[Neuroimaging] temporal filtering and confounds when loading with nilearn niftimasker

[bthirion]

On 12/12/2016 23:34, Gael Varoquaux wrote:
> That's a tricky problem, because the question is more: what do people
> expect is done?
>
> The way nilearn currently does it is to first remove the confounds, and
> then filter the resulting signal. I cannot really see a drawback of doing
> it that way. To take the case that you are describing below, the
> low-frequency of the CSF would be removed from the final signal.
>
> Now, IMHO, the right way to do things would be to express the frequency
> filter in a cosine basis and concatenate the confounds. AFAIK this is how
> SPM does it. We'd like to do it this way, and have an issue to do it:
> https://github.com/nilearn/nilearn/issues/1011
> However we haven't found time so far.
>
> If you write the equations (they are a bit horrible), the way we do
> things, the way you propose to do things, and the way I think that they
> should be done all vary slightly. I cannot put an intuition on what the
> differences are, though. Of course if the frequency filter and the
> confounds are orthogonal (the confounds have no energy in the filtered
> frequencies bands), they are equivalent.
>
I think that we should do either (simultaneous regression) or (band-pass 
then regression on band-passed confounds).
B