Hi David, One of the stages in feat_model uses an up-sampled version of the input data, and the EVs, before convolution is applied - this is potentially the reason behind the differences that you are seeing. Cheers, Paul On 9 May 2015 at 00:21, David Parker <[log in to unmask]> wrote: > Hello Everyone, > > I'm trying to understand how FSL creates its regresors from a 3 column > stimulus file by recreating them myself. > > To do this, I'm comparing my own regressors to those created in FSL from a > 3 column stimulus file, colvolved with the Double-Gamma HRF, at zero phase > and without temporal filtering. > > I've read elsewhere in the forums that the double gamma HRF used by fsl is > the same one used by SPM, so I'm using that at 20 Hz as my convolution > kernel. I'm creating a 20Hz box car array from the 3 column stimulus file, > and colvolving it with this HRF. After correcting for some temporal > offset, when I compare this to the regressor created in FSL, I'm still > seeing subtle differences. (FSL's regressor in blue, mine is in green) > > These differences may seem small, but it's important that I understand > exactly what FSL is doing. > > > I'm wondering if anyone can shed some light onto what FSL is actually > doing to create these regressors, as in, is there any filtering going on > that we're not aware of, or is there some subtle difference between their > HRF kernel and the one I'm using? Is the fsl HRF actually saved anywhere > that I could access directly? (I did try to look at this by creating a > regressor with a single impulse, and the resulting regressors were the same > to a very small margin of error, much less than above). So, any help in > this would be appreciated. Thanks! >