Re: [Marsbar-users] contrasts comparison

[Matthew B. <mat...@gm...>]

Hi,

> I am reading the marsbar devel tutorial. At the end of the tutorial, two
> contrasts are compared (p.30, "You can see that the contrast value =96 wh=
ich
> is proportional to the change in signal for a single event =96 is greater=
 for
> run 3 than for run 1"). My question is: how can I test if those two
> contrasts are significantly different?
>
> I believe that one answer is to run a single model with run1 and run3 and
> test the contrast [1 -1] ([Run1 Run3]).

Yes - that's right.

> However, if the ROI for run1 is different than the ROI for run3, how coul=
d I
> compare those two contrasts? In this case, I wouldn't be able to test the=
m
> inside the same model (suppose I would like to test if left amygadala
> activation is greater than right amygadala activation for a single patien=
t).

Well, for a single patient, that is a problem, because you'd expect
the variances and so on to be different between the two ROIs.  You
could get round this by creating a new time course for the two
sessions, two ROIs, in matlab for example (extract for session 1, save
to variable, extract to session 2, save to variable, concatenate, then
import data from Data menu.

Why are the ROIs different?  If you've selected the ROIs on the basis
of the activation for that session, you can easily run into problems
of bias.

> A second unrelated question is: For answering the question "does area A o=
n
> average activate more for condition 1 than condition 2", marsbar calculat=
es
> a new summary time course for each ROI and uses this time course to estim=
ate
> the model. Would it be correct to calculate the mean of the contrast valu=
es
> for every voxel inside the ROI (calculated in spm con_*.img files) and us=
e
> this value as an estimate of the average activation in the ROI? How this
> procedure differs from marsbar method?

It's very little different - it's only that, if you use marsbar for
the first level analyses, you can adapt the model correctly to the ROI
autocorrelation - but your suggestion is a reasonable shortcut...

Matthew