Dear Lisette,
Yes, the tutorial is popular still, but for a valid inference you should have the correct error terms and the correct degrees of freedom, which would not be the case for main effect group in your model (which wasn't considered that much back then).
As long as you have two levels for the within-subject factors (no matter how many factors you have), it's very simple to come up with valid statistics, as you can set up the necessary contrasts on single-subject level (averaging across the factor levels, differential contrasts). Your first single-subject T contrast would thus be something like [0.5 0.5], your second single-subject T contrast would be something like [1 -1].
Then go with two two-sample t-tests on group level. The first model should be based on the con_0001 images from above, within that model,
- F contrast [0.5 0.5] reflects average activations (averaged across groups and conditions)
- F contrast [1 -1] reflects main effect group
The second two-sample t-test would be based on the con_0002 images, within that model,
- F contrast [0.5 0.5] reflects main effect condition
- F contrast [1 -1] reflects the interaction group x condition
Instead of the F contrasts you could also go with T contrasts. As the T contrasts in SPM are one-sided you have to define additional contrasts with inverted signs to test for the "other" direction. For a "proper" two-sided T contrast you would have to go with two separate one-sided T contrasts thresholded at an appropriate voxel threshold (e.g. both thresholded at .001/2 instead of 2x .001), although hardly anyone does so.
With more within-subject factors with two levels you would just go with more T contrasts on single-subject level (reflecting the differences between level 1 and level 2), which would then be forwarded into another two-sample t-test on group level. E.g. if it were a 2x2x2 ANOVA with two within-subject factors, you would need three two-sample t-tests, if it were a 2x2x2x2 ANOVA with three within-subject factors, you would need four two-sample t-tests and so on.
This can be easily adapted to between-subject factors with more than two levels - in that case, you wouldn't go with a series of two-sample t-tests on group level but a series of One-way ANOVAs ("between-subject") - and to designs with more than one between-subject factor - in that case, you would go with a series of Full factorials.
The only less straightforward mixed designs are those with within-subject factors with more than two levels, but that's another issue ;-)
Best
Helmut
|