Print

Print


Dear Stanford,

> After searching the archive I still have 3 questions about SPM's default
> masking threshold.  I apologize if these are naive questions, but I hope
> someone can answer them.
>
> 1)  Why is the default masking threshold value set at 0.8 (i.e
> defaults.mask.thresh = 0.8)?  Is there a significance to the value 0.8?

As you probably know, the idea behind the masking is to only estimate
a model in voxels that seem to have decent signal in them.  Because
signal in the brain is higher than signal outside of the brain (except
in the case of artifacts...), SPM uses a proportional threshold to
identify these voxels.  So defaults.mask.thresh = 0.8 will only
estimate the model in voxels whose mean value is at least 80% of the
global signal.

I don't know where the 0.8 value came from, but I assume it was an
empirical observation that it seemed to work well.  In my experience,
anyway, this certainly seems to do a fairly decent job of correctly
identifying voxels with decent signal.


> 2)  Is there a  down-side to lowering this threshold (other than including
> voxels with low/noisy signal and voxels that are not part of the brain)?

Not really—that's exactly the downside. :)  (Which should probably not
be underestimated!)


> 3)  Is there a minimum threshold value that would still yield principled
> results?

I don't think so—this will depend on your data.  As you lower the
threshold, more and more low-signal (including non-brain) voxels will
be included.  I think these are interpreted on a continuum.  If you
have a voxel in the brain with high signal, this is fairly easily
interpreted (and I think this is basically what the standard threshold
gives you).  You could imagine lowering the threshold to something
like 0.1.  In this case, let's say you get a voxel outside the
brain—this is easy to interpret in that it is unlikely to be
informative about neural activity, but then the usefulness of
including it is not clear.  You might also have a voxel technically
inside the brain, but with very low signal (say, in orbital frontal
cortex, or some other area prone to signal dropout).  In this case, I
would also wonder about the utility of including such a voxel.  That
is, even if you get a "significant" result, I don't know how
meaningful this is if the signal is very low.  In a group study I
would expect these results to be inconsistent and thus probably not
carry through, but it's still a bit odd.  I suppose what I'm saying is
that if it were my study, I would probably have to do a fair bit to
convince myself that significantly lowering the threshold would really
be providing useful information.

Hope this helps!

Best regards,
Jonathan

-- 
Dr. Jonathan Peelle
Department of Neurology
University of Pennsylvania
3 West Gates
3400 Spruce Street
Philadelphia, PA 19104
USA
http://jonathanpeelle.net/