> So regarding the first question: the number of trials for one condition is indeed (in fact
> exactly) three times more than the other condition (9 vs. 27).

Ideally it would be nice to account for the different variances across conditions, but that's currently not feasible.

> I've heard FSL can handle this correctly - is there a way for AFNI to handle this?

The modeling capability in FLAME 1/2 of FSL is exactly the same as in 3dMEMA: the only legitimate approach in incorporating the reliability information for the effect estimate is to model one condition (or the contrast between conditions or the linear combination of multiple conditions).

> I'm thinking that ASL-scan 1 minus ASL-scan 2 could be a between subjects variable.

Are you saying that ASL is a within-subject variable because it varies between the two scans? And it does not vary across conditions, right?

Gang