Generates multi-paradigm predictions from a fitted joint_model:
dynamic survival probabilities from the Cox model, binary event probabilities from the
status model (both raw and calibrated), and expected duration from the time model.
Usage
# S3 method for class 'joint_model'
predict(object, new_data = NULL, eval_time = NULL, type = NULL, ...)Arguments
- object
A
joint_modelobject.- new_data
Optional new data frame to predict upon. If
NULL, predicts on training data.- eval_time
Horizon times for survival probabilities. Defaults to the model's
eval_time.- type
Optional single prediction type, to return only that column as a one-column tibble:
"survival","time"(or"numeric"),"prob"(calibrated status probability),"status","linear_pred","risk_score","stack_linear_pred"or"stack_risk_score".NULL(the default) or"all"returns every column.- ...
Not used; checked with
check_dots_empty.
Value
A tibble with columns:
- .pred_survival
Nested list of survival probability curves over
eval_time.- .pred_status
Predicted event probability from the status model.
- .pred_status_calibrated
Calibrated event probability (if calibration was enabled).
- .pred_time
Predicted duration from the time model.
- .pred_linear_pred
Linear predictor from the Cox model (higher = longer survival).
- .pred_risk_score
Relative hazard risk score from the Cox model.
- .pred_stack_linear_pred
Linear predictor from the stacked meta-learner (if
engine = "stacks").- .pred_stack_risk_score
Relative hazard risk score from the stacked meta-learner (if
engine = "stacks").
Details
Predictors are re-derived from new_data with forge
against the model's training blueprint, so factor levels and dummy columns match
training exactly, even if new_data does not exhibit every level.
The result has one row per row of new_data, also for start/stop data, and each
row is predicted from its own covariates, taken as constant from time 0. That is not how a
subject's survival is scored along their start/stop covariate path; see
cv_joint_model (argument covariates).