missForest

missForest is a nonparametric imputation method for mixed-type tabular data in R. It handles numeric and categorical variables simultaneously by iteratively training random forests to predict missing entries from the observed ones. No explicit modeling assumptions, no matrix factorizations—just strong predictive baselines that work well out of the box.

• Works with any mix of numeric and factor columns
• Captures nonlinearities and interactions
• Reports out-of-bag (OOB) imputation error (NRMSE/PFC)
• Supports parallel execution (per-variable or per-forest)
• Two forest backends: ranger (default) and randomForest (legacy/compat)

The package also includes utilities to measure imputation error, generate missingness for experiments, and inspect variable types.

Installation

# CRAN (recommended)
install.packages("missForest")

# Development version (from GitHub)
# install.packages("remotes")
remotes::install_github("stekhoven/missForest")

Quick start

library(missForest)

# Example data
data(iris)

# Introduce ~20% MCAR missingness
set.seed(81)
iris_mis <- prodNA(iris, noNA = 0.20)

# Impute with default backend (ranger)
imp <- missForest(iris_mis, xtrue = iris, verbose = TRUE)

# Imputed data
head(imp$ximp)

# Estimated OOB errors (NRMSE for numeric, PFC for factors)
imp$OOBerror

# True error if xtrue was provided (for benchmarking only)
imp$error

Choosing a backend

# Legacy behavior using randomForest
imp_rf <- missForest(iris_mis, backend = "randomForest")

# Explicitly use ranger with limited threads
imp_rg <- missForest(iris_mis, backend = "ranger", num.threads = 2)

Parallelization

Two modes are available via parallelize:

• "variables": build forests for different variables in parallel (register a foreach backend).
• "forests": parallelize within a single variable’s forest (ranger threads; or foreach sub-forests for randomForest).

# Not run:
# library(doParallel)
# registerDoParallel(2)
# imp_vars <- missForest(iris_mis, parallelize = "variables", verbose = TRUE)
# imp_fors <- missForest(iris_mis, parallelize = "forests",  verbose = TRUE, num.threads = 2)

API overview

missForest(xmis, ...)

Core imputation function.

Key arguments:

• xmis — data frame/matrix with missing values (columns must be numeric or factor).
• maxiter — maximum iterations (default 10).
• ntree — trees per forest (default 100).
• mtry — variables tried at each split (default sqrt(p)).
• nodesize — length-2 numeric: minimum node size for c(numeric, factor). Default c(5, 1).
• variablewise — return per-variable OOB error if TRUE.
• parallelize — "no", "variables", or "forests".
• num.threads — threads for ranger (ignored by randomForest).
• backend — "ranger" (default) or "randomForest".
• xtrue — optional complete data for benchmarking (adds $error).

Some argument mappings for backend = "ranger":

• ntree → num.trees
• nodesize → min.bucket (separately for regression/classification; default c(5,1))
• sampsize (counts) → sample.fraction (fractions; overall or per-class)
• classwt → class.weights
• cutoff handled by fitting probability forests and post-thresholding

Utilities

• mixError(ximp, xmis, xtrue) — computes NRMSE (numeric) and PFC (factor) over true missing entries.
• nrmse(ximp, xmis, xtrue) — NRMSE for continuous-only data.
• prodNA(x, noNA = 0.1) — injects MCAR missingness into a data frame.
• varClass(x) — returns "numeric"/"factor" per column.

Tips & best practices

• Convert character columns to factors before calling missForest.

• For wide data, consider parallelize = "variables". For deep/expensive trees, consider parallelize = "forests".

• Set a seed for quasi-reproducible results:

  set.seed(123); imp <- missForest(x)

• You can lower ntree during prototyping to speed up iteration.

Citation

If you use missForest, please cite:

• Stekhoven, D. J. & Bühlmann, P. (2012). MissForest—nonparametric missing value imputation for mixed-type data. Bioinformatics, 28(1), 112–118. https://doi.org/10.1093/bioinformatics/btr597

You can also cite the package:

citation("missForest")

Contributing

Issues and pull requests are welcome. Please include a minimal reproducible example when reporting bugs. For performance discussions, share small benchmarks and session info.

License

GPL (≥ 2)

Contact

Daniel J. Stekhoven — stekhoven@stat.math.ethz.ch