NEWS

---

SuperSurv 0.1.9 (2026-10-02)

Changes since the previous CRAN release, 0.1.7.

Backward-Compatibility Summary

The default survival transformation remains exponential. The following changes can affect existing analyses or function calls.

Fitting and prediction

• Direct simplex optimization can change ensemble weights relative to normalized nonnegative least squares (NNLS).
• Corrected tied-event calibration can change calibrated survival predictions and ensemble weights on data with tied event times.
• Right-continuous censoring endpoint corrections can change fitted weights at grid points coinciding with observed censoring times.
• Models can now be fitted without supplying prediction data or prediction times. Existing calls that supply both arguments remain supported.
• Product-limit conversion is opt-in; exponential conversion remains the default.
• metalearner = "entropy" remains temporarily available with a deprecation warning. Use "brier" or "logloss" for new analyses.
• Single-patient and single-time forest predictions retain their documented matrix dimensions.

Evaluation and interpretation

• Corrected reverse Kaplan-Meier tie handling can change evaluation metrics on data with tied event and censoring times.
• RMST point contrasts remain available. The standard error, confidence interval, and p-value are no longer returned; inference = TRUE now produces an error.
• The optional SHAP backend is now kernelshap, and explain_kernel() requires an explicit eval_time. Explanations target the fitted ensemble's event probability at that time rather than a composite of native learner scores. The output table structure is retained, but earlier numerical SHAP results are not comparable. Recompute explanations with an explicitly chosen horizon.

Documentation

• Shortened vignette builds for routine CRAN checks. The final vignette-runtime changes do not alter the fitting, prediction, or evaluation code.