Bootstrap sampling

Generates bootstrap replicates. "simple" resamples rows independently; "block" is a moving-block bootstrap, which preserves the serial dependence in ordered data by concatenating ceiling(n / block_length) independently chosen blocks per replicate.

Usage

design_bootstrap(
  n_replicates = 1000L,
  n = NULL,
  method = c("simple", "block"),
  block_length = NULL
)

Arguments

n_replicates

Number of replicates to generate.

n

Rows per replicate. NULL uses nrow(data), the standard nonparametric bootstrap.

method

"simple" or "block".

block_length

Block length for method = "block", and an error for method = "simple", where it would have no effect. NULL uses floor(nrow(data)^(1/3)), a common rule of thumb.

Value

A design object, for use with draw().

Details

draw() returns all replicates in one data frame with a leading .replicate column. Split them with split(out, out$.replicate).

Not a survey design

The bootstrap resamples the data you already have to approximate the sampling distribution of a statistic (Efron 1979). It is not a probability sample of a finite population, so it has no inclusion probabilities and cannot be passed to ht_total(). For the design-based variance of a total or mean, use the estimators there; for a statistic they do not cover, such as a median, this is the tool. Blocks preserve short-range dependence in ordered data (Künsch 1989).

References

Efron, B. (1979). Bootstrap methods: another look at the jackknife. Annals of Statistics, 7, 1–26.

Künsch, H. R. (1989). The jackknife and the bootstrap for general stationary observations. Annals of Statistics, 17, 1217–1241.

See also

draw()

Other designs: design_certainty(), design_cluster(), design_multistage(), design_reservoir(), design_simple(), design_spatial(), design_spread(), design_stratified(), design_systematic(), design_temporal(), design_weighted()

Examples

df <- data.frame(id = 1:100, value = (1:100) / 10)

reps <- draw(df, design_bootstrap(n_replicates = 5, n = 20), seed = 1)
table(reps$.replicate)
#> 
#>  1  2  3  4  5 
#> 20 20 20 20 20 

# A statistic per replicate
vapply(split(reps, reps$.replicate), function(r) mean(r$value), numeric(1))
#>     1     2     3     4     5 
#> 5.450 5.285 4.395 4.735 5.570