||Get Public Help
||a new version is out and there have been quite a lot of changes. some new features are:
- the data structure of objects generated by DR_data() has been changed. it is no longer a list, but now a numeric matrix of class "DirichletRegData" with lots of attributes. this permits you to store the processed data in the original data.frame to keep your workspace tidy, e.g.,
> ArcticLake$Y <- DR_data(ArcticLake[, 1:3])
so you can fit a model like:
> DirichReg(Y ~ depth, ArcticLake)
note that `free floating' dependent variables (i.e., not part of the data.frame) can also be used, as in prior versions.
- formula processing is now handled by the package `Formula' which is much more robust than my `homemade' routines in earlier versions and has the advantage that
- there is an update() method for Dirichlet regression models. so if you fit a model like
> res1 <- DirichReg(Y ~ depth + I(depth^2), ArcticLake)
you can easily omit, e.g., the quadratic term for `sand' by typing
> res2 <- update(res1, . ~ . -I(depth^2))
instead of having to specify a completely new model like
> DirichReg(Y ~ depth | depth + I(depth^2) | depth + I(depth^2), ArcticLake)
therefore, model selection is made a lot easier.
- i did my best to complete the documentation which was a little fragmentary for some topics.
- optimization settings have been tweaked and estimation is now both, faster, and more robust.
- the package is now, by default, byte-compiled and therefore even faster.
- some new methods for the class "DirichletRegModel" have been implemented.
- DirichReg() has a couple of useful new features (subset, weights, etc.).
- lots more ... please consult the documentation and have a look at the examples.
||A new version has been released on R-Forge and it features lots of changes, fixes and improvements (especially since the last post of version 0.4-0 on 2012-04-03).
One central improvement was to implement all time-critical routines in C (called via .Call). In this version, log-likelihood and gradient functions are finally mostly calculated using compiled C code. While working on the transition from R to C, some routines were optimized for performance.
Another important feature was the implementation of a drop1 method for the models. This greatly eases backward-selection – add1 and finally something like a stepAIC method (as in MASS) are planned to automatize the process.
To expand the package with confidence (i.e., change stuff without breaking existing functions), there is a test-suite (right now for the common model) using the great package testthat, that checks results, etc.
A working paper was published and citation information was updated. The vignette is the working paper’s full code.
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. URL: http://epub.wu.ac.at/4077/
Finally, there have been lots and lots of small changes, fixes, etc.
Thanks for your interest and if you find any bugs or experience unexpected behavior, please don’t hesitate to contact me.
||Version 0.6-1, which has been thoroughly tested, is now available for productive use on R-forge and CRAN.
Performance has been boosted considerably and the drop1() method hopefully is a useful tool to ease (backward) model selection.
Thanks for your interest in the package – if you find any bugs or experience unexpected behavior, please don’t hesitate to contact me.