Is it recommended to apply a band pass filter to resting state data with about 2/3mm resolution? I’m not sure if the recommended approach is not to use any filter at all, to use only high-pass filter or use both high- and low-pass filter. I do not have physio recordings.
Thank you very much!
I don’t see why the voxel dimension would affect choosing to bandpass filter resting state data (which is temporal bandpassing).
Note that if you do decide to bandpass, then you will lose a lot of degrees of freedom in your data; for a 2s TR, bandpassing to be between 0.01-0.1 Hz will reduce your degrees of freedom in the time series by about 60% (and that is before even considering other steps such as motion regression, baseline drift, etc.).
I’m using 3TProject to regress out fixed components from resting state data and I use a bandpass filter 0.01 0.1 (TR=2):
I get this output:
++ setting up stopband frequency mask
- Block #0: 250 time points – 147 stopband regressors
++ 1 Blocks * 3 polynomials – 3 polort regressors
- – 65 other fixed ort regressors
++ 250 retained time points MINUS 215 regressors ==> 35 D.O.F. left
++ no -mask option ==> processing all 123410 voxels in dataset
++ Compute pseudo-inverse of fixed orts
I have two questions:
- What are the minimum D.O.F. acceptable for resting state data?
- reading the afni_proc.py documentation was very helpful but I’m confused with the recommendation given in Resting state note ~2~.
“Resting state data should be processed with physio recordings (for typical single-echo EPI data). Without such recordings, bandpassing is currently considered as the default.”
but then below that it seems that the recommendation is not to bandpass, because we end up loosing 60% of D.O.F (with TR of 2). I have no physio recordings, so do you still recommend bandpass filtering the resting state data?
Re. #1: “What are the minimum D.O.F. acceptable for resting state data?”
That’s a million dollar question! (Or at least an R21 one…) Given that most people don’t report this valuable/necessary bit of information in their studies, I don’t believe there is consensus in the field (well, is there consensus about anything in the field??). I think many people have ignored the degree of freedom loss with bandpassing, and it is really a problem in interpreting results-- people say “I had 200 time points and censored only 5 due to motion”, but they leave out that they probably removed 60% or more degrees of freedom from bandpassing alone (and since each censored time point is one degree of freedom, it would be like censoring an additional 120 time points!!!). Sooooo, I don’t know, precisely, unfortunately.
Re. #2: typically, people do bandpassing for the purported reason of removing breathing+cardiac frequencies from the data-- these tend to be >0.1 Hz (however, some can get aliased back into the lower range, depending on the TR…). So, the point of that statement in the help file is: if you want to get rid of breathing/heart rate effects, perhaps measuring them more directly would be better, because then you include a couple regressors in your model (i.e., reduce your DFs by a couple), rather than the much blunter bandpassing (reducing DFs by 60% or more, typically). If you are doing surgery, use a scalpel, not a chainsaw (if you can help it!).
Re. #2b: do I recommend bandpassing: welllllllllllllllllllllll, it’s hard to say. There are a fair number of studies that have shown that higher frequency information does contain useful signal (it probably has to, but these have shown there it is a sizeable amount, useful for things like seedbased correlation, ICA, etc.); in particular, see Gohel & Biswal 2015, and other work by them; Chen & Glover have also written a couple useful papers on the topic.
Off the record (no one reads the internet, right?), I would thing that not bandpassing might be a better way to go: you reduce your information content SO much with bandpassing; people say “bandpassing makes my data more consistent/reproducible”, but one should note that multiplying all time points by 0 would also make results quite reproducible… OK, that might be going too far, but hopefully the trade-offs are a bit more apparent.
AFNI: complicating analyses with pesky (but important) technical considerations since 1994!
Thanks Paul for your (always) helpful response! You should turn it into a paper ;), many important points that I believe many of us are unaware of.
“I think many people have ignored the degree of freedom loss with bandpassing, and it is really a problem in interpreting results”
This comes to the crux of the matter. I suppose this is more a statistical rather than afni question, but how important really is to have less D.O.F.? If I am reading your response correctly, it should matter a lot, because you are saying that one D.O.F. less will be equivalent to removing 1 time point from the dataset?
Is this equivalent to the idea that including a regressor of zeros and ones (eg for censored volumes) amounts to deleting that volume from the dataset? Do bandpass regressors also consist of zeros and ones? I’m asking this because the output file from 3dTProject contains the same amount of volumes as the original dataset, so no volume seems to be deleted (or zeroed).
“AFNI: complicating analyses with pesky (but important) technical considerations since 1994!”
Re. DF (degree of freedom) counting: yes, you are correct. Adding a regressor to a general linear model means using up one degree of freedom. Censoring can be done (as you note) by including a column of all zeros except for a 1 at the time point to be censored-- hence, censoring using 1 DF. Interestingly to bandpass a given frequency, you have to include both a sin() and cos() form of it in the model, so that uses up 2 DFs. That is part of the reason why bandpassing is so expensive, in terms of DFs. And indeed, bandpassing won’t zero-out a time point (but you might see some apparent ‘ringing’ effects).
(To your point on a paper-- there is one in progress. It has taken a while to progress, because of so many topics involved, but slowly, slowly…)
I am processing a dataset these days and I was trying to figure out whether to bandpass or not and this discussion was very helpful. I have one main concern:
Considering that I dont have physio recordings, just like samw, what would be a good justification when a reviewer complains about the cardiac and respiratory noise left in the data? It’s more of a strategy question but I am asking because I have been through that once before with the reviewers and I couldn’t convince them using the D.O.F argument? Could you point to some literature may be or perhaps phrase the argument in a way that would be more convincing?