% Signal Change vs. mean or baseline? ; FIR + GLM; NaNs

Hello, 
     I was hoping someone may be able to clarify for me a couple of details
about how MarsBar works that I gathered from previous postings.  Could
someone could either confirm or clarify the following statements?  

1) I think MarsBar calculates % Signal Change as change compared to the mean
of the session, NOT compared to fixation trials (or other "baseline" trials
that are not explicitly entered into the design matrix).  Is this true?  If
so, is there any option to change the way % Signal Change is calculated such
that it reflects change from fixation?

2) I believe MarsBar FIR responses must in some way interact with the SPM
GLM above and beyond simply querying the trial onsets.  I was originally
under the impression that the FIR response calculated was derived solely
from the raw data and the trial onsets.  However, when I compared MarsBar
timecourses from 2 GLMs that differed only with respect to the way the event
duration was specified (i.e. in one GLM, the event duration was set to 0s,
whereas the other was set to 2s), the MarsBar timecourses were different. 
To my understanding, if MarsBar only used the raw data and the trial onsets
to calculate the timecourses, the two timecourses would be identical.  

3) I am less clear on this last point -- how MarsBar handles NaNs.  If
MarsBar calculates timecourses within an ROI from the RAW data (smoothed
analyze SPM images), then all of my ROIs for all subjects have data within
them, because no brain mask was applied to my raw data.  However, the SPM
beta and con outputs were applied with the default brain mask, thus causing
some voxels on the edge of the brain to become NaNs.  Does this same issue
thus pertain to MarsBar?  Are NaNs simply omitted from the ROI or are they
converted to 0?