Multiple Comparison Correction Method for Spatially Dependent Correlation Landscapes

Hello everyone,

I have a two-parameter model (let’s call them a and b). For each point (a,b) and for each participant, I am calculating a similarity measure between modeled and empirical metric. Then, for each point (a,b), I compute Spearman’s Rho value and the corresponding p-value (from an approximate permutation test) for the correlation between the similarity measure value and some behavioral variable. This results in a landscape of correlation values across the grid defined by a and b.

The question is how to approach multiple comparison correction in this setting, considering the following factors:

  • There is spatial codependency between similar points (mainly along diagonals, where close points produce similar simulated data; this generally creates three subplanes). The Rho values are almost identical in these regions, leading to similar p-values.
  • This is an exploratory study, so I don’t require a very strict Type I error correction, but some level of correction is still necessary.

I was considering some kind of random-field correction, but perhaps there is a better and simpler approach. There are 110 (a,b) points in total, with n=45 pairs. The Rho values are approximately +0.4 in two separate regions and -0.4 in one region.

I believe that it is kind of popular problem in neuroimaging so I would really appreciate any suggestions.

Thank you.

I think that you should rely on non-parametric / permutation testing of the statistics, which would automatically account for the level of correlation. You cannot be only partially rigorous…
Best,
Bertrand

Dear Bertrand,

Thank you for answer. I ended up with using BH FDR.