Hi Bart

> I have a (kind of) sparse-scanning design: Acquisition time = 2.4 seconds,
> then a silent gap of 1.6 seconds, and then the next 2.4 sec scan starts.
> Short auditory stimuli (+/- 1000 ms) are presented in the silent gap (at 2.6
> seconds after the start of the scan).
>
> So I guess this makes TA=2.4 and TR= 4 (interscan interval)

Right.


> I apply an ascending scan sequence with 46 slices and I applied slice
> correction to the middle slice (in time and space), which is slice 23 (this
> slice is scanned around time 1.2 seconds from the beginning of the TR).

A quick note here, maybe others can comment more: because slicetiming
correction uses data from other scans to interpolate the data, it's
less accurate at longer TRs.  Also, I'm not sure off the top of my
head if it copes gracefully with sparse designs (i.e. mismatch between
TR and TA).  Hopefully someone else can comment more on this; however,
as a matter of course, I never use slicetiming correction on sparse
designs.  (Actually, as you probably know, there's some debate on
whether it should be used at all, but that's another story...)


> As I understood it would be advisable if microtime onset fMRI_T0 coincides
> with the reference slice (thus at time 1.2 seconds). Based on Rik Hensons
> mail (dating from 2002!) I understood that microtime resolution fMRI_T
> refers to the number of time-bins per TR and not per TA.

Yes.


>  Consequently, I
> calculated fMRI_T0 as follows:
>
>             fMRI_T0 = fMRI_T * TA/TR * 1/2
>             fMRI_T0 = fMRI_T * 2.4/4 * ½
>             fMRI_T0 = fMRI_T * .30
> To operationalize this with integer values I chose to put fMRI_T = 20 and
> fMRI_T0 = 6.

Yes, exactly.


> Hence I specified the design in scans with an interscan interval of 4 sec.

Sounds good.



> To specify the stimulus onsets I reasoned as follows:
>
> The stimulus always starts 2.6 sec after the start of the scan (2.6/4 =
> 0.65)
>
> Zero-point in time fMRI_T0, however, was put at 1.2 sec or at scan 1.2/4=
> 0.30
>
> This implies all scan timings are shifted with 0.30 scan (or 1.2 seconds).
> Hence, the start of a stimulus is at time-point scan 0.35 (instead of 0.65)
>
> Is this correct so far?

I would just leave the times at 0.65; maybe others have more of an
opinion on this when using slicetiming correction.  However, I
generally try to specify the events actually when they occur in the
scan (as you've correctly calculated at 0.65).



> A next question: Stimuli are presented in blocks of 4 (acoustically
> different) stimuli that belong to the same (phonological) condition. Since I
> am not interested in the influence of this acoustical variation, I intend to
> analyze them as a block design. In view of the above timing considerations,
> I specified that every block starts at scan x.35 (the timing of a first
> stimulus in a block) and had a duration of 3.35 scans (i.e. until the start
> of the ‘fifth/next’ scan acquisition). Instead of a duration of 4 scans I
> specified 3.35 scans with the idea of ‘cutting out’ the noise of the next
> scan acquisition (since we are interested in auditory perception). Is this
> reasoning correct or am I cutting out relevant signal as such?

That seems fine to me.  I suspect it wouldn't make a huge difference
in practice, due to the somewhat sluggish timecourse of the
response...I.e. once you convolve your 3.35-scan length block with an
HRF, it's going to overlap the next scan, anyway (and of course, all
the scans that occur DURING the 3.35-scan block mean you can't really
get rid of this influence anyway; you just rely on the fact that it's
consistent across conditions).


> Given that the stimuli only take about 1000 ms and the whole TR is 4000 ms:
> would it be better to model the data as an event-related design, so that
> individual stimulus timings are more exactly specified? Would this be an
> advantage over a blocked design?

One answer is to try it both ways and look at the model fit to see
which seems like it does a better job. :)  In general, though, I would
say that you want your model to reflect what is really going on in the
neural processes you are interested in.  In your case, if the
processes you are interested in are really only occurring during the 1
sec presentation, I would suspect that an event-related design will do
a better job of capturing them.


> Finally: does it make sense to apply slice timing correction in a blocked
> design or should I better drop the slice timing correction and put fMRI_T0
> and fMRI_T simply to 1 and 16?

The slicetiming issue is independent of the T/T0 issue; you can adjust
one without changing the other (although of course, if you're doing
both, it makes sense to have them agree).  I would skip slicetiming,
but set T and T0 as you've done above.  In theory setting T0 to the
middle of the acquisition (as you've done) should be slightly more
accurate than always having it at the beginning; in practice, I
wouldn't expect it to make a huge difference.

Hope this helps!  Good luck.

Jonathan