Package {missknn}

missknn: Fast Masked K-Nearest Neighbor Imputation

Description

Fast masked k-nearest neighbor imputation for mixed-type tabular data. Distances are computed only on the variables jointly observed by the receiver and the donor, and missing values are filled by local regression (numeric) or weighted vote (categorical) over the nearest donors. Returns completed data for single imputation or a list of completed data sets for multiple imputation.

Author(s)

Maintainer: Imad El Badisy elbadisyimad@gmail.com

See Also

Useful links:

• https://CRAN.R-project.org/package=missknn

• Report bugs at https://github.com/ielbadisy/missknn/issues

Fast masked KNN imputation

Description

'missknn()' performs single imputation or multiple imputation for tabular data using masked k-nearest neighbor distances computed only on jointly observed variables.

Usage

missknn(
  data,
  k = 5L,
  m = 1L,
  scale = TRUE,
  add_indicator = FALSE,
  seed = NULL,
  weights = c("distance", "uniform"),
  numeric_estimator = c("regression", "mean"),
  ridge = 1e-04,
  donor_cap = 2000L,
  max_iter = 1L,
  parallel_cores = missknn_default_cores(),
  progress = FALSE
)

Arguments

Details

Let X = (x_{ij}) be an n \times p data matrix and let R_{ij} be the missingness indicator:

R_{ij} = 1 \; \text{if } x_{ij} \text{ is observed}, \qquad R_{ij} = 0 \; \text{if } x_{ij} \text{ is missing}.

For a receiver row 'i', a donor row 'l', and a target variable 't', the masked distance is computed on the shared observed set

S_{ilt} = O_i \cap O_l \setminus \{t\}.

The mixed-data masked distance is

d^2_t(i,l) = \frac{\sum_{j \in S_{ilt}} w_j \, delta_j(i,l)} {\sum_{j \in S_{ilt}} w_j},

where 'delta_j(i,l)' is squared difference for numeric variables and an indicator of inequality for categorical variables.

Single imputation uses a distance-weighted estimator from the 'k' nearest donors. Multiple imputation samples a donor from the same 'k' nearest donors using normalized inverse-distance probabilities.

'missknn()' is an imputer only: it returns completed data. Analysis and pooling of multiply imputed data (fitting a model on each completed data set and combining the fits with Rubin's rules) are left to dedicated packages such as 'mimar' (for example 'mimar::pool_lm()' on a list of 'lm()' fits). Tabular handling (input coercion, column classes, type restoration) uses the 'basetable' package. Only 'NA' cells are imputed.

Value

With 'm = 1', a completed 'data.frame' with the same columns and column types as 'data' (plus '<name>_missing' indicator columns when 'add_indicator = TRUE'). With 'm >= 2', a list of 'm' such data frames with class 'c("mimar_mi", "list")'. Either carries a '"tuning"' attribute: a data frame with one row per variable giving its number of 'missing' values, the tuned number of donors 'k', the 'estimator' ('"regression"', '"mean"', or '"vote"' for categorical variables) and whether the variable was filled from its global mean or mode ('global', no neighbor signal found).

Examples

set.seed(1)
x <- data.frame(
  a = c(1, 2, NA, 4, 5),
  b = c(2, NA, 3, 4, 6),
  g = factor(c("x", "x", "y", NA, "y"))
)
missknn(x, k = 2)

# multiple imputation: fit on each completed data set
imps <- missknn(x, k = 2, m = 3, seed = 1, parallel_cores = 1L)
fits <- lapply(imps, function(d) lm(a ~ b, data = d))
sapply(fits, coef)