Abstract
Head movements, or other global nuisance signals, can be a severe problem in FMRI analyses, yielding artefactual activations. To reduce this problem, the first step of data preprocessing is spatial alignment of the collected brain volumes over time, and the voxel wise removal of BOLD signal components correlated with the nuisance signals. We investigated the influence of the removal of nuisance signals on GLM and support vector machine analyses by creating simulated data sets and removing nuisance regressors of varying correlation with the stimulus time course. We report that for both types of analyses, false positive and false negative rates increased with increasing similarity between regressor and stimulus. Additionally, crossvalidated classification performance became ever more strongly biased downward as the correlation between nuisance regressor and stimulus increased, down to a performance level of 0%, where every instance was misclassified in crossvalidation. n the other hand, when the nuisance regressor was uncorrelated with the stimulus, classification performance was artefactually biased upward when a small number of time points was used. Overall, these results highlight the problematic nature of any signal that correlates with the stimulus pattern in FMRI experiments. Head motion is a particularly relevant example, but other signals could include respiration, heartbeats, and eye movements. These problems are particularly serious in the context of multivoxel analyses, which - due to their high sensitivity - are also especially sensitive to global nuisance signals as well as biases introduced by their attempted removal.
Funded by The Pew Charitable Trusts.