Thanks, Joe.
Indeed, you are correct about the contrasts and I have been rather worried
about the potential of bias associated with the HRF. The total number of
errors was not large, but given the relative scarcity of Stop trials (~30 in
total), even a few errors can significantly bias results.
I think the reviewers are concerned with a group x Stop trial performance
difference. Thus, I may re-analyze my data again, this time only modeling
correct and incorrect Stop trials and creating two contrasts, [1 1] and [1
-1] to model stop v go and correct stop v incorrect stop, respectively. Any
thoughts?
Best,
David
_______________________
David Glahn, Ph.D.
Department of Psychiatry and Research Imaging Center
University of Texas Health Science Center at San Antonio
Voice (210) 567-5508
Fax (210) 567-1291
-----Original Message-----
From: Joseph Devlin [mailto:[log in to unmask]]
Sent: Friday, August 20, 2004 9:26 AM
To: [log in to unmask]
Subject: Re: [FSL] go-no go
Dear David,
I assume the same contrasts were being looked at across the analyses? For
instance, if you are comparing Go to NoGo trials, in the initial analyses
the contrasts would simply be either [1] or [-1], depending on the direction
your interested in. Assuming you are only really interested in correct
trials, that contrast would change to either [-1 0 1 0] or [1 0 -1 0]
assuming an order of correct, incorrect, correct, incorrect. Is that right?
I realise there may be group comparisons as well but I'm only asking about
the simple main effects within group, for the moment. By modeling all of
these trials, your design presumably became rank deficient, whereas it
wouldn't have been previously. So in the second analyses, it is important
to use contrasts which add to zero rather than things like [1] or [-1] --
that could certainly lead to strange z-values.
There is another issue in the design you mentioned -- namely if the ISI
matches the TR then there is an intrinsic (and strong) bias when sampling
the HRFs that can dramatically affect the results. By changing the contrast
to an actual comparison on EVs, this can either over or underestimate
effects. In other words, the new contrasts may underestimate the actual
BOLD differences between conditions.
Finally, depending on the number of errors subjects made, the modeling could
be telling you that most of the results in the original analysis were driven
by errors (presumably this is what you reviewers are worried about). Were
there significant error rates and did these differ across groups?
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]
|