Fast Stratified Sampling from a data.table like object

This function may be used to draw a stratified random sample (with or without replacement) of rows from a data.table or similar object. It additionally includes options on how to handle cases in which the number to be drawn is less than the number of actual rows when using replace=FALSE, which makes it useful in exact matching.

Usage

stratified_sample(data, n, strata, replace=FALSE,
                  max_replace=NULL, if_lt_n="stop")

Arguments

data: A data.table like object that should be sampled from. May also be any object that can be coerced to be a data.table, such as a data.frame or a tibble.
n: A named numeric vector containing values > 0, specifying the number of rows that should be sampled for each value in strata.
strata: A single character string specifying the integer, character or factor variable representing the strata that should be sampled from in data. Only supports a single variable. If data should be sampled from strata of multiple variables, the user needs to first create a single variable from these multiple variables as shown in the examples.
replace: Either TRUE or FALSE (default). Specifies whether the sampling should be performed with or without replacement.
max_replace: Either NULL (default) or a named numeric vector containing values > 0, specifying the number of times that rows in each value of strata may be replace when using replace=TRUE. Ignored if replace=FALSE. Keeping this argument at NULL results in no upper limit on replacement of rows per strata.
if_lt_n: Must be either "stop", "warn" or "nothing". Controls whether to throw an error, a warning or silently accept when the number of rows in a stratum specified by strata is smaller than the number that should be sampled from it and replace=FALSE is used.

Details

Internally splits the data.table into parts as defined by the strata variable and then uses the sample.int function to sample rows from these strata. By using only data.table functions and sample.int instead of sample, this function is very efficient even for very large data.tables and many possible strata.

Value

Returns a single data.table containing the sampled rows.

References

Lohr, Sharon L. (2010). Sampling: Design and Analysis. Bd. 2. Boston: Cengage Learning.

Author

Robin Denz

Examples

library(MatchTime)
library(data.table)
library(fastmatch)

set.seed(344)

data <- data.table(A=stats::rnorm(n=100),
                   B=sample(c("A", "B", "C"), size=100, replace=TRUE),
                   strat=sample(c("1", "2", "3"), size=100, replace=TRUE))

n <- c(10, 25, 31)
names(n) <- c("2", "1", "3")

# sample (without replacement):
# - 25 rows from strat=="1"
# - 10 rows from strat=="2"
# - 31 rows from strat=="3"
samp <- stratified_sample(data, n=n, strata="strat")

# take the same stratified sample with replacement
samp <- stratified_sample(data, n=n, strata="strat", replace=TRUE)

Fast Stratified Sampling from a `data.table` like object