Steve,
>Is this heading in the right direction or are your issues something
>different?
This is exactly the issue that I'm struggling with.
Let's say you have 4 regressors: EV1=cue1, EV2=working memory, EV3=cue2,
and EV4=response. EV1, EV2, and EV3 are temporally continuous and the
convolution makes them collinear. Brain Area 1 (BA1) is significantly
correlated to EV1, BA2 is correlated to EV2, and BA3 is correlated to EV3.
BA1, BA2, and BA3 are non-overlapping.
1. What can we say about the estimability of EV1, EV2, and EV3? It seems
to me this should be related to the variance of the residual between each
EV and it's reduced data.
2. With regard to false positives: Assume that cue1 lasts a very short
time (a few hundred milliseconds) and the delay period/working memory is
equally short. I don't doubt that we can detect a change in amplitude
that's correlated with EV1 and EV2, but there may be enough noise in the
data that variance associated with EV1 is attributed to EV2 and vice
versa. So, it's not that I'm afraid of not seeing activation but rather
that the activation is attributed to the wrong regressor. How confident
can I be that regressors parse the variance correctly?
3. What can we say about the function of BA1, BA2, and BA3? That is, can I
make the statement that BA1 activity represents cue encoding, while BA2
activity represents working memory? Or must I always do a contrast between
EV1 and EV2, such that the [1, -1] contrast represents cue encoding while
the [-1, 1] contrast represents working memory? If the latter, what about
EV3 and EV4... do I need to include them in the contrast as well?
thanks a lot!
jack
|