Hi Matthew, anyone else interested,
I just stumbled across this comment --
"although the RFT maths gives us a Bonferroni-like correction, it is
not the same as a Bonferroni correction. It is easy to show that the
RFT correction is better than a Bonferroni correction, by simulation.
Using the code in the
http://imaging.mrc-cbu.cam.ac.uk/scripts/randomtalk.m, you can repeat
the creation of smoothed random images many times, and show that the
RFT threshold of 4.06 does indeed give you about 5 images in 100 with
a significant Z score peak."
-- from Matthew Brett's Wiki http://tinyurl.com/2etkhz
(http://imaging.mrc-cbu.cam.ac.uk/imaging/PrinciplesRandomFields#head-504893e8afe62f1e3e8aaf3cb368a1d389261ef5)
But I remembered seeing in the code for SPM (v2 or 5), that the
*minimum* critical threshold of RF and Bonferroni is used:
spm_uc.m
And that the minimum p-value is reported:
spm_P.m
Which seems to conflict with the above comment, since if RF is more
accurate, taking Bonferroni where it is less strict will inflate the
type I error rate above the nominal value. No?
Does anyone have any more info on this? E.g. a ref justifying the SPM
minimum approach? I should admit here that I haven't read all the
Worsley papers like I should have done, so may be missing something
obvious... If I am missing something though, perhaps a note added to
the Wiki and/or the SPM code referring to it would help others?
Best,
Ged.
|