> basic question but no so easy answer:
> I have a 2 tasks x 3 stimulus types task design. The two tasks are
> performed on separate sessions and the stimuli are the same across
> ssessions ...
>
> To compare the tasks, my reviewer proposes me to do a two-sample t-test
> ! Well, it is the same subjects (repeated design) and my choice was a
> paired t test .. but his/her argument is that it is separate sessions
> for the two tasks :-\
>
> In practice I would say that the two-sample will reduce the power as the
> between variance will be added to the model. In the same time, I can set
> the sphericity correction and specify that measures are correlated ...
> but in this case what SPM does ? I'm not sure ..
Cyril -
If your samples truly are repeated measures (which sessions would appear to be),
then your reviewer is simply wrong.
A paired t-test is simply a two-sample t-test with the subject effects added.
There would seem to be little reason for not modelling the subject effects,
since doing so should remove a large source of error (between-subject variability),
and so give more power for your task effect.
There is no need for nonsphericity correction (at least for correlated errors)
whenever one only has 2 conditions, because there is only one off-diagonal
error component (though you could allow for different variances - ie different
on-diagonal terms - eg if you had a priori reason for thinking the two conditions
differed in variability).
Rik
|