We are performing site maintenance. Some features may not not be available at this time.

Dear expert,

I have a question about the jackknife sensitivity analysis: what would be the criterion for deciding that an activation cluster has been 'replicated'? If you leave one study out it will likely affect the size of the cluster and the location of the peak voxel.

I have used the criterion: "at least one voxel of the cluster after leaving one study out lies in the original cluster of the analysis that includes all studies'. Is there a better (more rational) criterion?

Thanks and best,

Jurriaan

Dear Jurrian,

I feel that this is an interesting question that should be further investigated, but I think that your criterion seems good.

With the best wishes,

Joaquim

Dear all,

I use the same criterion for a Jackknife-analysis as the one stated by Jurriaan. This worked fine and previous results seemed reasonable (varying between 50-100% for different clusters).

However, I just ran a Jackknife-analysis for a meta-analysis of approx 60 studies (n1~1500; n2~1200). General thresholding procedure was p<.005, SDM-Z = 1 and k>=20. Surprisingly, JK results singled out one (!) particular cluster of the main analysis (SDM-Z = -1.6, k =49). This cluster was not significant in any of the JK combinations (0 of 60). We checked several possibilities to understand this 'odd behavior', e.g., validating our script again, looking at the individual statistical maps of the JK analysis, and re-calculation of the main analysis. After checking these options, we can say for sure that this cluster is indeed not significant in any of the individual JK results (even when excluding studies with very small sample sizes [n=6 per group]). Still this cluster is significant when including all studies.

It seems unlikely that the JK analysis and the main analysis diverge that much. Thus, I would be glad if anyone has a suggestion what is going on here. And importantly, this 'odd behavior' shows only for this one cluster. The rest of the results is fine.

Thank you very much in advance!!!

Dear Marco,

To explore the behavior, you could look if the difference between the main analysis and the jackknife analyses is in the sdm-z of the (peak) voxel, in the cluster extent, and / or in the sdm-z of the statistical threshold, and how large are these differences. There is a possibility that the sdm-z values are really similar but in the main analysis they reach the statistical threshold just for a very small amount, while in the jackknife analyses they fail to reach the threshold just for a very small amount. I agree this behavior seems odd, but as the number of meta-analyses increases, the possibility of a few unlikely outcomes also increases...

Best wishes,

Joaquim