Dynamic prediction accuracy metrics for fitted joint models
DynPredAcc(
seed = 100,
object,
landmark.time = NULL,
horizon.time = NULL,
obs.time = NULL,
method = c("Laplace", "GH"),
quadpoint = NULL,
maxiter = NULL,
n.cv = 3,
quantile.width = 0.25,
opt = c("nlminb", "optim"),
LOCF = FALSE,
LOCFcovariate = NULL,
clongdata = NULL,
metrics = c("AUC", "Cindex", "Brier Score", "MAE", "MAEQ"),
cpu.cores = 1,
...
)An object of class DynPredAcc, returned as a list containing:
jm.classThe class of the fitted joint model.
n.cvThe number of cross-validation folds.
landmark.timeThe landmark time used for dynamic prediction.
horizon.timeThe vector of horizon times used for evaluation.
methodThe dynamic prediction approximation method used, if applicable.
quadpointThe number of quadrature points used, if applicable.
CompetingRiskLogical value indicating whether the fitted model accounts for competing risks.
seedThe random seed used for cross-validation.
metricsThe requested evaluation metrics.
quantile.widthThe width of quantile groups used for quantile-based calibration summaries.
AUC.cvA list of fold-specific time-dependent AUC estimates,
returned when "AUC" is requested.
Cindex.cvA list of fold-specific concordance index estimates,
returned when "Cindex" is requested.
Brier.cvA list of fold-specific inverse-probability-of-
censoring weighted Brier scores, returned when "Brier Score" is
requested.
MAE.cvA list of fold-specific inverse-probability-of-
censoring weighted mean absolute errors, returned when "MAE" is
requested.
MAEQ.cvA list of fold-specific quantile-based calibration
summaries, returned when "MAEQ" is requested.
A numeric value used to set the random seed for cross-validation.
Default is 100.
A fitted object of class jmcs, JMMLSM, or
mvjmcs.
A numeric value specifying the landmark time at which dynamic prediction begins.
A numeric vector of future times at which predicted probabilities are evaluated.
A character string specifying the longitudinal time variable in the longitudinal data.
A character string specifying the approximation method used for
dynamic prediction. Available options are "Laplace" and "GH".
This argument is not used for objects of class mvjmcs.
Number of Gauss--Hermite quadrature points used when
method = "GH". If NULL, the value stored in object is
used.
Maximum number of EM iterations allowed when refitting the
model within each cross-validation fold. If NULL, the model-specific
default is used.
Number of cross-validation folds. Default is 3.
Numeric value specifying the width of the quantile
groups used for "MAEQ" summaries. Default is 0.25. The
reciprocal of quantile.width must be an integer.
Optimization method used when refitting the model in each fold.
Available options are "nlminb" and "optim".
Logical value indicating whether the last-observation-carried-
forward approach is applied for time-dependent survival covariates during
prediction. Default is FALSE.
A character vector specifying the time-dependent
survival covariates to be updated by last observation carried forward when
LOCF = TRUE. Default is NULL.
A long-format data frame containing time-dependent survival
covariates used when LOCF = TRUE. Default is NULL.
A character vector specifying which prediction accuracy
metrics to compute. Available options are "AUC", "Cindex",
"Brier Score", "MAE", and "MAEQ". Default is
c("AUC", "Cindex", "Brier Score", "MAE", "MAEQ").
a number of cpu cores used for parallel computing.
Further arguments passed to model-specific methods.
Computes dynamic prediction accuracy measures for fitted joint models of
class jmcs, JMMLSM, or mvjmcs using grouped
cross-validation.
At each cross-validation fold, the function refits the joint model on the training set, obtains subject-specific dynamic predictions for validation subjects at the requested horizon times, and evaluates the requested prediction accuracy metrics.
Supported metrics include:
"AUC"Time-dependent area under the ROC curve.
"Cindex"Concordance index for dynamic predictions.
"Brier Score"Inverse-probability-of-censoring weighted Brier score.
"MAE"Inverse-probability-of-censoring weighted mean absolute error.
"MAEQ"Quantile-based calibration summaries comparing empirical and predicted cumulative incidence rates or survival probabilities.
For competing-risks models, event-specific cumulative incidence predictions are evaluated. For single-event models, conditional survival probabilities are evaluated.
The function performs grouped cross-validation at the subject level. Within each fold, the model is refit on the training data, and dynamic predictions are computed for validation subjects who remain under observation beyond the landmark time and who have longitudinal observations observed up to that time. For single-event models, prediction accuracy is based on conditional survival probabilities. For competing-risks models, prediction accuracy is based on event-specific cumulative incidence functions.
jmcs, JMMLSM, mvjmcs,
survfitjmcs, survfitJMMLSM, survfitmvjmcs