Skip to contents

Sample covariates using multivariate imputation using chained equations (mice)

sample_covariates_mice.Rd

In contrast to sampling methods bootstrap and NHANES, categorical covariates need to be specified using the cat_covs argument, otherwise they will be treated as continuous. A warning is raised for numeric covariates that look like categorical codes (whole numbers with at most 3 distinct values, e.g. SEX = 0/1) but are not listed in cat_covs.

Usage

sample_covariates_mice(
  data,
  cat_covs = NULL,
  conditional = NULL,
  n_subjects = nrow(data),
  cont_method = c("pmm", "norm"),
  replicates = 1,
  seed = NULL,
  ...
)

Arguments

data

data.frame (n x p) containing the original, observed, time-invariant covariates (ID should not be included) that will be used to inform the imputation.

cat_covs

character vector containing the names of the categorical covariates in orgCovs.

conditional

list with conditional limits for sampled population. For continuous covariates, specify a numeric vector of length 2 giving the c(min, max) range, e.g. list("WT" = c(40, 60), "BMI" = c(15, 25)). For categorical covariates (those listed in cat_covs), specify a character vector of the allowed category values, e.g. list("SEX" = c("F")).

n_subjects

number of simulated subjects, default is the number of subjects in the data.

cont_method

method used to impute continuous covariates, either pmm (predictive mean matching, default) or norm (Bayesian linear regression).

replicates

number of multiple imputations replicates to sample. Default is 1.

seed

integer random seed passed to set.seed() for reproducibility. Default NULL does not set a seed.

...

additional arguments passed to the internal MICE implementation, e.g. maxit (number of iterations, default 5) or donors (number of donors for predictive mean matching, default 5).

Value

data.frame with the simulated covariates, with n_subjects * m rows and p columns

Details

missing values in data must be coded as NA.

The imputation uses a minimal, internal re-implementation of the subset of the mice package that is needed here (predictive mean matching or Bayesian linear regression for continuous covariates, logistic regression for binary and polytomous regression for other categorical covariates). This avoids a dependency on mice and its dependency tree.