Dear Istvan

> dear SPM group,
>
> here a naive question about a group analysis of fMRI data over which
> we - some people in our fMRI groups - strain our brains for a few days
> now.  the main question is how to generate contrasts on the first
> level (as INDEPENDENT measures vs as ONE mean-contrast per subject),
> and how to process those constrasts on the second level for the
> following admittedly simple scenario:
>
>        2 groups, g1 & g2, each with 13 subjects.
>        2 sessions A & B per subject (sometimes just one session).
>        A & B are not identical as order of stimuli is different
>        stimuli are regular tones & mismatch tones (rare)
>        focus of interest on effect of mismatch tones

I don't understand why modelling A & B in different matrices as it
corresponds to 2 repeated sessions (not independent).
For repeated measures,  your design seems OK.

>         fixed effect analysis:
>          together for both sessions A & B, treating sessions A & B as
> ONE contrast:
>        - first regressor for effect of tones
>        - second regressor for effect of mismatch tones (parametric)
>        - creating one subject specific constrast file for a) effect
>          of tones and for b) effect of mismatch for both sessions A &
>          B together.
>          in contrast manager:  [1 0 1 0 0]  and  [0 1 0 1 0].
>
>        random effect analysis RFX:
>        - for simple groups effects: pooling 13 subject specific
>           contrast files of group g1 or g2 for 1-sample t-test
>        - for group comparison: pooling 26 (g1 & g2) subject specific
>           contrast files for 2-sample t-test

You can also perform a two-sample t-test on differences, e.g. 1st level
[-1 1 -1 1] and 2sd level gp1 vs gp2

> analysis 2 (INDEPENDENT contrasts):
>        fixed effect analysis:
>          separately for sessions A & B, treating sessions as
>          INDEPENDENT measures:
>        - creation of two session specific contrast files for a)
>           effect of tones and for b) effect of mismatch, for A or B
>           independently.
>          in contrast manager:  A:  [1 0 0]  and  [0 1 0]
>                                B:  [1 0 0]  and  [0 1 0]
>
>        random effect analysis RFX:
>        - for simple groups effects: pooling 26 contrast files for
>           sessions A & B for group g1 or g2.
>        - for group comparison: pooling 52 contrast files for sessions
>           A & B for both groups g1 & g2.

In this case, at the 1st level (if A1sess B1sess A2sess B2sess)
contrasts should be [1 0 0 0], [0 1 0 0], [0 0 1 0] & [0 0 0 1]. You can
assess the effect of the session. However, pooling the data of the 1st
and 3st contrast should give you the same result as in the 1st analysis
(approximately). However, if you don't want to look at the effect of the
session, I think the first model is better as the error should be
smaller in this model than in the second model.

Best,
Cyril