Hi Mark,
Can I just follow up on your comments about motion correction and including
motion parameters as covariates? One thing that I think is important to
mention is that "motion correction" is a misnomer -- the process actually
realigns the 3D volumes to minimize the effects of minor head
movements. It does not "correct for motion" for all the reasons Mark
said. As a result, there are still motion effects in one's data set after
realignment.
In my experience, including the estimates motion params as covariates of no
interest is often helpful in improving statistical sensitivity despite the
caveats Mark mentioned. But the main thing to consider is whether the
motion is correlated with your experimental paradigm. If it is, then these
covariates may remove signal that *may be* related to your experimental
manipulation. As the two are correlated, however, there is no way to tell
for certain where the activation is coming from (motion or experiment or
both). Personally, when that happens in one of my subjects, I throw the
data from that subject (or session) out. BTW, one easy/informal way to
check for correlated motion is to look at the covariance matrix provided in
the FEAT web report.
With that in mind, I can't see any good reason why one would ever 1) not
realign data but only include motion estimates as covariates or 2) only
realign without including motion estimates in the model. It seems to me
that together the two are more powerful than either alone (assuming the
data aren't confounded by stimulus-correlated motion for which there is no
solution). And even together, they still don't "correct for motion" as the
process introduces additional smoothing, residual non-linear effects
remain, etc.
Is this a fair summary or am I still missing some issues?
Joe
Joseph Devlin, Ph. D.
FMRIB, Dept. of Clinical Neurology
University of Oxford
John Radcliffe Hospital
Headley Way, Headington
Oxford OX3 9DU, U.K.
Phone: +44 (0)1865 222 738
Fax: +44 (0)1865 222 717
Email: [log in to unmask]
|