IPCW Log-Loss and Integrated IPCW Log-Loss

Evaluates the two-term inverse-probability-of-censoring weighted log-loss. By default, censoring survival is estimated using marginal reverse Kaplan-Meier. Conditional or externally estimated censoring survival values can instead be supplied explicitly. Probability clipping is applied only inside logarithms.

Usage

eval_logloss(
  time,
  event,
  S_mat,
  times,
  tmin = min(times),
  tmax = max(times),
  ipcw_floor = 1e-06,
  eps = 1e-10,
  ipcw_cap = Inf,
  G_T_left = NULL,
  G_times = NULL
)

Arguments

time

Numeric vector of observed follow-up times.

event

Numeric vector of event indicators (1 = event, 0 = censored).

S_mat

A numeric matrix of predicted survival probabilities (rows = observations, columns = time points).

times

Numeric vector of evaluation times matching the columns of S_mat.

tmin

Numeric. Lower bound for IBS integration. Defaults to min(times).

tmax

Numeric. Upper bound for IBS integration. Defaults to max(times).

ipcw_floor

Positive lower bound applied to censoring survival before inversion.

eps

Probability clipping constant in (0, 0.5).

ipcw_cap

Largest IPCW contribution. Use Inf for no cap.

G_T_left

Optional numeric vector containing subject-specific \(G(T_i-\mid X_i)\) values. Supply together with G_times.

G_times

Optional numeric matrix with the dimensions of S_mat containing \(G(t\mid X_i)\), or a vector aligned with times for a common censoring survival curve. Supply together with G_T_left.

Value

A list containing logloss_scores, integrated_logloss, times, and diagnostics. Diagnostics report the effective denominator threshold, intervention counts, weight quantiles, effective sample sizes, and the fraction of weighted loss contributed by capped rows, separately for failure and survivor terms.