The aim of this document is to keep track of the changes made to the different versions of the `R`

package `ptmixed`

.

The numbering of package versions follows the convention a.b.c, where a and b are non-negative integers and c is a positive integer. When minor changes are made to the package, a and b are kept fixed and only c is increased. Major changes to the package, instead, are made apparent by changing a or b.

Each section of this document corresponds to a major change in the package - in other words, within a section you will find all those package versions a.b.x where a and b are fixed whereas x = 1, 2, 3, … Each subsection corresponds to a specific package version.

- Released: June 2020
- The article describing the Poisson-Tweedie GLMM has been accepted for publication in
*Statistical Modelling*! You can read the preprint of the article on arXiv (arXiv:2004.11193). - Added link to arXiv preprint in package description and vignette
- Added file with citation info
- Added further arguments to
`make.spaghetti()`

;`cex.lab`

argument fixed

- Released: April 2020
- Adapted code so that it runs both for balanced and unbalanced datasets (previously balanced design was assumed)
- Fixed problem with visualization of vignette on CRAN page

- Released: April 2020
- Added vignette to illustrate the functionalities of the package
- Simplified syntax of
`ptmixed()`

,`ptglm()`

,`nbmixed()`

and`nbglm()`

(wrt the`id`

and`offset`

arguments).`ranef()`

function updated accordingly - Added example dataset
`df1`

, used in the`ptmixed()`

and`nbmixed()`

help pages. Examples in help pages revised - Added
`simulate_ptglmm()`

function, to be used for illustration purposes (in the vignettes) - Added
`pmf()`

function to visualize the pmf of a discrete variable `make.spaghetti()`

: fixed minor bug in that arose when the`col`

argument was specified + added`legend.inset`

argument

- Released: February 2020
- Added the possibility to use Laplace approximation, which is a special case of the adaptive Gauss Hermite quadrature method where just 1 quadrature point is used (simply set
`npoints = 1`

in`ptmixed()`

or`nbmixed()`

). Note: use of the Laplace is not recommended, because it is less accurate than the adaptive GH, results in lower convergence rates and can yield biased parameter estimates! We recommend using a sufficient number of quadrature points (5 typically produces a good likelihood approximation) - Added
`make.spaghetti()`

function to create a spaghetti plot / trajectory plot to visualize longitudinal data - Added example dataset
`df1`

- Added
`silent`

argument to`summary.ptglmm()`

. Furthermore, printed output table with parameter estimates and Wald test is now presented with at most 4 decimals - Fixed bug that caused
`ptglm()`

and`nbglm()`

to print detailed optimization info also when`trace = T`

- Released: January 2020
- Changed class check within
`wald.test()`

to prevent problems with future`R`

release (4.0.0) - Fixed bug that occurred when
`freq.updates = 1`

was set in`ptmixed()`

and`nbmixed()`

- Released: October 2019
- Computation of starting values for
`ptmixed()`

and`nbmixed()`

improved - Added
`wald.test()`

function for computation of the multivariate Wald test - Added checks on
`maxit[1] == 0`

within`ptglm()`

and`nbglm()`

so as to make it possible to skip BFGS optimization and go straight to Nelder-Mead - Help files revised and improved
- Added extra checks in
`summary.ptglmm()`

and`summary.ptglm()`

(to verify that the smallest eigenvalue is not too small)

- Released: September 2019
- Added
`ptglm()`

function for the estimation of a Poisson-Tweedie GLM - Added
`nbmixed()`

and`nbglm()`

functions for the estimation of negative binomial GLMM and GLM using the Poisson-Tweedie parametrization (negative binomial: a = 0) - The package now comprises two classes:
`ptglmm`

for objects obtained from`ptmixed()`

and`nbmixed()`

, and`ptglm`

for objects obtained from`ptglm()`

and`nbglm()`

. Summary functions for objects of both classes have been implemented `min.var.init`

argument added to`ptmixed()`

- Released: July 2019
- Added function to compute the empirical Bayes estimates of the random intercept
- Class name of
`ptmixed()`

output changed from`ptmm`

to`ptglmm`

- Corrected typo in
`summary.ptglmm()`

function (the MLE of the dispersion parameter was wrongly called “deviance” instead of dispersion in the previous versions) - Added NEWS file

- Released: June 2019
- This is a major update aimed at speeding up the maximization of the loglikelihood. When
`ptmixed()`

is called, it first attempts to maximize the loglikelihood with the Nelder-Mead algorithm and then, if this fails, with the BFGS algorithm. Until version 0.0.4 the quadrature points were updated at every iteration for both Nelder-Mead and BFGS. Starting from this version, when Nelder-Mead is called it is possible to update the positioning of the quadrature points every*n*iterations by setting the`freq.updates`

argument equal to*n*. Default is set to`freq.updates = 200`

(this typically makes the optimization about 10 times faster than when`freq.updates = 1`

)

- Released: May 2019
`ptmixed()`

now outputs extra information (number of quadrature points used, initial values, warnings)- A mistake in the computation of the GH quadrature points was introduced from version 0.0.2. This has been fixed in this version

- Released: May 2019
- Fixed a typo in message on initial loglikelihood value that is displayed when
`trace = T`

in`ptmixed()`

function - Added exceptions and warnings for the case that
`maxit[1]`

and/or`maxit[2]`

are set = 0

- Released: Apr. 2019
- Function
`ptmixed()`

does not require the specification of a`time`

argument any more `maxit`

argument default value in function`ptmixed()`

increased to c(1e4, 100)- Internal function that computes starting values improved
- Added warning with indication that a simpler Poisson mixed model may fit the data sufficiently well
- Added warning when initial estimate of the variance parameter is < 0.001

- Released: Feb. 2019
- First version of the package