Dear Afraim,
> Dear All,
>
> Returning back to this topic, does anyone know the rational
> behind including the raw realignment parameters, as is?
>
The inclusion of the movement parameters in the design matrix is
motivated by the observation that even after realignment there is
considerable "movement related" variance in the data. If you do a
singular value decomposition of your data after realignment you will
find that some of the eigenvariates (and of then the first ones) will be
almost perfect linear combinations of the estimated movement parameters.
You can e.g. read about this in Friston et al., MRM 35:346-355.
>
> What I mean to say is that any motion related signal change is
> more likely to be a function of the rate of movement as opposed
> to just the XYZ head position (i.e the realignment
> parameters). I have tried using the differential of these
> parameters (the dif() function) for example, either
> justifiably or unjustifiably, but with 'better' results!
>
This really depends on what you belive is the main reason for the
"failure" of the realignment. If the effects of motion was perfectly
described by the rigid body model, and interpolation was perfect, then
we would expect no movement related variance after realignment.
If we belive the the main "problem" with the realignment is inperfect
interpolation we would expect sines and cosines of translation of a
given voxel (note that this means voxelwise correction) as described in
Grootonk et al., NeuroImage 11:49-57.
If we believe the main problem with realignment is that it assumes all
movement to occurr between aqcuisition of volumes (the slice-to-volume
problem) then the temporal derivatives of the realignment parameters is
probably a good idea.
If we believe that the main problem is that susceptibility-by-movement
interactions are neglected then the realignment parameters themselves is
probably a good idea (and possibly the squares and cross products).
etc.
>
> There just seems to be no no limit to how complex one's model.
> If you assume for example that there is a rapidly attainable
> limit to any %signal change arising from any degree of movement
> then the log(dif()) may be an even more realistic representation
> (providing you avoid -ve values of course).
>
The models can indeed get quite complex. log(dif()) might be a good
idea, although it sounds a bit ad hoc.
>
> A quick look at the global means also reveals movement effects
> when these are significant such that any global scaling will also
> provide some indirect motion correction. If this is then taken
> in addition to whatever spin history correction is applied at the
> realignment stage (providing you have chosen to 'adjust
> for sampling errors'), the whole thing becomes very very messy
> indeed.
>
> It is for these reasons that I wonder what the most sensible
> or better still, the most evidence based strategy really is?
> (Assuming no stimulus-locked motion for the moment...)
>
I personally believe the most sensible strategy would be to try and
correct the underlying causes by extending the realignment model to
incorporporate the effects I indicated above (and possibly yet some)
rather than to just regress out the variance. It is a bit tricky though.
>
> Any references on this topic would be most welcome.
>
See above.
Good luck Jesper
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|