Hi everyone,
I’m working on using a searchlight decoder to predict a continuous experimental variable. I came across the paper by Stelzer, J., Chen, Y., & Turner, R. (2013), which recommends using a permutation + bootstrap approach for multiple comparison correction. However, when dealing with regression problems, shuffling the labels (dependent variable) violates the exchangeability assumption of permutation tests.
How should I approach permutation testing in the context of regression? One idea I have is to use a symmetrical score function, like Pearson correlation, and then apply a sign-flipping permutation test. Does this seem like a valid approach?
Thanks a lot for your insights!