Hi David,
I don't think using randomise gains you anything. If a subject has no
responses of a given type, then that subject simply can't contribute any
information for *any* contrast involving that EV. Your only solution is
to construct higher level analyses using only the subjects that have
non-empty EVs on all the EVs that contribute to the contrast.
Note that if you have a single EV contrast such as 1 0 0, then including
any subject with an empty EV1 will result in a blank higher level
contrast. This is somewhat easy to detect because of the blank map.
However, if you have a contrast such as -1 0 1, then if only EV1 or EV3
is missing (but not both) in individual subjects, you'll get a non-blank
group map, but the result is still not interpretable since the data
wasn't available to form the true difference contrast for each subject.
This case is more insidious because the group map would be non-blank,
yet it isn't truly testing the hypothesis that you want to test.
cheers,
-MH
On Thu, 2012-04-19 at 10:35 +0100, David Soto wrote:
> Hi,
>
> I know the issue of having empty EVs for run subjects run has been addressed
> but still cannot figure out my case
> Each subject received a single fMRI run and the trials are classified in 3 EVs
> as function of the subjective visibility of the visual target (1- not seen, 2 maybe, 3- sure).
>
> I want to do a linear contrast eg. -1 0 1 across the EVs.
>
> The thing is that some subjects dont have one of the EVs. Given there is a single fMRI run
> this is a bit worrying...
>
> Would still be OK to use the empty EVs ? would the higher level analyses with the linear contrast
> still work in FEAT?
>
> I was thinking or re-labelling my EVs in 2 levels rather than 3 but this is not ideal....
>
> any other sugestions please? would perhaps randomise offer a better approach?
>
> Thanks for any input,
>
> regards, David
|