krishna Miyapuram wrote:
> Thanks for your reply. i do agree that the two
> contrasts are equivalent in terms of t- or F-
> statistics. But when i take the con images to a second
> level analysis, in order to compare with another
> condition that has fewer sessions, would not the
> condition with more number of sessions dominate.
Hi Krishna,
I'm forwarding this to the list again, as I'm worried that I'm talking
even more nonsense than usual...
Mauro's reply sounded much more sensible than mine, has it answered
your question? Incidentally, the post from Will Penny he referenced
wasn't actually sent on Christmas day, but on the 19th (assuming I've
found the right one)
http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind05&L=spm&O=A&P=468686
So my understanding is that Mauro's suggested contrasts should be
okay, even though the overall design is no longer balanced. As for
your original A and B contrasts, I'm afraid I'm not sure... For what's
it's worth, I think FSL (see e.g. the paper below) uses both the
contrasts of parameter estimates and the covariances from the lower
levels when estimating the higher level models. I believe this might
be the right answer to your question, since then, although the
contrast images will scale with overall multiplication of the contrast
vector, so to will the covariance, so again, I would expect the
overall results to be the same. However, I don't think SPM takes the
covariances up to higher levels, so I would guess in this case,
contrasts such as your original A and B will give different results.
http://dx.doi.org/10.1016/S1053-8119(03)00435-X
If they differ, I'm not sure if the answer is to prefer A or B (B
perhaps sounds right) or instead to avoid such contrasts, in favour of
examples such as Mauro's apparently more balanced ones.
I hope someone corrects/clarifies all this... and if not, that it is
of some small help!
Ged.
|