Research implementation of a nonnegative optimism-corrected scale
corrected_pair.RdReuses frozen ridge preprocessing and ridge paths. For centered training data, computes the scale as the positive part of the inner product of the response with the unscaled fit minus estimated variance times the selected ridge diagonal sum, divided by the squared norm of the unscaled fit. Zero fitted vectors receive scale zero. This is a plug-in correction, not an unbiased risk estimate for the adaptively selected estimator.
Usage
corrected_pair(X, y, grid, foldid, unit_id, variance_fun = estimate_variance)
estimate_variance(X, y, unit_id, subset = "full")
variance_folds(n, unit_id, subset = "full")Arguments
- X
Finite numeric design matrix.
- y
Finite response vector.
- grid
List containing positive ridge penalties and threshold fractions, starting at zero.
- foldid
Saved outer cross-validation folds.
- unit_id
Frozen replication identifier.
- variance_fun
Training-subset variance estimator, replaceable for validation.
- subset
Deterministic namespace for full training or an outer training subset.
- n
Number of observations in a variance-estimation subset, at least 20.
Value
The two selected fits, a shared CV surface and variance diagnostics; or variance diagnostics or deterministic fold IDs for the respective helpers.
Details
Variance uses ten-fold cross-validated Lasso at lambda.min, with residual sum of squares divided by max(1,n-selected size-1), matching the intercept adjustment in the supplied edition-4 code. It is recomputed within each outer training subset. The original outer folds and data seeds are unchanged. The fixed-threshold ridge diagonal expression is used as a plug-in term when thresholds depend on the response. No new risk dominance claim is made.