Added methods for is_discrete() and
is_continuous() for the new distributions3
objects.
Replaced deprecated C function finite() with
isfinite() (again).
Added support for distributions3
workflows for censored and truncated normal, logistic, and Student’s t
distributions: CensoredNormal(),
TruncatedNormal(), CensoredLogistic(),
TruncatedLogistic(), CensoredStudentsT(),
TruncatedStudentsT(). See the corresponding manual pages
for examples illustrating how to work with the distributions in
practice, for computing moments, probabilities, densities, simulating
random values, etc.
Added prodist() method for extracting the
distributions3 objects above from fitted crch
objects, either in-sample or out-of-sample.
Bug fix in the computation of the mean of censored or truncated logistic distributions with large (or infinite) censoring/truncation points.
finite() with
isfinite().Added argument type to crch() which can
be set to "crps" for parameter estimation with minimum CRPS
instead of maximum likelihood.
Added S3 method for crps() from
scoringRules for crch objects.
Improvements for the predict() method:
"parameter", "density",
"probability", and "crps".type = "response" now the expected value and not
the location parameter is returned (not equal for censored and truncated
distributions). For better backward compatibility, the default type is
set to "location".Added pit(), rootogram(), and
simulate() methods for crch objects.
Changed argument names mean and sd to
location and scale in logistic and Student’s t
distribution functions
Added new function crch.stabsel() for stability
selection based on
crch.boost.fit(). Some S3 methods for the returned class
stabsel.crch are also provided.
New release accompanying the R Journal paper: Heteroscedastic Censored and
Truncated Regression with crch by Messner, Mayr, and Zeileis. See
also citation("crch").
Added estfun() method for crch
objects
The crch() function now supports coefficient
optimization by boosting to automatically select the most relevant input
variables in high-dimensional data settings. Extractor and plotting
functions for corresponding crch.boost objects are also
available.
Transferred functions to estimate density, distribution, score, and Hessian matrices to C code to accelerate coefficient optimization.
Added option to crch() to avoid computation of
covariance matrix.
Added left and right arguments to
predict.crch() and predict.crch.boost() to
allow quantile predictions for non-constant censoring or truncation
points.
Added model.matrix() and model.frame()
methods for crch objects
Bug fix in predict.crch(): In previous versions
predictions for models with other link functions than the log gave wrong
results
Added vignette to introduce the crch() function with
some theoretical background and an illustrating example:
vignette("crch", package = "crch")
The crch() function now also supports truncated
responses. Furthermore added a wrapper function trch() to
fit truncated regression models.
crch(): Analytical gradients and Hessian matrices
are provided for most models to speed up maximum likelihood optimization
(not available for Student’s t distribution with degrees of freedom
estimation).
crch(): For the scale model a link function can now
be specified ("log", "identity", or
"quadratic"). In previous version only the log was
supported.
Added functions for probability density, cumulative distribution, random numbers, and quantiles for censored and truncated normal, logistic, and Student’s t distributions.
The residuals() method for crch objects
now also provides quantile residuals (Dunn and Smyth 1996).
Added update() method for crch
objects.
citation("crch") for the accompanying manuscripts. Note
that the interface of both crch() and hxlr()
is still under development and might change in future versions of the
package.