Dear Scott,
Please note that averaged DCMs should *not* be
used for model comparisons. The model comparison
routine makes use of both observed and predicted
data. For convenience, the averaging code in DCM
has so far stored the results in a copy of the
DCM file of the first subject, keeping his/her
observed data. I understand that it may seem
plausible to perform comparisons on averaged
models, but mathematically this does not work
within the present setting. I have now changed
the code in SPM such that subject-specific
information is removed from averaged DCMs and the
use of averaged DCMs for comparisons is
prevented. This code will be available with the next update of SPM.
Best wishes,
Klaas
----- Ursprüngliche Mail ----
Von: Scott Fairhall <[log in to unmask]>
An: [log in to unmask]
Gesendet: Mittwoch, den 1. November 2006, 04:46:06 Uhr
Betreff: Subject-order selection effects in DCM_Average
Hi Klaas,
we met when you visited Zurich earlier this year
and a couple of times in Florence. I hope life is treating you well.
I've recently been looking into the
spm_dcm_average function. Perhaps I am misusing
it but I've noticed that when one averages across
Condition A, selecting subjects 1 through 5 and
then 'compares' this to the average again across
Condition A but this time selecting subjects *5
through 1* you get different likelihoods for what
should be two identical models.
I have a suspicion that the average posteriors
(or whatever is used in the comparison option) is
actually just from the first-selected subject. As
if you compare model A to model B using either
the single subject or the average of five, you get the same result.
Let me know what you think.
All the very best
Scott
|