jm() now allows for zero-correlations constraints in the covariance matrix of the random effects. When the mixed models provided in the Mixed_objects argument have been fitted assuming a diagonal matrix for the random effects, this will also be assumed in the joint model (in previous versions, this was ignored). In addition, the new argument which_independent can be used to specify which longitudinal outcomes are to be assumed independent.
jm() can fit joint models with a combination of interval-censored data and competing risks (e.g., one of the the competing events is interval-censored and the other(s) not).
A bug in the predict() method causing low AUC values has been corrected.
The time-varying ROC and AUC now allow to correct for censoring in the interval Tstart to Thoriz using inverse probability of censoring weighting. The default remains model-based weights.
Portable implementation of parallel computing.
function area() has gained the argument time_window that specifies the window of integrating the linear predictor of the corresponding longitudinal outcome.
Function tvBrier() has gained the argument integrated for calculating the integrated Brier score.
Function tvBrier() has gained the argument type_weights and now also allows to correct for censoring in the interval Tstart to Thoriz using inverse probability of censoring weighting. The default remains model-based weights.
The new function tvEPCE() calculates the time-varying expected predictive cross-entropy.
This version supports Super Learning for optimizing predictions using cross-validation and a library of joint models. In that regard, the new function create_folds() can be used to split a dataset in V-folds of training and test datasets. More information can be found in the corresponding vignette.
Weak informative priors are now used for the fixed-effects of the mixed-effects models.
Several improvements in various internal functions.
Dynamic predictions for competing risks data can now be computed. An example is given in the Competing Risks vignette.
Function jm() can now fit joint models with a recurrent event process with or without a terminating event. The model accommodates discontinuous risk intervals, and the time can be defined in terms of the gap or calendar timescale. An example is given in the Recurrent Events vignette.
Added the function tvBrier() for calculating time-varying Brier score for fitted joint models. Currently, only right-censored data are supported.
Added the functions calibration_plot() and calibration_metrics() for calculating time-varying calibration plot and calibration metrics for fitted joint models. Currently, only right-censored data are supported.
Added new section in the vignette for Dynamic Prediction (available on the website of the package) to showcase the use of the functions mentioned above.
Improved the plot method for dynamic predictions.
Several bug corrections.
Added a predict() method for jm objects and a corresponding plot() for objects of class predict_jm for calculating and displaying predictions from joint models. Currently, only standard survival models are covered. Future versions will include predictions from competing risks and multi-state models.
Added the functions tvROC() and tvAUC() for calculating time-varying Receiver Operating Characteristic (ROC) curves and the areas under the ROC curves for fitted joint models. Currently, only right-censored data are supported.
Added a vignette (available on the website of the package) to explain how (dynamic) predictions are calculated in the package.
Added two vignettes (available on the website of the package) to showcase joint models with competing risks and joint models with non-Gaussian longitudinal outcomes.
Simplified syntax and additional options for specifying transformation functions of functional forms.
The slope() function has gained two new arguments, eps and direction. This allows calculating the difference of the longitudinal profile over a user-specified interval.
parallel::clusterSetRNGStream() in jm_fit() for distributing the seed in the workers.floor() in the C++ code.