Hi Joseph,
Thanks for your help and quick reply. I was just a little confused about
1 thing:

"The one thing I would add is to guarantee that the main
effect of I>N is not being driven solely by either the first or second
half, you'd need to do some form of inclusive masking -- that is, mask
[1
1] with [1 0] and [0 1] at some arbitrary threshold. BTW, it's probably
best to use a voxel threshold rather than a cluster threshold for this.

Then [1 -1] would yield I trials with more activation in the first half
than second (it would not include N trials)."

Wouldn't masking [1 1] with [1 0] yield a map of I voxels across the
entire task that is masked so that only those in the first half appear
as active? If so, wouldn't this yield the same map as that of the simple
(unmasked) [1 0] contrast?

From my understanding, If I wanted to find I voxels that were more
active in the first compared to the second half, wouldn't I mask [1 -1]
with the [1 1] contrast? I interpret this as asking "of I voxels active
across the entire task, which are more active in the first compared to
the second half?"

Thanks again!
Alex




-----Original Message-----
From: Joseph Devlin [mailto:[log in to unmask]]
Sent: Monday, June 07, 2004 3:53 PM
To: [log in to unmask]
Subject: Re: [FSL] FEAT setup

Dear Alex,

>Assuming this is OK and I have only these 2 EVs, I assume the following
>contrasts would represent (take EV1 as the first-half regressor & EV2
as
>the second-half):
>EV1 EV2
>1 0 I>N for first half only
>0 1 I>N for second half only
>1 1 I>N for the entire task

Your design looks fine assuming that the box-car is modeling the
interference condition and that the neutral condition is left as an
implicit baseline. The one thing I would add is to guarantee that the
main
effect of I>N is not being driven solely by either the first or second
half, you'd need to do some form of inclusive masking -- that is, mask
[1
1] with [1 0] and [0 1] at some arbitrary threshold. BTW, it's probably
best to use a voxel threshold rather than a cluster threshold for this.

Then [1 -1] would yield I trials with more activation in the first half
than second (it would not include N trials).

>Finally, is the choice of HRF convolution merely one of preference?
>Are their optimal convolutions for particular designs?
>Does it matter in a boxcar design?

Yes, the HRF choice is arbitrary unless you have some reason to choose
specific parameters (time-to-peak, dispersion, etc). The optimal
convolution will depend on the subject and brain region and probably can
not be known a priori. So you could you a less constrained basis set
but
then you end up using F-tests and potentially over-fitting. There are
two
recent papers in NeuroImage by Woolrich et al and Handwerker et al
relevant
to this. And yes, it does still matter with box-car designs but much
less
than in event-related designs.

Hope this helps.

Joe

--------------------
Joseph T. Devlin, Ph. D.
FMRIB Centre, Dept. of Clinical Neurology
University of Oxford
John Radcliffe Hospital
Headley Way, Headington
Oxford OX3 9DU
Phone: 01865 222 738
Email: [log in to unmask]