Word has come down that the best way to go is likely to first use '3dttest++ -Clustsim' option; and yes, I have capitalized the letter on purpose-- see the help output from 3dttest++:
-clustsim = With this option, after the commanded t-tests are done, then:
(a) the residuals from '-resid' are used with '-randomsign' to
simulate about 10000 null 3D results, and then
(b) 3dClustSim is run with those to generate cluster-threshold
tables, and then
(c) 3drefit is used to pack those tables into the main output
dataset, and then
(d) the temporary files created in this process are deleted.
The goal is to provide a method for cluster-level statistical
inference in the output dataset, to be used with the AFNI GUI
Clusterize controls.
++ If you want to keep the 3dClustSim table .1D files, use this
option in the form '-Clustsim'. If you want to keep ALL the
temporary files, use '-CLUSTSIM'.
++ Since the simulations are done with '-toz' active, the program
also turns on the '-toz' option for your output dataset. This
means that the output statistics will be z-scores, not t-values.
++ '-clustsim' will not work with less than 7 datasets in each
input set -- in particular, it doesn't work with '-singletonA'.
-->>++ '-clustsim' runs step (a) in multiple jobs, for speed. By
default, it tries to auto-detect the number of CPUs on the system
and uses that many separate jobs. If you put a positive integer
immediately following the option, as in '-clustsim 12', it will
instead use that many jobs (e.g., 12). This capability is to
be used when the CPU count is not auto-detected correctly.
-->>++ It is important to use the proper '-mask' option with '-clustsim'.
Otherwise, the statistics of the clustering will be skewed (badly).
---==>>> PLEASE NOTE: This option has been tested for 1- and 2-sample
---==>>> unpaired and paired tests vs. resting state data -- to see if the
---==>>> false alarm rate (FAR) was near the nominal 5% level (it was).
---==>>> The FAR for the covariate effects (as opposed to the main effect)
---==>>> is still somewhat biased away from the 5% level :(
on the set of FA maps (or any other scalar map).
This is following the method described in "The Present- A NonParametric Approach to ClusterSize Thresholding" on page 8 here:
[
biorxiv.org]
The FPR prettiness of the output results is also described there, in Fig. 4.
You will want to run this on a computing cluster, if at all possible.
--pt (under advisement from G Chen)