Dear Joaquim,

I am calculating a neuroimaging meta-analysis using SDM and have a question about the statistical thresholding procedures in SDM. Based on your validation work, it is recommended to use an uncorrected p-value of .005 along with an additional threshold of Z>1 (Radua et al., 2012).

I stumbled upon an online discussion that this procedure might severely inflate the number of false positives. What is your take on this and do you intend to implement FWE or FDR-based correction methods in the near future?

Thank you very much in advance,
Greg

Dear Greg,

this is an interesting question.

There is an issue with p-values derived from current peak-based meta-analytic methods. In these methods, the randomization tests swap blocks of voxels, for what:

a) the null hypothesis would be that the effect size of a voxel is the same that the effect sizes of the remaining voxels. This hypothesis differs from the hypothesis that we would like to test, i.e., that the effect size of the voxel is zero. Note that if all voxels had a similarly high and statistically significant activation, the p-values in the meta-analyses would be not significant, given that no voxel would have a significantly higher activation than the remaining voxels.

b) even if the tests swap blocks of voxels instead of individual voxels, the spatial structure is lost in the randomizations, which probably modifies the resulting p-values.

Thus, I think that these p-values (either uncorrected or corrected) should be only understood as indicators of what the "true" p-values might be.

To overcome to some degree this issue, I recommend to use the above set of thresholds because it approximates the results of correctly thresholded mega-analyses of the same data.

With the best regards,

Joaquim