Machine Learning Can Solve the Reproducibility Crisis by Supplanting Reductionist Statistics

Abstract

Resolving the "replication crisis" is a top priority of the scientific community now. "Reproducibility" is claimed as a central tenet of science and the estimated economic and social burden is huge. Numerous proposals have been made. Still, there lacks not only an established solution but even an agreement on whether there exists a "crisis" or not. Here, by questioning the philosophical foundations of our study designs and analyses, I trace back the "crisis" to reductionist ontologies and methodologies ingrained in the modern statistical methods which have dominated biological, medical, psychological, and social sciences for a century. The crisis is not our inability to "reproduce results" but that we expect to be able to "reproduce results" despite neglecting almost all individual-level and contextual variables of complex processes.