Cluster Gate Discussion: 9/14/2016:
Meeting help by Lisa Eyler, Tom Liu, and Alan Simmons.
Brief description of the issues and approaches:
Discussion points of ethical considerations in the face of clustering concerns:
Ethical Issues Raised by ClusterGate.docx
The paper that started it all:
The SPM response:
The AFNI response:
The FSL response: ....
The HBM response:
Suggestions for AFNI users:
- Use ACF: gaussian estimates of data were wrong.
- Use a higher voxel threshold: p<.05 doesn't cut it. p<.001 is cleaner but really hard for group analysis.
- Use 3dttest++ -clustsim: it's the only permutations test we have.
- Do a-priori ROI analysis: we don't make new NEO subscales with each new sample.
- Consider using
randomise: it's pretty flexible but you have to
learn how to talk to it.
So what's next:
- R has good options:
lmPerm give permutations that give options.
Network analysis doesn't have these same concerns: it has concerns but not the same ones.
PRONTO: Sometimes similarity between datapoints is good.
Final thoughts, data cleaning and data alignment is still paramount. If you are putting optimal data in you will get the truth of your data. If that truth is underpowered or null then that is the truth and no data threshold will fix that just make sure you don't miss it with noisy data.
Note. thanks to Colm Connolly for feedback.