New working paper "Extended-Support Beta Regression for [0, 1] Responses" by Ioannis Kosmidis and Achim Zeileis in the arXiv.org E-Print Archive, doi:10.48550/arXiv.2409.07233.
New package web page (via altdoc/quarto) at
https://topmodels.R-Forge.R-project.org/betareg/
Extended functionality of predict() method for betareg objects and
enhanced the corresponding documentation, see ?predict.betareg.
Turned vignette("betareg", package = "betareg") and
vignette("betareg-ext", package = "betareg") from Sweave into Quarto
vignettes. Some improvements/updates in the text.
Major extension in betareg(): In addition to classic beta regression for
responses in the open interval (0, 1), extended-support beta regression is
added which can model responses in the closed interval [0, 1] (i.e., including
boundary observations at 0 and/or 1). This is accomplished by adding two new
response distributions: The extended-support beta distribution ("xbeta")
leverages an underlying symmetric four-parameter beta distribution with
exceedence parameter nu to obtain support [-nu, 1 + nu] that is subsequently
censored to [0, 1] in order to obtain point masses at the boundary values 0
and 1. The extended-support beta mixture distribution ("xbetax") is a
continuous mixture of extended-support beta distributions where the exceedence
parameter follows an exponential distribution with mean nu (rather than a
fixed value of nu). The latter "xbetax" specification is used by default
in case of boundary observations at 0 and/or 1. The "xbeta" specification
with fixed nu is mostly for testing and debugging purposes.
Quantile residuals are added to the residuals() method for betareg objects.
They are easy to compute and have good distributional properties. Hence,
they are the new default residuals.
Bug fix in pseudo.r.squared computation for weighted models where previously
the weights were erroneously ignored (reported by Ray Tayek).
Bug fixes in betatree(): Split points were computed incorrectly due to wrong
sign of the log-likelihood (reported by Se-Wan Jeong). And trees with only
intercepts for both mu and phi could not be fitted (reported by Ludwig
Hothorn).
betatree() the "xlevels" attribute from partykit::mob is now correctly
stored in $levels (rather than $xlevels) of the returned object.IGNORE_RDIFF flags in some examples in order to avoid showing
diffs due to small numeric deviations in some checks (especially on CRAN).suppressWarnings(RNGversion("3.5.0")) in those places where set.seed()
was used to assure exactly reproducible results from R 3.6.0 onwards.sctest() method for betatree objects when
strucchange package is loaded.The betatree() function now uses the new mob() implementation from the
partykit package (instead of the old party package). The user interface
essentially remained the same but now many more options are available through
the new mob() function. The returned model object is now inheriting from
modelparty/party.
Included grDevices in Imports.
Fixed model.frame() method for betareg objects which do not store the
model frame in $model.
betamix() gained arguments weights (case weights for observations) and
offset (for the mean linear predictor).
The Formula package is now only in Imports but not Depends (see below).
Method FLXgetModelmatrix for FLXMRbeta objects modified due to
changes in flexmix 2.3.12.
For some datasets betareg() would just "hang" because dbeta() "hangs"
for certain extreme parameter combinations (in current R versions).
betareg() now tries to catch these cases in order to avoid the problem.
Depends/Imports/Suggests have been rearranged to conform with current
CRAN check policies. This is the last version of betareg to have the
Formula package in Depends - from the next version onwards it will
only be in Imports.
The predict() method gained support for type = "quantile", so that
quantiles of the response distribution can be predicted.
The Formula package is now not only in the list of dependencies
but is also imported in the NAMESPACE, in order to facilitate
importing betareg in other packages.
.Call()-ing logit link functions directly, instead
use elements of make.link("logit").citation("betareg"). The paper presents
the recently introduced features: bias correction/reduction in
betareg(), recursive partitioning via betatree(), and finite
mixture modeling via betamix(). See also vignette("betareg-ext", package = "betareg") for the vignette version within the package.Formula interface for betamix() changed to allow for three parts
in the right hand side where the third part relates to the
concomitant variables.
Modified the internal structure of vignettes/tests. The original vignettes are now moved to the vignettes directory, containing also .Rout.save files. Similarly, an .Rout.save for the examples is added in the tests directory.
Support bias-corrected (BC) and bias-reduced (BR) maximum likelihood
estimation of beta regressions. See the type argument of betareg().
To enable BC/BR, an additional Fisher scoring iteration was added
that (by default) also enhances the usual ML results.
