boot_scrnp: Compute the bootstrap-corrected estimator of SCRNP.

In nlpred: Estimators of Non-Linear Cross-Validated Risks Optimized for Small Samples

Description

This estimator is computed by re-sampling with replacement (i.e., bootstrap sampling) from the data. The SCRNP is computed for the learner trained on the full data. The SCRNP is then computed for the learner trained on each bootstrap sample. The average difference between the full data-trained learner and the bootstrap-trained learner is computed to estimate the bias in the full-data-estimated SCRNP. The final estimate of SCRNP is given by the difference in the full-data SCRNP and the estimated bias.

Usage

boot_scrnp(
  Y,
  X,
  B = 200,
  learner = "glm_wrapper",
  sens = 0.95,
  correct632 = FALSE,
  ...
)

Arguments

Y	A numeric vector of outcomes, assume to equal 0 or 1.
X	A data.frame of variables for prediction.
B	The number of bootstrap samples.
learner	A wrapper that implements the desired method for building a prediction algorithm. See ?glm_wrapper or read the package vignette for more information on formatting learners.
sens	The sensitivity constraint to use.
correct632	A boolean indicating whether to use the .632 correction.
...	Other options, not currently used.

Value

A list with $scrnp the bootstrap-corrected estimate of SCRNP and $n_valid_boot as the number of bootstrap of bootstrap samples where learner successfully executed.

Examples

# simulate data
X <- data.frame(x1 = rnorm(50))
Y <- rbinom(50, 1, plogis(X$x1))
# compute bootstrap estimate of scrnp for logistic regression
# use small B for fast run
boot <- boot_scrnp(Y = Y, X = X, B = 25, learner = "glm_wrapper")

nlpred documentation built on Feb. 24, 2020, 1:11 a.m.