C-Routines are now symbols prefixed with C_ (e.g., C_wght_LL_grad_alternative).
.onUnload now also unloads the DLL.
Changed call to structure() from the deprecated special name .Dim to dim.
Implemented CRAN-Notes.
summary.DirichletRegModel now returns a summary_DirichletRegModel-class object without printing it (see https://github.com/maiermarco/DirichletReg/issues/3).
Moved rmarkdown to Suggests to comply with new version of knitr (see https://github.com/yihui/knitr/issues/1864).
Fixed: Cutoff in formula-deparsing removed.
Removed rgl dependency (package will be loaded if necessary).
Moved to GitHub.
Instead of producing an error, density functions now return NaN with a warning if any element in alpha is <= 0.
Fixed an error in the ternary plot (bottom axis tick labels were printed in reversed order; pointed out by Emilio A. Laca).
Imports functions from default packages, as required by the new CRAN-check.
Changed the Description field to eliminate the note “Malformed Description field: should contain one or more complete sentences.” when checking the package.
Re-derived gradients/Hessians, optimized C-code, and tweaked computation of starting values which leads to considerably better performance.
Fixed some bugs in drop1.
Expanded testthat checks.
All likelihood and gradient functions are now written in C (all called via .Call() now instead of .C()) which leads to a considerable gain in speed.
A drop1 method for Dirichlet regression models was added.
As it is still experimental and will probably change, use it with care.
Options such as scope will be added in one of the next releases.
Fixed a bug in the print method for confidence intervals and one in confint.DirichletRegModel().
Expanded the testthat check- and test-suite.
Known issue: If you have collinear (aliased) terms, the estimation will fail.
This will be handled automatically in subsequent releases, but for now, please remove the respective terms.
If you fit a model and it says something like: Error in prepareFixed(start = start, activePar = NULL, fixed = fixed): At least one parameter must not be fixed using argument 'fixed' you most likely have collinear terms or “empty” combinations of interaction terms.
Fixed checking functions in tests/testthat.
Fixed a bug when using the subset argument.
Added tolerance for normalization check to DR_data.
moved NEWS to the new fancy NEWS.Rd file/format.
Added the possibility to do quick analyses and transforming data “on the fly”, like DirichReg(DR_data(A) ~ 1).
However this is only intended for quick checking purposes and may be removed at any time.
Transformation in DR_data is now not only TRUE/FALSE, but, by default, a small numeric value to avoid troubles with floating point numbers close to 0 or 1.
Time-critical routines were implemented in C (pure R versions are available, see ?ddirichlet).
anova.DirichletRegModel now invisibly returns results as an object that is printed by a method.
Optimized estimation routines.
Fixed a bug in the predict method.
Started development of a comprehensive test-suite using testthat.
Published a working paper on the package:
Maier, M. J. (2014). DirichletReg: Dirichlet Regression for Compositional Data in R. Research Report Series / Department of Statistics and Mathematics, 125. WU Vienna University of Economics and Business, Vienna. doi:10.57938/ad3142d3-2fcd-4c37-aec6-8e0bd7d077e1
Added vignette with code to the working paper.
Added citation info.
The trafo Argument of DR_data has been changed, because it has lead to problems in practical applications when numbers very close to 0 or 1 were present.
DR_data checks for negative values and generates an appropriate error message.
DR_data has been made more robust in the presence of NAs.
Data structure generated by DR_data has changed – the new objects can now be integrated into data frames.
Formula processing is now handled by the package Formula.
New methods have been implemented, especially for the class DirichletRegModel.
The documentation is now quite complete.
Some speed improvements could be achieved.
Lots of minor (invisible) changes.
Added the analytical Gradient and Hessian for both parametrizations.
Optimization: preliminary results by BFGS that become starting values for Newton-Raphson optimization computing the final results.
Implemented some residuals
Updated help entries