SimDesign (version 2.0.1)

add_missing: Add missing values to a vector given a MCAR, MAR, or MNAR scheme

Description

Given an input vector, replace elements of this vector with missing values according to some scheme. Default method replaces input values with a MCAR scheme (where on average 10% of the values will be replaced with NAs). MAR and MNAR are supported by replacing the default FUN argument.

Usage

add_missing(y, fun = function(y, rate = 0.1, ...) rep(rate, length(y)), ...)

Arguments

y

an input vector that should contain missing data in the form of NA's

fun

a user defined function indicating the missing data mechanism for each element in y. Function must return a vector of probability values with the length equal to the length of y. Each value in the returned vector indicates the probability that the respective element in y will be replaced with NA. Function must contain the argument y, representing the input vector, however any number of additional arguments can be included

...

additional arguments to be passed to FUN

Value

the input vector y with the sampled NA values (according to the FUN scheme)

Details

Given an input vector y, and other relevant variables inside (X) and outside (Z) the data-set, the three types of missingness are:

MCAR

Missing completely at random (MCAR). This is realized by randomly sampling the values of the input vector (y) irrespective of the possible values in X and Z. Therefore missing values are randomly sampled and do not depend on any data characteristics and are truly random

MAR

Missing at random (MAR). This is realized when values in the dataset (X) predict the missing data mechanism in y; conceptually this is equivalent to \(P(y = NA | X)\). This requires the user to define a custom missing data function

MNAR

Missing not at random (MNAR). This is similar to MAR except that the missing mechanism comes from the value of y itself or from variables outside the working dataset; conceptually this is equivalent to \(P(y = NA | X, Z, y)\). This requires the user to define a custom missing data function

References

Sigal, M. J., & Chalmers, R. P. (2016). Play it again: Teaching statistics with Monte Carlo simulation. Journal of Statistics Education, 24(3), 136-156. 10.1080/10691898.2016.1246953

Examples

# NOT RUN {
set.seed(1)
y <- rnorm(1000)

## 10% missing rate with default FUN
head(ymiss <- add_missing(y), 10)

## 50% missing with default FUN
head(ymiss <- add_missing(y, rate = .5), 10)

## missing values only when female and low
X <- data.frame(group = sample(c('male', 'female'), 1000, replace=TRUE),
                level = sample(c('high', 'low'), 1000, replace=TRUE))
head(X)

fun <- function(y, X, ...){
    p <- rep(0, length(y))
    p[X$group == 'female' & X$level == 'low'] <- .2
    p
}

ymiss <- add_missing(y, X, fun=fun)
tail(cbind(ymiss, X), 10)

## missingness as a function of elements in X (i.e., a type of MAR)
fun <- function(y, X){
   # missingness with a logistic regression approach
   df <- data.frame(y, X)
   mm <- model.matrix(y ~ group + level, df)
   cfs <- c(-5, 2, 3) #intercept, group, and level coefs
   z <- cfs %*% t(mm)
   plogis(z)
}

ymiss <- add_missing(y, X, fun=fun)
tail(cbind(ymiss, X), 10)

## missing values when y elements are large (i.e., a type of MNAR)
fun <- function(y) ifelse(abs(y) > 1, .4, 0)
ymiss <- add_missing(y, fun=fun)
tail(cbind(y, ymiss), 10)

# }
# NOT RUN {
# }