I believe '-rlt' detrends for each individual run before concatenation occurs.
We completely agree with what you said above, and were inclined to believe this prior to looking at our data:
"If that was the case, a second detrending during deconvolution would basically fit the baseline with a horizontal line."
..but in looking at our data we see that it must be fitting the baseline to an "almost" horizontal line because our data are different if we used '-rlt'ed and non'-rlt'ed data. There is not a large difference, but an obvious one none-the-less.
The question is whether, in your opinion, the data that is "detrended" twice, via both 3dTcat's '-rlt' and 3dDeconvolves '-polort', produces a model that is less representative of the true BOLD response (even if it is just a small degree of misrepresentation)?
Again thank you so much for your time!
Jeremy