Different results from searchlight and ROI analyses

Hi Martin,

I did a whole-brain searchlight analysis, which yielded a significant cluster at voxel-p < 0.005. To show the decoding accuracy of that region, I tried an ROI analysis within a spherical mask based on the peak coordinate from the searchlight analysis. However, the result from the ROI analysis was insignificant (p = 0.08). Is the ROI analysis more conservative than the searchlight analysis in general? What should I do in this case? Thank you so much!


Hi @frodo,

obviously not @Martin but hopefully still somewhat helpful.

I’m not quite sure why you need to run an ROI analyses post-hoc to obtain the decoding accuracy of that region? If you run a searchlight analyses the output should be a searchlight map entailing the accuracy value for each voxel, ie searchlight center, no? With that, you could simply extract the values from that region and average them to obtain the respective average accuracy.

Furthermore, as far as I understand these things using that region in an ROI approach might a) be slightly double-dipping as you use the same data to create the ROI and then for analyses (and you already know that this regions contains information based on the searchlight outcome) and b) not directly relate to the searchlight outcomes. Regarding the latter you’re asking different questions in an ROI vs. searchlight approach. The former evaluates if the entire voxel pattern of a given region contains information regarding the classification task at hand while the searchlight uses a small sphere around a given voxel to evaluate this, travelling through the specified searchlight mask. Thus, the voxel pattern deviates more or less prominently from sphere to sphere and with it the respective outcome metric. Therefore, you could also observe a difference re overall outcome metric and significance when comparing searchlight and ROI results.

HTH, Cheers, Peer

Hi Frodo,

First of all: Thanks a lot, Peer, for chiming in!

I don’t have much to add really. One crucial difference that may explain differences in results is that we ask people to smooth searchlight results after completion. This is something not done with ROIs. Smoothing effectively makes the distribution of accuracies more normal and thus makes it easier to conform to the distributional assumptions of your statistical test.

If you didn’t smooth and did what I think you did, then you can explain the difference in the fact that you ran a cluster analysis which only means that at least one data point in the cluster is truly significant but it doesn’t tell you which one, and it need not coincide with the peak in that cluster.

And finally, perhaps the shape and size of the sphere is different if you are using non-isotropic voxels.

Anyway, hope you’ll figure this out.