Clustering in AFNI - Updates Feb 2017

During the past year, the topic of clustering and false positives rates in FMRI group analysis has been widely discussed both in research circles and in the popular press. A number of very strong claims were made in the initial paper by Eklund et al (2016, PNAS), both about previous methods in general and about AFNI in specific.

These are all important issues for our group. We have addressed them, as well as new developments within AFNI, in two recently accepted articles in PNAS and Brain Connectivity. Links to several PDF items can be found here:

They are, in order of brevity to prolixity:

  1. an “Executive Summary” of the main points in item 4
  2. the Letter accepted in PNAS (a direct reply to Eklund et al)
  3. slides outlining the situation and new developments (presented yesterday in the AFNI bootcamp)
  4. the paper accepted in Brain Connectivity (quite lengthy)

Further changes to AFNI are in the works, which are hinted at in items 3 & 4 (the ETAC method). In the near future, changes will be made to AFNI to discourage the use of the “classic” (old) clustering method, which does not control the False Positive Rate well in many cases.

Links to the article pre-prints on the arXiv:

Summary of the BC paper [1 page] –

Brain Connectivity [long] –

PNAS [short] –

And the slides –

And the final versions of these articles are now available online from the respective journals:

The (long) Brain Connectivity one:
(and re. this one, please do check out the “executive summary” version, linked above, as an overview of the many topics discussed)

and the shorter PNAS Letter:


Greetings fellow imagers–

The Eklund group showed with resting data that false positive rates are not well contained when traditional assumptions are made about the spatial smoothness of BOLD fMRI data when block-based analyses are conducted–and that this is less so for event-related analyses. This leads us to ask about the implications for resting fMRI connectivity analyses. It’s challenging to apply the logic of Eklund et al. in this case because their approach hinges on finding spurious block / event-related effects in resting data (NB: there is room here for speculation about natural fluctuations in thinking/emoting that could account for what Eklund considers spurious). Specifically, we can’t apply the Eklund approach to determining false-positive rates resulting from resting fMRI connectivity analysis given that there are no naturally generated ‘connectivity free’ data in which spuriously high levels of connectivity could be detected. Perhaps there have been attempts with simulated data? This is not to say that we should continue to make classic assumptions in the case of resting fMRI functional connectivity analyses. But given that false positive rates vary as a function of block versus event-related approaches, it’s not clear, either, that the false-positive rates they report are due entirely to heavy tails in the spatial distribution of BOLD data that should, then, be assumed across all manner of fMRI analyses. Moving forward, we will make this assumption, however, until we have reason to do otherwise. Thoughts on this topic or, perhaps, alternative courses of action to recommend?

Many thanks!


You ask difficult questions, Paul. I’m planning on addressing this to some extent in the future, although my enthusiasm for clustering is about equal to 1/(first Skewe’s number).

Hi Bob,

Thank you for your discussion and the Brain Connectivity paper from your team. I think your paper puts fMRI clustering analysis in a well balanced perspective.

Thanks for maintaining AFNI and continuing to improve it.