Hi Branka,
> Epoch as a series of events:
> I have a (parametric) block design. With 6 conditions, 9 s baseline
> and 18 s blocks, TR = 3 (TA = 2.53), each block contains 8 stimuli (2
> s on, 250 ms fixation). (conditions repeated 2x per session, with 4
> consecutive sessions per subject)
>
>
> Q1: Would it be valid to use the Mechelli et al, 2003 approach, and
> model this design as a series of events (rather than an epoch) given
> that the events of interest do not coincide with the TR. (I think it
> was a similar case in the data set presented by Mechelli et al)
There is no principle difference between epoch- and event-related
studies, so yes it would be valid.
>
> Q2: Should I do a slice timing correction (given non-coinciding
> acquisitions and events).
The slice-timing correction is _not_ intendend to "deal with"
non-coinciding acquisitions and events (which isn't a problem to begin
with), but rather with the fact that different parts of the brain is
imaged at different points in time. Let us say you use a descending
acquisistion with a TR of 4 seconds. In that case the bottom-most slice
is acquired ~4 seconds after the time when the top-most slice was
acquired. Hence, a model that is "good" for the topmost slice will be 4
seconds "too early" for the bottom slice. That is the problem that
slice-timing correction tries to deal with.
The important thing is that you tailor the model to the middle slice so
that no part of the brain is more than +-0.5 TR off. Slice-timing might
also help a little, and it will never do any harm.
> Effects of stimulus type:
> There are 4 stimulus types within each epoch (2x2 factors). I didn˙t
> design the study with the aim of investigating these effects (this is
> my first fMRI experiment, so I was trying to keep it manageable), but
> if analysis allows me I would like to.
>
>
> I have 2 ideas as to how I could do this.
> 1) If I model each epoch as a series of events, I could model each
> stimulus type separately.
>
>
> Q3: Would this be a valid approach?
It would be a valid approach. Chances are though that it is a quite
inefficient approach (i.e. that there is a small chance of you finding
anything). For a rapid event related presentation the most efficient
design is if you present the different event types randomly (i.e. at
each SOA there is an equal chance of all event types, regardless of what
the previous types were). If in the other hand (in contradistinction)
they are always presented e.g. like 12341234, where each number
represent a certain type of event, your chances will be limited. It
would still be valid though.
> 2) The other idea that I had was to include stimulus type as a
> parametric modulation, but then I would be stuck with all 4 stimulus
> types together. I guess I could still look at the interaction effects,
> and at least be able to state that stimulus type did/did not have an
> effect.
I assume you mean you would put in a regressor (in addition to your
blocks) where each stimulus type was given a weight of 1 or -1 such that
they added up to the interaction you are interested in. This is also
valid, but suffers from the same potential problems if the ordering of
the events is not random. Also it is less flexible than your first
suggestion. I would go with the previous option.
Good luck Jesper
|