Major revision of the package, refactoring some crucial infrastructure.
Rootograms and other visualizations are in topmodels (currently on R-Forge),
see https://topmodels.R-Forge.R-project.org/articles/topmodels.html.
Distribution functions (d/p/q/r) are in distributions3 (on CRAN),
see https://www.zeileis.org/news/user2022/.
Also, data sets SerumPotassium and VolcanoHeights are now in topmodels.
Improved predict() method with more type of predictions, including
moments (type = "mean" or "variance"), "quantile", "density",
cumulative distribution function ("probability") etc.
By default, the prediction is computed for the "full" outcome model
(e.g., hurdle or zero-inflation) but can also be just the "count" or
"zero" component or the distribution "truncated" at zero.
Bug fix in dhpois() and dhnbinom() when x is a vector and pi is a
scalar (reported by Andrea Gilardi).
Try to auto-detect columns in regressor matrices that are aliased, i.e.,
whose coefficients cannot be estimated. This is based on pivot and rank
from qr() and can detect, e.g., constant or linearly dependent columns.
Catch errors when inverting the Hessian from optim, returning an NA
matrix of correct dimension instead.
Conditionally register autoplot.rootogram() method.
New d/p/q/r/s functions for all combinations of Poisson vs. negative
binomial (NB) and zero-truncated vs. zero-inflated vs. hurdle. The parameter
names are all greek letters: lambda for Poisson mean, mu for NB mean,
theta for NB dispersion, pi for zero-modification probability (either
inflation or hurdle crossing). The theta parameter may also be called
size for consistency with base R's nbinom family.
New generic function pit() for extracting the probability integral
transform F(y) where F is the predicted cumulative density function
and y the observed response.
Quantile residuals are now computed based on the pit() extractor.
New graphics function pithist() for PIT histograms based on pit().
New data set OralHealthNL accompanying Hofstetter et al. (2016):
"Modeling Caries Experience: Advantages of the Use of the Hurdle Model",
Caries Research, 50(6), 517-526.
doi:10.1159/000448197
New d/p/q/r/s functions for the zero-truncated Poisson distribution
where the parameter can either be specified in terms of the zero-truncated
mean or the untruncated mean lambda.
New ztpoisson() family object for the estimation of zero-truncated
Poisson regression model via glm().
New qresiduals() generic for computing (randomized) quantile residuals
along with methods for various objects. (This is somewhat more flexible
than statmod::qresiduals() which is not generic.)
A Q-Q plot based on quantile residuals is available in the new function
qqrplot().
The hurdle() function now allows for the restriction of the theta
parameter for the negative binomial distribution across the censored
zero and truncated count components. The restriction is applied if
dist = "negbin", zero.dist = "negbin", and separate = FALSE.
Bug fix in the computation of the standard error of log(theta)
in zeroinfl() when with weighted data. The weights were used
for the estimation of log(theta) but the standard error was not
scaled accordingly.
style and scale are now added as attributes to rootogram
objects so that they can be re-used in the c() and + methods.New data sets SerumPotassium, TakeoverBids, VolcanoHeights.
New function disptest() providing several (score) tests for over/underdispersion.
New rootogram() function with a wide range methods for creating rootograms
based on various fitted model objects (fitdistr, glm, hurdle,
zeroinfl, zerotrunc, ...)
FLXMRnegbin() driver for estimating mixtures of negative binomial models
using flexmix.
MBnegbin(), MBztnegbin(), MBztpoisson() and MBbinomial() families for estimating
boosted components of hurdle models using mboost.
Ported zeroinfl() and hurdle() along with corresponding vignette
from pscl to countreg.
Added zerotrunc() for estimating zero-truncated count regressions.