Hi Ed,
that's the beauty of mixed-effects modelling. Your analysis looks OK -
can't see anything wrong in the way you've set it up - problem is your
extremely low number of degrees of freedom - that will reduce z-scores
dramatically and also cause the higher-level estimates of the
random-effects contribution to be very poorly conditioned (read: not
very accurate). You might want to think about not modelling the
between-scan variation and just treat it as a fixed-effects problem.
cheers
christian
On 16 Apr 2004, at 02:57, Ed Vessel wrote:
> Hi folks -
>
> I have a simple experiment with a block type design (16 second blocks)
> in
> which I show four types of images plus periods of fixation, 4 times in
> a
> scan. I then do another scan of the same conditions, with a different
> block
> ordering.
>
> I analyzed each scan individually using four EV's and creating a
> contrast for
> each image type, plus some contrasts I am interested in. I got a lot
> of very
> high z values in my contrasts (like as high as 13) which were
> consistent in
> both scans, so I thought that combining the scans would give me a good
> boost
> in power and reduction in noise.
>
> However, when I combine the two runs in a higher level analysis (by
> selecting
> both lower level feat directories as inputs and selecting all my
> contrasts),
> the resulting maps are consistent with what I saw before, but the max z
> values are now down around 2-2.5! This is true both for the contrasts
> of
> single EV's (i.e. 1 for one EV and 0 for the rest in the low level
> analysis)
> and for contrasts between EV's (i.e. 1 -1, or in one case, 3 -1 -1 -1).
>
> Am I combining these two runs incorrectly? Or is it expected for the
> z values
> to go down so drastically? here is my higher level GLM setup (its
> super
> simple):
>
> Number of EV's: 1
> Number of groups: 1
>
> Group EV1
> Inp 1 1 1
> Inp 2 1 1
>
> Contrasts 1
> EV1
> C1 Combine 1
>
> This is, of course, run on each of the lower level contrasts. I had
> expected
> an "average" of my two runs, which is basically what I get, but with
> much
> reduced z scores. Should I be including an EV for "run" or something
> to soak
> of variance between the runs?
>
> thanks,
> Ed
>
> --
> Ed Vessel
> New York University [log in to unmask]
> Center for Neural Science
> 4 Washington Place, Suite 809
> http://cns.nyu.edu/~vessel
> New York, NY 10003
> (212) 998-3908
>
--
Christian F. Beckmann
Oxford University Centre for Functional
Magnetic Resonance Imaging of the Brain,
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK
Email: [log in to unmask] - http://www.fmrib.ox.ac.uk/~beckmann/
Phone: +44(0)1865 222782 Fax: +44(0)1865 222717
|