I don't think it's a good idea to remove linear trends before using 3dDeconvolve. It's not necessary, and besides, 3dDeconvolve does a better job of it. When 3dDeconvolve removes a baseline linear trend, it also allows your stimuli to explain some of the linear trend. Imagine that you have several different tasks in a single run that you're trying to compare, and one of the tasks precedes another. If you remove a linear trend calculated across all the points, you might be removing information about the activation as well. Usually these effects are negligible, but in extreme cases, you could get truly incorrect results using this procedure.