- new function
`externVar`

to perform a secondary regression analysis after the estimation of a primary latent class model - new argument
`pprior`

in hlme, lcmm, multlcmm and Jointlcmm to fix the probability to belong to each latent class - packages survival, parallel, mvtnorm, randtoolbox, marqLevAlg, doParallel, numDeriv are now listed in Imports rather than in Depends
- no subject-specific predictions in multlcmm with ordinal outcomes
- corrections in mpjlcmm
- correction in predictL without random effects
- correction in epoce and predictY.Jointlcmm
- use of R’s random number generator in Fortran codes
- use double precision rather than real(kind=8) in Fortran

- all vignettes excepted the introduction vignette (now renamed lcmm.Rmd) are removed from the CRAN version because of too long check time.
- We now provide a website at https://CecileProust-Lima.github.io/lcmm

- new vignette
`Joint latent class model with Jointlcmm`

- new vignette
`Multivariate latent class model with mpjlcmm`

- new argument
`pprior`

in the`hlme`

function - new argument
`computeDiscrete`

in the`lcmm`

function `mpjlcmm`

can be used with a mix of hlme/lcmm/multlcmm objects`summarytable`

and`summaryplot`

implement two versions of ICL criterion- new output
`levels`

in all estimating functions - new output
`varRE`

in`hlme`

- check the convergence of the initial model when using B=random()
- random parameters are generated with rmnvorm instead of using the Cholesky transformation
`permut`

,`cuminc`

,`VarCov`

,`coef`

,`vcov`

functions are available for mpjlcmm objects- corrections in
`mpjlcmm`

, especially with competing risks - correction in residuals for Jointlcmm models
- bug fixed when using
`posfix`

and`partialH`

simultaneously - correction in the likelihood for mutlcmm models
- bug fixed in
`predictClass`

and`predictRE`

when using splines - verbose=FALSE by default

- the model’s estimation is now available in parallel mode!
- The optimization relies on the parallelized marqLevAlg R package.
- models with latent classes (ng>1) require initial values
- the
`hlme`

function has now a pprior argument - the
`mpjlcmm`

function can be used without a time-to-event model - the
`summary`

functions now shorten the parameters names - the log-likelihood functions are now exported
- bug fixed in
`mpjlcmm`

when no random effect is included - bug fixed in
`Jointlcmm`

with Weibull hazards and competing risks - bug fixed in
`permut`

when used on`Jointlcmm`

objects with competing risks - correction of the outputs of
`multlcmm`

models

- the multlcmm function is now available for ordinal outcomes (link=“thresholds”) providing a longitudinal IRT model!
- new vignette
`Dynamic IRT with multlcmm`

- new dataset simdataHADS
- new function
`simulate`

to simulate a dataset from a hlme, lcmm, multlcmm or Jointlcmm model - new functions
`ItemInfo`

and`plot.ItemInfo`

to compute and plot Fisher information for ordinal outcomes - new argument
`var.time`

in the hlme, lcmm, multlcmm and Jointlcmm functions (used in plot(, which=“fit”); issue #91) - fix CRAN error with as.vector.data.frame
- correction in the
`permut`

function (transformation parameters were not updated) - add envir=parent.frame() in permut and gridsearch to enable the use of these functions in a parallel setting
- fix bug in the estimation functions with infinite posterior probabilities
- the
`gridsearch`

function now checks that the initial model converged (ie minit$conv=1) - the
`fixef`

and`ranef`

function are now imported from the nlme package

- new functions
`predictClass`

,`predictRE`

and`summaryplot`

- ICL computation in
`summaryplot`

- use of
`rmvnorm`

in`multlcmm`

to generate random initial values `maxiter`

is used in the estimation of the final model in`gridsearch`

- fix bug in
`cuminc`

without covariates - fix bug in the check for numeric type for argument
`subject`

