Dear FSL experts,
I have some questions regarding the eigenvalue plot from melodic after doing a group ica. Some time ago I already posted a thread about the rapid drop in eigenvalues, which was answered to indicate one massive component suggesting that there might be something badly wrong in my data.
However, recently I have done a similar analysis as before and repeated the concat group ICA several times, once decomposing the data into 20, 30 and 40 components, respectively. I noticed, that the eigenvalue plot looked nice in the 30 components solution (relatively asymptotic curve) and again showed a rapid drop (after about the first 2-3 components) in the 20 and 40 components solution. For each run the identical data set was used (just simply altered the “-d” option in the melodic command line). So, how can this happen? I would have expected to find the same eigenvalue plots for each run independent of the number of components, given that the same data set was used.
Further, I thought that this plot (at least the line named “ordered Eigenvalues”) resembles a “scree-plot” and had the idea to use this “scree-plot” as an aid to determine the optimal number of components. However, the y-axis of this graphic displays the “% of explained variance” if I got it right. Therefore, my question is, where to find the concrete eigenvalues (someone suggested to look for the “melodic_pcaD” , but this file contains just 1 value..so I think this is not the correct file). For now, I have the impression that neither of my decompositions (20, 30 or 40 components) give “clean” networks. Most of the time, some (not all) seem to be merged with artifacts or other networks. It seems that the criterion whether to select 20, 30, 40, or x components is rather arbitrary, so the “scree-plot” looks to me like a possible alternative to select the number of components.
Kind regards and many thanks in advance for your answer,
Stefan Lenzen
|