Searchlight scores

Dear reader,

I am currently applying seachlight-analysis (nilean.decoding) to my fMRI-data which involves two experimental conditions. I am wondering of what nature the resulting searchlight.scores_ are. Intuitively, I assumed they are classification accuracies of the form #correct / #false. On one of my data-sets the obtained values seemed to confirm this assumption but on another, I get scores close to zero with a maximum of .13. Since chance-level is .5 something seems wrong. I am currently using svc as the estimator, so I thought the searchlight.scores_ may be the SVM scores but then I would expect to obtain negative scores as well (or maybe only one of the classes is returned?).
I am grateful if somebody can shed light on the matter!

Best, Marius

Hello Marius,

The searchlight.scores_ are of the same shape as your process_mask_img :

  • Outside this mask, scores are 0.
  • For a voxel i inside this mask, its score, searchlight.scores_[i], is the average classification score for all voxels in the searchlight centered on i

This score is accuracy by default for a SVC unless you overrode it with by using the scoring argument.

Without more details on your decoding setup and replication code it’s difficult to explain the inconsistencies you observe. If you want to investigate more, could you provide some code to replicate this ?

1 Like

Dear tbazeille,

thank you for your answer. For now I switched to another package to do the analysis. I will come back to the issue and post a follow-up, soon.