Testing Loss and Gain of Statistical Power
Surface quality can be tested by tracking change in statistical power between groupstats results of the ADNI60 subjects. The test diff_power, currently located in utils/test/diff_power, locates surfaces regions with the largest change in power. Given a source and reference groupstats test (i.e. thickness of controls vs ADs in dev and stable6), diff_power does the following:
Subtracts the stable6 sig.mgh from the dev sig.mgh (producing sigdiff.mgh) and masks any vertices that did not gain or lose significance (default threshold: p=0.05)
- Computes clusters from the masked sigdiff.mgh and locates the vertex of peak change (in either direction) for each cluster
For the largest nc clusters, finds the difference in thickness (between the dev and stable6) for each individual subject at the peak vertex
For the top ns subjects with the greatest thickness difference, maps the peak vertex from fsaverage to the dev and stable6 white and pial surfaces and computes the distance between the two
The results are stored in a summary.log file (in a diffpower output dir underneath the contrast directory), and a script called inspect.sh is created to automatically point the user to the offending surface regions
Here is an example of the cluster information saved in:
================================================================================ fsaverage cluster: 1 avg change in power: -1.7761 peak change in power: -4.6965 peak vertex ID: 6560 peak diff avg diff peak peak white pial subj thickness thickness src-vtx ref-vtx vtx-dist vtx-dist -------------------------------------------------------------------------------- 0657 -2.4819 -2.5324 94065 101722 8.9340 9.5627 0934 -0.4215 -0.1048 94278 93622 0.8505 0.9646 1251 0.3980 0.1895 102093 102358 0.5470 0.6754
The average diff thickness column represents the average difference within the cluster.
To view each surface region detailed in the report, run the inspect.sh saved to the diffpower output directory, which will walk through each vertex of interest in freeview.
note: inspect.sh just links to the most recent recons stored in the freesurfer subject tree; diff power doesn't save any subject data. So visually inspecting the results of diff_power on anything but the most recent groupstats test will be misleading.
running diff power
diff_power will be added to the ADNI60 testing cronjob that runs weekly on launchpad, but if it needs to be run manually, here is the usage:
required: --src <dir> source groupstats directory --ref <dir> reference groupstats directory --contrast <name> contrast name as listed in group.fsgd (e.g. control-ad) --test <var> test variable (e.g. thickness) optional: --thresh <float> -log(P) significance threshold (default: 1.301) --smooth <int> specify smoothing (default: 00)