Balanced bootstrap resampling strategy

Usage

balanced_boot(k, split, type = c("DA", "R"), probs = NULL)

Arguments

k: Integer. Number of bootstrap resamples.
split: Numeric. Fraction of samples drawn for the training set (e.g. 2/3). Sampling is performed with replacement.
type: Character. Either "DA" (classification) or "R" (regression).
probs: Numeric vector of quantile probabilities used to stratify continuous Y when type = "R".

Value

A named list with elements:

train: List of integer vectors containing training set indices for each resampling iteration.
strategy: Character string indicating the resampling strategy.
n: Integer. Number of samples in the dataset.
seed: Integer. Random seed used to generate the resampling splits, ensuring reproducibility.

Details

Generates k bootstrap samples (training sets sampled with replacement). The remaining samples (the out-of-bag set) can be used as a test set.

Balancing ensures equal representation of strata in the training data:

type = "DA": Class labels define the strata, and sampling is balanced across classes.
type = "R": The response is discretised into bins using quantiles defined by probs, and each bin contributes equally to the training set.

Examples

n <- 100
# bivariate outcome
thr <- 1.5
Y <- c(rnorm(80, thr-3, 0.3), rnorm(20, thr+3, 0.3))  # unbalanced low/high outcome
mean(Y>thr)
#> [1] 0.2

cv_k <- kfold(k = 10)
cv_boot <- balanced_boot(k = 10, split = 2/3, type = "R", probs = c(0, 0.8, 1))

k_inst <- metabom8:::.arg_check_cv(cv_pars=cv_k, model_type='R', n=n, Y_prepped=cbind(Y))
b_inst <- metabom8:::.arg_check_cv(cv_pars=cv_boot, model_type='R', n=n, Y_prepped=cbind(Y))

# balanced splits: proportion above global median stays ~0.5
q80 <- quantile(Y, 0.8)
round(sapply(k_inst$train, function(i) mean(Y[i] > q80)), 2) # resembles original Y distr.
#>  [1] 0.20 0.22 0.21 0.22 0.17 0.18 0.21 0.20 0.20 0.19
round(sapply(b_inst$train, function(i) mean(Y[i] > q80)), 2) # balanced strata (low/high)
#>  [1] 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5