To: spm@jiscmail.ac.uk
Subject: FDR & large fixed effects analyses
Date: Fri, 13 Jul 2001 11:30:50 -0400
From: "Thomas E. Nichols"
Hi,
If you have tried to use FDR with large fixed effects analyses and
had trouble with SPM/matlab hanging, please read this.
It appears that a core Matlab 5 function, betainc, has trouble with large
vectors in certain situations. In particular, to create a p-value from a
T statistic, spm_Tcdf calls betainc; when the degrees of freedom of the T
are very large, and a large vector of T values has been submitted betainc,
the function fails to converge. Since there is no max-iteration check,
the function basically falls into an infinite loop.
However, Matlab 6 does not have this problem, seemingly due to
a single line change to betainc, or, precisely, to a function that
betainc calls, betacore.m. Line 25 of betacore.m in ML5 reads
while any(abs(y(:)-yold(:)) > 4*eps*abs(y(:)))
while line 25 of betacore.m in ML6 reads
while any(abs(y(:)-yold(:)) > 10*eps*abs(y(:)))
Apparently, the convergence criterion of 4*eps is too strict, and
just increasing to 10*eps fixes the problem.
This problem has never been exposed before because SPM only determines
p-values for a few voxels at a time, and doesn't usually convert a
whole image of T's into p's.
In conclusion:
If you are running Matlab 5 and have a problem with FDR hanging with
large df analyses, you probably need to modify betacore.m to match
it's ML6 version.
While Matlab is hung, hit control-C and see if the problem is in the
function betacore. If it is, Matlab will tell you the exact path to
betacore.m; copy this file to some common matlab directory, and make
the above change (change 4 to 10). You then should be all set.
-Tom
PS: Thanks for Matthew Brett for finding this problem.
-- Thomas Nichols -------------------- Department of Biostatistics
http://www.sph.umich.edu/~nichols University of Michigan
nichols@umich.edu 1420 Washington Heights
-------------------------------------- Ann Arbor, MI 48109-2029