> Except for fixation, there was only one on period for each
> task. Given that, can we use a 3dANOVA3 -type 5 to
> analyze the data or must we use a 3dANOVA2 -type 1?

Only one period for each task would not have any effect on group analysis because you would carry over one regression coefficient per task to group analysis level anyway. What it really matters is for deconvolution/regression at individual subject analysis level. If you can manage to run 3dDeconvolve, there should have no problem analyzing at group level with 3-way ANOVA. There is one limitation with 3dANOVA3 though: it currently does not allow 2nd-order contracts. So if you might have to consider using the following Matlab package if you are interested any 2nd-order contrasts:

[afni.nimh.nih.gov]

> Given that state induction is so long, is it possible for
> baseline to be longer or shorter than the other time
> periods in the deconvolve?

I am not so sure what your question is here. Is your "baseline" a real task instead of "fixation"?

> We are also interested in comparing the early and late
> time points of state induction. To do that, is it more
> appropriate to censor out the middle time points in state
> induction or to divide them up into smaller "chunks" even
> though they are not independent tasks?

Censoring would create some discontinuties along time axis which can lead to other problems. You might consider the approach recommended in the following thread by modeling IRF of each block if you have enough time points:

[afni.nimh.nih.gov]

Gang