Most of the previous replies on this subject described adding regressors to the design
matrix to get rid of bad outliers. An alternate approach to get rid of outliers is
interpolation, which may have certain advantages.
Interpolation replaces outlier data points with an average of the nearest valid data points.
This method removes the effects of outliers from the desired estimates, the same as
adding a regressor. But instead of changing the design matrix, interpolation changes the
data going into the design matrix. Interpolation is easy to automate, does not involve
custom changing of the design matrix for each subject, and can be applied per volume,
per slice or per voxel as opposed to applying a regressor to an entire volume.
While interpolation can affect the statistical maps, the effect seems to be minor. In my
experiments, the effects were far less than the algorithm choice of smoothing kernel,
high pass filter cutoff, or whether or not to use motion regressors. From an overall view,
the interpolation approximation seems to be a small component of the total noise budget
relative to scanner thermal noise, head motion, spontaneous deep breaths, cardiac and
respiratory pulsatility, magnetic susceptibility drift, HRF modeling errors, etc.
For troublesome subject data, it seems the primary goal should be to catch ALL the
outliers, whether or not they were due to head motion. One implementation of this
approach is the ArtRepair Toolbox available at the SPM Extensions website. The program
suite has a visualization option to easily review all the data for outliers. The program also
can automatically detect large outliers from head motion and other causes, and apply an
interpolation repair to the data. Documentation is available at the website.
While interpolation is not discussed often in the fMRI literature, it is common in other
image and signal processing domains. We've found it useful for analyzing fMRI data sets
from children and clinical subjects.