Nonparametric Missing Value Imputation using Random Forest (ranger by default)

Description

The missForest package provides nonparametric missing-value imputation for mixed-type data (continuous and categorical). It models each variable with missingness using random forests that learn complex interactions and nonlinear relations and returns out-of-bag (OOB) error estimates. The default backend is ranger for speed and scalability, with an optional legacy randomForest backend for backward compatibility. Parallelization is supported either across variables (via foreach/doRNG) or within forests (via ranger threads).

Details

The main function is missForest, which iteratively imputes missing entries by fitting per-variable random forests to the currently imputed data matrix. The implementation now defaults to a ranger-based backend while preserving the original randomForest-based behavior via the backend argument. See missForest for arguments, details on stopping criteria, OOB error reporting (NRMSE for numeric variables and PFC for factors), and parallel options.

Author(s)

Daniel J. Stekhoven [aut, cre]

References

Stekhoven DJ, Bühlmann P (2012). “MissForest — nonparametric missing value imputation for mixed-type data.” Bioinformatics, 28(1), 112–118. doi:10.1093/bioinformatics/btr597.

See Also

missForest, mixError, prodNA

Nonparametric Missing Value Imputation using Random Forests (ranger or randomForest)

Description

missForest imputes missing values for mixed-type data (numeric and categorical). It models complex interactions and nonlinear relations and returns an out-of-bag (OOB) imputation error estimate. It supports parallel execution and offers two backends: ranger (default) and randomForest (legacy/compatibility).

Usage

missForest(xmis, maxiter = 10, ntree = 100, variablewise = FALSE,
           decreasing = FALSE, verbose = FALSE,
           mtry = floor(sqrt(ncol(xmis))), replace = TRUE,
           classwt = NULL, cutoff = NULL, strata = NULL,
           sampsize = NULL, nodesize = NULL, maxnodes = NULL,
           xtrue = NA, parallelize = c("no", "variables", "forests"),
           num.threads = NULL, backend = c("ranger", "randomForest"))

Arguments

Details

Algorithm. The method iteratively imputes each variable with missing values by fitting a random forest on the observed part of that variable and the current imputations of all other variables. After each iteration, the difference between the current and previous imputed matrices is computed separately for numeric and factor columns. The stopping rule is met once both differences have increased at least once (or only the present type increases if there is only one type). In that case, the previous imputation (before the increase) is returned. Otherwise, the process stops at maxiter.

Backends. With backend = "ranger", arguments are mapped as:

• ntree -> num.trees

• nodesize (numeric/factor) -> min.bucket for regression/classification, respectively (defaults used here are c(5, 1)).

• sampsize (counts) -> sample.fraction (overall or per-class fractions).

• classwt -> class.weights.

• cutoff: emulated via probability forests and post-thresholding.

• maxnodes: no direct equivalent in ranger (ignored).

The reported OOB error uses ranger's $prediction.error (MSE for numeric, error rate for factors), except when cutoff is used: in that case, the misclassification rate is computed by applying the cutoffs to OOB class probabilities.

Parallelization. Two modes are available via parallelize:

• "variables": different variables are imputed in parallel using foreach; per-variable ranger calls use num.threads = 1.

• "forests": a single variable’s forest is built using ranger multithreading (controlled by num.threads) or, for "randomForest", by combining sub-forests via foreach.

Make sure you have registered a parallel backend if you choose a parallel mode.

See the vignette for further examples and discussion.

Value

Author(s)

Daniel J. Stekhoven [aut, cre]

References

Stekhoven DJ, Bühlmann P (2012). “MissForest — nonparametric missing value imputation for mixed-type data.” Bioinformatics, 28(1), 112–118. doi:10.1093/bioinformatics/btr597.

See Also

mixError, prodNA, randomForest, ranger

Examples

## Mixed-type imputation on iris:
data(iris)
set.seed(81)
iris.mis <- prodNA(iris, noNA = 0.2)

## Default: ranger backend
imp_rg <- missForest(iris.mis, xtrue = iris, verbose = TRUE)
imp_rg$OOBerror
imp_rg$error  # requires xtrue

## Legacy behavior: randomForest backend
imp_rf <- missForest(iris.mis, backend = "randomForest", verbose = TRUE)

## Parallel examples (register a backend first, e.g., doParallel):
## 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)  # used by ranger
## End(Not run)

Compute Imputation Error for Mixed-type Data

Description

mixError computes imputation error for mixed-type data given the imputed matrix (ximp), the original matrix with missing values (xmis), and the complete ground truth (xtrue). It reports the normalized root mean squared error (NRMSE) for numeric variables and the proportion of falsely classified entries (PFC) for factor variables.

