> the model has columns for each factor and the interaction of group and session
To make sure: While you *can* go with three main effects group, session, subject plus the interaction group x session during "Main effects & Interactions" you *do not have to*. You can also just go with main effect subject plus the interaction group x session (the same holds if it were two within-subject factors and/or if you have more than two levels). They lead to the same results.
The first design matrix corresponds to an overparameterised model, the second design matrix corresponds to a cell means model. This is because adding the interaction to the model does not result in a single column (which would be sufficient to model the two-way interaction) but rather, four columns, each of them dummy-coding one of the "cells" (G1Sess1, G1Sess2, G2Sess1, G2Sess2). These columns
G1Sess1 G1Sess2 G2Sess1 G2Sess2
1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1
can be rotated though, resulting in a factor effects model:
Constant "Main effect group" "Main effect session" Interaction
1 1 1 1
1 1 -1 -1
1 -1 1 -1
1 -1 -1 1
One could come up with other design matrices as well, but that's another issue. In any case, this means that when adding the interaction in the GUI one does *not* just add the interaction term to the model, but rather, a constant term, main effect A, main effect B, plus the interaction term.
SPM seems to favour the cell means approach over the factor effects approach as the former is more straightforward to combine with the usual single-subject models, in which separate condition regressors often already correspond to certain cells - in that case you can just take the beta images and don't have to set up contrasts on single-subject level at all. Note however that one could also come up with different single-subject models, instead of coding two factor levels A1, A2 with separate regressors one could also go with a regressor reflecting (A1 + A2)/2 (= average) and A1 - A2 (= difference).
Best
Helmut
|