How to threshold maps?

[adelavega]

How do we recommend to users how to threshold and cluster correct their results?
Goal is to make it easy for users to do an end-to-end analysis but theres no clear option

Some options to recommend:

• Randomize
• EasyThresh
• NiLearn TFCE
• AriBrain

[adelavega]

Idea: Add a documentation page that explains all the options, or wrap some of these tools to add to FitLins

[poldrack]

That seems like a good menu of options, I agree that an explainer on the pros/cons could be useful cheers rp

On Aug 7, 2025, at 10:37 AM, Alejandro de la Vega ***@***.***> wrote: adelavega created an issue (bids-standard/stats-models#96) <#96> How do we recommend to users how to threshold and cluster correct their results? Goal is to make it easy for users to do an end-to-end analysis but theres no clear option Some options to recommend: Randomize EasyThresh NiLearn TFCE AriBrain — Reply to this email directly, view it on GitHub <#96>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAGUVECQBC6MXUEBVFJ4KXT3MOFFLAVCNFSM6AAAAACDLVI5X2VHI2DSMVQWIX3LMV43ASLTON2WKOZTGMYDCMZWGAYDONA>. You are receiving this because you are subscribed to this thread.

— Russell A. Poldrack Albert Ray Lang Professor and Department Chair, Department of Psychology Professor (by courtesy), Department of Psychiatry and Behavioral Sciences Associate Director, Stanford Data Science Director, SDS Center for Open and Reproducible Science Building 420 Stanford University Stanford, CA 94305 ***@***.*** ***@***.***> http://www.poldracklab.org/

[jmumford]

This is a multipurpose notebook notebook to give us material that can be repurposed for stat map thresholding suggestions and so Poldrack lab members can use it to choose their own strategy with whatever pipeline they are using. I was thinking a version of the table at the top might be useful? It would need some editing, but the information is there. If you want me to add this to the documentation, just let me know how I can access that.

It isn’t intended as a tutorial for each method, since that can become overly complex, especially for permutation tests. That said, two methods that work directly with z-stat maps and integrate well with the BIDS Stats Models are easythresh and voxelwise fdr correction using nilearn. Both have simpler implementations that can be described concisely.

It is a bit odd that the two TFCE methods are producing different results. TFCE has several nuanced parameter settings that the original developers tuned specifically for fMRI data, so differences there could explain the mismatch. In any case, I would be more comfortable recommending TFCE after further testing, optimization, and broader adoption.

I have started to look at ARIbrain and I think the R version is the most useful, but that would need more explanation on how to use it and I figured we could leave it out for now as a recommended approach. We can always add it in later.

[yarikoptic]

and what about "gofigure" per @mrneont et al? https://afni.nimh.nih.gov/pub/dist/doc/htmldoc/codex/fmri/2025_TaylorEtal.html

[jmumford]

I'll check it out (gofigure), thanks!

[mrneont]

Thanks for mentioning that, @yarikoptic . Happy to chat more about transparent thresholding at any point.

And I could also suggest a couple other group analysis/clustering options if that is helpful.