A0302
Title: Deep nonparametric conditional independence tests for images
Authors: Marco Simnacher - Humboldt-Universität zu Berlin (Germany)
Xiangnan Xu - Humboldt University of Berlin (Germany)
Hani Park - Hasso-Plattner-Institut (Germany)
Christoph Lippert - Hasso-Plattner-Institut (Germany)
Sonja Greven - Humboldt University of Berlin (Germany) [presenting]
Abstract: Conditional independence tests (CITs) are often used to study the relationship between health outcomes and exposures or genetic information, adjusting for potential confounders. While complex health outcomes are increasingly studied using medical imaging, there is a significant gap in the literature regarding CITs for an image and a scalar, given a confounding vector. To fill this gap, we introduce novel deep nonparametric CITs (DNCITs), which integrate nonparametric CITs for vector-valued data with embedding maps to extract feature representations from images. We show the validity of DNCITs under conditions on the embedding map that are fulfilled, particularly for learning schemes such as sample splitting, transfer learning, and (un)conditional unsupervised learning. We discuss the transferability of existing nonparametric CITs to our situation and propose tests adapted to our setting. We provide an implementation for these DNCITs in an R package and study their validity, power and sensitivity to different factors in an extensive simulation study. We apply the DNCITs to study the dependence between brain MRI scans and behavioral traits, given confounders, in healthy individuals in the UK Biobank, to shed new light on previous mixed results from several personality neuroscience studies with our more powerful tests on a larger dataset.
