Question about RSA searchlight target/chunk division

[Flo Martinez Addiego]

Dear all,

I hope you're doing well. 

I have a question about the target/chunk assignment for RSA searchlight.

I have two nearly-identical scripts intended to do whole-brain searchlight RSA and give me very similar whole-brain maps (t-test of Fisher-transformed RSA values against zero), but also give me very different histograms of those values across the brain. I've tried to check everything carefully, but I'm not able to figure out where the differences may be coming in, so I'm hoping someone may be able to give me more insight.

Setup - 

Four runs per subject

- First two: hand

- Last two: foot

- Each has draw and tool conditions separately

I average within body part x condition to get four samples per subject (draw/tool x hand/foot)

Code 1 -

I assign the target by unique action (draw  = 1, tool = 2), chunks are 1 for hand runs, 2 for foot runs. I calculate the average per target/chunk

 cosmo_disp(ds)
.samples:     4x123700
.sa.chunks:  [1;2;1;2 ]                                                            

.targets: [1;1;2;2];

Code 2 -

I give target a unique value, even across runs. draw hand = 1, tool hand = 1, draw foot = 2, tool foot = 2. Chunks here = 1. I calculate the average within target

cosmo_disp(dsSub)
.samples:   4x123700 (same as code 1, just reordered)
.sa.chunks: [1;1;1;1 ]

.targets = [1;2;3;4]

Additional notes -                                                            

+ The values inside ds.samples are the same, just reordered. 

+ The target dsm is created after this, so it's not predetermined, and follows the same logic

+ In both cases, I use cosmo_target_dsm_corr_measure with 'center_data' = true and no regress_dsm

+ I apply atanh to the resulting correlation values.

The problem - 

Despite using the same data and logic, the histogram of RSA values is:

- bimodal and saturated in Code 1, but 

- smooth and unimodal in Code 2. 

The final maps are almost identical in spatial distribution, so it's only the distribution shape of the RSA values that are different.

I've read through the documentation and can't figure out why the sample order or chunk/target structure (when aligned) would give me different results. I'm wondering if there's a subtle interaction with sample ordering, distance calculation, or vectorization I may be missing.

Any ideas as to why this may be happening?

Thank you in advance for your help, I appreciate it.

Best wishes,

Flo

[Nick Oosterhof]

Greetings,

With N classes (unique target values), there are N*(N-1)/2 pairs between them. If N=2 (in your “code 1”) there is only 1 pair, which is insufficient to get a meaningful correlation with a target matrix.