I’m new using SPM to conduct second-level analysis for searchlight, and I hope someone could help me out on this question.
In SPM12, the typical second-level analysis includes individual’s con_000* files which include contrast information from the 1st-level analysis. So I’m wondering how to specify the contrast (or is it necessary to specify) in the group-level searchlight analysis? There’s a lack of information about the pipeline to set up this analysis on SPM, so I hope anyone who has experience with SPM could provide me some guidance. Thanks!
it’s really easy, just take the images and put them in the SPM second level analysis. In SPM, you use con-images, because they reflect the combination of several runs or specify a contrast. But SPM just calculates a T-test on them (or ANOVA). So if you want to know in your decoding analysis if results are significant and you want to use a T-test, then you just take your decoding results to that level. Our default output is accuracy_minus_chance, i.e. you can just pass the images as they are.
One catch: often people run analyses on unnormalized or unsmoothed results (I recommend that myself). If your analyses were done on unnormalized results, then you want to normalize the accuracy maps using the normalization parameters from SPM and then smooth your results slightly. I would recommend smoothing, because it improves the assumptions of the T-test and of Gaussian random field theory, can increase your SNR, and can reduce the effects of inaccuracies in the normalization procedure. (I would even smooth data again if you had used smoothed data). Smoothing kernel FWHM for me is always the rule of thumb about 2x the voxel size.
Thank you very much, this is extremely helpful! I actually used normalized and smoothed data for the searchlight, as there’s a small sample size in my dataset, and I think smoothed data might increase the SNR for the interested conditions…thank you for suggesting smoothing the result again before feeding them into the 2nd-level analysis.
Just a follow-up question on this, does any T contrast need to be specified (i.e., 1 for accuracy and -1 for chance) after the model estimation? Thank you!