Hi Megan,

The timing shift should be negative, -0.831, it seems. Apply this offset to the stimulus times.

See [afni.nimh.nih.gov] .

Keep in mind that 3dTshift is not really shifting the timing of the data, it is sampling the data onto a new timing grid. It is more like 3dresample in the time domain. This analogy is useful because the intention of 3dresample is not to move the volume, it is merely to sample it on another grid. The volume is supposed to be in the same place in space.

This applies to 3dTshift, too. Each time series is resampled onto a different time grid, but the intention is not to "move the signal". The input to 3dTshift has data sample all across the TR. The output is supposed to be as if the entire volume were sampled at just one time point (per TR).

Since 3dDeconvolve interprets the time series as if it were sampled at the start of each TR, but now the time series values are really sampled at the mid-TR, 3dDeconvolve is effectively shifting the data earlier, and so the timing needs a corresponding shift.

Anyway, see the the example in the linked post, that is more clear than theoretical descriptions. But still, it is predicated on the "resample" point of view.

The TR/2 as just an approximation is described there, too.

- rick