Predictive mean matching (PMM) is an imputation technique introduced by Donald . Rubin in 1987. This imputation method aims to maintain the natural variability of the data and avoid implausible imputations that can occur with other univariate imputation methods.
Usage
pmean.match(
data,
family = "AUTO",
robust = FALSE,
k = 3,
char_to_factor = FALSE,
seed = NULL,
verbose = FALSE
)Arguments
- data
a numeric matrix or data frame of at least 2 columns.
- family
the distribution family of your observations. The family arguments defaults to 'AUTO'; and it will automatically select a distribution family (gaussian, binomial, multinomial) based on the type of variable (numeric or factor). The distribution family dictates the regression model used (lm,glm, multinom). However, the user can change the family argument to match his response variable distribution and the function will adapt to this input by using the generalized linear model or beta regression.
- robust
logical indicated whether to use robust estimation methods or ignore them. If set to 'TRUE', the function will make use of robust linear and generalized linear models to make its prediction.
- k
numeric vector indicating number of nearest neighbors to extract for imputation. Currently k defaults to 3 but can be changed.
- char_to_factor
transform character variable to unordered factor variable
- seed
numeric vector used for reproducible results. Used to sample the same predicted value over time.
- verbose
verbose error handling
Details
How's predictive mean matching different from conditional mean imputation(CMI)?
PMM is a combination of CMI and HDI. Predictive mean matching (PMM) uses regression on observed variables to estimate missing values, like CMI, however, PMM will also fill in the missing value by randomly sampling observed values whose predicted values are closest to the predicted values of the missing observation. This is currently done using the nearest neighbor approach with a set number of neighbors (3) but can be changed depending on your data.
Examples
set.seed(123)
data <- data.frame(x1 = stats::rnorm(100),x2 = stats::rnorm(100),y = stats::rnorm(100))
data$x1[sample(1:100, 20)] <- NA
data$x2[sample(1:100, 15)] <- NA
data$y[sample(1:100, 10)] <- NA
fact_dat <- data.frame(data, c = gl(5,20))
matched_data <- pmean.match(fact_dat, robust = TRUE)
summary(matched_data)
#> x1 x2 y c
#> Min. :-2.3092 Min. :-2.05325 Min. :-1.75653 1:20
#> 1st Qu.:-0.4136 1st Qu.:-0.80110 1st Qu.:-0.53131 2:20
#> Median : 0.2173 Median :-0.24669 Median : 0.07063 3:20
#> Mean : 0.2115 Mean :-0.08539 Mean : 0.12123 4:20
#> 3rd Qu.: 0.7904 3rd Qu.: 0.53973 3rd Qu.: 0.69013 5:20
#> Max. : 2.1873 Max. : 3.24104 Max. : 2.19881