Optimistic Bias in the Evaluation of Statistical Methods: Illustrations and Possible Solutions
MCML Authors
Christina Sauer (née Nießl)
Dr.
* Former Member
Abstract
Christina Sauer (née Nießl)
Dr.
* Former Member
Abstract
This dissertation investigates optimistic bias in statistical benchmark studies, showing how design choices, preprocessing decisions, and dataset selection can unintentionally favor certain methods. Across four contributions, it analyzes sources of bias, demonstrates how benchmark outcomes can be manipulated through methodological flexibility, and proposes strategies for more transparent, systematic, and reliable evaluation practices. (Shortened.)
BibTeXKey: Sau25