Likelihood-based and Bayesian methods for various compound Poisson linear models based on Zhang, Yanwei (2013) < https://link.springer.com/article/10.1007/s11222-012-9343-7>.

CHANGES:

o ‘gini’ now can handle a single score o Fixed bug when supplying initial values

CHANGES:

o 'predict' method failed when 'cpglm' is rank deficient. Fixed this. Now returns the prediction but with a warning. o 'predict' method for 'cpglmm' defaults to type = 'response'. o Fix issues when specifying initial values for 'bcplm' o Add author info for the 'amer' functions o Update vignettes

CHANGES:

o update to be compatible with R 3.x (Prior versions failed)

CHANGES:

o fix bug related to sp2d

CHANGES:

o remove dependency on lme4 which is now completely re-written

CHANGES:

o remove dependency on amer which now is withdrawn from CRAN

NEW FEATURES

o Add vignettes o Add function 'zcpglm' which implements a zero-inflated compound Poisson generalized linear model o Add function 'gini' which computes the Gini indices that enable robust model comparison involving compound Poisson distributions

CHANGES:

o Add a new data set 'AutoClaim' o Set the max number of terms in computing the compound Poisson density using series evaluation method. o Correct bugs in predicting 'cpglm'. When the new data set has fewer factor levels, the old method produces wrong predictions. o 'cpglm' has an argument 'optimizer' that allows users to select optimization routines. o Function 'bcplm' now implements MCMC methods for both GLM and mixed models. This is a combination of the old functions 'bcpglm' and 'bcpglmm' o Produce model summary for Bayesian estimates (from 'bcplm') o Provide methods 'fixef' and 'VarCorr' for class 'bcplm' o The tuning procedure in MCMC is now based a method described in Browne and Draper (2005) o MCMC now only implements univariate M-H within Gibbs sampling. Block update for the fixed effects are removed. o Remove the latent variable approach for Bayesian estimation o Remove method 'mcmcsamp' o Functions 'getF' and 'plotF' (copied from 'amer') now works for additive models to extract and plot fitted smoothing effects o Fix bugs in quadrature estimation of mixed models due to code inherited from lme4. This will only affect models with multiple random effects per level. o Fix the underflow issue in the quadrature estimation. o Implement the PQL method to generate initial values in 'cpglmm'.

NEW FEATURES

o Methods 'mcmcsamp' now is not available for 'cpglm' and 'cpglmm' objects as a convenient way to perform MCMC simulations. o Function 'cpglmm' handles additive models in a similar way as the package 'amer'. o Big data capability is added to function 'cpglm', which uses the bounded memory regression facility from the package 'biglm'. o Function 'cpglmm' has an additional argument 'optimizer' that allows the users to choose which optimization routine to be used. The package 'minqa' is now imported for the use 'bobyqa'. o Function 'cpglmm' now implements adaptive Gauss-Hermite quadrature method for models with a single grouping factor. o Function 'bcpglmm' now implements an additional latent variable approach.

CHANGES: (user-visible) o The MCEM method is now completely removed from 'cpglm' o In 'cpglmm', the Laplace approximated loglikelihood seems to have left out the dispersion parameter for one term, resulting a larger than expected variance component estimate. This is now fixed and it is more consistent with the quadrature estimate. o Add method 'predict' for 'cpglm' and 'cpglmm', which computes the predicted values for a new data set, but not the prediction errors. o In 'cpglmm', fix a bug in specifying 'offset' o In 'cpglmm', 'sigmaML' is updated after fitting the model so that the 'postVar' option in 'ranef' in 'lme4' can be used now. o 'weights' was not reflected in the update of the deviance. This is fixed now. o In 'cpglmm', 'vcov' now computes variances for 'phi' and 'p'

(not user-visible) o register native routines in initialization

NEW FEATURES

o Function 'bcpglmm' is added that handles Bayesian mixed-effect models using MCMC simulations.

CHANGES: (user-visible) o create the class 'cplm' as a fundamental structure in the package, and define utility methods for it o replace the 'pstart', 'phistart' and 'betastart' arguments by a single argument 'inits' in most functions o combine the documentation for all classes and methods

NEW FEATURES

o Function 'cpglmm' is added that handles mixed-effect models using
Laplace approximations. This is based on the R package 'lme4'.

o Function 'bcpglm' now has a second method to fit Bayesian compound
Poisson GLM using direct Tweedie density approximation.
o Function 'bcpglm' also has a tuning phase that automatically updates
the scale parameter in the proposal distribution.
o The profile likelihood method in 'cpglm' is now automated

CHANGES: (user-visible) o Prior distribution of the dispersion parameter in 'bcpglm' is changed to be Uniform, specified in the argument 'bound.phi' o 'bcpglm' has another argument 'method' that allows users to choose from the latent variable approach or direct density evaluation o An insurance example 'insLoss' is added in 'bcpglm' o Remove the 'digits' parameter in control in 'cpglm' as the profile likelihood method is automated now o MCEM method in 'cpglm' simplifies the process of increase sample size. The old time-consuming method of estimating approximate covariance matrix is removed. So the 'alpha' parameter in control is removed. o The default method in 'cpglm' is now set to be 'profile' o Remove the 'summary' slot in 'bcpglm' o The profile method in 'cpglm' now returns covariance estimate for the dispersion and index parameter

(not user-visible): o 'bcpglm' replaces ARMS with M-H update. Now the dependency on the ARMS functions is eliminated o 'bcpglm' now generates starting values using 'cpglm' o Simplify rejection sampling of latent variables (now twice faster)

NEW FEATURES

o The package now implements MCMC methods for Bayesian compound
Poisson GLM in the function "bcpglm" with the use of
latent variables.

o The R package "coda" is imported so that a large number of functions
and methods defined there are now directly applicable to
the simulation results from "bcpglm" to help diagnose convergence
and summarize posterior inference.

CHANGES: (user-visible) o Various methods defined for the class "bcplm" and "bcpglm" (not user-visible, all in C code): o Change the use of "R_alloc" in "lbfgsb" to "Calloc" and "Free" (the old function eats up memory quickly) o Simplify rejection sampling of latent variables (now twice faster)

CHANGES (not user-visible, all in C code): o Fix a bug in rejection sampling of the latent variable o Fix a bug in specifying weights o Divide cpglm_str into three parts, one for data and parameters, one for latent variable, and one for EM related

NEW FEATURES

o Add a wrapper of the profile likelihood approach to the "cpglm" function that runs automatically to generate estimate of the index parameter to arbitrary accuracy.

CHANGES:

o The MCEM algorithm is now implemented in pure C code o Remove the restriction on the "weights" argument (but not tested) o Add "beta.step" in "control" to allow skips in the update of beta o Allow "link" to be both character and numeric o Force coercion of argument type before callings the C function - thanks Mikel Esnaola Acebes for pointing out this bug o Re-write "summary" and "show" function to produce statistical test output automatically o Revise "residuals" to allow different types of residuals to be computed o Add methods for "formula", "AIC", "deviance", "model.matrix", "terms" o Output now returns "deviance", "aic" and "model.frame" o Tracing info from MCEM tidied up by showing only the dispersion, the index parameter, and the sample size (if necessary) o Fix bug in the definition of "[[", add methods for "["