Hi Joseph,
Some ringing will occur with censoring if you just drop samples and use FFT-based methods. Such filters do not play well with censored data because they assume the time series were uniformly sampled. Our preferred approach is to do the filtering and censoring with the same model using linear regression. That is the approach used by afni_proc.py.
For your use of 3dBandpass, I take it you are replacing censored points with the average of the neighbors, as Carp suggested in his commentary in Neuroimage, and that is OK too.
In general, unless you are censoring a considerable chunk of your data, the differences due to filtering should have little effect on your final group comparison.
cheers,
Ziad
