rsample bootstraps() in R: Bootstrap Resampling Splits

The rsample bootstraps() function in R builds bootstrap resamples by drawing rows with replacement, so you can estimate model variance, build confidence intervals, or train ensemble models on small datasets.

bootstraps(df)                              # default 25 resamples
bootstraps(df, times = 1000)                # large bootstrap for CIs
bootstraps(df, strata = y)                  # stratified bootstrap
bootstraps(df, apparent = TRUE)             # add apparent (original) split
analysis(boots$splits[[1]])                 # in-bag rows of resample 1
assessment(boots$splits[[1]])               # out-of-bag (OOB) rows
set.seed(123); bootstraps(df, times = 100)  # reproducible resamples

What bootstraps() does

bootstraps() draws resamples of the same size as the input data, with replacement. It is part of the rsample package, the resampling engine in the tidymodels ecosystem. Each resample uses bootstrap sampling: rows are drawn at random with replacement from the original data, so the analysis set is the same length as the input but contains duplicate rows. On average about 63.2 percent of the original rows appear at least once; the remaining 36.8 percent form the out-of-bag (OOB) assessment set.

The function returns a tibble with a splits list-column of rsplit objects and an id column labelling each resample. It does not copy the underlying data. Each rsplit stores integer row indices into the original frame, so a thousand bootstrap resamples of a million-row table is still a small object. Use bootstrap resampling to estimate the sampling variance of a statistic, build percentile confidence intervals, or feed fit_resamples() when cross-validation is impractical on very small data.

Syntax and arguments

The signature has one required argument and five tuning knobs.

The arguments that matter in practice:

• data: the data frame or tibble to resample.
• times: the number of bootstrap resamples. The default 25 is fine for quick checks; use 1000 or more for percentile confidence intervals.
• strata: a column name. Resamples are drawn within each stratum so class or quantile proportions are preserved.
• breaks: when strata is numeric, the number of quantile bins used to stratify.
• apparent: if TRUE, appends one extra row to the result where both analysis and assessment sets are the full original data. Used by some bias correction methods.

bootstraps() examples

Basic 25-sample bootstrap

Call bootstraps() on a data frame and print the result to see the resample structure. The default times = 25 produces a tibble with 25 rows, one per resample.

Extract analysis and assessment sets

Each split has an analysis() set with duplicates and an assessment() set of out-of-bag rows. The two helpers return ordinary data frames you can feed to any modeling function.

Increase times for confidence intervals

Set times to a large number when you want a percentile confidence interval. A common rule of thumb is at least 1,000 resamples for a 95 percent interval and 10,000 for the tails of a 99 percent interval.

Stratify by a class column

Use strata to keep the outcome distribution balanced across every resample. Without stratification a small minority class can be absent from some bootstraps, breaking models that need all classes present.

bootstraps() vs other resampling functions

bootstraps() draws with replacement; the alternatives partition without replacement or sample with constraints. Pick the function that matches the estimator you need.

A typical bootstrap workflow looks like: bootstraps() returns the resamples, you fit a statistic or model on each analysis() set, then aggregate the per-resample statistics into a distribution. The OOB rows from assessment() are useful for honest accuracy estimates that mirror random forest OOB error.

Common pitfalls

Three mistakes account for most bootstraps() bugs.

• Forgetting set.seed(). bootstraps() shuffles and samples with replacement before partitioning. Without set.seed() before the call, every run produces different resamples and your bootstrap intervals drift. Set the seed in the same script, right before the resample.
• Using too few resamples for a confidence interval. The default times = 25 is meant for quick checks and fit_resamples() plumbing tests, not statistics. Percentile confidence intervals need at least 1,000 resamples; tail quantiles of a 99 percent interval need 10,000 or more.
• Treating duplicates as a bug. A bootstrap resample is supposed to contain repeated rows. sum(duplicated(analysis(split))) will be non-zero by design. The OOB assessment set is the unique held-out portion.

Try it yourself

Related rsample functions

bootstraps() is the workhorse for resampling with replacement; these functions extend or pair with it.

• analysis() and assessment(): extract the in-bag and out-of-bag data frames from a split object.
• initial_split(): carve off a final testing set before bootstrapping the training portion.
• vfold_cv(): build k-fold cross-validation splits when you need every row held out exactly once.
• group_bootstraps(): bootstrap whole groups together for clustered or hierarchical data.
• int_pctl(): compute percentile confidence intervals from a bootstrapped tibble of statistics.

See the rsample bootstraps reference on the tidymodels site for the full argument list.

FAQ

What is the default number of resamples in bootstraps()?

The default times argument is 25, which builds 25 bootstrap resamples. That is enough for quick sanity checks and for plumbing fit_resamples() calls, but it is far too few for a stable percentile confidence interval. For inference, use at least 1,000 resamples; for the tails of a 99 percent interval or for bias correction, use 5,000 to 10,000.

How do I make bootstraps() reproducible?

Call set.seed() with any integer immediately before bootstraps(). The function draws row indices with replacement using R's random number generator, so a fixed seed guarantees the same resamples every time the script runs. Keep the seed and the resample call together. Re-running with the same seed and the same R version reproduces the exact in-bag and out-of-bag assignments.

What is the difference between bootstraps() and vfold_cv()?

bootstraps() samples rows with replacement, so the analysis set is the same size as the original data and contains duplicates. vfold_cv() partitions rows without replacement, so every row appears in the held-out set exactly once across the v folds. Use bootstrap resampling for variance estimates and confidence intervals on small data, and use cross-validation for stable performance estimates during model tuning.

What does the apparent argument do?

Setting apparent = TRUE appends one extra row to the resulting tibble where both the analysis set and the assessment set are the full original data. Some bias correction methods, including the 0.632+ estimator, use this "apparent" split alongside the bootstrap resamples to compare in-sample fit to out-of-sample fit. Most everyday bootstrap workflows can leave it at the default FALSE.

Can bootstraps() handle time series data?

Not directly. bootstraps() ignores row order and samples uniformly with replacement, which breaks the temporal dependence a forecasting model needs. For chronological resamples use rolling_origin(). If you need a bootstrap-style estimator on time series, look at block bootstrap methods in the boot package or tsboot(), which preserve local autocorrelation by resampling contiguous blocks.