The Decoding Toolbox (TDT) - 2nd level statistics


#1

Hi everyone,

I’m a new TDT user, and I have a very basic question: what kind of result measure from TDT can I/should I input a SPM 2nd level analysis in order to compute group-level statistics (on whole-brain maps obtained from a searchlight analysis)?

I was initially considering simply passing on the accuracy_minus_chance maps (smoothed), but I’m running a leave-one-run-out CV (with 5 runs), and I have only one example of each class in each run, so the corresponding accuracy map takes quite discrete values, and I’m not sure if it’s compatible with SPM assumptions (even after smoothing the accuracy maps). Alternatively, I guess I could use AUC_minus_chance maps…
What do you think?

(Also, does running a searchlight decoding -even with a small radius- make sense with so few examples in my training and testing sets?)

Thanks in advance for your help!

Emma


#2

Hi Emma,

Sorry for the late reply, please add the tag “tdt” to your posts of you want me to reply faster :slight_smile:

Yes, the accuracy values are indeed quite discrete! But smoothing your results afterwards helps with the assumptions required for SPM.

I would give it a try with a 2nd-level analysis that way! Also, bear in mind that 2nd-level analyses of decoding results will in fact just test a fixed effect. It’s what most people do though, but it’s good to know the assumptions of what you test statistically and the conclusions you can draw. See this paper.

Hope that helps!
Martin