Hi Scott, list,
try adding N-1 session constants (1-s for scans belonging to a session,
0 otherwise) when you have N sessions. Together with the default
constant regressor (all 1-s), your model spans the entire space of
possible different session-averges in your data. Adding N
session-constants to your model would make your design over specified
due to multi-collinearities, since the sum of the 5 regressors explain
the same variance as the default constant. The inv(X) does not exist
(and the GLM is not solvable). Or, less formal: you can envision I guess
that when having 3 sessions, 2 session specific constants and an overall
constant can explain any jump in baseline over sessions that you can
imagine.
I tried that in a similar case, and it works. There is a problem now
though with high-pass filtering, since the DCT regressors modeling
low-frequency drifts (invisibale in the interface, but in your model
anyway) are then calculated over the entire data set, as if it where
recorded at once, e.g. spm does not know it are separate sessions, which
is not correct. You'd have to add them manually for each session
separately. I am not sure whether modelling serial correlations (the
AR(1) + w model for autoregression) is now appropratiate, perhaps
someone else could comment on that one?
Regards,
Bas
Scott Fairhall wrote:
>Hello all,
>
>To facilitate DCM, I have concatenated a series of five sessions into a
>single series. There are obvious (though not terrible) effects of inter-
>session motion apparent in the series.
>
>It struck me today that perhaps if I set up five session constants (as in
>the 'proper' session design, which would have ten regressors) for the
>concatenated series (so a total of six regressors) this may alleviate the
>problem somewhat. Then again, perhaps not…
>
>My specific questions are two:
>1. Have I correctly understood the role of the session constant? and might
>this help reduce the variance seen over the two scans which lie on the
>cusp of two sessions?
>2. Is there a way to implement this directly or be altering the SPM.mat
>data structure? (when I use 'regressors' the results are not quite right
>and I would need to remove/change the default constant also).
>
>Thanks is advance
>Scott
>
>
|