Siobhan/Lee -
If you think the neural processes engaged by your
trials last more than ~1s, then you should
model them as epochs (ie duration>0), rather
than events (duration 0). This is because the predicted BOLD response, after convolving your "neural model" with
your "haemodynamic model" (eg, SPM's
canonical HRF), will differ in shape
for events >~1s.
So if you think that the neural processes of
interest in your trials last until the response is made, then choose durations equal to
the RT.
Regarding the use of the partial derivatives of
the HRF, the dispersion derivative will be unable to
capture such large changes in shape that arise for neural durations
>~2s.
Including derivatives can change the results
(of a T-test on the canonical HRF) if their regressors are correlated with that
for the HRF. The basis functions themselves are orthogonalised by SPM, but there
are situations where the regressors become correlated after convolving the basis
functions with the "neural model": 1) when there is a non-random ordering of
different event-types, 2) when there is large undersampling of the response (ie,
the effective sampling rate is too low), 3) when the neural model is an epoch of
>1s rather than an event.
The fact that your results are changing
dramatically with the inclusion of the derivative suggests that one of the above
applies.
Bas is correct that collinearity can arise for
one of the above reasons, though I would add that collinearity is usually
minimal, even for rapid ER designs, for true events (duration
0) whose types are randomly ordered (as is common).
Note however that such collinearity is NOT a
problem, if you use F-tests to make inferences - ie tests that include all the
basis functions (canonical HRF + derivatives). This is because such tests
"include" the "shared" variance (whereas T-tests only test the "unique" variance
associated with a regressor).
Finally, if you
want to stick with T-tests on the canonical HRF, but you intend to model nonzero
durations for your trials, you should probably not include the derivatives.
Nonzero durations introduce correlation between the regressors (point 3 above),
as was noted by Jesper:
Hope this helps
Rik
----------------------------------------
Dr Richard
Henson
MRC Cognition & Brain Sciences Unit
15 Chaucer
Road
Cambridge
CB2 2EF, UK
Tel: +44 (0)1223 355 294 x522
Fax: +44 (0)1223 359 062
----- Original Message -----
Sent: Wednesday, October 26, 2005 9:06
AM
Subject: Re: [SPM] Time & dispersion
derivatives
Hi Lee,
did you check the collinearity and
multi-collinearity of your design? With
rapid ER-fMRI design there is a
real chance that events from trial n+1
explain variance caused by events
for trial n, especially when using Taylor
series expansions. That might
explain the dramatically different findings, it
would suggest your design
is sub-optimal.
Kind regards,
Bas
Op woensdag 26 oktober
2005 04:11, schreef Siobhan M. Hoscheidt:
> SPM Users,
> I would
like some opinions regarding the optimal way to analyse a
>
dataset. The study is a 2x2 event-related design, crossing memory
type
> (semantic, episodic) with spatial content (spatial,
nonspatial). Each
> trial includes a memory question that the
subject reads and then responds
> to, lasting a total of 8
seconds. Subjects respond to the question with a
> button press,
usually 3-4 secs after the onset of the trial. The time
> period
after the button press, based on their rt for each trial, is
> specified
as a separate "wait" condition and is not of interest. There is
>
a separate control condition (reading a string of nonwords) that also
>
requires a button press after about 3-4 secs.
>
> I'm currently
analysing the data in spm99 as event-related, with 0 duration
> and no
global scaling, and got some small but reliable regions of
> activation
when comparing the various conditions to the control
> condition.
I then re-analysed the data, also using 0 duration and no
> global
scaling, but adding time and dispersion derivatives. Now when I
do
> a t-test on the canonical HRF component, I get a very different
pattern of
> activations that are much "messier".
>
> I'd
appreciate hearing a) how people would deal with long trials (3-4
>
secs). Would you specify 0 duration or use the rt as the
specified
> duration for each trial? And b) Why does including the
time and dispersion
> derivatives change the t-test results for the
canonical HRF so
> dramatically? Does anyone have suggestions on
the best way to analyse this
> dataset?
>
> Any guidance or
advice on the matter would be greatly appreciated.
> Thanks,
>
Siobhan Hoscheidt
> Lee Ryan
--
--------------------------------------------
Dr. S.F.W.
Neggers
dept. of Psychonomics,Helmholtz Institute
Utrecht
University
Heidelberglaan 2
3584 CS, Utrecht, room 17.09
the
Netherlands
Tel: (+31) 30 253 4582 Fax: (+31) 30 2534511
E-mail: [log in to unmask]
Web: http://www.fss.uu.nl/psn/web/people/personal/neggers
--------------------------------------------