Splitting Methods

library(splithalfr)

This vignette demonstrates the methods of splitting data that are supported by the splithalfr. Each splitting method is illustrated by calling by_split with the right arguments, printing to the terminal what data is in each of the two parts produced by a split. For a comprehensive review of each splitting method, see Pronk et al. (2021).

Example data

We’ll use this example dataset with eight trials of one participant, each trial having a condition and rt variable.

ds <- data.frame(
  participant = rep(1, 8),
  condition = rep(c("a", "b"), each = 4),
  rt = 100 * 1 : 8
)

First-second splitting

First-second splitting assigns trials of the first half of rows to one part and trials of the second half of rows to the other (Green et al., 2016; Webb, Shavelson, & Haertel, 1996; Williams & Kaufmann, 2012). For this splitting method, set method to first_second.

dummy = by_split(
  ds,
  ds$participant,
  method = "first_second",
  function(ds) { print(ds); },
  ncores = 1,
  verbose = F
)

Odd-even splitting

Odd-even splitting assigns trials with an odd row number to one part and trials with an even row number to the other (Green et al., 2016; Webb, Shavelson, & Haertel, 1996; Williams & Kaufmann, 2012). For this splitting method, set method to odd_even.

dummy = by_split(
  ds,
  ds$participant,
  method = "odd_even",
  function(ds) { print(ds); },
  ncores = 1,
  verbose = F
)

Permutated splitting

Permutated splitting is also known as random splitting (Kopp, Lange, & Steinke, 2021), bootstrapped splitting (Parsons, Kruijt, & Fox, 2019) and random sample of split halves (Williams & Kaufmann, 2012). It assigns trials to each part via random sampling without replacement. This splitting method is the default, but you can make it explicit by setting method to random. In practice, random splits are averaged over many replications, but for illustration we’re only printing one.

dummy = by_split(
  ds,
  ds$participant,
  method = "random",
  replications = 1,
  function(ds) { print(ds); },
  ncores = 1,
  verbose = F
)

Monte Carlo splitting

Monte Carlo splitting assigns trials to each part by sampling with replacement (Williams & Kaufmann, 2012). For constructing parts that are of any length, use the split_p argument and set replace to TRUE. The example below constructs two parts of the same length as the original dataset by setting split_p to 1.

dummy = by_split(
  ds,
  ds$participant,
  method = "random",
  replace = TRUE,
  split_p = 1,
  replications = 1,
  function(ds) { print(ds); },
  ncores = 1,
  verbose = F
)

Stratified splitting

If a split is stratified by a variable, then trials are separately assigned to each part for each level of that variable (Green et al., 2016). For example, if splits are stratified by ds$condition, the trials with condition a and b are split separately. Stratification can be used in combination with any of the methods above. For illustration we combine it with first-second splitting

dummy = by_split(
  ds,
  ds$participant,
  method = "first_second",
  stratification = ds$condition,
  function(ds) { print(ds); },
  ncores = 1,
  verbose = F
)

Subsampled splitting

In a subsampled split, a subset of the trials is randomly sampled without replacement and then split (see the supplementary materials of Hedge, Powell, & Sumner, 2018). Sub-sampling only works well with splitting methods that uses random sampling (permutated and Monte Carlo). Since the sub-sampling procedure already randomizes the trials selected for splitting, splitting methods that assign trials to part based on their row number, such as first-second and odd-even, should give results that are similar to permutated splitting. Any stratifications are applied both to the sub-sampling and splitting.

dummy = by_split(
  ds,
  ds$participant,
  method = "random",
  stratification = ds$condition,
  subsample_p = 0.5,
  function(ds) { print(ds); },
  ncores = 1,
  verbose = F
)