A1630
Title: Transforming heavy tailed data improves the power and validity of inference
Authors: Samuel Davenport - University of California, San Diego (United States) [presenting]
Abstract: The purpose is to demonstrate that transforming heavy-tailed data leads to improvements in both the power and validity of inference, with a particular focus on applications in neuroimaging in which the noise can be very heavy-tailed. Validity is improved because the rate of convergence of the CLT is accelerated, leaving valid p-values under the null. Instead, a power gain occurs because the transformation can increase the value of Cohens d. It is shown that transformations can also be combined with sign-flipping to infer the null distribution of the transformed data, ensuring validity and providing a power boost. The approach is validated using gold standard resting state fMRI null simulations and task fMRI datasets used from the Human Connectome Project to illustrate the improvements in power.
