-----------------------------------------------------------------------------------------
Packages for the computation of optimally robust estimators
-----------------------------------------------------------------------------------------
We would like to announce the availability on CRAN (with possibly a
minor delay until on every mirror) of new versions of our packages for
the computation of optimally robust estimators; i.e., "RandVar",
"ROptEst", "RobLox" as well as a new package
"RobAStBase" (not yet:
ROptRegTS and RobRex).
-----------------------------------------------------------------------------------------
Devel versions on R-forge
-----------------------------------------------------------------------------------------
The development of these packages is under r-forge project RobASt
(Robust Asymptotic Statistics):
http://r-forge.r-project.org/projects/robast/
http://robast.r-forge.r-project.org/
If you find this project interesting and would like to collaborate, you
are warmly welcome.
We look forward to receiving questions, comments and suggestions.
Matthias Kohl
Peter Ruckdeschel
-----------------------------------------------------------------------------------------
RandVar - Implementation of random variables (version 0.6.3)
-----------------------------------------------------------------------------------------
The package RandVar which includes an S4 implementation of random
variables together with the packages distr, distrEx and distrMod form
the basis of our packages on robust statistics.
-----------------------------------------------------------------------------------------
RobAStBase - Robust Asymptotic Statistics (version 0.1.0)
-----------------------------------------------------------------------------------------
This is a new package including some necessary S4 class infrastructure
like neighborhoods, influence curves and robust models.
-----------------------------------------------------------------------------------------
ROptEst - Optimally robust estimation (version 0.6.0)
-----------------------------------------------------------------------------------------
This is the main package for the optimally robust estimation in smoothly
(L2-differentiable) parametric models [optimal in the sense of the
shrinking neighborhood setup]. By using S4 classes and methods
the implementation so far covers the optimally robust estimation for
all(!) smoothly (L2-differentiable/differentiable in quadratic mean)
parametric models which are based on a univariate distribution. Many
well-known parametric (in particular, exponential) families (Binomial,
Poission, Normal, Gamma, Gumbel, ...) are L2-differentiable.
We include several
+neighborhood types (convex contamination, total variation)
+risks (MSE, Hampel, overshoot/undershoot),
+bias-types (symmetric, one-sided, asymmetric)
+norms (unstandardized, self-standardized, information-standardized)
for all these models.
After installation you find a folder "scripts" in the package
directory
which includes many example scripts.
As the computation of optimally robust estimators involves several
steps, we -- in this new version -- added an interface function
"roptest" which can be used to perform all steps via one function.
-----------------------------------------------------------------------------------------
RobLox - Optimally robust influence curves for location and scale
(version 0.6.0)
-----------------------------------------------------------------------------------------
This package includes functions for the computation of many well known
influence curves (e.g., Huber-, Hampel-, Tukey-, Andrews-type) for
normal location and scale in the framework of our asymptotic setup.
Moreover, (and for us, more importantly) it includes the functions
"roblox", "rowRoblox" and "colRoblox" which can be
used to compute
optimally robust estimators in case of normal location and scale. These
functions are optimized for speed and can be applied to large scale
problems like for instance gene expression data. Using rowRobLox the
computation for a 50000 x 20 matrix takes about 2 sec. on a Centrino Duo
with 1.66 GHz. As a comparison (all on the same system): using apply and
huberM (robustbase), resp. huber (MASS) takes about 168 sec. resp 197
sec., using apply and roblox takes about 16 minutes and using apply and
roptest (ROptEst) takes about 1 month.
-----------------------------------------------------------------------------------------
ROptRegTS - Optimally robust estimation for regression-type models
RobRex - Optimally robust influence curves for regression and scale
-----------------------------------------------------------------------------------------
These two packages which provide S4 classes and methods for the
computation of optimally robust estimators in regression-type models are
not yet adapted to the new implementation. If you are interested in
working with these packages you have to use the old versions of the
above packages which we are pleased to provide on request (the sources
can also be found in the CRAN archives). But, of course, we will try to
update these packages as soon as possible.
--
Dr. Matthias Kohl
www.stamats.de
_______________________________________________
R-packages mailing list
R-packages at r-project.org
https://stat.ethz.ch/mailman/listinfo/r-packages
