Dear Martin,
thanks for the opportunity to clarify this (if you hadn't asked I would
probably have been forced to "fake" this question at some stage). There seem
to be some confusion.
> if i use the unwarp function in SPM2, how much sense does it make to
> include the realigment parameters as confounds in the design matrix. In the
> SPM2 FIL batch file, altough realigment pars. are used as covariates to
> demonstrate how a covariate can be included, the comment is that including
> realignment pars. would render unwarping somewhat unnecessary. I also
> vaguely remember a statement on the SPM list that unwarping would render
> the inclusion of realignment pars. unnecessary.
> On the other hand in a recent Paper by Rick Henson (cereb. cortex, 2003)
> unwarping was performed and the realigment pars were used as a covariate.
> Therefore my question is what are the advantages/disadvantages of including
> realigment pars. as confounds when unwapring is performed, and is there any
> general advice what strategy is advisable?
There are four main causes for residual (after realignment) movement related
variance. These are
i) Movement-by-susceptibility-distortion interaction
ii) Movement-by-susceptibility-dropout interaction
iii) Spin-history effects
iv) Inter-volume movement (the slice to vol problem)
Of these I believe (i) to be the dominating in most data sets, and it is also
this variance that Unwarp attempts to remove.
If you include the realignment parameters you will remove (to a first
approximation) variance due to all these causes.
Hence, if you include the realignment parameters there IS NO REASON to do
Unwarp first.
Why then would you consider using Unwarp?
Let's say you have task correlated movement (casual or not). Then there are
five causes for movement CORRELATED (N.B. not related) variance. These are
then
i) Movement-by-susceptibility-distortion interaction
ii) Movement-by-susceptibility-dropout interaction
iii) Spin-history effects
iv) Inter-volume movement (the slice to vol problem)
v) True activations
Again, if you include the realignment parameters you will remove (to a first
approximation) variance due to all these causes (including the true
activations).
Also again, if you use Unwarp you will only (well mainly) remove variance due
to (i) and your activations will be saved.
So,
If you include movement parameters you will
i) Remove almost all movement related variance->higher specificity, which is
good.
ii) Risk removing your activations for task correlated movemnt->lower
sensitvity, which is bad.
If you use Unwarp you will
i) Remove not all of the movement related variance->lower specificity, which
is bad.
ii) Not remove any of your activations->higher sensitivity, which is bad.
Doing both is not wrong per se, just a bit silly, and the results will be very
similar to if you had just included the movement parameters.
Hence, using Unwarp versus including realignment parameters is really a trade
off, and it isn't easy to make a blanket recommendation.
I would personally favour Unwarp, but that is because I find the approach more
elegant and because the bloke who wrote it is such a nice person. So I guess I
am a bit biased.
Good luck Jesper
|