Dear Maria,
As you have a mixed-design ANOVA I would suggest to construct as many contrasts as possible on single-subject level and forward these into two-sample t-tests (in your case) to ensure correct error terms and statistical validity. The procedure is maybe confusing (two different models for an ANOVA), but it reflects partioned error terms, which is also what we usually rely on for behavioral analyses with programs like SPSS. See Henson & Penny (2005, http://www.fil.ion.ucl.ac.uk/~wpenny/publications/rik_anova.pdf ) or Henson (2015, http://www.mrc-cbu.cam.ac.uk/wp-content/uploads/2015/03/Henson_EN_15_ANOVA.pdf ) on that issue.
Assuming your single-subject models include two sessions (pre and post), then two contrasts would be required, one for the average task effect (e.g. [1/2 ... 1/2 ...] with ... indicating zero-padding, averaging across pre condition 1 and post condition 1), and one for the differenctial task effect (e.g. [1/2 ... -1/2 ...]).
You would then set up two different two-sample t-tests,
#1 for the average task effect based on e.g. the con_0001 images: within that model set up F contrast [1 1] for average task activations, set up [1 -1] for main effect group.
#2 for the differential task effect based on e.g. the con_0002 images: within that model set up F contrast [1 1] for main effect time, set up [1 -1] for interaction group x time.
If you prefer T contrasts you would have to use the same vectors but also set up contrasts "into the other direction".
Alternatively, you could also work with GLM flex for group models, this package allows to set up ANOVAs with partioned error terms. Results should be identical to those obtained with the procedure described above.
Best
Helmut
|