Skip to contents

Reproducing the studies

Source: vignettes/articles/reproducibility.Rmd
reproducibility.Rmd

Package usage is not the article protocol

The general package API fits and predicts with SGLASSO. The article-specific scripts retain study designs, seeds, targets, grids, comparator settings, numerical validation and checkpoint rules. Substituting package defaults does not reproduce all calculations of a recorded study.

Resource Scope
paper_codes/ Published Gaussian SGLASSO article, cited below
logistic_paper_codes/ Logistic manuscript; study-specific code and settings

The published Gaussian study is: Yüzbaşı, B. and Cao, J. (2026). Collinear Groupwise Selection via Scaled Group Lasso. The American Statistician, 1–23. Advance online publication, 17 September 2026. doi:10.1080/00031305.2026.2709494.

The logistic publication directory contains code/settings only: no measurements, fitted models, figures, manuscript PDFs or result archives. The package’s own documentation images are separate illustration assets.

Logistic study: staged execution

  1. Obtain the exact separately distributed CoRSIVSZ file for the external application. Simulation fitting needs no measurement matrix.
  2. Run the code manifest verification and the small local validation suite.
  3. Explicitly launch the selected full study in your own terminal, using a new output directory outside the immutable code folder.
  4. Resume with the identical settings and input paths plus --resume.
  5. Verify final manifests, complete task counts and numerical gates before interpreting results. A completion marker alone is insufficient.

These commands are examples for the code-only distribution. Replace /absolute/... paths; they are not executed by this guide.

cd /absolute/path/to/sglasso/logistic_paper_codes
export OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1
export BLIS_NUM_THREADS=1 VECLIB_MAXIMUM_THREADS=1 NUMEXPR_NUM_THREADS=1
Rscript --vanilla reproduce.R --mode=verify
Rscript --vanilla reproduce.R --mode=validate \
  --data=/absolute/CoRSIVSZ_v1.rds --output=/absolute/new_checks

The run-code README gives full production, progress, resume and verification commands, resources and dependency requirements. New runs record their own code/data signatures and software versions. The research scripts require Unix fork support; this limitation is distinct from the general package API.

What remains fixed

The simulation comprises eight scenarios with 50 repetitions each and six audit variants per task. The independent external study uses development-only five-fold CV followed by refitting; the external cohort is not used to select tuning parameters. AUC and MCC inference uses paired, class-stratified draws from fixed predictions. It does not refit models or tune the classification threshold. The run code retains all six variants while the report displays five methods: SGLASSO, Group ENET, Group Lasso, Group MCP and Group SCAD.

Evidence and limits

Local validation is not a promise of convergence on every future dataset, proof of statistical superiority or validation on every operating system. Do not change tolerances or remove failed numerical gates to obtain a desired result. Preserve failure records, fix code locally and validate a new version.

No full simulation, CV study or bootstrap study is executed by this website build. It uses small synthetic documentation examples and completed outputs for illustration. Hosting the separately distributed CoRSIVSZ data file remains a separate step; the original measurements are publicly available in GEO.