Just as subjects are samples from a hypothetical population pool, so trials in an FMRI experiment are instantiations from a condition category. If cross-subject variability should be properly accounted for, why would cross-trial variability not be on an equal footing? In addition to improving generalizability, capturing trial-level effects may also gain estimation accuracy and statistical sensitivity. The analytical tool at the whole-brain voxel level is already available: 3dLMEr.
The subject line asks a question, but I don’t see the answer (“yes” or “no”) in the posting!
–anonymous
Hi anonymous, if the question whets your appetite, the answer is only a click away.
You mean, read a paper? Yikes. That would be, like, a looooot of tweets…
Is this paper available on TikTok?
–anonymous
Hi Gang,
I’m very interested in modeling trial as a random factor. Do I understand correctly from the bioRxiv paper that the implementation would be to use 3dREMLfit with one regressor for each trial, and then use 3dLMEr with each trial as an input file for one massive model? As in (following your example):
3dLMEr -prefix LME
…
Subj Emotion Trial InputFile
s1 pos T1 s1_decon+tlrc’[1]’
s1 pos T2 s1_decon+tlrc’[2]’
s1 pos T3 s1_decon+tlrc’[3]’
…
s1 neu T1 s1_decon+tlrc’[x]’
s1 neu T2 s1_decon+tlrc’[y]’
…
s2 pos T1 s2_decon+tlrc’[1]’ …
Thank you,
nmuncy
nmuncy,
Do I understand correctly from the bioRxiv paper that the implementation would be to use 3dREMLfit with
one regressor for each trial, and then use 3dLMEr with each trial as an input file for one massive model?
Yes, that’s correct. The output from 3dDeconvolve would be largely fine too. With your example, the model can be something like
3dLMEr -prefix LME
-model ‘Emotion+(1|Trial)+(1|Subj)’
…