kfold.brmsfit: K-Fold Cross-Validation

Description

Perform exact K-fold cross-validation by refitting the model \(K\) times each leaving out one-\(K\)th of the original data. Folds can be run in parallel using the future package.

Usage

# S3 method for brmsfit
kfold(
  x,
  ...,
  K = 10,
  Ksub = NULL,
  folds = NULL,
  group = NULL,
  exact_loo = NULL,
  compare = TRUE,
  resp = NULL,
  model_names = NULL,
  save_fits = FALSE,
  recompile = NULL,
  future_args = list()
)

Value

kfold returns an object that has a similar structure as the objects returned by the loo and waic methods and can be used with the same post-processing functions.

Arguments

x: A brmsfit object.
...: Further arguments passed to brm.
K: The number of subsets of equal (if possible) size into which the data will be partitioned for performing \(K\)-fold cross-validation. The model is refit K times, each time leaving out one of the K subsets. If K is equal to the total number of observations in the data then \(K\)-fold cross-validation is equivalent to exact leave-one-out cross-validation.
Ksub: Optional number of subsets (of those subsets defined by K) to be evaluated. If NULL (the default), \(K\)-fold cross-validation will be performed on all subsets. If Ksub is a single integer, Ksub subsets (out of all K) subsets will be randomly chosen. If Ksub consists of multiple integers or a one-dimensional array (created via as.array) potentially of length one, the corresponding subsets will be used. This argument is primarily useful, if evaluation of all subsets is infeasible for some reason.
folds: Determines how the subsets are being constructed. Possible values are NULL (the default), "stratified", "grouped", or "loo". May also be a vector of length equal to the number of observations in the data. Alters the way group is handled. More information is provided in the 'Details' section.
group: Optional name of a grouping variable or factor in the model. What exactly is done with this variable depends on argument folds. More information is provided in the 'Details' section.
exact_loo: Deprecated! Please use folds = "loo" instead.
compare: A flag indicating if the information criteria of the models should be compared to each other via loo_compare.
resp: Optional names of response variables. If specified, predictions are performed only for the specified response variables.
model_names: If NULL (the default) will use model names derived from deparsing the call. Otherwise will use the passed values as model names.
save_fits: If TRUE, a component fits is added to the returned object to store the cross-validated brmsfit objects and the indices of the omitted observations for each fold. Defaults to FALSE.
recompile: Logical, indicating whether the Stan model should be recompiled. This may be necessary if you are running reloo on another machine than the one used to fit the model.
future_args: A list of further arguments passed to future for additional control over parallel execution if activated.

Details

The kfold function performs exact \(K\)-fold cross-validation. First the data are partitioned into \(K\) folds (i.e. subsets) of equal (or as close to equal as possible) size by default. Then the model is refit \(K\) times, each time leaving out one of the K subsets. If \(K\) is equal to the total number of observations in the data then \(K\)-fold cross-validation is equivalent to exact leave-one-out cross-validation (to which loo is an efficient approximation). The compare_ic function is also compatible with the objects returned by kfold.

The subsets can be constructed in multiple different ways:

If both folds and group are NULL, the subsets are randomly chosen so that they have equal (or as close to equal as possible) size.
If folds is NULL but group is specified, the data is split up into subsets, each time omitting all observations of one of the factor levels, while ignoring argument K.
If folds = "stratified" the subsets are stratified after group using loo::kfold_split_stratified.
If folds = "grouped" the subsets are split by group using loo::kfold_split_grouped.
If folds = "loo" exact leave-one-out cross-validation will be performed and K will be ignored. Further, if group is specified, all observations corresponding to the factor level of the currently predicted single value are omitted. Thus, in this case, the predicted values are only a subset of the omitted ones.
If folds is a numeric vector, it must contain one element per observation in the data. Each element of the vector is an integer in 1:K indicating to which of the K folds the corresponding observation belongs. There are some convenience functions available in the loo package that create integer vectors to use for this purpose (see the Examples section below and also the kfold-helpers page).

When running kfold on a brmsfit created with the cmdstanr backend in a different R session, several recompilations will be triggered because by default, cmdstanr writes the model executable to a temporary directory. To avoid that, set option "cmdstanr_write_stan_file_dir" to a nontemporary path of your choice before creating the original brmsfit (see section 'Examples' below).

Examples

Run this code

if (FALSE) {
fit1 <- brm(count ~ zAge + zBase * Trt + (1|patient) + (1|obs),
           data = epilepsy, family = poisson())
# throws warning about some pareto k estimates being too high
(loo1 <- loo(fit1))
# perform 10-fold cross validation
(kfold1 <- kfold(fit1, chains = 1))

# use the future package for parallelization
library(future)
plan(multiprocess)
kfold(fit1, chains = 1)

## to avoid recompilations when running kfold() on a 'cmdstanr'-backend fit
## in a fresh R session, set option 'cmdstanr_write_stan_file_dir' before
## creating the initial 'brmsfit'
## CAUTION: the following code creates some files in the current working
## directory: two 'model_.stan' files, one 'model_(.exe)'
## executable, and one 'fit_cmdstanr_.rds' file
set.seed(7)
fname <- paste0("fit_cmdstanr_", sample.int(.Machine$integer.max, 1))
options(cmdstanr_write_stan_file_dir = getwd())
fit_cmdstanr <- brm(rate ~ conc + state,
                    data = Puromycin,
                    backend = "cmdstanr",
                    file = fname)
# now restart the R session and run the following (after attaching 'brms')
set.seed(7)
fname <- paste0("fit_cmdstanr_", sample.int(.Machine$integer.max, 1))
fit_cmdstanr <- brm(rate ~ conc + state,
                    data = Puromycin,
                    backend = "cmdstanr",
                    file = fname)
kfold_cmdstanr <- kfold(fit_cmdstanr, K = 2)
}