Optimism-corrected scaling for soft-thresholded ridge.
This RcppArmadillo companion depends on the matching softridge engine. It jointly tunes corrected SSTR and the zero-threshold Ridge-c0 control using the same observations, outer folds and ridge-penalty candidates.
Install softridge first, then install this subdirectory:
remotes::install_github("byuzbasi/softridge", subdir = "companion/scaleCorrection",
upgrade = "never")See the shared getting-started guide, worked scaling example, and function reference.
Corrected scale and matched control
For a centered training response and the fitted vector from thresholded ridge, the scale is the nonnegative part of the response-fit inner product minus estimated variance times the selected ridge diagonal sum, divided by the squared norm of the fitted vector. A zero fitted vector receives scale zero. The returned SSTR_c0 and Ridge_c0 records contain intercept-first coefficient vectors, selected penalties, scales and CV scores.
The variance estimate uses training-only ten-fold CV-Lasso and records its fold digest. This plug-in correction and response-adaptive tuning do not inherit a fixed-parameter unbiased-risk guarantee.
| Return component | Meaning |
|---|---|
SSTR_c0 |
Jointly tuned ridge penalty and threshold with corrected scaling |
Ridge_c0 |
Zero-threshold control tuned over the same ridge penalties |
cv_error |
Mean validation squared-error surface |
variance |
Training-subset variance diagnostics and fold digests |
foldid, grid
|
User-supplied outer folds and candidate definitions |
The methods and fixed data-processing rules are those of the frozen research implementation, version 0.1.0. This package is distributed on GitHub, not CRAN.