Conjunction analysis (old topic I know...)

AFNI Message Board
1

Hi all (and especially Gang ;-),

I know this has been discussed before, and there were also some notes about it on Gang’s page (which do not seem to exist anymore), but I wonder what is the current preferred method in afni land for detecting regions that are significantly activated by two contrasts — (cond.A - cond.B) and (cond.C - cond.D) — and whose levels of activation are not significantly different from one another — i.e. (cond.A - cond.B) is not significantly different from (cond.C - cond.D). I remember a while ago the standard procedure was to create a simple intersection map of two already thresholded and clustered maps, but that does not answer exactly the same question. Would it make sense to further multiply this intersection map by the brain-wise complement mask of the thresholded and clustered map for the contrast [(cond.A - cond.B) - (cond.C - cond.D)] (that is, the mask of non-significant voxels for the latter contrast)?

Also, what would be a sensible choice for the threshold of the maps to be intersected? Is there any principled ground for it to be less stringent than when assessing a single contrast?

thanks for any comment
giuseppe

2

Greetings Giuseppe!

there were also some notes about it on Gang's page (which do not seem to exist anymore)

Yes, there were an old webpage about conjunction analysis and an old me too, but the page has been turned. Currently I tend to think that the conventional conjunction analysis is full of issues:

  1. inefficient modeling at the whole-brain voxel level,
  2. artificial dichotomization,
  3. spatial ambiguity,
  4. vulnerability to misinterpretations.

Would it make sense to further multiply this intersection map by the brain-wise complement mask
of the thresholded and clustered map for the contrast [(cond.A - cond.B) - (cond.C - cond.D)] (that is,
the mask of non-significant voxels for the latter contrast)?

This is very problematic (the issue 4) above): the absence of strong evidence for an effect cannot be interpreted as the evidence for the absence of the effect.

To avoid these issues, it would be more appropriate to directly define those regions of interest and go from there. These thoughts are discussed here: Thoughts about modeling and result reporting

If you prefer the traditional approach, here is the old page with a lot of yellow stains:
https://sscc.nimh.nih.gov/sscc/gangc/ConjAna.html

3

Hello Gang,

thanks for the swift reply and yes, I can see the issues you pointed out with the conventional approach. I had actually printed out your paper a while ago, but hadn’t been able to read it yet: this gives me the right priority to do it.

hope to see you at the next HBM

giuseppe