AFNI and Clustering: False Positive Rates Redux


There has been a lot of discussion around a recent paper by Eklund et al. (2016) investigating false positive rates (FPRs) in the fMRI field. There were several important points raised there, in particular strong evidence that the standardly-used assumption of fMRI noise having a spatial Gaussian ACF is quite flawed. In response to this useful work, Bob has been working on some new clustering methods for AFNI-- he presented a start of this work at OHBM, and there are further descriptions and results of simulation shown here:

Additionally, there is discussion of the effects of the now-famous bug in 3dClustSim that was also reported in Eklund et al. (2016). A quantitative comparison is presented of the earlier “buggy” version of 3dClustSim and the fixed version (that has been in AFNI since May, 2015) using the same simulations. Importantly, the effects of the bug on FPRs was quite small-- see the linked draft for more details.

And there are even other interesting points discussed, including implications of all the above results on experimental design and parameter choices during analysis. Enjoy!


great work guys, awesome!

Awesome! Thanks for passing along! Looking forward to the new changes!