Jack,
There are basically two questions here, the effects on the parameter
estimates and on the statistics:
- Tim has already emailed that, because of the orthogonality, the
inclusion of the temporal derivative does not change the parameter
estimate for the original EV of interest
- the statistics for the EV of interest will change, simply because the
temporal derivative will explain additional bits of variance that in
the absence of the temp. der. end up in the residual noise - lowering
the stats.
You can potentially use a hypothesis test which includes both
parameters simply by treating the EV and the temporal derivative as a
set of basis functions and using an F test. I think Calhoun et. al. did
something slightly different, but the F test, potentially combined with
post-threshold masking by the Z for the 1 0 contrast over the EV of
interest, would be my choice if I was really interested in the combined
effect. Unfortunately, this (F + contrast masking at every level) gets
quite messy and complicated. IMHO the increase in power typically does
not justify the hassle (and inconvenience of ending up with a
two-tailed test).
I also would argue that sometimes you do not want to combine the EV and
the temporal derivative in a single test, because they encode different
things: the PE for the EV of interest is the average amplitude
modulation of the effect of interest while the PE for the temp.
derivative encodes the average amount of mis-specification in the
temporal domain. If you combine the two into one single hypothesis
test, any post-thresholded 'blob' can appear for one of two reasons (i)
because the effect of interest was there _and/or_ (ii) the design was
badly specified.
I think you are much better off by doing what Serge suggested (Z
contrast over the temporal derivative) to 'learn'
about the amount of mis-specification in the temporal domain and then
use a better hrf model instead...
cheers
christian
On 27 May 2004, at 06:31, Jack Grinband wrote:
> Hi All,
> I just wanted to confirm that FEAT only includes the non-derivative
> portion of the model in its
> hypothesis testing and that the parameter estimate associated with
> temporal derivative is not
> passed up to the group level. Is this correct?
>
> If the derivative portion explains a significant portion of the
> variance, you can get an artifactual
> decrease in the non-derivative parameter estimate? Is this right?
>
> I recently read a paper (Calhoun et al, 2004) which suggested using a
> hypothesis test that includes
> the parameter estimate for the derivative term. Does anyone have any
> opinion on this?
>
> Finally, is the PE for the derivative saved during the FEAT analysis?
> Where could I find it?
> thanks,
>
> jack
>
--
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
|