with tibbles - fix bug in
`predictY`

with hlme object when the dataset is named “x” - fix bug in the
`update`

function when the model has unestimated parameters (posfix) - fix bug in
`hlme`

when posterior probabilities are NA - fix bug in
`plot`

with option which=“fit” (observations at the maximum time measurement where not systematically included) - correction in the outputs (ppi and resid) of the
`mpjlcmm`

function

- event variable in joint models can be logical
- bug fixed in
`Jointlcmm`

with prior when there are missing data - bug fixed in
`mpjlcmm`

: initial values were badly modified (with at least 3 dimensions) - small bugs fixed in
`predictY`

with median=TRUE

- parallel implementation of
`gridsearch`

function. Thanks to Raphael Peter for his suggestion. - add
`condRE_Y`

option in`predictYcond`

- add
`median`

options in`predictY`

- corrections in
`Jointlcmm`

,`multlcmm`

and`mpjlcmm`

when prior is specified - bugs fixed in some prediction functions
- small bugs fixed in the summary when some parameters are not estimated
- bug fixed in
`VarExpl`

with models including BM or AR - bug fixed in
`update.mpjlcmm`

(variance matrix was not correct) - manage infinite ppi in
`hlme`

- correction of epsY type, URL in vignettes, data statements position

- new function
`mpjlcmm`

for estimating joint latent class models with multiple markers and/or latent processes - various post-fit functions for
`mpjlcmm`

objects - new functions
`permut`

and`xclass`

- creation of vignettes, thanks to Samy Youbi for his help
- variable
`subject`

must be numeric - in plot(which=‘fit’), time intervals do not depend on subset
- add score test result in summarytable
- bug fixed in
`lcmm`

with prior - bug fixed in
`Jointlcmm`

with infinite score test - bug fixed in
`dynpred`

with TimeDepVar

- bug in summary when the model did not converge
- bug in dynpred when draws=TRUE and only 1 horizon or 1 landmark, or when o covariates are included in the survival model, or when using factor
- bug in Jointlcmm when using B=m1
- bug in plot.predictY with CI
- bug in Jointlcmm when B=random(m1)

- shades in plot.predictlink/L/Y
- subset in plot, which=“fit”

- Small bugs identified and solved in multlcmm

- Small bugs identified and solved in multlcmm, predictY and predictL

The package uses lazydata to automatically load the datasets of the package.

`jlcmm`

and`mlcmm`

are shortcuts for functions`Jointlcmm`

and`multlcmm`

, respectively.Function

`gridsearch`

provides an automatic grid of departures for reducing the odds of converging towards a local maximum.Initial values can be randomly generated from a model with 1 class (called m1 in next example) with option B=random(m1) in hlme, lcmm, multlcmm and Jointlcmm.

Functions

`hlme`

,`lcmm`

,`multlcmm`

,`Jointlcmm`

now include a posfix option to specify parameters that should not be estimated.Functions

`lcmm`

,`multlcmm`

,`Jointlcmm`

now include a partialH option to restrict the computation of the inverse of the Hessian matrix to a submatrixFunctions

`hlme`

,`lcmm`

,`multlcmm`

,`Jointlcmm`

now allow optional vector B to be an estimated model (with G=1) to reduce calculation time of initial values.Bug identified and solved in calculation of subject-specific predictions in

`hlme`

,`lcmm`

,`multlcmm`

and`Jointlcmm`

when cor is not NULL.Bug identified and solved in the calculation of confidence bands for individual dynamic predictions in dynpred with draws=T.

Bug identified and solved in the calculation of the explained variance for multlcmm objects when cor is not NULL.

Function plot now includes a which=“fit” option to plot observed and predicted trajectories stemming from a hlme, lcmm, Jointlcmm or multlcmm object.

