Nested Cross-Validation for Joint Models — nested_cv_joint

Scores the joint_model pipeline on the outer splits of an nested_cv object. As in cv_joint_model, each outer split's model is fit on its analysis set and scored on its assessment set one row per subject, and the coxnet model of a start/stop outcome is scored along each subject's covariate path (covariates), as cv_coxnet does. The inner resamples are not used to fit anything: the penalty of each outer split's coxnet model is tuned by the internal cross-validation of joint_model() (unless penalty is given), and the status and time models choose theirs by cv.glmnet.

Usage

nested_cv_joint_model(
  object,
  outcome = survival::Surv(time, status) ~ .,
  subject_id = NULL,
  engine = c("glmnet", "baguette", "stacks"),
  calibration = TRUE,
  parallel = FALSE,
  covariates = c("path", "baseline"),
  check_subject_overlap = TRUE,
  tune_method = c("none", "race_anova", "race_win_loss"),
  checkpoint_dir = NULL,
  shard = NULL,
  checkpoint_mismatch = c("recompute", "overwrite", "error"),
  ...
)

Arguments

object: An rsample::nested_cv object.
outcome: Survival formula with a Surv() outcome.
subject_id: Subject identifier column. Required for start/stop outcomes.
engine: Modeling engine: "glmnet", "baguette", or "stacks" (the "baguette" models plus a stored Cox meta-learner whose output predict() adds as extra columns; see joint_model).
calibration: Logical; whether to calibrate status predictions (default TRUE).
parallel: Logical; whether to run outer splits in parallel.
covariates: "path" or "baseline"; see cv_joint_model.
check_subject_overlap: If TRUE (the default), stop when an outer split, or an inner resample, puts the same subject in both its analysis and its assessment set; see Details. FALSE skips the check.
tune_method: Method for tuning the Cox component penalty on the inner resamples: "none" (default, uses internal CV of joint_model()), "race_anova" (tune_race_anova racing), or "race_win_loss" (tune_race_win_loss racing, which needs BradleyTerry2). With racing, the penalty found on each outer split's inner resamples replaces any penalty given in ....
checkpoint_dir: Optional directory path where per-split checkpoints are stored.
shard: Optional shard specification (e.g. c(1, 2)) for distributed computation.
checkpoint_mismatch: Action on mismatch: "recompute", "overwrite", or "error".
...: Additional arguments passed to joint_model.

Value

An S3 object of class c("nested_cv_joint_model", "tbl_df"): outer_id, model, and ibs, one row per outer split and model.

Details

A counting-process (start/stop) outcome needs subject_id, and the outer splits and every inner resample are checked before anything is fit: if a subject is in both the analysis and the assessment set of one of them, the function stops, as cv_joint_model does for supplied resamples. Build object with grouped resampling at both levels, e.g. nested_cv(data, outside = group_vfold_cv(group = subject_id), inside = group_vfold_cv(group = subject_id)). Out-of-bag assessment sets of bootstraps pass. check_subject_overlap = FALSE skips the check.