Paul Mazaika wrote:
>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.
To be painfully thorough, it's probably best to do both -- interpolation to
minimize the damage to neighboring volumes, and modeling the bad volumes out
to make sure that whatever you end up with in that volume won't affect your
parameter estimates at all (and to make sure your df is properly penalized).
To be sure, for typical BOLD data once you interpolate the missing/bad
volumes, modeling out those volumes probably makes very little practical
difference. But all else being equal, better to model them out.
For what it's worth, we have a little tool to interpolate out bad volumes
using either linear or cubic spline interpolation. (Cubic spline because it
was readily available in the GSL library.) It doesn't do a lot (especially
compared to the toolkit Paul mentioned), but it's easy to insert into your
toolchain. It can be downloaded from:
http://voxbo.org/index.php/vbinterpolate
It might be good to know what interpolation algorithm actually does the best
job of replacing a missing BOLD volume and/or series of volumes. Does
anyone know offhand?
dan
|