Dear Eric, Bruno and Russ
The difference most probably comes from the convolution
of your box car with the haemodynamic response.
What Eric wrote would be perfectly exact with "true"
box car. After convolution, the space of two complementary
box car functions is no longer completely redundant (because
of the begining of the first "on" period, I guess)
This should explain the difference found with the degrees of
freedom. I would simply test one model against the other using
an F contrast.
hope this helps
JB
> Bruno.
>
> > we have read the question by Russ Poldrack concerning contrast
> > definition:
> >
> > http://www.mailbase.ac.uk/lists/spm/1999-06/0064.html
> >
> > We have the very same questions and there seems to exist no
> > reply to it so far. Can anyone help us?
>
>
> I'll take a stab at this one in terms of the general question
> posed by Russ:
>
> "1) I used spm99b to analyze an fmri dataset with 4 conditions. I tried it
> in two ways: specifying all four conditions explicitly, or specifying
> three conditions and leaving one unspecified (as an implicit baseline
> condition).
>
> I then did the analysis either by specifying a contrast (1 0 0 -1) in the
> first model or (1 0 0) in the second model (where the fourth condition was
> the implicit baseline). The glass brains look about the same but the
> statistics are appreciably different, e.g., one cluster changed in size
> from 240 to 202 voxels; I'm guessing that this has to do with the slight
> different in dF between the two models.
>
> It seems to me that the model with all four conditions specified is more
> appropriate than the implicit-baseline model, particularly if one wants to
> go on and do comparisons of the non-baseline conditions - this is based
> upon the intuition that these comparisons will necessarily include the
> implicit baseline in that model in a way that could lead to misleading
> results. Is that correct, or are there other reasons to use an implicit
> baseline beyond the 2-condition situation?"
>
>
> I cannot explain why the four and three condition models yielded
> different results. The degrees of freedom should not, I believe, be
> different between them, seeming to rule this out as an explanation.
> The four condition model and the three condition model (both with
> intercept terms) should be equivalent in terms of their ability to test
> all possible comparisons between conditions. The difference between them
> is essentially that in the four condition model, SPM uses a required
> additional linear constraint to allow it to uniquely solve for all four
> condition coefficients, while in the three condition model no such
> constraint is required. Neverthless, in both models, the entire
> 3-dimensional space of contrasts is spanned. That is, any contrast
> in which the four condition weights sum to zero can be estimated. In the
> four condition model this constraint on the contrasts must be explicit,
> while it is implicit in every contrast of the three condition model. I
> look forward to being corrected on any or all of these ideas, as well as
> learning why Russ might have seen a difference between these two models.
>
>
> Sincerely,
> Eric
>
>
>
> Eric Zarahn
> University of Pennsylvania
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|