Statistical modeling of functional MRI data

After measured fMRI data has been preprocessed to remove artefacts related to image distortions, head motion and signal drifts, main data analysis usually begins with building a statistical model. This model is used to infer which regions of the brain were involved in specific task conditions. Usually, condition-related effects within voxels or regions are first quantified and statistically compared in individual subjects, followed by a second stage comparison across subjects. This article focuses on this essential aspect of data analysis, which may produce the main results of a study or may serve as the basis for additional analyses such as multi-voxel pattern analyses. After covering basic statistical concepts, the general linear model will be introduced as the main tool for univariate (voxelwise) statistical modeling. It will be demonstrated how the general linear model can be used to model block and event-related designs. Furthermore, the assumptions of general linear modeling will be described, including methods to ensure proper voxelwise statistical results. This involves addressing serial correlations in voxel time courses and by correcting for massive univariate testing. Since researchers are usually interested in generalizing results to the population level, basic principles of group level analysis will be covered. This analysis is typically formulated as a second-level random-effects model that fits group parameters to parameter estimates obtained from each participant in the first-level (voxel time course) analysis. Finally selected alternative analysis approaches including non-parametric and Bayesian statistics will be briefly described and their advantages and disadvantages with respect to conventional general linear modeling will be discussed.