It seem to be a common practice in signal processing to decimate/downsample signals after low-pass filtering (removal of high frequency data). Is there any reason not to do this (other than one needs to pay attention to the new “effective TR”) for BOLD/fMRI data?
This could save a lot of space in context of storing intermediate results of short TR multi-band data.
When doing that, you lose degrees of freedom, which is bad for you statistics. You actually need to resort to some tricks to compute the remaining dofs after such procedures.
If you’re not using parametric statistics, that may not be important.
Thanks for your comment.
I think the effective reduction of degrees of freedom happens when you do low-pass filtering even if you don’t decimate/downsample the data afterwards. Decimation/downsampling might actually lead to more accurate correspondence between number of time points and degrees of freedom.
You’re right --at least if your low-pass filtering operator is a projector, e.g. it is done by selecting frequencies in Fourier space-- but concretely,you need to be careful to get the actual dofs accurately.
Note that since at least the Woolrich 2001 paper, the consensus for task fMRI is that low-pass filtering (hence the subsequent downsampling) should be avoided, and instead a good AR model of the noise should be used.
I agree that in other settings it can be a sensible thing to do.