Corrected SSTR applies a nonnegative optimism correction to the LS
scale. corrected_pair() selects a thresholded fit and the
zero-threshold Ridge control on the same outer folds and ridge
penalties. This control isolates the empirical contribution of
thresholding within the scaled family.
library(softridge)
library(scaleCorrection)
set.seed(25)
X <- matrix(rnorm(60 * 100), 60, 100)
y <- drop(X[, 1:3] %*% c(2, -1.5, 1) + rnorm(60))
grid <- hr_grid(X, y, nk = 3, ng = 5)
foldid <- rep(1:3, length.out = nrow(X))
pair <- corrected_pair(X, y,
grid = list(k = grid$k, fraction = grid$fraction),
foldid = foldid, unit_id = "showcase-25")
pair$SSTR_c0[c("k", "gamma", "scale", "size")]
pair$Ridge_c0[c("k", "gamma", "scale", "size")]
prediction <- pair$SSTR_c0$coef[1] + X %*% pair$SSTR_c0$coef[-1]The same example is executable as
tools/smoke-companion.R from the repository root after
installing both packages. It is intentionally bounded. Its variance
estimator uses ten-fold CV-Lasso in each supplied training subset; this
is additional to the three outer folds in this demonstration.
Returned quantities
SSTR_c0 and Ridge_c0 contain
intercept-first coef, ridge penalty k,
threshold gamma, scale, selected size and validation score.
cv_error stores the shared mean-squared validation surface.
variance records training-only variance estimates and fold
hashes. Supplied foldid and grid are returned
for traceability.
The correction uses the selected ridge diagonal sum as a plug-in optimism term. It is a practical extension of established scaling ideas, including Hagiwara’s orthogonal-model work. An adaptive scale and selected tuning parameters do not inherit the fixed-tuning unbiased-risk statement. See risk and targets for the separate quantities.