[DirichletReg-News] DirichletReg 0.6-1 on R-Forge and CRAN
Marco Maier
marco.maier at wu.ac.at
Fri Dec 12 12:20:03 CET 2014
Dear DirichletReg list members,
I am happy to announce that a new release of DirichletReg is now available on R-forge and CRAN. Version 0.6-0 was a sort of “safety net” where a quite comprehensive test-suite was implemented (using Hadley Wickham’s “testthat”) and I have tried to move code to C, switching from the (kind of) deprecated .C() interface to .Call(). The package’s results are tested against values computed in Wolfram Mathematica.
Since the previous version, I have completely re-derived and simplified all analytic things (gradients, Hessians), implemented them in C and the gain in speed is considerable. Also, a new strategy to find starting values for the alternative model really boosted performance (in my test scenario about 3x faster).
Due to more performant model fitting, I have implemented a drop1() method that I find very useful. I have done some analyses using the package myself, and model selection was terribly tedious with manually fitting and comparing models. Now it is much easier and drop1 preserves the terms’ hierarchy, that is, if there is an interaction x:y, neither x nor y will be removed/tested. Still, there is a message when you use the function for the first time in a session, because, although I have tested it in many scenarios, there may be settings when it crashes – please report any strange behavior, so I can check/fix it.
I regard version 0.6-1 as a stable release – with all the tests I can confidently put it online for everyone to use it productively. Aside from potential bug-fixing business, I plan to concentrate on my dissertation now (of which this package is a part), so the next logical steps like an add1-method and consequently a method to perform automated model selection (like stepAIC in MASS) are mid-term goals.
As always, I have fixed some nasty bugs here and there – please don’t hesitate to contact me if you encounter problems or get suspicious results.
Thanks + kind regards,
Marco
Mag. Marco J. Maier
Institut für Statistik und Mathematik
Institute for Statistics and Mathematics
WU
Wirtschaftsuniversität Wien
Vienna University of Economics and Business
Building D4, Level 4
Welthandelsplatz 1
1020 Wien
Austria
Tel: +43-1-313 36-4335
Fax: +43-1-313 36-90-4335
E-Mail: marco.maier at wu.ac.at
wu.ac.at
http://statmath.wu.ac.at/~maier
More information about the DirichletReg-News
mailing list