A0655
Title: Large-scale importance selection of heteroscedastic units
Authors: Bowen Gang - Fudan University (China) [presenting]
Abstract: Choosing candidates to whom a limited set of resources will be distributed is a pervasive dilemma. In multiple testing procedures that can be used to choose such candidates, power is traditionally defined as the number or proportion of correctly selected non-null hypotheses. We propose a generalized power that allows researchers to better select more desirable testing units and propose a specific formulation to capture not only if a unit has been correctly categorized as null or alternative but also to better reward the detection of larger effect sizes. Our new empirical Bayes multiple testing framework rewards discovering not just significant but large effects while controlling type I error. Hence, the selection process is better able to incorporate effect size into selection. We provide theoretical guarantees for FDR control and power optimization as well as numeric evidence for the utility of a generalized power.
