Reply-To: | | [log in to unmask][log in to unmask]> wrote:
> List and Vladimir, > > I am analyzing scalp-frequency EEG data. I have a question about > appropriate "cluster forming" or voxel-wise significance thresholds in SPM > EEG. This is an oft-discussed topic on the listserv and I've collected some > opinions about it which I will review before asking my question. In > particular, Vladimir wrote: > > ----------------------- > Vladimir: You can either use peak-level correction and then FWE = 0.05 is > a common choice or you choose cluster-level correction, in which case you > would start with an uncorrected threshold of usually 0.05 or 0.01 and see > what clusters are significant. You can then plot just the significant > clusters by using the size of the smallest significant cluster as the > extent threshold. > > (https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=spm;395b1b78.1308) > ----------------------- > > It was similarly discussed in another thread: > > ----------------------- > Howard: I was under the impression that the standard/ default setting of > this was 0.001, at least this certainly seems to be the case in the fMRI > domain. > > Vladimir: I don't think there is any convention that the cluster-forming > threshold should be p= 0.001 and even if there is, there is no mathematical > reason behind it. Different cluster-forming thresholds make your analysis > sensitive to different kinds of effects, that is why cluster-forming > threshold should not be set based on the data. A higher threshold (in terms > of p-value) would mean larger clusters with weaker effects inside, and a > lower threshold vice-versa. As I said I think p<0.05 or p<0.01 are quite > reasonable choices for sensor-level EEG or TF data because we often have > the intuition that large chunks of significant voxels are a true > physiological effect and cluster-level correction just quantifies that > intutition saying how large is large enough. > > (https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=spm;accb28ee.1308) > ----------------------- > > So the distillation of the recommendation as I've found it so far is: > 1.) p-uncorrected < .05 or .01 should be fine. > 2.) don't set the the cluster forming threshold based on the data. > > However, getting to my particular issue, using p<.05 or .01 has often led > to the formation of superclusters in my data. By supercluster, I mean > visually distinct „môo |