 |
 |
Dear
Doctor Henson and other SPMers
Thank
you for you quick reply which is very helpfull for me. It has clarified a lot
of things in my mind. I have now more practical questions:
1)
If I choose to introduce null events,
should I determine the same probability occurence for this null event as for
events of interest ? For exemple A:0,33 ; B:0,33 ; C(null event): 0,33 or with
another probability occurence? For exemple A:0,4 ; B:0,4 ; C(null event): 0,2.
Note that in fact I would have more than two events of interest but I just
wanted to simplify the situation.
2)
I believe that if I choose equal
probability I could only construct a stationnary stochastic design and not a
dynamic stochastic design ? Am I right?
3)
If I’m right, which
probabilities should I enter for five conditions (4 trial types + 1 null event)
when I want to build a design matrix of a dynamic stochastic design with the
stochastic design option (yes/no) ?
4)
My last question concern the single SOA
which is asking me when I click “yes” to the design stochastic
option. I don’t know why SMP ask me to enter a SOA because I’m precisely
asking to him to constrcuct a serie of onsets? Is this SOA refer to the minimum SOA between
events?
Thank
you in advance for help me to clarifying these questions.
Pierre Gagnepain
GIP Cyceron
Inserm E0218-Université de Caen
Boulevard Henri Becquerel
BP 5229 14074 CAEN
France
Tel +33(0)2.31.47.02.60
-----Message d'origine-----
De : Rik Henson FIL
[mailto:[log in to unmask]]
Envoyé : mardi 1 février 2005
12:59
À : gagnepain;
[log in to unmask]
Objet : Re: [SPM] Priming
design questions
Pierre -
This is an interesting question.
From the point of view of a single
voxel, you will always be more sensitive to
a difference between A and B when
you use a shorter SOA, assuming that
1. A and B are randomly ordered, 2.
the total scan time is unchanged, and
3. the BOLD responses to successive
events summate in a linear fashion
(though nonlinearity - ie saturation
- is likely for SOAs < 10s, for SOAs > 2s,
such saturation is typicaly small in
size, and outweighed by the increased
efficiency of shorter SOAs)
From the point of view of
whole-brain statistical comparisons however, you
might be more "sensitive"
to A-B differences if you restrict such comparisons
to regions sensitive to A+B vs
baseline (ie if you add null events in order to
estimate A+B), by virtue of having a
smaller multiple comparisons problem.
In order words, you can employ a
less stringent p-value correction, based on
an SVC for regions activated for
A+B, rather than based on a whole-brain correction.
Note that, for this increased
"sensitivity", you need to assume that the regions
that show a A-B difference (eg
priming) will also be generally activated for
A and B stimuli versus your specific
interstimulus baseline. This may be a
reasonable assumption for priming
(see, eg, Henson, 2003), but beware that
you will not be sensitive to
regions that show a difference between A and B,
yet are only activated for one of A
or B (e.g, if A/B are the first/second presentation
of a stimulus, a region
sensitive to explicit memory may be activated for B but not
at all for A, and therefore unlikely
to be activated for A+B, and therefore unlikely to
be identified when restricting your
analysis to regions sensitive to the mean of A/B ).
In other words, there is no such
thing as a free lunch.
Rik