Skip to contents

Add Censoring Weights to Survival Predictions for yardstick

Source: R/surv_helpers.R
add_graf_weights.Rd

Fills the `.weight_censored` column of each element of a `.pred` list-column with [graf_weights()], so the data can go straight to `yardstick::brier_survival()`, `brier_survival_integrated()`, or `roc_auc_survival()`. Use it to score any model's per-subject survival predictions, including models fit to start/stop data.

Usage

add_graf_weights(
  data,
  truth = ".truth",
  estimate = ".pred",
  censoring,
  trunc = 0.05
)

Arguments

data

A data frame with one row per subject.

truth

Column of `data` holding right-censored `Surv` truth.

estimate

List-column of data frames with `.eval_time` and `.pred_survival`, one data frame per row, all with the same evaluation times.

censoring

A [censoring_km()] object, estimated on the training data.

trunc

Lower bound for \(G\), capping weights at `1 / trunc` (default 0.05, as in tidymodels' `parsnip`).

Value

`data`, with `.weight_censored` added to (or replaced in) every data frame in `estimate`.

See also

[surv_subject_truth()], [censoring_km()]

Examples

if (requireNamespace("survival", quietly = TRUE) &&
    requireNamespace("yardstick", quietly = TRUE)) {
  train <- survival::Surv(c(2, 4, 6, 8), c(0, 1, 0, 1))
  scored <- tibble::tibble(
    .truth = survival::Surv(c(5, 7), c(1, 0)),
    .pred = list(
      tibble::tibble(.eval_time = c(3, 6), .pred_survival = c(0.9, 0.5)),
      tibble::tibble(.eval_time = c(3, 6), .pred_survival = c(0.8, 0.6))
    )
  )
  scored <- add_graf_weights(scored, censoring = censoring_km(train))
  yardstick::brier_survival(scored, truth = .truth, .pred)
}
#> # A tibble: 2 × 4
#>   .metric        .estimator .eval_time .estimate
#>   <chr>          <chr>           <dbl>     <dbl>
#> 1 brier_survival standard            3    0.0333
#> 2 brier_survival standard            6    0.38