Version 4.2-5

The most important change in this version of* gamlss* is the way that the standard errors are calculated. In previous version the *vcov() *function was calculated using a final iteration to a non-linear maximisation procedure. This procedure failed in a lot of occasions and the result was that the reported standard errors were the ones given from the individuals distribution parameter fits of the algorithm which did not take account of the intercorrelation of the distribution parameters. In this version a new function *gen.likelihood() *is indroduced. This function calculates the likelihood function of the model, so that the numerical calculation of the Hessian matrix of the full model can be achieved easily. For fully parametric models this should produce accurate standard errors based on the observed information matrix. For models with non-parametric terms does non at the moment take account of the variability of the smoothing non-parametric terms, so a warning is produced.

The following are the changes in version 4.2.4:

1. gamlss:

- i) the function
*lms()*has an extra argument*method.pb*added on. This allows the use of the GAIC method of estimating the smoothing in pb() parameters. It was found that for large data sets the local maximum likelihood (which is the default method of estimation) produces too wiggly centiles. The penalty k is taken from the argument*k*and it is the same as the one selecting the distribution. - ii) the function
*rqres.plot()*is fixed and improved to allow worm plot as well as QQ-plots - iii) The
*Rsq()*function is introduced. It uses the generalised R-squared of Nagelkerke, (1991). - iv) The
*term.plot()*function has now*se=TRUE*and*ylim=”common*” as defaults. - v) the function
*gen.likelihood()*is introduced. It creates the log-likelihood of a fitted model. It is used by vcov() and it should be more reliable than previous versions of vcov(). - vi) The
*gamlss()*function now saves the offset’s for all parameters. This is needed for*vcov()*and*gen.likelihood()*and allows the*vcov()*to work with offsets. - vii) The
*wp()*function allows now two explanatory variables including factors (see the examples in help). Also it was found that if the model is very bad some of the values of the worm plot could be Inf so if*line=TRUE*the function was failing when the cubic fit in lm() was used. Now if there are Inf values, it does not fit the lm() model. - viii)
*summary()*now allow the coefficients and standard errors to saved in a table. - ix)
*acfResid()*is introduced to plot the act function for power function of the residuals, r, r^2, r^3 and r^4. - x) the method
*confint()*is introduced for gamlss objects so confidence intervals for the beta parameters can be produced - xi)
*attach()*was taken out from the functions*findhyper(), fitDist(), gamlssML(), histDis() , lms(),**par.plot(), prof.term() and wp()*to comply with the new regulation of R

2. gamlss.tr:

- i) A new argument
*“varying”*is introduce in all the truncated functions. It allows the distribution of the response variable to have different truncated points at the observational level. (Note that before the truncated parameters had to be the same for all observations)

Mike:

Do you have a generalized Poisson distribution in gamlss yet? Of all the count models it has been missing in the past. I am forced to use the VGAM’s vglm function for GP modeling, which is fine, but I would rather have it in gamlss, as well (hopefully) zero-inflated, truncated, and censored GP models. I’ve programmed these using Stata, but I prefer using gamlss for these models.

Also, thank you for using a direct parameterization of the dispersion parameter for the negative binomial. This is consistent with all commercial packages and I think common sense. Because of this I always recommend gamlss for NB modeling in R over glm.nb to those who read my books and articles. gamlss is the preferred software in R for NB modeling. Joseph Hilbe j.m.hilbe(at)gmail.com

Dear Joseph

We will try to put GP as soon as possible

Thanks

Mikis