Axel wrote,
>Looks like we have to use the threshold plugin to create a
>mask. We tried to select a mask created with 3dAutomask, but Permutation
>doesn't seem to recognize it, even though it's also a binary intensity fim.
This failure to inter-operate with other methods of mask generation is an
artefact of my rather solipsistic approach to engineering scientific software:
I often prefer to do it myself so that I know exactly how it works. This is a
fine strategy when what one wants is a piece of software for one's own work,
but starts to suck when one decides actually to release the code.
>Threshold does not seem to allow us any adjustments (such as –dilate in
>3dAutomask)
Yeah, it worked for my purposes without dilation or erosion, so I never added
such options.
>So I wonder whether Permutation can be cajoled
>into accepting 3dAutomask outputs?
Undoubtedly so, though you happen to have caught me on holiday so I may be a
bit slow getting to it. The change is likely very simple.
Just so that I know exactly what you're working with, it might help if I could
have access to an example of the 3dAutomask-generated mask data set that you're
trying to use. That way I'll be able to verify that everything works. If you
could put a copy of it up for me to retrieve by http or ftp, that would be
useful.
Note that, as Doug Ward very correctly admonished in this forum this past
spring, permutation testing without accounting for coloured noise can produce
artificially low tail probabilities, and thus a bias towards false positive
results. Based on the autocorrelations that I and others have seen at B0=1.5T
and TR=3s, I'm comfortable applying uncorrected permutation testing to data
collected with these parameters. I'm now starting to acquire data at 3T and
1.1s, though, and I certainly would not be comfortable applying uncorrected
permutation testing at this high field strength and low TR.
RSN I'll be implementing Ed Bullmore's (HBM 2001; 12(2):61-78) wavelet
transformation as a method of accounting for coloured noise. I've been meaning
to get to this for over a year now, but the date has kept slipping as other
work has come up. (Programming for which I'm actually getting paid tends to
take priority.) I will *have* to get to it sometime this fall, though, since
I'll be needing it for my own data.
In addition to the wavelet method, I plan to implement another version of the
permutation test which is more suited to event-related designs, and which
sidesteps the problem of coloured noise by permuting the time series of
experimental events rather than the time series of BOLD signal measurements
(Ardekani & al., Cognitive Brain Research 2002, 14(3):347-356; Raz & al.,
NeuroImage 2003, 19(2):226-232). Again, this has been on the back burner while
I work on science and on other, paid programming projects. But I will need it
for my own use sooner or later.
>The second question concerns group statistics based on permutation tests in
>individual subjects.
Well, keep in mind that what you're getting out of the AFNI permutation test is
a three-dimensional matrix of z-scores. (Call it an "SPM{z}" if you like --
especially if you're trying to get something published.) Actually what the
permutation test computes are tail probabilities, but mapping these to z-scores
makes them easier to store within AFNI, and also easier to manipulate: in the
z-score domain you can do things like averaging within a ROI and it'll work,
since a sum of normally distributed variables is normally distributed. Anyway,
the point is that once you have this within-subject matrix of z-scores, you can
manipulate it in the same way that you might manipulate a parametrically
derived statistical map -- spatial transformations, mixed-effects analyses, &c.
You're quite right in observing that it would be wasteful to base a group test
simply on those within-subject results that reach a particular threshold
probability, as such a strategy risks discarding useful information from
subthreshold activations. My own approach has been to average the z-scores
within each individually mapped region of interest and then to use the
resulting regional z-scores as the per-subject variables within a mixed-effects
analysis -- e.g. throw all the regional z-scores into an analysis of variance.
Individual mapping is, of course, time-consuming, and more automated methods
(e.g. transforming to Talairach space) can be substituted, albeit with some
loss of statistical power.
--
Matthew Belmonte <belmonte@mit.edu>
Autism Research Centre
Section of Developmental Psychiatry
University of Cambridge
Douglas House
18b Trumpington Road
Cambridge CB2 2AH, UK
