Dear Stephen and colleagues,
those reading the posts regularily (hopefully mainly Steve !!) may
remember issues reported by Pavel Hok on 19th of November 2009 on dtri
domain errors. The Olomouc group did not suceeded in solving those
problems and I came across it right now. My analysis involves one
leasioned patient against five healthy volunteers (quite standard
block paradigm) and I have plenty of those errors as well.
Right now, on a local fMRI meeting, I met nearly all Czech fMRI teams
and we sat around the source codes of FSL 3.3.11 and 4.1.5 and looked
around the ndtri() function calls in both versions. We found some
discrepancies and discovered some "new" calls in newest fsl. And this
might be a source of our bug. I am not a mathematics expert, but based
on info on http://search.cpan.org/~rgarton/Statistics-SDT-0.035/lib/Statistics/SDT.pm
I think the ndtri error is caused by the inverse-phi function called
with zero or one (I do not know whether you are using the code from
http://cpansearch.perl.org/src/RKOBES/Math-Cephes-0.47/libmd/ndtri.c
). The value range is checked in fslmaths.cc, but (I checked only the
greped ndtri lines) probably not in:
misc_c/ptoz.cc
miscmaths/base2z.cc
miscmaths/t2z.cc
miscmaths/miscprob.cc
miscmaths/f2z.cc
In my opinion, some of those calls could be a source of exception
error thrown by ndtri ... Please, Steve, could you have a look on the
code ?
The similar issue was "kind of solved" in
http://www.biac.duke.edu/forums/post.asp?method=TopicQuote&TOPIC_ID=1176&FORUM_ID=7
- but I do not have any NANs in first-level outputs (and neither has
the Olomouc team).
Besides, I discovered a bug (again one ;-) ) - if I run Feat in SGE,
corregistration fails - probably some thread is not well synchronized
and does not wait for the preceeding step.
Thank you for any help
with greetings from Olomouc
Jiří Keller, M.D.
Charles University, Prague
PS.: Please delete the pending copies ....
|