Dear William,
this is a very interesting question! I'm no statistican, but hopefully your thread attracts attention if there are some replies :-)
As far as I know 1) the regressors are not orthogonalised during multiple regression (in contrast to the parametric modulators for single-subject models) and 2) the estimates are based on the orthogonal components of the regressors / reflect the unique contributions. The common contribution is unmodelled (?). This might be problematic if your two regressors of interest correlate with the same region (?).
I'm not sure whether this is true though. At least it would make sense that you get larger / higher activations with a single regressor, as it would now explain the "common contribution" as well.
To avoid multicollinearity, you might also drop some of the regressors (like age A and education B, the high correlation might result from scanning children and adding grades as a regressor, which is highly correlated with age in general). For the remaining ones, you could try and orthogonalise them yourself outside of SPM and then forward the results to SPM.
I'm interested to hear some more opinions!
Best,
Helmut
|