the two-sample t-test would be the right model if both samples were
independenlty drawn from their respective population. But if sampling has not
been independent - eg controls were pairwise matched with patients for age,
gender, education or something like that - then the paired t-test would be
the "right" model and a two-sample t-test would be "wrong".
You can enter the "effect of interest" by typing "3:N" into the line "columns
for reduced design", where N is the number of pairs in the t-test. This would
exclude the all columns from your design matrix except the first two, which
are modelling the different conditions.
However this F-contrast would not help to find out what you want, since the
parameter estimates of the F-contrast are mean-corrected for each pair of
data. This means for significant voxels you will always get a positive and a
negative estimate although it could be true that both conditions activated
(i.e. have a positive estimate in the respective one-sample t-test).
In order to get meaningful parameter estimates you can either extract the
values from the data-images and plot these separatly for the groups or you
calculate an ANOVA with two groups and not independent data and define a
model without mean corrections and then plot the estimates from the
F-contrast.
Good luck,
Thilo
On Wednesday 27 August 2008 12:30, cyril pernet wrote:
> re Henrike
>
> > to compare patients and control we used a paired t-test.
>
> you have to use a two-sample t-test - paired is for the same subjects
> tested twice (in time or conditions)
>
> > Each pair consisted of a patient and his/her matched control. In the
> > contrast "patients > controls" we found activation differences in the
> > amygdala. To find out why these differences exist (Both groups show
> > activation, but the patients show more activation? The patients show
> > activation, but the controls deactivation? Both show deactivation, but
> > the controls more deactivation?) I would normally define an additional
> > F-contrast "Effect of interest" ([1 0; 0 1]) and look at the parameter
> > estimates. SPM does not allow me to do that? Why is it not possible in
> > paired t-tests?
>
> well I guess I don't need to answer this since this is the wrong model ...
>
> C
--
Thilo Kellermann
Department of Psychiatry and Psychotherapy
RWTH Aachen University
Pauwelsstr. 30
52074 Aachen
Tel.: +49 (0)241 / 8089977
Fax.: +49 (0)241 / 8082401
E-Mail: [log in to unmask]
|