Apologies for long email - forgot to reply to the list. Here is are the
emails in case anyone else finds them useful.
Joel
-----------------------------------------------
"I think the contrasts [1 1 1 1 etc] and [-1 -1 -1 etc] show me this."
Yes. Those essentially look at perfusion as a function of hormone
level, averaged over all subjects, so your interpretation is correct.
"A 'useful' effects of interest might show where the activation
correlates either exclusively overall positively OR overall negatively."
That sounds more like a conjunction: finding a place where all the
subjects activate positively. It's like a logical "AND".
"However, might the 'effects of interest' explored in the attached image
show a correlation driven by something more random, like subjects 1-4
correlating positvely and subjects 5-10 correlating negatively?"
Yes.
It's an F-test looking at the null hypothesis that _all_ the
coefficients in the first 10 columns are zero. Meaning, if there's a
large enough deviation of one or more of those coefficients from zero,
then the test shows significance ("activation"). But as you say, some
of the subjects could be positive, some could be negative; in the
"effects of interest" F-test, they don't "cancel each other out" and
instead would add up towards significance. Kind of like squaring each
and then adding together (metaphor only).
I think the effects of interest contrast that SPM creates for all or
almost all designs is supposed to be like the ANOVAs in statistical
textbooks. In those texts, you're supposed to conduct an overall F-test
first, to see if anything interesting is happening. If the test fails
(i.e., nothing is significant), you have to stop. If the ANOVA is
significant, then you can go on to "post-hoc" t-tests. These are the
same as the t-tests usually run in SPM contrasts. (By "usually" I mean
that most people doing SPM or related neuroimaging analysis packages are
usually concerned with t-tests/t-contrasts, not F.) In addition, one
thing that the classical stats books do, but which neuroimagers don't
appear to do, is to do a multiple comparison correction. Not over
voxels, but over the number of comparisons. (Names associated with
tests that attempt to deal with that multiple comparison problem are
e.g. Tukey and Scheffe.) So, that's the reason for the existence of the
"effects of interest" F-contrast. (Which many people ignore, as far as
I can tell.)
Best,
Stephen J. Fromm, PhD
Contractor, NIMH/MAP
(301) 451-9265
-----------------------------------------------
Stephen,
Thanks for your answers. I have attached some small screen shots of the
design & contrasts, in case you're able to add some further comment. The
first 10 columns are the 'scores' for each scan for each of the 10
subjects, and the next 10 are dummy(?) variables representing each
subject.
I know the GLM doesn't know what question I want to an answer to so I
guess the most pertinent questions would be:
* where does the activation correlate positively overall with subject
score?
* where does the activation correlate negatively overall with subject
score?
I think the contrasts [1 1 1 1 etc] and [-1 -1 -1 etc] show me this.
A 'useful' effects of interest might show where the activation
correlates either exclusively overall positively OR overall negatively.
However, might the "effects of interest" explored in the attached image
show a correlation driven by something more random, like subjects 1-4
correlating positvely and subjects 5-10 correlating negatively?
Its possible I'm making some wrong assumptions or the model is wrong so
your comments would be helpful. I would also like to be clear whether
'effects of interest' in this example design would always be useful...
Regards
Joel
-----------------------------------------------
I don't have that particular design type in front of me, but:
(1) As an F-test, it looks at both tails of the distribution. So
something is significant if it is really "large", in either the positive
or negative directions. (In contrast to the t-tests (t-contrasts), which
in SPM are only one-tailed on the positive side.)
(2) What the particular F-test (F-contrast) looks at depends on your
model. Usually SPM sets up the "effects of interest" F-contrast so that
it looks addresses the null hypothesis "none of the effects of interest
are different from zero." You can see what these effects are by
inspecting the matrix.
Stephen J. Fromm
-----------------------------------------------
> *Subject:* Effects of interest & Correlates
> *From:* Joel Dunn <[log in to unmask]>
> *Reply-To:* Joel Dunn <[log in to unmask]>
> *Date:* Mon, 30 Jun 2008 12:37:25 +0100
> *Content-Type:* text/plain
>
>
> Hi,
> I would like some clarification on what "effects of interest" means when
> looking at a single factor correlating with activation within subjects.
>
> Using SPM2 I have a design with one measured factor (e.g. hormone level)
> correlated with PET perfusion images within subjects (10 subjects, 12 scans
> per subject).
>
> When looking at the SPM2 default F-test results, some regions appear
> significant that do not appear when looking at the linear contrasts for
> positive correlations ([1 1 1 1 1 ...]) or negative correlations ([-1 -1 -1
> ...]). Even if you drop the significance thresholds, the linear contrasts
> are stuggling to find anything in those areas.
>
> Is this "effect of interest" F-test looking for absolute significant
> correlations that are either ALL positive or ALL negative, or is it, in this
> case, (a less useful?) test that might find significance where, for example
> half of the subjects positively correlate & half negatively correlate?
>
> Or does the "effect of interest" in multisubject/scan correlation mean
> something else entirely?
>
> Help appreciated.
> Cheers
> Joel
|