Become an expert in R — Interactive courses, Cheat Sheets, certificates and more!
Get Started for Free

impute_hotdeck

Hot deck imputation


Description

Hot-deck imputation methods include random and sequential hot deck, k-nearest neighbours imputation and predictive mean matching.

Usage

impute_rhd(
  dat,
  formula,
  pool = c("complete", "univariate", "multivariate"),
  prob,
  backend = getOption("simputation.hdbackend", default = c("simputation", "VIM")),
  ...
)

impute_shd(
  dat,
  formula,
  pool = c("complete", "univariate", "multivariate"),
  order = c("locf", "nocb"),
  backend = getOption("simputation.hdbackend", default = c("simputation", "VIM")),
  ...
)

impute_pmm(
  dat,
  formula,
  predictor = impute_lm,
  pool = c("complete", "univariate", "multivariate"),
  ...
)

impute_knn(
  dat,
  formula,
  pool = c("complete", "univariate", "multivariate"),
  k = 5,
  backend = getOption("simputation.hdbackend", default = c("simputation", "VIM")),
  ...
)

Arguments

dat

[data.frame], with variables to be imputed and their predictors.

formula

[formula] imputation model description (see Details below).

pool

[character] Specify donor pool when backend="simputation"

  • "complete". Only records for which the variables on the left-hand-side of the model formula are complete are used as donors. If a record has multiple missings, all imputations are taken from a single donor.

  • "univariate". Imputed variables are treated one by one and independently so the order of variable imputation is unimportant. If a record has multiple missings, separate donors are drawn for each missing value.

  • "multivariate". A donor pool is created for each missing data pattern. If a record has multiple missings, all imputations are taken from a single donor.

prob

[numeric] Sampling probability weights (passed through to sample). Must be of length nrow(dat).

backend

[character] Choose the backend for imputation. If backend="VIM" the variables used to sort the data (in case of sequential hot deck) may not coincide with imputed variables.

...

further arguments passed to VIM::hotdeck if VIM is chosen as backend, otherwise they are passed to

  • order for impute_shd and backend="simputation"

  • VIM::hotdeck for impute_shd and impute_rhd when backend="VIM".

  • VIM:kNN for impute_knn when backend="VIM"

  • The predictor function for impute_pmm.

order

[character] Last Observation Carried Forward or Next Observarion Carried Backward. Only for backend="simputation"

predictor

[function] Imputation to use for predictive part in predictive mean matching. Any of the impute_ functions of this package (it makes no sense to use a hot-deck imputation).

k

[numeric] Number of nearest neighbours to draw the donor from.

Model specification

Formulas are of the form

IMPUTED_VARIABLES ~ MODEL_SPECIFICATION [ | GROUPING_VARIABLES ]

The left-hand-side of the formula object lists the variable or variables to be imputed. The interpretation of the independent variables on the right-hand-side depends on the imputation method.

  • impute_rhd Variables in MODEL_SPECIFICATION and/or GROUPING_VARIABLES are used to split the data set into groups prior to imputation. Use ~ 1 to specify that no grouping is to be applied.

  • impute_shd Variables in MODEL_SPECIFICATION are used to sort the data. When multiple variables are specified, each variable after the first serves as tie-breaker for the previous one.

  • impute_knn The predictors are used to determine Gower's distance between records (see gower_topn). This may include the variables to be imputed..

  • impute_pmm Predictive mean matching. The MODEL_SPECIFICATION is passed through to the predictor function.

If grouping variables are specified, the data set is split according to the values of those variables, and model estimation and imputation occur independently for each group.

Grouping using dplyr::group_by is also supported. If groups are defined in both the formula and using dplyr::group_by, the data is grouped by the union of grouping variables. Any missing value in one of the grouping variables results in an error.

Methodology

Random hot deck imputation with impute_rhd can be applied to numeric, categorical or mixed data. A missing value is copied from a sampled record. Optionally samples are taken within a group, or with non-uniform sampling probabilities. See Andridge and Little (2010) for an overview of hot deck imputation methods.

Sequential hot deck imputation with impute_rhd can be applied to numeric, categorical, or mixed data. The dataset is sorted using the ‘predictor variables’. Missing values or combinations thereof are copied from the previous record where the value(s) are available in the case of LOCF and from the next record in the case of NOCF.

Predictive mean matching with impute_pmm can be applied to numeric data. Missing values or combinations thereof are first imputed using a predictive model. Next, these predictions are replaced with observed (combinations of) values nearest to the prediction. The nearest value is the observed value with the smallest absolute deviation from the prediction.

K-nearest neighbour imputation with impute_knn can be applied to numeric, categorical, or mixed data. For each record containing missing values, the k most similar completed records are determined based on Gower's (1977) similarity coefficient. From these records the actual donor is sampled.

Using the VIM backend

The VIM package has efficient implementations of several popular imputation methods. In particular, its random and sequential hotdeck implementation is faster and more memory-efficient than that of the current package. Moreover, VIM offers more fine-grained control over the imputation process then simputation.

If you have this package installed, it can be used by setting backend="VIM" for functions supporting this option. Alternatively, one can set options(simputation.hdbackend="VIM") so it becomes the default.

Simputation will map the simputation call to a function in the VIM package. In particular:

  • impute_rhd is mapped to VIM::hotdeck where imputed variables are passed to the variable argument and the union of predictor and grouping variables are passed to domain_var. Extra arguments in ... are passed to VIM::hotdeck as well. Argument pool is ignored.

  • impute_shd is mapped to VIM::hotdeck where imputed variables are passed to the variable argument, predictor variables to ord_var and grouping variables to domain_var. Extra arguments in ... are passed to VIM::hotdeck as well. Arguments pool and order are ignored. In VIM the donor pool is determined on a per-variable basis, equivalent to setting pool="univariate" with the simputation backend. VIM is LOCF-based. Differences between simputation and VIM likely occurr when the sorting variables contain missings.

  • impute_knn is mapped to VIM::kNN where imputed variables are passed to variable, predictor variables are passed to dist_var and grouping variables are ignored with a message. Extra arguments in ... are passed to VIM::kNN as well. Argument pool is ignored. Note that simputation adheres stricktly to the Gower's original definition of the distance measure, while VIM uses a generalized variant that can take ordered factors into account.

By default, VIM's imputation functions add indicator variables to the original data to trace what values have been imputed. This is switched off by default for consistency with the rest of the simputation package, but it may be turned on again by setting imp_var=TRUE.

References

Andridge, R.R. and Little, R.J., 2010. A review of hot deck imputation for survey non-response. International statistical review, 78(1), pp.40-64.

Gower, J.C., 1971. A general coefficient of similarity and some of its properties. Biometrics, pp.857–871.

See Also

Other imputation: impute_cart(), impute_lm(), impute()


simputation

Simple Imputation

v0.2.6
GPL-3
Authors
Mark van der Loo [aut, cre]
Initial release

We don't support your browser anymore

Please choose more modern alternatives, such as Google Chrome or Mozilla Firefox.