 |
 |
Andi
don't you woory so much - if you take 120sec it will work all fine - model this and look at this using review design / explore regressors ..
> Dear Cyril, > Dear SPMList,
> First, thank you, Cyril, for your helpful reply and for the links.
> <the filter will let the frequencies higher than the nb you input to
pass
> .. so if you have a design with 30 sec alternation, each block
> appears at .03 Hz and you can think of the whole design - alternation of AB
blocks -
> at 0.016Hz (your F0); SPM default cutoff is 128 sec ie 0.0078, so you can
go up
> <(in freq) but not much - as you suggestd you could go down (in time) at
60 sec
> but ou have some border effects etc so maybe 100 sec (anyone got
<theoretical advice on
> this?)
> I see. With respect to border effects: I have thought about this for a
while:
> The most common advice concerning cut-off values I found so far is to
> apply 2 times the maximum on-off intervall, which is 120seconds in the
> example above. If I go with a cut off value between 60 seconds (which
> is -as you said -equal to the F0 in the alternating design above) to
> 119seconds, then the regressors and the filter matrix may be correlated
> to some degree. Due to this correlation, the filter is able to explain
> some of the variance that should actually be explained by the
> regressors alone and the effect of a factor may be missed. One may
> actually loose power instead of gaining some. I think, this is the
> reason, why it is commonly suggested to go with twice the maximum
> on-off intervall. It is avoided that the20regressors can be modelled by
> the filter matrix. As a newby, I am not sure, of course, so can anyone
> comfirm/refute this notion?
yep that's the whole point of these things you've reading about: don't cut where you have the max of power !!
> My problem is: I have 4 conditions in a balanced blocked design with 30
> seconds per block and when I follow the standard rule (twice the
> maximum on-off interval), I would have to go with a cut off value as
> high as 420 seconds, which means my signal is probably flooded by
> noise. Your link
> http://imaging.
mrc-cbu.cam.ac.uk/imaging/DesignEfficiency gives me
> some hope, though. It is stated there, that for contrasts that span all
> four conditions (e.g., [1 1 -1 -1] ), i can safely go with 128 seconds
> (i.e. the standard value). This implies that the optimal cut off value
> is dependent on contrast. I assume therefore, that the above
> standard rule would be appropiate only for those contrasts that contain
> only one regressor, e.g. [1 0 0 0 ]. Thus, even for contrasts that do
> not span all conditions (e.g., [1 -1 0 0], i could reduce the cut-off
> value of 420 seconds to 210 seconds. Intuitively, this feels right,
> because with [1 -1 0 0] I am interested in an effect, that oscillates
> at twice the frequency of e.g. [1 0 0 0]. Again, I am glad if anyone
> could refute or confirm this interpretation.
I'm not sure to follow here - prob Rick does .. one thing is that you may want also to look at other contrasts so don't cut to high .. an easy I'm thinking of these is what I want to contrast and how fast these conditions alternate taking together (for instance on/off blocks of 30s = max power at 1/60sec ; contrast on vs off)
good luck
cyril