> Dear SPMers,
> I would like to know if my modification of hrf is suitable for our event
> fMRI study analysis.
> We did an event related study about sentence comprehension. Our concern is
> the relation between the word order in a sentence and cerebral activation.
> Event length was 30 seconds and TR=1.
> There were two conditions (two types sentences), A and B, in a random order,
> 12 events for each (then 24 x 30s). The total scan time was 720 second and
> the total scan was 720 (actually the total number of the scans was 725 and
> the first 5 scan was not used).
> The stimuli sentences were consist of 3 words (e.g., John kissed Mary)
> presented word by word. Each event was 30s and the first 10s was rest. At
> 11s, 12s and 13s, the first, the second and the third word was presented.
> Then, to test the subjects surely read the stimuli, a question sentence was
> presented at 15s and the subjects responded by button press (the response
> occured around17-18s). The rest of the event (from 18-19s to 30s) was also
> rest (a small fixation was presented).
> We would like to see the activation againt the first word. (type A sentence
> is more difficult than type B to understand because of its word order)
> All our concern is the cerebral activation occured 11, 12, 13s.
> Then I changed spf_hrf
> delay of undershoot 16 -> 7
> length of kernel 32 -> 12
I do not quite understand the rationale for changing the shape of the synthetic
HRF. These parameters would give you a slightly squashed (shortish) HRF without
appreciable undershoot. Do you have any data from these specific subjects to
support this particular HRF?
> At fMRI models and estimation, I selected all 720 scans.
> condition =2
> SOA = variable
> I set the vector onset for the two conditions as
> for condition 1: 11, 101, 131..., 701
> for condition 2: 41, 71, 161, ..., 671
> Those are the onset time of the first words presentation (at 11s).
> For 12s, vector onset was set as, 12, 102, 132, ..., 702,
> and For 13s, it was set as, 13, 103, 133, ..., 703 for condition1 and did
> the same way for condition 2.
> The result seems reasonable. But I am not sure if this process is correct or
> Is this way of analyzing is wrong? Or are there any better ways?
If I understand you correct your primary concern is the difference between
responses to words from sentence category 1 and sentence category 2. So for
example you are interested in w1c2 - w1c1 (w1 for word 1, and c2 and c1 for
categories 1 and 2).
I can see a slight problem with you design/analysis. For each of your "events"
you have effectively bursts of three events with a 1sec spacing, followed by
another (variable length) burst and finally a last event of a different kind.
Since the spacing within the bursts are shortish compared to the HRF you will
have a rather poor determination of the response to each type. So in you first
analysis above where you model only the variance from the first word there will
be a strong contribution from the subsequent two words (and your attempt to
"squash" the HRF doesn't really address this).
Strictly speaking, in order to look at the difference w1c2-w1c1 you would need
to include also the other words in your model, i.e. you would need to use at
least 8 conditions in your model, six for each of the words in each category
and two to try to capture the variance from the question. Ideally you would
need to model the question sentence word-by-word as well. Having done that you
could look at w1c2-w1c1 after all variance explicable by the other words had
been regressed out.
The problem with that is that there will be very little of it left and you will
be very insensitive.
I think your best bet would be to (possibly) model the three words as separate
conditions, but assessing the category difference with an F-contrast. I think
the easiest way to do that is to use three conditions to model all three words
in BOTH categories, and another three conditions to model them in one of the
categories (arbitray). Your F-contrast should then compare the model with and
without the second set.
However, that still leaves you with the problem of removing the variance caused
by the question. I don't really see a good solution for that (perhaps someone
else does). If you haven't performed the entire study yet I would suggest
redesigning it such that you put the question well away (>10sec) from the
presentation of the target sentence.
The above has assumed that your main interest is category differences. If your
main concern is word order within a sentence, I think that is a very tricky
question. You could try modelling event-by-time interaction within each little
three word burst, only it would look very much like a temporal derivative and
would therefor be very difficult to interpret. Also, for any sentence presented
at anywhere near a natural pace there will surely be non-linear interactions
between successive "events" (words), and these would be hard to disentangle
from "true" position differences.
I don't really have a suggestion. I would love to hear if anyone has any ideas
about how to model something like this.
Good luck Jesper