The impact of certain methodological choices on multivariate analysis of fMRI data with support vector machines

Multivoxel pattern analysis of functional magnetic resonance imaging (fMRI) data is continuing to increase in popularity. Like all fMRI analyses, these analyses require extensive data processing and methodological choices, but the impact of these decisions on the final results is not always known. This study explores the impact of four methodological choices on analysis outcomes and introduces the technique of partitioning on random runs for characterizing temporal dependencies and evaluating partitioning methods. The analyses were performed on two fMRI data sets, which were repeatedly analyzed with support vector machines, varying the method of temporal compression, smoothing, voxel-wise detrending, and partitioning into training and testing sets. Smoothing sometimes slightly increased classification accuracy. Partitioning other than on the runs increased classification accuracy, and the random runs technique allowed us to attribute this improvement to the increased amount of training data, rather than to bias. The impact of the temporal compression and detrending methods varied so strongly with data set that general recommendations could not be drawn. These interactions suggest that, rather than searching for a universally superior set of methodological choices, researchers must carefully consider each choice in the context of each experiment.