Skip to contents

Changelog

Source: NEWS.md

gkwreg 2.1.1

CRAN release: 2025-11-15

gkwreg 2.1.0

Comparative Testing

  • Introduced a dedicated comparative test suite to validate gkwreg’s beta family implementation against the reference betareg package, ensuring numerical accuracy and reliability.

  • Confirmed statistical equivalence despite different internal parameterizations. Tests demonstrate that gkwreg’s shape-based (gamma, delta+1) approach produces equivalent statistical models to betareg’s mean-precision (mu, phi) approach.

  • Validated key outputs, showing that log-likelihood, AIC/BIC, fitted values, and predictions are virtually identical between the two packages when fitting the same beta regression model.

  • Successfully benchmarked gkwreg with family = "beta" as a robust and reliable alternative for beta regression, yielding the same inferential conclusions as the established betareg package.

  • Verified consistency across multiple scenarios, including controlled simulations with known parameters and real-world datasets (GasolineYield, FoodExpenditure), ensuring robust performance in diverse modeling contexts.

gkwreg 2.0.0

Major Changes

Package Restructuring

  • Complete package reformulation following JOSS reviewer feedback to reduce complexity and improve maintainability.

  • Distribution functions moved to separate package gkwdist: All d*, p*, q*, r* density/CDF/quantile/random generation functions have been extracted to the companion package gkwdist for cleaner namespace and reduced dependencies. The gkwreg package now focuses exclusively on regression modeling.

  • Univariate fitting functions removed: gkwfit(), gkwgof(), and gkwfitall() have been removed to maintain package focus on regression. Users needing univariate distribution fitting should use the gkwdist package directly or standard MLE approaches.

Simplified Interface

  • Introduced gkw_control(): All technical/optimization parameters (method, start, fixed, hessian, maxit, tolerances, etc.) are now consolidated in a dedicated control function following the glm.control() design pattern. This dramatically simplifies the main gkwreg() interface.

  • Removed arguments violating separation of concerns from gkwreg():

    • plot argument removed (use plot() method instead)
    • conf.level argument removed (use confint() method instead)
    • profile, submodels, npoints arguments removed (focused functionality)

  • Streamlined gkwreg() signature: Reduced from 15+ arguments to ~12 core arguments, with technical options delegated to control.

Complete S3 Method Implementation

Enhanced Diagnostics

  • Comprehensive plot.gkwreg() method with 6 diagnostic plot types:

    1. Residuals vs Observation Indices
    2. Cook’s Distance
    3. Generalized Leverage vs Fitted Values
    4. Residuals vs Linear Predictor
    5. Half-Normal Plot with Simulated Envelope
    6. Predicted vs Observed Values

  • Dual graphics system support: Base R graphics (default) or ggplot2 with automatic grid arrangement via gridExtra/ggpubr.

  • Advanced customization: Named-list interface for plot captions (partial customization without repeating all titles), theme control, sampling for large datasets.

Powerful Prediction

  • predict.gkwreg() with 9 prediction types:
    • "response", "variance", "link", "parameter"
    • Individual parameters: "alpha", "beta", "gamma", "delta", "lambda"
    • Distribution functions: "density", "probability", "quantile"
  • Element-wise and vectorized modes: Flexible evaluation via elementwise argument for CDF/PDF/quantile calculations.

Model Comparison Tools

  • Likelihood ratio tests: anova.gkwreg() for comparing nested models with automatic ordering and chi-squared tests; dedicated lrtest() function for pairwise comparisons.

  • Information criteria: AIC.gkwreg() and BIC.gkwreg() with multi-model comparison support returning data frames.

Documentation Improvements

  • Extensive Roxygen documentation for all exported functions with detailed examples, mathematical formulas, and usage guidance.

  • Updated README.md with comprehensive feature overview, quick start guide, advanced examples, and ecosystem comparison table.

  • NULL default intelligent behavior: Several arguments default to NULL triggering smart auto-configuration (e.g., sub.caption, ask, theme_fn in plot.gkwreg()).

Testing Framework

Comprehensive Test Suite Added

The package now includes a robust testing framework with 1000+ unit tests covering all major functionalities:

Core Function Testing
  • gkwreg(): 20 tests for model fitting, parameter estimation, formula handling, all distribution families, link functions, and convergence
  • predict.gkwreg(): 10 tests for predictions, including response means, densities, CDFs, quantiles, and parameter extraction
  • residuals.gkwreg(): 10 tests for all residual types (response, Pearson, deviance, quantile, standardized, working, partial)
  • fitted.gkwreg(): 10 tests for fitted value extraction and validation
S3 Methods Testing
Test Coverage Includes
  • ll 7 distribution families (GKw, BKw, KKw, EKw, MC, Kw, Beta)
  • Different link functions and scales
  • Edge cases and boundary conditions
  • Missing data handling (NA)
  • Subset and weight specifications
  • Large dataset performance
  • Error handling and input validation
  • Statistical correctness verification
  • Numerical accuracy checks
Testing Framework
  • Built with testthat package
  • Uses simulated data from gkwdist package
  • Tests with real datasets (GasolineYield, FoodExpenditure)
  • Reproducible with fixed random seeds

Minor Improvements

  • Link scaling support: Added link_scale argument to gkwreg() for controlling transformation intensity.

  • Performance optimizations: Intelligent caching, sampling support for diagnostics on large datasets, optional Hessian computation.

Breaking Changes: Version 2.0.0 introduces breaking changes. Code using gkwfit(), gkwgof(), gkwfitall(), or distribution functions (dgkw(), etc.) must be updated to use the gkwdist package or the new gkwreg() interface with gkw_control().