Survival Cutpoints at Several Landmark Times

For each landmark time, keeps the subjects still at risk (follow-up beyond the landmark), restarts their clock at the landmark, splits a feature into low and high groups and compares the survival of the two. One call replaces a copy of the analysis per landmark.

Usage

tbl_landmark_cutpoints(
  data,
  feature,
  time_col = "survival_time_min",
  status_col = "status",
  landmarks = c(15, 30, 60, 90, 120),
  method = c("surv_cutpoint", "median", "quantile"),
  probs = 0.75,
  min_prop = 0.1,
  id_col = NULL,
  measure_time_col = NULL
)

Arguments

data

A data frame, one row per subject, or long with `measure_time_col`.

feature

Numeric feature column.

time_col, status_col

Survival time (same unit as `landmarks`) and event indicator.

landmarks

Landmark times.

method

`"surv_cutpoint"`, `"median"` or `"quantile"`.

probs

Probability of the cut for `method = "quantile"` (default 0.75, so the top quarter is "high").

min_prop

Smallest share of subjects per group for `"surv_cutpoint"`.

id_col, measure_time_col

See Feature values; long data only.

Value

A tibble with one row per landmark: `landmark`, `n`, `events`, `cutpoint`, `n_low`, `n_high`, `hr`, `conf.low`, `conf.high`, `p.value`, `status` (`"ok"` or why the landmark was not analysed).

Feature values

By default `feature` is one value per subject (a baseline value). With `measure_time_col` (and `id_col`) the data are long: for each landmark the value used is each subject's last measurement at or before the landmark, so nothing measured after the landmark is used.

Cutpoints and p-values

`"surv_cutpoint"` picks the cutpoint with [km_cutpoint_search()] (smallest log-rank p-value among candidates with at least `min_prop` of the subjects in each group). That p-value is then optimistic, because the cut was chosen on the same data; use it for description, or validate the cut elsewhere. `"median"` and `"quantile"` (cut at `probs`) are not data-driven. The hazard ratio is high versus low, from a Cox model; the p-value is the log-rank test.

See also

[km_cutpoint_search()], [batch_cutpoint_survdiff()]

Examples

set.seed(1)
d <- data.frame(time = rexp(200, 1 / 80), status = rbinom(200, 1, 0.7), x = rnorm(200))
tbl_landmark_cutpoints(d, "x", "time", "status", landmarks = c(15, 30, 60), method = "median")
#> # A tibble: 3 × 11
#>   landmark     n events cutpoint n_low n_high    hr conf.low conf.high p.value
#>      <dbl> <int>  <int>    <dbl> <int>  <int> <dbl>    <dbl>     <dbl>   <dbl>
#> 1       15   177    122 -0.00831    89     88  1.16    0.811      1.66   0.412
#> 2       30   148    101 -0.0133     74     74  1.22    0.822      1.81   0.322
#> 3       60   107     71 -0.0250     54     53  1.15    0.717      1.85   0.557
#> # ℹ 1 more variable: status <chr>