Searchlight Cross-Decoding Yields "Whole Brain on Fire"

[Flo Martinez Addiego]

Hi everyone! 

I hope you're doing well. 

I have been trying to run cross-decoding for one of my multivariate analyses for a couple of months, and I am starting to spin myself in circles, so I'm hoping you might be able to help me.

I am running a whole-brain searchlight cross-decoding, and I have verified that my targets and my chunks are correct, I have checked that I am supposedly decoding what I think I am decoding, but I am getting a completely unrealistic result. Each individual subject has a decoding accuracy > 95% across the entire brain. Even when I do a group analysis (random maps, montecarlo simulation, max size cluster statistic with initial threshold of 0.001), this result holds. 

I would be very grateful if you had any suggestions or ideas on how to proceed.

Thanks in advance,

Flo

[Nick Oosterhof]

Greetings,

This does indeed sound unrealistic. 

Maybe you can approach it from the question: what can you do to make this effect weaker or go away? 

For example:

- do you get plausible results if you reduce power? run the analysis with fewer participants? fewer chunks in each participant?

- do you get similar results with a smaller searchlight radius?

- what about data in an ROI that is not supposed to contain information, for example from a ventricle or outside the brain? 

Also I wonder:

- are any potential confounds that may explain cross-decoding, for example movement of participants? 

- when making an uncorrected group map using cosmo_stat (no TFCE / monte carlo), do you see peaks of strongest statistical effect where you would expect them?

best,

Nick