Thanks for the link. It's really a nice interpretation about the article by Eklund et al. in PNAS, especially for those like me who is not an expert in statistics. Here I have two more questions.
First, before the publication of that PNAS article, I have just finished writing a manuscript, in which we used a cluster-defining threshold (CDT) of 0.005 and 3dClustSim (0.005 seems wildly used in 3dClustSim) to calculate the threshold, which is 130 voxels.
From your link and the PNAS paper, I understand that the p=0.01 and p=0.001 CDTs’ FWE with 3dClustSim are about 27% and 9%, respectively." My experiment is a 2(A and B)x2(X and Y) design. In other words, the regressors of interest contain AX, AY, BX, BY. We
are interested in contrasts "(AX>AY)," "(BX>BY)," and "(AX>AY)>(BX>BY)," particularly the last one. Suppose "(AX>AY)" yields activation in brain regions Q and S, and "(AX>AY)>(BX>BY)" yields S and T. For each of the 3 contrasts, since we use a CDT of 0.005,
its FWE with 3dClustSim should be about, say, 20%. If I perform a conjunction analysis of "(AX>AY)" and "(AX>AY)>(BX>BY)", i.e., to search for the overlap by directly using their group statistical map thresholded at 0.005 with k=130, and this conjunction analysis
reveals overlap in brain region S, can I say that the error rate was more adequately controlled? The rationale is that the error rate for the identification of region S should be about 20%*20%=4%, which is <0.05.
I'm just wondering whether, for people who already use 0.005 with 3dClustSim and are not willing to re-do all analyses, conjunction analysis could help.
The second question is, since the inflation of FWE rate observed in the PNAS paper is from the simulation of resting-state data, can the conclusions (e.g., we need to stick on a stringent CDT, such as 0.001; voxel-wise thresholding is generally better than
cluster-wise approach) be generalized to task-related fMRI data?
Thanks in advance.