I’m using TDT to perform mvpa using 5 ROIs.
Brief background: 4 conditions(A1B1, A1B2, A2B1, A2B2), correlation_classification because there are only 2 runs, cross-validation [1 -1 1 -1], output is ‘AUC_minus_chance’ w/ unbalanced_data = ‘ok’.
To my understanding, I have 2 ways to perform roi analysis. First, I could submit 5 masks, and set cfg.analysis = ‘searchlight’; and second I could also set cfg.analysis = ‘ROI’;, where the second (ROI) would give me separate image files per ROI mask.
My question is if I set cfg.analysis as ‘searchlight’ with 3mm radius would this take care of the multiple comparison problem?
In addition, if I set it as ‘ROI’ instead, how should I go about correcting for the problem?
On a side note, if I set CV [1 -1 1 -1], am I correct to think that I am essentially only looking at the the B1 vs B2, regardless of condition A? Is there a way to perform a 4-way analysis?
I am not sure what “cross-validation [1 -1 1 -1]” means. Do you mean you have four conditions and these your labels?
My suggestion is to use the output ‘signed_decision_values’. This gives you the difference in correlation between and within class and should be more sensitive, because it’s continuous.
The former would run a searchlight analysis across all 5 ROIs, the latter would run an ROI analysis. If you want to run a searchlight analysis in an ROI (which I don’t recommend because of edge-related issues), then you need to pass each ROI separately or make sure that the radius is so small that none of the searchlights overlap between ROIs. Say, you use V1 and V2. Then there would be overlap and I would be cautious with ROI-specific interpretations. Say, you use V1 and FFA. Then everything should be fine, but you still have edge-related issues, that searchlights at the edge of the ROIs are smaller.
Could you expand on this question? I’m not sure why 3mm or what the searchlight radius has to do with the multiple comparisons problem.
Again, not quite sure what you mean, but if you are referring to the labels, then it could still be using the interaction of condition A and condition B to do the classification. But I think this is nothing people usually worry about. If you want to do a multiclass classification analysis, then set the labels to [1 2 3 4]. This may not work with the correlation classifier, though.
If you want to study interaction effects between A and B, then you can’t use decoding and need to use an encoding approach, e.g. CV-MANOVA.
Thank you for getting back to me so quickly Martin.
Sorry for the confusion. I think I had some of the key concepts confused, but I think I understand it now.
Thanks again for clearing things up for me and the great toolbox!