However, in practice. SPM does contain an orthogonalization step. So if the design matrix is orthogonalized, this effects will be diminished right? Are the confound-regressors not orthogonalized?
I could be wrong, but I believe the orthogonalization is only within a condition with respect to derivatives and parametric modulators.
When does orthogonalization become a problem? I thought I read that the only problem about orthogonalization can occur when the regressors are not linearly independent anymore. But then the regressors might still be correlated... Can somebody comment on this?
As regressors become more correlated, it is harder to attribute the variance to one regressor or the other. In these cases, the error for that regressor increases making the significance lower. The effect of this at the second level is more variability between subjects.
Correlated regressors are only a problem when there is complete correlation. In these cases, the contrasts cannot be estimate for each regressor independently (e.g. paired t-test contrast for first condition is not estimable). As long as there is some unique variance in each condition, you can estimate it; but as I said earlier, the error for that regressor increases.
Hope this helps.
Thank you!
Debby
