Cox Screen of Longitudinal Summary Statistics

For each feature, summarises every subject's repeated measurements (see [transform_longitudinal_features()]) and fits one univariate Cox model per summary, with the summary as the only covariate. Answers "does the way a marker moves over time, not only its level, predict survival?".

Usage

tbl_longitudinal_cox_screen(
  data,
  id_col = "subject_id",
  time_col = "tstart",
  status_col = "status",
  features,
  metrics = c("pct_change_directional", "variance", "excursion", "sd"),
  surv_time_col = NULL,
  metadata = NULL,
  conf_level = 0.95
)

Arguments

data: A data frame in long format.
id_col, time_col: Subject and measurement-time columns.
status_col: Event indicator (1 = event, 0 = censored; or logical).
features: Numeric feature columns.
metrics: Any of `"pct_change_directional"`, `"variance"`, `"excursion"`, `"sd"`.
surv_time_col: Optional follow-up time column (see above).
metadata: Optional data frame with `columns` and optionally `labels` and `domain`; adds `label` and `domain` to the result.
conf_level: Confidence level of the hazard ratio.

Value

A tibble with `feature`, `label`, `domain` (only with `metadata`), `metric`, `n`, `events`, `hr`, `conf.low`, `conf.high`, `p.value`, `status` (`"ok"` or why no model was fitted).

Survival time and status

Each subject has one follow-up time and one status. `surv_time_col` gives the time; if it is `NULL`, the subject's largest `time_col` is used (the last measurement). The status is the one on the subject's last row (by `time_col`). A subject whose status is missing is dropped.

Metrics

`"pct_change_directional"`: percent change from the first to the last value, signed (positive = rose). `"variance"` and `"sd"`: within-subject. `"excursion"`: `max - min`, never negative. Hazard ratios are per one unit of the summary, so they depend on its scale; standardise the features first if the ratios should be compared across features.

Examples

set.seed(1)
d <- data.frame(id = rep(1:40, each = 4), t = rep(c(0, 5, 15, 30), 40))
d$hr <- rnorm(160, 80, 10)
d$status <- rep(rbinom(40, 1, 0.5), each = 4)
tbl_longitudinal_cox_screen(d, "id", "t", "status", "hr")
#> # A tibble: 4 × 9
#>   feature metric                n events    hr conf.low conf.high p.value status
#>   <chr>   <chr>             <int>  <int> <dbl>    <dbl>     <dbl>   <dbl> <chr> 
#> 1 hr      pct_change_direc…    40     19 1.01     0.975      1.04   0.742 ok    
#> 2 hr      variance             40     19 0.998    0.990      1.01   0.535 ok    
#> 3 hr      excursion            40     19 0.973    0.917      1.03   0.371 ok    
#> 4 hr      sd                   40     19 0.933    0.814      1.07   0.322 ok