Dear Joaquim,
I'm wondering whether, in a meta-analysis that only uses T-maps, using T-maps with partial brain coverage may bias the results (because non-covered voxels are attributed an effect size of 0).
If yes, I'm wondering whether these potential biases only affect the voxels that are typically not covered by all studies (e.g. in the periphery of the brain) or whether they also affect voxels that are covered by all studies.
And in general, would you say that it's "safer" to only include whole-brain coverage studies?
Thank you very much for your help.
Best,
Guillaume

Dear Guillaume,

the brain coverage should be the same in all studies. Otherwise, as you say, the effect size in partially covered regions may be wrongly understood as null in some studies. This might bias the effect-size downwards in these regions and potentially bias the p-values downwards in the remaining regions.

Of course the problem may be considered negligible if the brain coverage is not identical but very similar across studies. However, if substantial differences exist in brain coverage, you should either discard the studies with different brain coverage, or restrict the meta-analysis to a mask that matches the brain covered by all studies.

With the best wishes,

Joaquim

Dear Joaquim,
thanks a lot for your response, that makes sense to me.
However, I'm still wondering about the following. If we consider that missing voxels have a normal distribution of T-values centered around 0, my understanding is that assigning them a value of 0 would not necessarily bias the p-values downwards in the remaining regions. Let's imagine the following situation, in which I'm interested in significantly positive effects in region A covered by all studies, while some voxels in region B are not covered by all studies. If missing voxels in region B would have mostly had negative T-values had they been measured, then my understanding is that including these voxels would have actually improved the significance of the positive effect in region A. And by assigning them an effect size of 0, I'm now biasing p-values upwards in region A. In contrast, if voxels in region B would have mostly had positive T-values, then I would indeed bias p-values downwards in region A.
Thus I'm still wondering whether assigning null effect sizes to missing voxels necessarily runs the risk of biasing p-values downwards.
I hope I'm explaining this clearly...
Thanks again for your insights!
Guillaume

Dear Guillaume,

I think I see your point, but please note that the statistically significance is separately assessed for positive and for negative effects. When assessing positive effects in region A, it doesn't matter if effects in region B are null or are negative. Thus, when assessing positive effects in region B, there would be no problem if the effects in region B in those studies that do not report region B were negative instead of null. Conversely, there would be a problem if the latter effects were positive.

Let's imagine a very simplified scenario in which we only have two studies that find the same effects (which can only be -1, 0 or +1), and we say that effects in a region are significant if they are higher than in the other region. Both studies find and report +1 in study A.
- Situation 1: Both studies find +1 in region B. Here one study would not report +1 and thus the meta-analytic effects in region B would be +0.5, for what the effects in region A would be falsely higher than in region B, and thus falsely significant.
- Situation 2: Both studies find 0 in region B. No problem.
- Situation 2: Both studies find -1 in region B. Again, one study would not report -1 and thus the meta-analytic effects in region B would be -0.5, but this would not modify the statistical significance of region A.

Hope this helps,

Joaquim