Title: Multiple Imputation by Super Learning
Version: 1.0.0
Description: Performs multiple imputation of missing data using an ensemble super learner built with the tidymodels framework. For each incomplete column, a stacked ensemble of candidate learners is trained on a bootstrap sample of the observed data and used to generate imputations via predictive mean matching (continuous), probability draws (binary), or cumulative probability draws (categorical). Supports parallelism across imputed datasets via the future framework.
License: MIT + file LICENSE
URL: https://github.com/JustinManjourides/misl
BugReports: https://github.com/JustinManjourides/misl/issues
Encoding: UTF-8
RoxygenNote: 7.3.3
Depends: R (≥ 4.1.0)
Imports: dplyr (≥ 1.1.0), future.apply (≥ 1.11.0), parsnip (≥ 1.2.0), recipes (≥ 1.0.0), rsample (≥ 1.2.0), stacks (≥ 1.0.0), stats, tibble (≥ 3.2.0), tidyr (≥ 1.3.0), tune (≥ 1.2.0), utils, workflows (≥ 1.1.0)
Suggests: earth (≥ 5.3.0), future (≥ 1.33.0), knitr, ranger (≥ 0.16.0), rmarkdown, testthat (≥ 3.0.0), xgboost (≥ 1.7.0)
VignetteBuilder: knitr
Config/testthat/edition: 3
NeedsCompilation: no
Packaged: 2026-03-26 02:34:01 UTC; j.manjourides
Author: Justin Manjourides [aut, cre]
Maintainer: Justin Manjourides <j.manjourides@northeastern.edu>
Repository: CRAN
Date/Publication: 2026-03-30 18:10:02 UTC

Fit a stacked super learner ensemble

Description

Fit a stacked super learner ensemble

Usage

.fit_super_learner(
  train_data,
  full_data,
  xvars,
  yvar,
  outcome_type,
  learner_names,
  cv_folds = 5
)

Arguments

cv_folds

Integer number of cross-validation folds used when stacking multiple learners. Ignored when only a single learner is supplied.

Value

Named list with $boot (fit on bootstrap sample) and $full (fit on full observed data; NULL unless continuous).

Validate the input dataset before imputation

Description

Validate the input dataset before imputation

Usage

check_dataset(dataset)

Arguments

dataset

The object passed to misl().

Determine the outcome type of a column

Description

Determine the outcome type of a column

Usage

check_datatype(x)

Arguments

x

A vector (one column from the dataset).

Value

One of "categorical", "binomial", or "continuous".

List available learners for MISL imputation

Description

Displays the learners available for use in misl(), optionally filtered by outcome type and/or whether the required backend package is installed.

Usage

list_learners(outcome_type = "all", installed_only = FALSE)

Arguments

outcome_type

One of "continuous", "binomial", "categorical", or "all" (default).

installed_only

If TRUE, only learners whose backend package is already installed are returned. Default FALSE.

Value

A tibble with columns learner, description, package, installed, and outcome-type support flags (when outcome_type = "all").

Examples

list_learners()
list_learners("continuous")
list_learners("categorical", installed_only = TRUE)

MISL: Multiple Imputation by Super Learning

Description

Imputes missing values using multiple imputation by super learning.

Usage

misl(
  dataset,
  m = 5,
  maxit = 5,
  seed = NA,
  con_method = c("glm", "rand_forest", "boost_tree"),
  bin_method = c("glm", "rand_forest", "boost_tree"),
  cat_method = c("rand_forest", "boost_tree"),
  cv_folds = 5,
  ignore_predictors = NA,
  quiet = TRUE
)

Arguments

dataset

A dataframe or matrix containing the incomplete data. Missing values are represented with NA.

m

The number of multiply imputed datasets to create. Default 5.

maxit

The number of iterations per imputed dataset. Default 5.

seed

Integer seed for reproducibility, or NA to skip. Default NA.

con_method

Character vector of learner IDs for continuous columns. Default c("glm", "rand_forest", "boost_tree").

bin_method

Character vector of learner IDs for binary columns (values must be 0/1/NA). Default c("glm", "rand_forest", "boost_tree").

cat_method

Character vector of learner IDs for categorical columns. Default c("rand_forest", "boost_tree").

cv_folds

Integer number of cross-validation folds used when stacking multiple learners. Reducing this (e.g. to 3) speeds up computation at a small cost to ensemble accuracy. Default 5. Ignored when only a single learner is supplied.

ignore_predictors

Character vector of column names to exclude as predictors. Default NA.

quiet

Suppress console progress messages. Default TRUE.

Details

Supported *_method values and their required packages:

Use list_learners() to explore available options.

Value

A list of m named lists, each with:

datasets

A fully imputed tibble.

trace

A long-format tibble of mean/sd trace statistics per iteration, for convergence inspection.

Parallelism

Imputation across the m datasets is parallelised via future.apply. To enable parallel execution, set a future plan before calling misl(): 

library(future)
plan(multisession, workers = 4)
result <- misl(data, m = 5)
plan(sequential)

The inner cross-validation fits (used for stacking) run sequentially within each worker to avoid over-subscribing cores.

Examples

# Small self-contained example
set.seed(1)
n <- 100
demo_data <- data.frame(
  x1 = rnorm(n),
  x2 = rnorm(n),
  y  = rnorm(n)
)
demo_data[sample(n, 10), "y"] <- NA

misl_imp <- misl(demo_data, m = 2, maxit = 2, con_method = "glm")