Functions and methods for imputing missing values (NA) in tables and list patterned after the tidyverse approach of 'dplyr' and 'rlang'; works with data.tables as well.
Comprehensive Library For Handling Missing Values
tidyimpute is tidtverse/dplyr compliant toolkit for imputing missing
values (NA) values in list-like and table-like structures including data.tables.
It had two goals: 1) extend existing
na.* functions from the stats packages
and 2) provide dplyr/tidyverse compliant methods for tables and lists.
This package is based on the handy na.tools package which provides tools for working with missing values in vectors.
impute_*family of functions for table- or list-based imputations.
library(devtools) install_github( "decisionpatterns/tidyimport")
There are four types of imputation methods. They are distinguished by how the replacement values are calculated. Each is described below as well as describing each of the methods used.
In "constant" imputation methods, missing values are replaced by an
a priori selected constant value. The vector containingmissing values
is not used to calculate the replacement value. These take the form:
impute_constant- Impute with a constant
(Impute using function(s) of the target variable; When imputing in a table this is also called column-based imputation since the values used to derive the imputed come from the single column alone.)
In "univariate" replacement methods, values are calculated using
only the target vector, ie the one containing the missing values. The functions
for performing the imputation are nominally univariate summary functions.
Generally, the ordering of the vector does not affect imputed values. In general,
one value is used to replace all missing values (
NA) for a variable.
impute_median- median value
impute_quantile- quantile value
impute_sample- randomly sampled value via bootstrap.
Ordered Univariate (Coming Soon)
(Impute using function(s) of the target variable. Variable ordering relevant. This is a super class of the previous column-based imputation.)
In "ordered univariate" methods, replacement valuse are calculated from the vector that is assumed to be ordered. These types are very often used with time-series data. (Many of these functions are taken from or patterned after functions in the zoo package.)
impute_loess- loess smoother, assumes values are ordered
impute_locf- last observation carried forward, assumes ordered
impute_nocb- next observation carried backwards, assumes ordered
Multivariate (Coming Soon)
(Impute with multiple variables from the same observation. In tables, this is also called row-based imputation because imputed values derive from other measurement for the same observation. )
In "Multivariate" imputation, any value from the same row (observation) can be
used to derive the replacement value. This is generally implemented as a model
traing from the data with
var ~ ...
impute_predict- use a model
impute_by_group- use by-group imputation
Generalized (Coming Soon)
(Impute with column and rows.)
impute_restore- restore NAs to the vector; remembering replacement
impute_toggle- toggle between
NAand replacement values
tbl <- data.frame( col_1 = letters[1:3], col_2=c(1,NA_real_,3), col_3=3:1) impute( tbl, 2) impute_mean( tbl )
NA_explicit_as an exported constant for explicit categorical values.
na_explicit) to add levels for values if they do not already exist.
ifelsebecause of edge cases
ifelseand prevent recycling