Scaling and nuisance regression

Hi all,

Forgive me for a potentially naive question. I am preprocessing multiple runs that should be scaled so they are comparable across subjects and runs. Per convention, percent signal change will do. I’ve run all of the steps, slice time correction, deobliquing, and motion correction. I think input this time series into 3dDeconvolve to regress out motion parameters, as well as CSF and WM signal.

I then submit the -errts output to 3dmerge for smoothing/blurring.

The resulting time series seems reasonable, but is not scaled to a mean of 100 and thus PSC. When I scale to 100, the resulting data is very messy and seems as though something has went wrong. (see attached)

Is it potentially the order in which I’ve scaled and regressed out nuisances?


Hi Andrew,

The errts dataset has a mean of zero since there are
constant terms in the linear regression. It would not be
appropriate to use this dataset in scaling (at least not
for computation of the mean).

I take it you are not using for the analysis.
Even if you choose not to do so, consider running it just
for comparison, and see how steps like scaling are

For example, consider: AFNI_data6/FT_analysis/

  • rick

Thank you Rick!

So to clarify, scaling should be done before nuisance regression. I should not then assume that if NO SCALING is done that the errts output is comparable across subjects.