"tough love for fMRI"

I lifted the title from a great post on Chris Chambers' blog NeuroChambers: "Tough love for fMRI: questions and possible solutions". He has a very succinct summary of a lot of the problems the field has been struggling with: "MRI = v expensive method + chronically under-powered designs + intense publication pressure + lack of data sharing = huge fraud incentive." And not just an incentive to outright fraud, but also to practices that fall on the spectrum of questionable research practices.

I happened to read Chris Chambers' post the same day I read "Cluster-extent based thresholding in fMRI analyses: Pitfalls and recommendations", a new NeuroImage paper by Woo, Krishnan, and Wager, and I must say the combination is disheartening.

The paper describes mass-univariate analysis interpretation, but there are a lot of parallels between the problems in spatial specificity they point out and those that can occur in searchlight or ROI-based MVPA. In short: it is not appropriate to interpret sub-regions when an analysis produces a large cluster that spans several regions; some of the cluster voxels are significant, but you can't say which ones. You can't say that an area is "activated" simply because it is part of a larger activated area that passed a cluster threshold - it could be that only adjacent regions (other parts of the cluster) are actually activated.

The authors make several practical suggestions, one of which is to individually color-code each cluster that passed the cluster-extent thresholding (so you'd plot a group of blue voxels, another of red voxels, etc), rather than the common practice of plotting all surviving voxels with the same heatmap color scale. They further recommend "that figure legends and captions explicitly state that the true activation location and extent within each significant cluster cannot be determined." This pushes interpretation (properly) to the cluster level, paralleling  ROI-based MVPA, which presents descriptions of the properties of ROIs, not the voxels they contain.

This (and other advice in the paper) is very practical, but I suspect people will resist: it is a lot harder to come up with a compelling story (and get a blob in the "right spot") when clusters are properly interpreted. But this is the sort of "tough love" that is needed for fMRI.


Woo CW, Krishnan A, & Wager TD (2014). Cluster-extent based thresholding in fMRI analyses: Pitfalls and recommendations. NeuroImage PMID: 24412399