> 1. When is it appropriate to use TENT/GAM/BLOCK basis functions
> (I'm primarily looking at the visual cortex in this study; 1s on, 15s fixation,
> with a response in the first second of those 15s)? I ask because I'm sure
> I've seen all 3 used in similar situations. I'm leaning toward using a TENT
> function here - probably TENT(0,16,9).

It all depends on your research goal. Since you have stimuli each of which lasts for 1s, your choice is between TENT and GAM. If the BOLD response shape of particular interest to you, TENT might be appealing. Or you can try out both, and see they pan out at both individual and group level.

> 2. Concerning detrending, is it appropriate to model the baseline (in
> deconvolution, via -stim_base) as just the motion correction parameters,
> or should I be looking to fit in the 15s fixation period somewhere as well
> (and if so, how)?

Use the option -polort to model the slow drift, and it will automatically account for the trend as part of the baseline. Read the help about -polort for more information.

> 3. Considering detrending again, how does one remove just the linear
> trend (and not the entire baseline), as Gang suggested to me here: ?
>
> Just to be clear, I want the data detrended so that it is as "clean" as
> possible for a different analysis step (ROC, which will be run in Matlab).

If you really want to remove the linear trend even if there are higher orders of trend, I guess you could do that with option -select in 3dSynthesize (digits under -select?). Alternatively you can use 3dcalc to directly subtract the linear trend based on the results from 3dDeconvolve/3dREMLfit.

Gang