Letting the LaxKAT Out of the Bag: A Powerful Kernel Test for Neuroimaging Studies

• Activity Number: 26 - Imaging Speed Session
• Type: Contributed
• Date/Time: Sunday, August 8, 2021 : 1:30 PM to 3:20 PM
• Sponsor: Section on Statistics in Imaging
• Author(s): Jeremy Samuel Rubin* and Simon N Vandekar and Lior Rennert and Mackenzie Edmonson and Russell Shinohara
• Companies: University of Pennsylvania and Vanderbilt University and Clemson University and University of Pennsylvania and University of Pennsylvania
• Keywords: High-dimensional hypothesis testing; Score test

Abstract

Biomedical research areas including genomics and neuroimaging often have a number of independent variables that is much greater than the sample size. The sequence kernel association test (KAT) and sum of powered scores tests (SPU) can offer improved power in this feature setting; however, power is significantly reduced in the presence of a large number of unassociated independent variables. We propose the Linear Maximal KAT (LaxKAT), which maximizes the KAT test statistic over a subspace of linear kernels to increase power. A permutation testing scheme is used to estimate the null distribution of the LaxKAT statistic and perform hypothesis testing. We find that this test has greater power over the SKAT and SPU tests for signal distributions with varying signs and maintains competitive power for signal distributions with constant signs. The LaxKAT also controls the type I error across different sample sizes and signal distributions. We further assess the performance of the LaxKAT relative to the SPU and SKAT tests when applied to detect predictors of memory impairment in cortical thickness measurements from the Alzheimer’s Disease Neuroimaging Initiative study (ADNI).