Usage

mixError(ximp, xmis, xtrue)

Arguments

Value

A named vector with the imputation error(s):

• NRMSE: normalized root mean squared error computed over the numeric entries that were missing in xmis.

• PFC: proportion of falsely classified entries computed over the factor entries that were missing in xmis.

If only one type (numeric or factor) is present among the missing entries, only the corresponding error is returned.

Note

Columns are treated by their R classes: numeric metrics are computed for numeric columns and classification metrics for factor columns. Character columns should be converted to factors beforehand.

This function is used internally by missForest when a complete matrix xtrue is supplied.

Author(s)

Daniel J. Stekhoven [aut, cre]

References

Stekhoven DJ, Bühlmann P (2012). “MissForest — nonparametric missing value imputation for mixed-type data.” Bioinformatics, 28(1), 112–118. doi:10.1093/bioinformatics/btr597.

For the NRMSE notion in imputation benchmarking: Oba S, Sato M, Takemasa I, Monden M, Matsubara K, Ishii S (2003). “A Bayesian missing value estimation method for gene expression profile data.” Bioinformatics, 19(16), 2088–2096.

See Also

missForest, nrmse

Examples

## Mixed-type error computation on iris:
data(iris)

## Introduce missingness:
set.seed(81)
iris.mis <- prodNA(iris, noNA = 0.2)

## Impute:
iris.imp <- missForest(iris.mis)

## Compute the true imputation error:
err.imp <- mixError(iris.imp$ximp, iris.mis, iris)
err.imp

Normalized root mean squared error

Description

nrmse computes the normalized root mean squared error (NRMSE) for a given complete data matrix xtrue, an imputed matrix ximp, and the corresponding matrix with missing values xmis.

Usage

nrmse(ximp, xmis, xtrue)

Arguments

Details

The NRMSE is computed over the entries that were missing in xmis and are numeric in xtrue / ximp, using

\mathrm{NRMSE} = \sqrt{\frac{\mathrm{mean}\{(X_{\mathrm{true}} - X_{\mathrm{imp}})^2\}} {\mathrm{var}(X_{\mathrm{true}})}}\,,

where \mathrm{mean} and \mathrm{var} are the empirical mean and variance computed over the continuous missing entries only. This measure is intended for continuous data; for categorical or mixed-type data, see mixError.

Value

A numeric scalar: the normalized root mean squared error.

Note

This function is used internally by mixError.

Author(s)

Daniel J. Stekhoven [aut, cre]

References

Oba S, Sato M, Takemasa I, Monden M, Matsubara K, Ishii S (2003). “A Bayesian missing value estimation method for gene expression profile data.” Bioinformatics, 19(16), 2088–2096.

See Also

mixError, missForest

Examples

## Simple numeric example
set.seed(1)
xtrue <- matrix(rnorm(100), ncol = 5)
xmis  <- xtrue
xmis[sample(length(xmis), 10)] <- NA
ximp  <- xmis
ximp[is.na(ximp)] <- rowMeans(ximp, na.rm = TRUE)[row(ximp)[is.na(ximp)]]

nrmse(ximp, xmis, xtrue)

Introduce Missing Values Completely at Random (MCAR)

Description

prodNA artificially introduces missing values by deleting entries completely at random (MCAR) up to a specified proportion.

Usage

prodNA(x, noNA = 0.1)

Arguments

Details

Missingness is introduced independently and uniformly over all cells, i.e., Missing Completely At Random (MCAR). No structure by row/column or variable type is imposed.

For reproducibility, call set.seed before prodNA.

Value

An object of the same base type as x (data frame or matrix) with approximately noNA proportion of its entries set to NA.

Author(s)

Daniel J. Stekhoven [aut, cre]

See Also

missForest, mixError

Examples

data(iris)

## Introduce 5% MCAR missingness into the iris data set:
set.seed(81)
iris.mis <- prodNA(iris, noNA = 0.05)
summary(iris.mis)

## Higher missingness:
set.seed(81)
iris.mis.20 <- prodNA(iris, noNA = 0.20)
mean(is.na(as.matrix(iris.mis.20)))

Extract Variable Types from a Data Frame

Description

varClass returns the variable types of a data frame. It is used internally in several functions of the missForest package.

Usage

varClass(x)

Arguments

Value

A character vector of length p, where p is the number of columns in x. Entries are "numeric" for continuous variables and "factor" for categorical variables.

Note

This function is used internally by missForest and mixError.

Author(s)

Daniel J. Stekhoven [aut, cre]

See Also

missForest, mixError, nrmse

Examples

data(iris)
varClass(iris)

## We have four continuous and one categorical variable.