compared to pre-registered replications
The original estimates have large baises. Meta-analysis divides the mean bias by two. The PEESE and PEESEpos methods brings the bias further down by 40%. The FATPETPEESEpos method brings the bias further down by 50%. The other criteria tell a braodly similar story. This data makes the FATPETPEESEpos estimators the best estimators so far to estimate true treatment effects using the published record, with selection models a close second. p-curving does not manage to improve on bais correction beyond what standard meta-analysis can achieve.
If you have comments on this project, please post them on the corresponding GitHub comment page.
If you want to alter this project, please do so on its GitHub repo.