Evaluate regional access under a whole-brain pattern fit

Usage

local_performance(result, regions, independent_roi = NULL)

Arguments

result: A pattern_global_result retaining fold fits.
regions: Named list of input-column positions or logical masks. These are dataset matrix columns, not voxel IDs. Regions may overlap.
independent_roi: Optional named list of fold-resolved pattern_ledgers from independent ROI models on exactly the same rows and folds. Truth, target type, partition, and baseline must agree.

Value

A table with region, metric, whole_brain, and local_restricted columns (and independent_roi if supplied). Regional pooled and fold-resolved ledgers are retained as attributes.

Details

Each region uses the whole-brain task representation and the exact marginal covariance Psi_RR; neither patterns nor noise are refitted. All preprocessing is columnwise. Repeated assessments are averaged per observation before scoring, just as for the whole-brain ledger. R-squared uses fold training means. This measures locally accessible information under the shared model, not optimal performance of a separately fitted ROI. Finite-sample regional accuracy need not be below whole-brain accuracy.