Manually assigning chunks

Hi experts,

My experiment only consists of 2 longer runs, but instead of correlation, I would like to run svm and conduct leave-one-run-out CV. Would it be appropriate to divide each run into 3 chunks (giving me 6 total chunks per condition)?


Hi Erin,

The idea is good. There are two issues I can see with this approach. First, the runs are not really independent because of the temporal autocorrelation. If you split up the runs in smaller chunks it should in principle be ok (this is more of a problem with trial-wise analyses in a fast event-related design).

The second problem, though, is that there is an effect of run which may bias your analyses. There are two possible ways to go ahead. Either you do a split-half (e.g. split each run in 3 and do a leave-3-our-CV), or you always leave out two chunks, one from each run. Since there is the additional effect of time within each run, I would only leave out the first chunk in both runs, then the second chunk in both, etc. Effectively, this is a stratified leave-2-our-CV.

Hope this helps!

Oh, and another alternative would be to run crossnobis distance. This is an encoding analysis and should give you rather high power. You could do this in a leave-one-run-out approach. There is a template for this type of analysis in our toolbox.


Thank you so much for the detailed explanation. I will try them out instead!
And again, thanks for developing a such a wonderful toolbox!