Function

`predictlink`

replaces deprecated function`link.confint`

Function

`plot`

gathers deprecated functions`plot.linkfunction`

,`plot.baselinerisk`

,`plot.survival`

,`plot.fit`

together

The function

`Jointlcmm`

now allows competing risks data for the survival part and is also available for non-Gaussian longitudinal data. All existing methods for Jointlcmm objects (except EPOCE and Diffepoce functions) are adapted to the new framework.Functions

`link.confint`

,`plot.linkfunction`

,`predictL`

are now available for Jointlcmm objects.The new functions

`incidcum`

and`plot.incidcum`

respectively compute and plot the cumulative incidence associated to each competing event for Jointlcmm object.The new function

`fitY`

computes the marginal predicted values of longitudinal outcomes in their natural scale for lcmm or multlcmm objects.Bug identified and solved in

`dynpred`

function when used with a joint model assuming proportional hazards between latent classes.The Makevars file now allows compilation of the package with parallel make.

- bug solved regarding installation problem with parallel make

The new functions

`dynpred`

and`plot.dynpred`

respectively compute and plot individual dynamic predictions obtained from a joint latent class model estimated by Jointlcmm.The new function

`VarCovRE`

computes the standard errors of the parameters of variance-covariance of the random effects for a hlme, lcmm, Jointlcmm or multlcmm objectThe new function

`WaldMult`

computes multivariate Wald tests and Wald tests for combinations of parameters from hlme, lcmm, Jointlcmm or multlcmm objectThe new function

`VarExpl`

computes the percentages of variance explained by the linear regression for a hlme, lcmm, Jointlclmm or multlcmm objectThe new functions

`estimates`

and`VarCov`

get respectively all parameters estimated and their variance-covariance matrix for a hlme, lcmm, Jointlcmm or multlcmm objectFunction

`summary`

now returns the table containing the results about the fixed effects in the longitudinal modelAll plots consider now the … options

Functions plot.linkfunction and plot.predict have now an add argument

Function multlcmm now allows “splines” or “Splines” specification for the link functions

Functions

`lcmm`

and`multlcmm`

now compute the transformations even if the maximum number of iterations is reached without convergencebug identified and solved in multlcmm when the response variables are not integers

bug identified and solved in multlcmm when using contrast

bug identified and solved in plot.linkfunction for the y axes positions

bug identified and solved in hlme, lcmm, Jointlcmm and multlcmm when including interactions in

`mixture`

.

The new function

`multlcmm`

now estimates latent process mixed models for multivariate curvilinear longitudinal outcomes (with link functions: linear, beta or splines). Various post-fit computation and output functions are also available including plot.linkfunction, predictY, predictL, etcAll the functions hlme, lcmm, Jointlcmm include a

`cor`

option for including a brownian motion or a first-order autoregressive error process in addition to the independent errors of measurementbug identified and solved in predictL, predictY and plot.predict when used with factor covariate

- bug identified and solved in predictY.lcmm when used with a
`splines`

link function and an outcome with minimum value not at 0

The function

`predictY`

now computes the predicted values (possibly class-specific) of the longitudinal outcome not only from a lcmm object but also from a hlme or a Jointlcmm object for a specified profile of covariates.bug identified and solved in predictY.lcmm when used with a

`threshold`

link function and a Monte Carlo method

missing data handled in hlme, lcmm and Jointlcmm using

`na.action`

with attributes 1 for`na.omit`

or 2 for`na.fail`

The new function

`predictY.lcmm`

computes predicted values of a lcmm object in the natural outcome scale for a specified profile of covariates, and also provides confidence bands using a Monte Carlo method.bugs in epoce computation solved (with splines baseline risk function, and/or NaN values under solaris system)

bug identified and solved in summary functions regarding the labels of covariate effects in peculiar cases

improved variable specification in the estimating functions Jointlcmm, lcmm and hlme with

- categorical variables using factor()
- variables entered as functions using I()
- interaction terms using “*” and “:”

computation of the predictive accuracy measure EPOCE from a Jointlcmm object either on the training data or on external data (post-fit functions epoce and Diffepoce)

for discrete outcomes, lcmm function now computates the posterior discrete log-likelihood and the universal approximate cross-validation criterion (UACV)

Jointlcmm now includes two parameterizations of I-splines and piecewise-constant baseline risks functions to ensure positive risks: either log/exp or sqrt/square (option logscale=).