A1345
Title: A general score-difference framework for independence testing
Authors: Wenhao Cui - Beihang University (China) [presenting]
Abstract: A general framework is developed for testing unconditional and conditional independence based on comparisons of predictive distributions using strictly proper scoring rules. The central idea is to characterize independence through the invariance of predictive performance, whereby under the null, augmenting the information set does not improve expected scores, whereas violations induce systematic score differences. This perspective yields a flexible class of distributional tests that accommodates discrete, continuous, and mixed data, and is naturally compatible with modern machine learning methods for estimating conditional distributions via cross-fitting. To address degeneracy in score-based criteria, a randomized construction is introduced that delivers asymptotically pivotal inference under weak temporal dependence. In addition, localized tests are developed that detect dependence in specific regions of the outcome distribution, capturing heterogeneous and tail-specific relationships. Simulation studies and empirical applications demonstrate the effectiveness of the proposed approach.
