pTFCE with or without FDR/FWE multiple-comparison correction?

[gjoue]

Hi,

This is a naive series of clarification questions, but I'd just like to confirm that I understand a few practical things about pTFCE, seeing as there is also a bit of variation in published articles that have used (p)TFCE with some articles reporting along the lines of, "p < 0.005 pTFCE corrected" while others, "p < 0.05, FWER corrected, pTFCE".

1a). Would it be a reasonable way to think about pTFCE as addressing the problem of over-inflated p's due to multiple comparisons (without applying FDR or FWE) -- and therefore, we can very well say "p < 0.005 pTFCE corrected"? I thought TFCE was actually an alternative to cluster-extent based approach to deal with multiple comparisons -- with AFNI's AlphaSim, we would usually report along the lines of "p<0.001 uncorr. voxelwise, corrected at p<0.05 on cluster size" -- so wouldn't applying FDR and FWER on top of pTFCE be overkill? When would you do pTFCE and FDR or FWE?

1b). Any rules of thumb/things to keep in mind/pitfalls to avoid for choosing the threshold level of pTFCE-enhanced images when not using FWE/FDER as well?

According to the Wiki on ptfce.r, "In general, it is valid to threshold your enhanced Z-score image with the same Z-score threshold you would have applied for the original (unenhanced) Z-score image. E.g. you can threshold the enhanced Z-score image for an uncorrected p<0.001 with the corresponding Z-score value of Z>3.1." -- what I read in this snippet is that not applying FDR/FWE on top of pTFCE is methodologically okay as long as we choose a threshold that is low enough. (In my case, it seems like a threshold of p<0.001 in the "p_uncorr" column in the "peak" section of the SPM results table may still not be low enough -- I have quite a few 1-vx activations in my pTFCE-enhanced image, which does not instill much confidence in me. However, looking at FDR/FWE (if I were to just read it straight off the SPM results table of the pTFCE-enhanced stat map) seems overly conservative.

2). To apply FDR or FWER, is it just a matter of reading off p_FWE-corr and q_FDR-corr in the "peak-level" section of the SPM results table of the pTFCE-enhanced contrasts?

In the 2019 NeuroImage paper, "pTFCE_vox" is decribed as "T-score maps...converted to Z-score maps, fed into the pTFCE algorithm and thresholded based on GRF theory, with a corrected threshold of p < 0.05" -- would this correspond to p<0.05 in the "p_FWE_corr" column in the SPM results table of the pTFCE-enhanced stat image?

On the Wiki for the SPM-toolbox version of pTFCE, it does say that the stats in the "peak-level" section of the SPM results table is correct (and just to ignore the "cluster-level" section). Is this because the FDR/FWE q/p values are based on the original stat map? Because otherwise it would seem to contradict the ptfce.r Wiki which says:
"Never use the enhanced Z-score map as an input for GRF-based multiple correction techniques. Instead, use the original Z-score map to compute the Z-score threshold and then use this threshold on the enhanced image."

3). On the Wiki for ptfce.r, should the following passage:

"Never use voxel-level thresholding on the pTFCE-enhanced image. It already incorporates the cluster information, thus thresholding it based on cluster-extent (of the original image) would be double-dipping."

actually say "Never use cluster-level thresholding on the pTFCE-enhanced image" ?

Thank you very much for your help (and for the toolbox!)!
Gina

[spisakt]

Hi Gina,

Thanks for your questions!

1a)

pTFCE is "enhancing" Z-scores, so that it looks at neighbourhood information and incorporates "our surprise" about the spatial extent of the activation. That is, is DOES NOT DO any kind of correction for multiple comparisons. Therefore I always recommend avoiding the phrase "pTFCE corrected" and say "pTFCE enhanced".
Correction for multiple comparisons are needed with pTFCE, as well. The good thing about the enhanced T/Z scores pTFCE provides that they can be simply thresholded with the FWER-threshold of the original (unenhanced) image, with the same guarantees of family-wise error rate.

Some examples:

• pTFCE enhanced, uncorrected, p<0.001
• pTFCE enhanced, FWER corrected, p<0.05

1b)

Uncorrected thresholding, even with low p-threshold, is generally to be avoided you can't build strong conclusions on that.
It can be, however, useful for visualization purposes if the activations are otherwise very small (see e.g. Fig 2. here), or when it seems appropriate to start some speculative discussion...
Using pTFCE is not different in all these than using the unenhanced images.

2)

While the cited text refers to the R-package, it is largely applicable to the SPM toolbox.
The matlab toolbox also makes sure that GRF-theory is applied on the original (unenhanced) images.
So, yes, peak-level p_FWE-corr values can and should be used.

Regarding the q_FDR-corr values, as reported by SPM: they should work as well, theoretically, but FDR has not been thoroughly validated yet with pTFCE. I recommend weighing this in when interpreting these values.

3)

Many thanks, this was a very confusing mistake in the docs. Fixed!

Best wishes,
Tamas