set.seed(15)
x <- sample(c("", LETTERS[1:10]), 25, TRUE, prob=c(.2, rep(.08, 10)))
split_match(x)
split_match(x, "C")
split_match(x, c("", "C"))
split_match(x, include = 0)
split_match(x, include = 1)
split_match(x, include = 2)
set.seed(15)
x <- sample(1:11, 25, TRUE, prob=c(.2, rep(.08, 10)))
split_match(x, 1)
Run the code above in your browser using DataLab