About the project
The project begins with a diagnostic approach, making use of simulated data to test whether new machine learning methods maintain their accuracy for variable selection under varying types of data dependencies. These findings are then validated through empirical applications to existing meta-analyses, while also reviewing how researchers currently handle such dependencies in practice. If problems are identified, the project aims to develop methodological refinements to address the specific limitations discovered, allowing meta-analysts for a more nuanced use of these selection methods with real-world data.
Background
Machine learning methods for moderator selection in meta-analysis were developed assuming clean data structures with independent or simply nested effect sizes. However, real meta-analytic datasets typically contain more complex dependencies. Whether these methods remain accurate under such realistic conditions has not been systematically examined.
