Dear Brigitte,

> 1) How can I have significative voxels in my conjunction (in SPM99)
> when some of the contrasts from the conjunction don't give any
> significant voxel??? I guess it has something to do with the lower t
> value in the conjunction but I'm not sure how this is calculated.

If you specify a p value for a conjunction of n SPMs the effective
threshold for each component SPM is p^(1/n) for uncorrected p values
and roughly this for corrected values.  In some instances the threshold
can become negative for each compoent SPM even though the conjunction p
is very small (especially for large n).

> 2) If I have a factor A with several levels, I know that if I want to
> say that all the levels activate the regions X I have to do a
> conjunction analyis.  However, does it have a sense to say that
> globally the factor A activates the regions X when compared to the
> control conditions (one contrast with 1 for all the conditions of my
> factor and -1 for all my controls). I know that this kind of analyse
> could reveal activations that are only due to one or two conditions but
> then again if I look at the plot of "effects of interest" for the
> concerned voxels the effects all go in the same direction, but are
> significative at a corrected level only for some individual contrasts.

It is perfectly OK to average with a single contrast.  This can be
construed as a main effect (the average of several simple main effects
each due to different verisions of A relative to their baseline).  Here
the versions (and respective baslines) would be thought of as levels of
another factor.

> I'm asking that for I have an interesting deactivation.  If I look at
> the effects of interest, it seems to be present in all the conditions
> of a factor and it comes out if I average all these conditions in one
> contrast.  However, if I look at the individual comparisions, it is
> only significant at a corrected level for some conditions (lets call
> them conditions D). I cannot say that this deactivation is specific
> only for these conditions as when I compare the conditions D to the
> other conditions where the deactivation was not significant there is no
> difference. Would a correct conclusion be that: there seem to be a
> deactivation in the region X for the factor A but this deactivation
> only reaches statistical signifiance in the conditions D? Or can I only
> talk of a deactivation in the conditions D?

The simplest way to report the results in an fatcorial design is to (i)
test for an interaction (i.e. any activation due to A that was
significantly different in a subset of pairs D relative to the
remainder).  If there is an interaction report the simple amin
effects.  If not then report the main effects.  In your case there was
no interaction so you should just report the overall main effect and
not worry about some simple main effects reaching a corrected level of
significance and others not.


> 3) How do you set up an interaction if you have a factor with more than
> 2 levels? How do you set up an interaction if you have a factor
> imbedded in another one?

One can test for any form of interaction with the appropriate
F-contrast however in practice poeple usually look at a subset of
conditions that conforms to a 2 x 2 layout and use a T-contrast to test
for the interaction in the usual way (e.g. 1 ... -1 ... -1 ... 1 with 0
for the remaining conditions.  This is repeated for the number of 2x2
perumations your deisgn offers.

> 4) When I set up my covariates spm asks me if I want them to interact
> with the subjects, the conditions or no interaction, what should I do?
> (remember I have condition centered covariates) With condition centered
> covariates, will spm compare the values of my covariate between (and
> not within) the subjects so I could have covariates with only one value
> per subject (and 0 for all the others), is this correct?

In fact SPM will do all the appropriate centering for you if you simply
enter a vector of covariates for all scans and request the appropriate
interaction.  Modelling interactions here simply means splitting the
covariates into subject or condition-specific columns and centering
within subject or condition.  As you require condition-specifc
regressions choose 'conditions'.

> 5) In the last mail, you said that for flipped vs non-flipped analyses
> I had to consider flipped images as coming from different subjects. As
> I work in PET I should thus do a multi-study analyse? Can't I just say
> the flipped images come from the same subject but are from a different
> condition?

No because you want to remove the main effect of hemisphere. Treat each
the flipped image as another subject.  This gives a better model that
accomodates both the subject effect and hemisphere effect as
confounds.  Hemisphere x condition effects can be tested using the
appropriate T-contrast.

With very best wishes - Karl


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%