On Thu, 4 Aug 2005 16:57:24 -0400, Xu, Ben (NIH/NINDS)
<[log in to unmask]> wrote:
>Hi,
>
>I hope someone can help us with the following questions.
>
>We are planning a motor-function related fMRI study with four event types
(A
>B C and D) which must occur in this sequence. During a scan, the subject
>will be holding a small device in a hand. Here are the events:
>
>- Event A: sudden increasing in the weight of the device;
>- Event B: holding the device with the added weight;
>- Event C: sudden decreasing in the weight of the device;
>- Event D: holding the device with the decreased weight
>
>The main interest of the study is in the following contrasts: 1) A vs B
and,
>2) A vs D
>
>Questions:
>
>1. Will an event-related fMRI design be appropriate for this study?
You can use an event-related design if you wish. The question is whether
it's appropriate for the question you're trying to study, including issues
like statistical power to detect an effect. Of course, the costs/benefits
of ER fMRI must be compared to other techniques.
>2. Is it possible to model event B or C which are non-transient events
>(i.e., will last for several seconds) in an event-related
> design?
You can specify that particular events have a non-zero duration. Whether
the model is really valid is another matter (see below).
>3. If question 2 is possible, how should the onset time of B and / or C be
>specified during analysis?
Presumably when they first occur.
Problem is, of course, that (at least as I read it) the onset of B is
coincident with event A (which has zero duration). There are two issues
here:
(1) Because A and B are always adjacent in time, it will be difficult to
untangle the corresponding BOLD response. It's not actually _impossible_,
because if you look at the modeled response (basically, A is represented
by a delta (spike) function convolved with the model, canonical
hemodynamic response function (HRF), whereas B is an indicator (or boxcar)
function convolved with the HRF), they're not precisely multiples of one
another. However, the statistical "efficiency" in untangling responses is
bound to be very low, and I imagine you have to worry more than usual
regarding how well the canonical HRF models the actual HRF.
(2) Modeling the neural responses as a delta function for A and an
indicator function with B (both of which get translated into BOLD
responses via convolution with the HRF, as above) involves some very
strong assumptions. Perhaps the response to A isn't just a relatively
brief spike, but rather a spike plus something longer lasting (if
transient). And for B, perhaps the actual neural response is steady for
the first half of the event, and then a gradual decline in the second
half. One doesn't really know (without external evidence), and the
problem is that choice of an incorrect model will bias the results.
The best thing to do in such situations is come up with a very clever
design that finesses such issues.
>Thanks for any advice!
>
>Ben
>
|