Integrated Brier Score of a Penalized Cox Model on Start/Stop Survival Data

Tunes an elastic-net Cox model on the analysis set of one resampling split and scores it on the assessment set by the integrated Brier score (IBS). Lower is better. A convenience wrapper around [cv_coxnet()] for data in counting-process (start/stop) layout, kept for existing analysis code; new code can call [cv_coxnet()] and [nested_cv_coxnet()] directly.

Usage

glmnet_IBS(
  object,
  alpha = 1,
  recipe,
  feature_names,
  time_data = NULL,
  formula = NULL,
  internal_folds = 5,
  id_col = "id",
  start_col = "tstart",
  stop_col = "tstop",
  status_col = "status",
  censoring_weights = c("ipcw", "none"),
  tune_engine = c("yardstick", "cv.glmnet"),
  prep_on = c("baseline", "all"),
  eval_time = NULL,
  metric = "brier_survival_integrated",
  rule = c("min", "1se"),
  covariates = c("path", "baseline"),
  failure_ibs = NA_real_,
  parallel = FALSE,
  parallel_over = c("resamples", "everything"),
  ...
)

Arguments

object: An `rsplit` (e.g. one element of `rsample` resamples).
alpha: Elastic net mixing parameter: 1 is lasso, 0 is ridge.
recipe: An unprepped [recipes::recipe()].
feature_names: Character vector of baked predictor column names, or a function that takes the baked analysis set and returns them – for recipes whose output columns vary by fold, such as `step_pca(threshold = )`.
time_data: Optional data frame with columns `start_col` and `stop_col`; its positive stop times are used as `eval_time` when that is `NULL`.
formula: Optional character string of `+`-separated feature names (or a one-sided formula) to restrict the model to a subset of `feature_names`. Defaults to all of them.
internal_folds: Number of subject-grouped folds used to choose the penalty.
id_col, start_col, stop_col, status_col: Column names of the subject identifier, interval start, interval stop, and event indicator (1 = event).
censoring_weights: Weighting scheme for Brier score calculation: `"ipcw"` (the default) uses inverse probability of censoring weighting via Graf weights (a proper scoring rule). `"none"` applies unweighted unit weights (1 for events and at-risk subjects, 0 for censored; legacy, emits a once-per-session warning; useful for historical reproducibility or sensitivity comparisons).
tune_engine: Cross-validation tuning engine: `"yardstick"` (default, uses subject-grouped folds via [cv_coxnet()]) or `"cv.glmnet"` (legacy, splits intervals across folds without subject grouping using [glmnet::cv.glmnet()]; emits a once-per-session warning, useful for historical reproducibility).
prep_on: `"baseline"` (default) preps `recipe` on each subject's first interval in the analysis set; `"all"` preps on every analysis row.
eval_time: Evaluation times for the IBS. Defaults to the stop times of `time_data`, or else deciles of the analysis-set event times.
metric: Name of the yardstick survival metric that chooses the penalty, e.g. `"brier_survival_integrated"` or `"concordance_survival"`.
rule: `"min"` or `"1se"`: report the penalty at `lambda.min` or `lambda.1se`.
covariates: `"path"` or `"baseline"`; see [cv_coxnet()].
failure_ibs: IBS reported when the model cannot be fit (default `NA`); a warning gives the reason.
parallel: Logical or character. When TRUE (or "resamples" / "everything"), evaluates internal CV folds in parallel using [furrr::future_map2()] for tune_engine = "yardstick" or [glmnet::cv.glmnet()]'s parallel mode. When FALSE (default), evaluates sequentially.
parallel_over: Tidymodels-compatible partition option: "resamples" (default) or "everything".
...: Further arguments passed to [glmnet::glmnet()], such as `cox.ties` (`"breslow"` or `"efron"`). glmnet 5.0 defaults to Breslow and 5.1 to Efron, so pass `cox.ties` to keep results stable across glmnet versions.

Value

A tibble with columns `IBS`, `lambda`, `term`, `estimate`, and `alpha`: one row per feature (including those the penalty set to 0). When the fit fails, a single row with `IBS = failure_ibs`, `lambda = NA`, and `alpha`.

Details

1. **Preprocessing.** `recipe` is prepped on the analysis set – only each subject's first interval when `prep_on = "baseline"` – and baked onto both sets. The recipe must keep `id_col`, `start_col`, `stop_col`, and `status_col` in its output (e.g. with an `"id variable"` role). 2. **Tuning.** [cv_coxnet()] chooses the penalty on the baked analysis set with `internal_folds` folds grouped by subject, by `metric` (the IBS by default). Earlier versions used [glmnet::cv.glmnet()], which chose the penalty by concordance on folds of rows, splitting subjects' intervals between folds. 3. **Scoring.** The model, refit on the whole analysis set, predicts each assessment-set subject's survival from the analysis-set Breslow baseline hazard along their covariate path (see [cv_coxnet()]), and is scored by [yardstick::brier_survival_integrated()] with inverse-probability-of- censoring (Graf) weights estimated on the analysis set.

References

Graf E, Schmoor C, Sauerbrei W, Schumacher M (1999). Assessment and comparison of prognostic classification schemes for survival data. *Statistics in Medicine*, 18(17-18), 2529-2545.

Examples

# \donttest{
if (requireNamespace("glmnet", quietly = TRUE) &&
    requireNamespace("yardstick", quietly = TRUE) &&
    requireNamespace("rsample", quietly = TRUE)) {
  set.seed(1)
  # Synthetic start/stop data: 60 subjects, up to 3 intervals each
  long <- do.call(rbind, lapply(1:60, function(i) {
    k <- sample(1:3, 1)
    stops <- c(5, 10, 20)[1:k]
    risk <- rnorm(1)
    data.frame(subject_id = i, tstart = c(0, head(stops, -1)), tstop = stops,
               status = c(rep(0, k - 1), rbinom(1, 1, plogis(risk))),
               x1 = risk + rnorm(k, sd = 0.2), x2 = rnorm(k))
  }))
  rec <- recipes::recipe(~ ., data = long) |>
    recipes::update_role(subject_id, tstart, tstop, status, new_role = "id variable") |>
    recipes::step_range(recipes::all_numeric_predictors())
  split <- rsample::group_initial_split(long, group = subject_id)
  glmnet_IBS(split, alpha = 0.5, recipe = rec, feature_names = c("x1", "x2"),
             time_data = unique(long[, c("tstart", "tstop")]), internal_folds = 3,
             id_col = "subject_id", cox.ties = "breslow")
}
#> # A tibble: 2 × 5
#>     IBS lambda term  estimate alpha
#>   <dbl>  <dbl> <chr>    <dbl> <dbl>
#> 1 0.112 0.0746 x1           0   0.5
#> 2 0.112 0.0746 x2           0   0.5
# }