eval_bscore: Integrated Brier Score
In bcjaeger/xgboost.surv: Extreme Gradient Boosting for Survival Analysis

Description Usage Arguments Value Examples

View source: R/evaluate.R

Integrated Brier Score

1
2
3

eval_bscore(surv_probs, label = NULL, time = NULL, status = NULL,
  cens_data = NULL, cens_model = "cox", eval_times = NULL,
  scale = TRUE)

`surv_probs`	a matrix of survival probabilities
`label`	a numeric vector of class `sgb_label`
`time`	(needed iff `label` is unspecified) a numeric vector of time values. All `time` values must be greater than 0.
`status`	(needed iff `label` is unspecified) a numeric vector of status values. All `status` values should be 0 or 1, with 1 indicating that the event of interest occurred and 0 indicating censoring.
`cens_data`	(optional) a `data.frame` or `tibble` containing a selection of variables that will be used to model censoring probability. The inverse probability of censoring will then be used to weight each observation's contribution go the integrated Brier score.
`cens_model`	Method for estimating inverse probability of censoring weights: `cox`: A semi-parametric Cox proportional hazard model is fitted to the censoring times `nonpar`: Nonparametric extension of the Kaplan-Meier for the censoring times using symmetric nearest neighborhoods – available for arbitrary many strata variables on the right hand side of argument formula but at most one continuous variable (see prodlim and neighborhood) `aalen`: The nonparametric Aalen additive model fitted to the censoring times. Requires the timereg package. Note: if `cens_data` is unspecified, the Kaplan-Meier estimator will be used to estimate inverse probability of censoring weights
`eval_times`	A numeric vector giving times at which to compute Brier scores, which will then be used collectively to compute the integrated Brier score.
`scale`	(`TRUE`/`FALSE`) if `TRUE`, a marginal Kaplan-Meier prediction model will be fitted to the given (`time`, `status`) values and used as a reference Brier score. Then, a scaled Brier score will be computed as 1 - (model-IBS / reference-IBS)

a numeric value. If scale = TRUE, return values of 0 indicate a non-informative model and values of 1 indicate a perfect model. If scale = FALSE, return values of 0 indicate a perfect model, and return values greater than 0 are difficult to interpret.

# prodlim needs to be loaded for pec functions to work.
library(prodlim)
library(survival)

set.seed(329)
# predict 1/2 for everyone, all the time
surv_probs <- matrix(
  data = rep(1/2, 1000),
  nrow = 100,
  ncol = 10
)

# make random time & status values
time = runif(100, min = 1, max = 10)
status = c(rep(1, 50), rep(0, 50))

# use all the possible times
eval_times = seq(1, 10, length.out = 10)

# censor data has one variable that is
# related to time values
cens_data = data.frame(x1 = time + rnorm(100, sd = 1/2))

# note that the Brier score is 1/4, as expected, when
# no adjustment is applied for censoring.
eval_bscore(
  surv_probs = surv_probs,
  label = sgb_label(time, status),
  eval_times = eval_times,
  cens_data = cens_data,
  cens_model = 'marginal',
  scale = FALSE
)

# Now adjust for censoring:
eval_bscore(
  surv_probs = surv_probs,
  label = sgb_label(time, status),
  eval_times = eval_times,
  cens_data = cens_data,
  cens_model = 'cox',
  scale = FALSE
)