On Wed, Feb 24, 2010 at 12:50 PM, Cornelius Werner
<[log in to unmask]> wrote:
> With r=0.948 for block designs, you probably do not need to go all the way of de- and reconvolving. But for ER, while I would say that r=0.709 is not too bad, there is still a difference. What is not clear to me, however, if they also convolved the psychological regressor (which has a somewhat unusual shape, by the way) as they did with the artificial neural vectors. If yes, then this should probably be the way to go, shouldn't it? Perhaps one could include the method proposed in Makni, Beckmann, Smith and Woolrich (2008, Neuroimage 42, 1381-96) in a not-too-future version of FSL?
It is also worth considering just how accurate the timing of "neural"
events is. I don't believe the Gitelman et al. paper looked at the
effect of introducing temporal noise to the neural events, and let's
face it; in most fMRI tasks we don't have a particularly precise
knowledge of the onsets or offsets of neural firing. What would be the
correlation between two event-related "neural interaction" regressors
in which the timing of the neural events was varied in line with our
degree of uncertainty? Would it be any better than the 0.709 reported
between neural and haemodynamic interaction vectors? Then there is the
related question of accuracy of the HRF used to perform the
deconvolution.
-Tom
|