3dGC usage question

Hi there,

I’d like to check with you my approach to using 3dGC. My data consists of people watching movies and we’ve got annotations for every word spoken (onset, duration and whether they are low or high context words). I’m interested in exploring whether there are any differences in connectivity for the TTG between the two types of words (both seed-to-target and target-to-seed). Already read https://afni.nimh.nih.gov/3dgc and https://afni.nimh.nih.gov/CD-CorrAna, which inspired the following steps:

After 3dDeconvolve to obtain an estimated IRF for two types of words, used 3dmaskave to average the IRF inside the TTG ROI. Then, I used waver to convolve my estimated IRF back with the time onsets of the words:

waver -dt 1 -FILE 1 TTG_high_face.1D -tstim cat movie_high_words_delay4_decon.1D -numout 5470 > high_convolved_ts.1D
waver -dt 1 -FILE 1 TTG_low_face.1D -tstim cat movie_low_words_delay4_decon.1D -numout 5470 > low_convolved_ts.1D

TTG_high_face.1D - the average IRF inside the TTG for high context words
TTG_low_face.1D - the average IRF inside the TTG for low context words

Words are not separated equally in time, as these are normal movies. Our time onset resolution goes to 1 decimal place - e.g. 84.5

Am I correct in saying that the above output from waver should be my seed timeseries that goes into 3dGC? My approach seems to differ from what I’ve read about in the PPI analysis, since I haven’t convolved a TS with any Gamma function and used the raw ts (after pre-processing) to do the first deconvolution. I wanted to check if I’m going down the right route.

Thanks for the help,


I don’t feel comfortable enough to recommend the modeling approach implemented 3dGC even though I wrote the program more than 10 years ago. As for PPI, check out the following video:


Another possibility is to model the effect associated with each occurrence of high/low word, and perform what is called beta series analysis in the literature.

Thanks Gang, this is helpful. Out of interest, why are you advising against using 3dGC? Is it to do with the temporal resolution of fMRI (we’ve got a TR=1s)?

Or is the implementation of 3dGC out of date?


Florin, I don’t have much experience with causality analysis of FMRI data, so I cannot make general recommendation about the methodology. It would be interesting if you could gain some insight out of 3dGC.