Hi,
> A while ago there were several statements that orthogonalization of main EVs wrt the motion regressors and EVs was unnecessary, since results obtained with and without orthogonalization would be the same. Forum 31412 for example. Are these biophysical or mathematical statements, or information about how the design matrix is constructed? Could you comment on this?
This is a mathematical statement, as the GLM attributes any "shared" signal due to correlations between EVs to the ones that are not part of the contrast of interest. Alternatively, the contrast of interest is only driven by signals over and above what can be matched to the remaining EVs (those EVs that do not feature in the contrast, including the confound EVs). So it does exactly the same as if you had orthogonalised the EVs featuring in the contrast wrt the others. Hence there is no need to do the orthogonalisation as the GLM does it automatically. This is a feature of the GLM, not of how the model is set up with the matrix, it is about how the statistical estimation is done within the GLM and how it treats shared or correlated signals/EVs.
> I've recently noticed that the confound regressors in my resulting design.mat (4.14; yes, old and in need of update) are not the same timecourses as those in the confound regressors file that I specified, even when filtering and HRF are turned off. The main model EVs in the design.mat do seem to be the same as the input. Its not due to z-transformation or de-meaning, I take care of that beforehand. Could you elaborate on what happens to the regressors in the design.mat? I have looked at the design matrix rules online, and am not sure if the confound regressors are treated the same as the motion regressors that may also be added.
Confound regressors are treated exactly as motion regressors. However, the highpass filter that is used in FSL is a local Gaussian-weighted linear fit, and so even with a very high cutoff period it will still remove all linear trends from the data. So it depends how you tried to "turn off" the highpass filter. If you tried to set a high cutoff period then it will still remove this linear trend, regardless of how high a value you chose. If you turned it off in the pre-stats tab then the difference between the confound EVs before and after should only be a constant. Either way though it is fine as no reasonable design should contain such a trend. If you are seeing a difference which is not a constant or a linear trend then let us know.
> Finally, in browsing source code in search of answers I note the terms "real EVs" "original EVs" and "motion parameters" but I don't see reference to "confound EVs". Could you tell me what the different groups include?
The confound EVs are treated the same as the motion parameters in the code too. In feat_model (where most of the processing is done) they both enter via the confound matrix text file option, but internally are called motion parameters, as that is the first thing the code supported and the principle is not different, so the variables have never been renamed. If you specify both a confound matrix and motion parameters, then they are simply concatenated together to form one big confound matrix before being passed into feat_model.
All the best,
Mark
|