Dear Stephan,
>we ran into a problem while defining contrasts using the contrast
>manager. Our design is 8 events/session (146 volumes/session,
>event-related design) with 5 sessions overall in one design matrix.
>Defining a new t-test contrast for the first event of all sessions ("1 0
>0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 0
>0 0") fails with error msg. "!invalid contrast". Using "1 -1 ...."
>instead surprisingly works. Has anyone a clue for this ? May it be that
>the design is to large ?
No, I don't think that there is any problem with the size of the
design matrix. I have often used much larger ones than this.
Presumably the parameter estimate for the first event is not uniquely
specified. This is the case if the first regressor can be reproduced
by a linear combination of other regressors. To check this, you can
look at the parameter estimability bar which appears below the design
matrix (after you have specified a contrast, you are offered the
design matrix as an option in the bottom left hand window). This can
be a little difficult to see when you have lots of regressors, but
none the less, you may well find that the bit of the bar which
immediately under your first regressor is grey instead of white. If
so, then your design matrix is overdetermined.
I must admit that it is a bit unusual in an event-related design that
you have co-linearity between regressors (although in epoch designs
it happens all the time). Is it possible that these events have been
modelled twice in your design matrix? The fact that a '1 -1 ....'
contrast is valid implies that the problem is that regressor 1 has
variance in common with regressor 2. Therefore they can be compared
validly, since the shared variance essentially drops out in the
comparison. Looking at the parameter estimate from just one of
them, though, is ambiguous. How much of the shared variance is
carried by this regressor is entirely arbitrary. If you can somehow
make sure that the common component is only modelled once, then you
may be able to use a '1 0 0 ...' contrast. But look out; the
interpretation can be complicated in this situation.
Best wishes,
Richard.
--
from: Dr Richard Perry,
Clinical Lecturer, Wellcome Department of Cognitive Neurology,
Institute of Neurology, Darwin Building, University College London,
Gower Street, London WC1E 6BT.
Tel: 0207 679 2187; e mail: [log in to unmask]
|