Examining the performance of trend surface models for inference on Functional Magnetic Resonance Imaging (fMRI) data

The current predominant approach to neuroimaging data analysis is to use voxels as units of computation in a mass univariate approach which does not appropriately account for the existing spatial correlation and is plagued by problems of multiple comparisons. Therefore, there is a need to explore alternative approaches for inference on neuroimaging data that accurately model spatial autocorrelation, potentially providing better type I error control and more sensitive inference. In this project we examine the performance of a trend surface modeling (TSM) approach that is based on a biologically relevant parcellation of the brain. We present our results from applying the TSM to both task fMRI and resting-state fMRI and compare the latter to the results from the parametric software, FSL. We demonstrate that the TSM provides better Type I error control, as well as sensitive inference on task data.