Dear James,
This is an old message, but I was also wondering why AFNI, FSL and SPM so often result in very different outcomes for pretty standard analyses. I think that the biggest difference between those packages is the way they model temporal autocorrelation. In AFNI each voxel time series is separately modeled with an ARMA(1,1) process, while in FSL the nonparametric autocorrelation parameters are spatially smoothed. I have compared these three different techniques (AFNI/FSL/SPM) for different TRs, including resting state data with assumed dummy experimental designs, and task data. The differences in results were particularly strong for short TRs, but also for normal TRs, like 2s (you had data with TR=2s), there were surprisingly big differences. A preprint is on arXiv:
https://arxiv.org/pdf/1711.09877.pdf
The general conclusion was that more accurate autocorrelation estimation substantially improves the sensitivity-specificity trade-off (less false positives or less false negatives). Given those autocorrelation modeling differences, I am not surprised with your AFNI/FSL discrepancy.
Best,
Wiktor Olszowy
|