Dear Priyantha,

I really don't have a lot of experience of the specification of stochastic
designs in SPM, so I hope someone more experienced will correct any mistakes I
make.


> Dear All,
>
> I am trying to desing an event related fMRI experiment   with 4 event types.
>
> I hope that some of you will read my description below and tell me whether
> the model that I have created is accurate ( within limits!)
>

> There are 4 event types. Each is 8000 ms long.
>
> In SPM 99 I specified a stochastic model as follows.
>
> TR = 3.9 sec
> scans per session = 90
> 4 conditions
> null events +
> SOA =2 ( this is something unclear to me. How does one select the SOA? is
> there a rule of the thumb?)
> relative freq = 1 1 1 1 1
> occurrence probability - modulated
> basis function = hrf with time derivatives ( because I am interested in t-
> images)
> no interactions and no user specified repressors.
>
> I will have about 10 sessions.
>
> These are the questions.
>
> 1.Is this a generally  OK model?
>

No, not precisely as it stands. If I understand you correctly each "event"
lasts for 8 seconds. The SOA (Stimulus Onset Asynchrony) denotes the time
between the onset of two events (any event, not necesarily the same event
type). In your specification you have specified for the next event to occurr
2*3.9=7.8 seconds after the previous, which strictly speaking means slightly
before the cessation of the last event.
Also, on a matter of terminology (and modelling) I am not really sure I would
call an 8 second "event" an event. More importantly, I would certainly not
model it as such. I think that generally speaking when an "event" lasts for
more than a couple (2) seconds the modelling will be improved by modelling them
as epochs with a specified length.

>
> 2.How does one select the SOA?
>

This is a tricky one in your case.

First of all it is generally a good idea to pick an SOA that is not an integer
no. of scans (SPM authors, perhaps the default is slightly unfortuitous?). This
is because a non-integer number will ensure an adequate oversampling of the HRF
(e.g. an SOA of 2.5 will ensure we sample the HRF with 2/3.9=0.51Hz rather than
0.25Hz as is the case for an SOA of 2).

Secondly, an event-related study is usually thought of as a sequence of very
brief events causing short (HRF length) "blips" in BOLD signal over the
"baseline". In this instance the SOA is ultimately limited downwards by concern
about non-linear responses. I think few people would be comfy with an SOA
shorter than 2 seconds because of this. The other factor determining the SOA is
the "design efficiency", i.e. how much power it will give in being able to
detect the effects. The efficiency for a given design will be a factor of SOA,
but also of contrast weight vector (i.e. are we interested in "main effect" of
an event or are we interested in "differential effects"). It is really all to
complex to treat here, but there is a paper by Josephs and Henson in
Proceedings of the Royal Society that provides in my opinion the best
explanation.

Very briefly, had your events really been events (i.e. not 8 second epochs) 7.8
seconds SOA with null-events would have been a pretty good choice.

Now, seeing you have these 8second "mini-epochs" I would suggest the following.
Use an 8 second SOA (2.05 scans is not bad from a sampling perspective) which
would basically turn this into an epoch-related study with a random order of
the "epoch-types" (including the possibility of run-on epochs of 16seconds
length). With this design it won't suffice to include "null-events", but you
would actually need to specify it as 5 trial types (the fifth being
"null-epochs") each modelled as box-cars convolved with the HRF (possibly
including the derivatives) of 8 seconds length.

>
> 3.once the scanning is done, how does one select the correct scans? Is there
> a trick? Or do  you just count them?
>

I am not sure I understand this question. Could you elaborate what the problem
is?

>
> 4.When I look at the temporal vectors that SPM generates I realize that for
> some sessions there are 38 events and for others 34. Is this correct?
>

What you see are the non-null-events and these could certainly vary. I suspect
though that the sum of non-null and null-events should be constant across
sessions (although I am not really sure what "modulated occurence probability"
means).

>
> 5.Is it possible to look at the time vector in matlab space in terms of
> scans and not in terms of seconds?
>

What you see in the graphical display is the event-train in terms of scans. In
SPM_fMRIDesMtx.mat you will find the variable Sess, and in Sess{i}.ons{j} you
will find the onset times of event type j in session i in terms of seconds. To
see this in terms of scans simply amounts to dividing that vector with the
repetition time, i.e. Sess{i}.ons{j}/3.9.

Good luck Jesper