A Novel, Variance Component-Based Method for Detecting Brain-Behavior Associations in Neuroimaging Data

The sequence kernel association test (SKAT) is a widely used and computationally efficient method in high-dimensional studies that tests for the joint effect of multiple predictors while accommodating covariates. However, the omnibus nature of the test hinders interpretation. We develop a new method called LaxKAT (linear maximum kernel association test) that can identify both global and local signal in high-dimensional data. The LaxKAT statistic maximizes the SKAT statistic over a pre-specified subspace of linear kernels. We demonstrate via simulations that it exhibits improved global and local power compared to previous methods while controlling the family-wise error rate (FWER). We also apply LaxKAT to neuroimaging data from cognitively normal controls in the Alzheimer's Disease Neuroimaging Initiative (ADNI) to identify brain regions exhibiting sex-specific differences in cortical thickness patterns.