pense 2.5.2

• Fix division by a value close to 0 on certain LAPACK implementations, leading to residuals without variation and a failure in RIS-CV.

pense 2.5.0

• Use Robust Information Sharing Cross-Validation (RIS-CV) by default for pense_cv() and adapense_cv(). Details about RIS-CV can be found in Kepplinger & Wei (2025).
• Add support for the optimal rho function (Maronna et al. 2018, Section 5.8.1), using the polynomial approximation from the robustbase package.
• Fix undefined behavior identified by the extended CRAN checks.

pense 2.2.2

• Fix build warnings from GCC 14.1.0 on CRAN

pense 2.2.1

• Fix build error on Windows using GCC version < 9.

pense 2.2.0

• Fix bug where argument max_solutions is not used correctly.
• Add new numerical coordinate-descent algorithms for LS-EN (en_cd_options()).
• Revamp the regularization path and how starting points are used and shared.
• Share starting points across CV folds and replications.
• Address changes in the upcoming Rcpp release.

pense 2.1.0

• Penalty loadings are now applied to both the L1 and L2 parts of the EN penalty. This will lead to different results for adaptive PENSE and other adaptive estimators when fitted with alpha < 1!
• PENSE and regularized M-estimators now accept multiple alpha values and automatic hyper-parameter selection will also choose the best alpha value.
• New support for specifying the more general “1-SE” rules for the penalization level as string. All methods which support lambda = "min" to extract the best fit, also support the syntax lambda = "{m}-se" to extract the most parsimonious fit within m standard-errors of the best fit.
• Adaptively choose the actual breakdown point based on the number of observations. The chosen breakdown point is close to the user-specified breakdown point, but avoids numerical instabilities in the S-loss and excessive computation time caused by these instabilities.
• Simplify the DAL algorithm to fully rely on linear algebra routines from the BLAS/LAPACK library linked to R. To improve the speed of the DAL algorithm, optimized BLAS/LAPACK libraries are recommended.
• Fix memory issues from edge-cases and OpenMP problems with Intel compilers

pense 2.0.3

• Fix a bug causing PENSE-Ridge, i.e., pense(..., alpha = 0), to take a long time to compute.
• Fix a compilation error on RHEL due to an error in the autoconf script.
• Fix problems in prediction_performance() related to the non-standard evaluation of objects.
• Also return standardized coefficients as std_beta and std_intercept. # pense 2.0.2
• Fix mishandling of response variables with a robust scale of 0, e.g., 0-inflated responses or responses with more than 50% identical values. # pense 2.0.1
• Add new functions for compute adaptive PENSE estimates (adapense() and adapense_cv()).
• Functions for fitting the model (pense(), adapense(), regmest(), etc.) are not estimating prediction performance via cross-validation anymore. This can now be done using the corresponding functions pense_cv(), adapense_cv(), and so on.
• New function prediction_performance() to summarize the prediction performance of several fits.
• The plot(), coef(), summary(), and predict() methods for cross-validated fits also implement the “one-standard-error rule” (with the “1” adjustable by the user).
• Decrease computation time for most problems.
• New ADMM algorithm for (weighted) elastic net problems with many observations and many predictors. The new algorithm can be selected with en_admm_options().
• Argument correct in pense(), pensem(), coef(), etc., is not supported anymore and will be ignored with a warning. All estimates are now uncorrected (i.e., correct=FALSE in previous versions of the package).
• Make interface more consistent and deprecate the following methods:
  • pensem() is now called pensem_cv().
  • initest_options() is replaced by enpy_options() using better naming of arguments.
  • en_options_aug_lars() and en_options_dal() are replaced by en_lars_options() and en_dal_options() for more consistent naming.
  • pense_options() and mstep_options() are superseded by mm_algorithm_options() and arguments specified in the calls to pense() and companions.
  • enpy() is replaced by enpy_initial_estimates() which has different default argument values.
• Deprecated functions can still be used (for now) with a warning.

pense 1.2.9

• Fix LTO warnings reported in CRAN checks
• Update autoconf script to address deprecation warnings in r-devel.

pense 1.2.5

• Fix compatibility of BLAS/LAPACK prototypes with RcppArmadillo 0.9.500.

pense 1.2.4

• Fix autoconf script.

pense 1.2.1

• Prepare for changes to the upcoming Rcpp (make compatible with STRICT_R_HEADERS)
• Fix a bug in computing PSCs when using the augmented ridge algorithm for EN.

pense 1.2.0

• Changed the internal scaling of the regularization parameter for pense and pensem. Note: The lambda values in this release are not the same as in previous releases!
• Fixed a bug when standardizing predictor variables with a MAD of 0 (thanks @hadjipantelis for reporting).
• The maximum value for the regularization parameter lambda is now chosen exactly.
• Fixed a bug when computing “exact” principal sensitivity components. # pense 1.0.8
• Fix error with robustbase-0.92-8 as reported by Martin Maechler.
• Fix undefined behavior in C++ code resulting in build error on Solaris (x86).
• Fix predict() function for pensem objects if computed from a fitted pense object.
• Always use delta and cc specified in pense_options() for the initial estimator. Remove delta and cc arguments from initest_options() and instead add them to enpy().
• Add further measure of the prediction performance (resid_size) to obj$cv_lambda_grid, where obj is of class pense or pensem. # pense 1.0.6:
• Initial stable release of the package.