New vignette("betareg-ext", package = "betareg") introducing BC/BR
estimation along with the recent additions beta regression trees and latent
class beta regression (aka finite mixture beta regression models).
Enabled fitting of beta regression models without coefficients in the mean equation.
Enabled usage of offsets in both parts of the model, i.e., one can use
betareg(y ~ x + offset(o1) | z + offset(o2)) which is also equivalent to
betareg(y ~ x | z + offset(o2), offset = o1), i.e., the offset
argument of betareg is employed for the mean equation only. Consequently,
betareg_object$offset is now a list with two elements (mean/precision).
Added warning and ad-hoc workaround in the starting value selection
of betareg.fit() for the precision model. If no valid starting value can be
obtained, a warning is issued and c(1, 0, ..., 0) is employed.
Added betareg_object$nobs in the return object containing the number
of observations with non-zero weights. Then nobs() can be used to extract
this and consequently BIC() can be used to compute the BIC.
New betatree() function for beta regression trees based
on model-based recursive partitioning. betatree() leverages
the mob() function from the party package. For enabling this
plug-in, a StatModel constructor betaReg() is provided
based on the modeltools package.
New betamix() function for latent class beta regression, or
finite mixture beta regression models. betamix() leverages the
flexmix() function from the flexmix package. For enabling this
plug-in, the driver FLXMRbeta() is provided.
Added tests/vignette-betareg.R based on the models fitted
in vignette("betareg", package = "betareg").
The "levels" element of a betareg object is now a list
with components "mean", "precision", and "full" to match
the "terms" of the object.
Improved data handling bug in predict() method.
?gleverage.Package now published in Journal of Statistical Software,
see https://www.jstatsoft.org/v34/i02/
and citation("betareg") within R.
Bug fix and improvements in gleverage() method for betareg
objects: Analytic second derivatives are now used and
variable dispersion models are handled correctly.
dbeta(..., log = TRUE) is now used for computing the
log-likelihood which is numerically more stable
than the previous hand-crafted version.
The starting values in the dispersion regression are now chosen differently, resulting in a somewhat more robust specification of starting values. The intercept is computed as described in Ferrari & Cribari-Neto (2004), plus a link transformation (if any). All further parameters (if any) are initially set to zero. See also the vignette for details.
Various documentation improvements, especially in the vignette.
New vignette (written by Francisco Cribari-Neto and Z)
introducing the package and replicating a range of
publications related to beta regression:
vignette("betareg", package = "betareg")
provides some theoretical background, a discussion of the
implementation and several hands-on examples.
Implemented an optional precision model, yielding
variable dispersion. The precision parameter phi may
depend on a linear predictor, as suggested by
Simas, Barreto-Souza, and Rocha (2010). In single part
formulas of type y ~ x1 + x2, phi is by default assumed to
be constant, i.e., an intercept plus identity link. But
it can be extended to y ~ x1 + x2 | z1 + z2 where phi
depends on z1 + z2, by default through a log link.
Allowed all link functions (in mean model) that are
available in make.link() for binary responses, and added
log-log link.
Added data and replication code for Smithson & Verkuilen
(2006, Psychological Methods). See ?ReadingSkills,
?MockJurors, ?StressAnxiety as well as the complete
replication code in demo("SmithsonVerkuilen2006").
Default in residuals() (as well as in the related plot()
and summary() components) is now to use standardized
weighted residuals 2 (type = "sweighted2").
Package betareg was orphaned on CRAN, Z took over
as maintainer, ended up re-writing the whole package.
The package still provides all functionality as before
but the interface is not fully backward-compatible.
betareg(): More standard formula-interface arguments;
betareg objects do not inherit from lm anymore.
betareg.fit(): Renamed from br.fit(), enhanced interface
with more arguments and returned information. Untested
support of weighted regressions is enabled.
betareg.control(): New function encapsulating control
of optim(), slightly modified default values.
anova() method was removed, use lrtest() from lmtest
package instead.
gen.lev.betareg() was changed to gleverage() method
(with new generic) and a bug in the method was fixed.
envelope.beta() was removed and is now included in
plot() method for betareg objects.
Datasets prater and pratergrouped were incorporated
into a single GasolineYield dataset.
New data set FoodExpenditure from Griffiths et al. (1993),
replicating second application from Ferrari and Cribari-Neto
(2004).
Added NAMESPACE.
The residuals() method now has three further types of
residuals suggested by Espinheira et al. (2008) who recommend
to use "standardized weighted residuals 2" (type = "sweighted2").
The default are Pearson (aka standardized) residuals.