search

poldracklab / fitlins

Fit Linear Models to BIDS Datasets
https://fitlins.readthedocs.io
Apache License 2.0
74 stars 31 forks source link

Nistats first level model scaling will include non-steady state images in the mean #267

Open Shotgunosine opened 3 years ago

Shotgunosine commented 3 years ago

It's probably too late for me to be posting this, but it might be an issue. Nistats first level model does voxel-wise scaling over time by default. (The default value is 0, but they catch 0 and treat it differently from False, so the default value is voxel-wise scaling over time). If the input scans to fitlins haven't had the non-steady state scans truncated and this is instead dealt with by censor columns (which I think is what's output by fmriprep if I remember correctly), then the non-steady state scans are still present when the scaling is calculated, which means that the interpretation of the coefficients will be percent signal change from the time series including the non-steady state scans. In theory, you'd be able to increase the values of the coefficients just be having a longer time series, since there'd be more steady state scans to dilute the impact of the non-steady state frame, lowering the mean, and thus increasing the scaled values. T-stats and such should be fine though. What do you think @adelavega?

Shotgunosine commented 3 years ago

On further reading, I suppose this is a duplicate of https://github.com/nilearn/nilearn/issues/2403

adelavega commented 3 years ago

Hey, I think this is a fair concern, although empirically i'd be curious how much it matters in practice. A lot of datasets that we're using in OpenNeuro already had a few scans chopped off prior to releasing the dataset, so the number of non steady state scans tends to be low (although not always zero, or at least according to fmriprep).

Otherwise, I think the other thread covers what can be done in nistats well...

Anyone have a sense how big of a deal this is?