mice.impute.rfcat {CALIBERrfimpute}

Impute categorical variables using Random Forest within MICE

Description

This method can be used to impute logical or factor variables (binary or >2 levels) in MICE by specifying method = 'rfcat'. It was developed independently from the mice.impute.rf algorithm of Doove et al., and differs from it in some respects.

Usage

mice.impute.rfcat(y, ry, x, ntree_cat = NULL,
    nodesize_cat = NULL, maxnodes_cat = NULL, ntree = NULL, ...)

Arguments

y

a logical or factor vector of observed values and missing values of the variable to be imputed.

ry

a logical vector stating whether y is observed or not.

x

a matrix of predictors to impute y.

ntree_cat

number of trees, default = 10.

A global option can be set thus: setRFoptions(ntree_cat=10).

nodesize_cat

minimum size of nodes, default = 1.

A global option can be set thus: setRFoptions(nodesize_cat=1). Smaller values of nodesize create finer, more precise trees but increase the computation time.

maxnodes_cat

maximum number of nodes, default NULL. If NULL the number of nodes is determined by number of observations and nodesize_cat.

ntree

an alternative argument for specifying the number of trees, over-ridden by ntree_cat. This is for consistency with the mice.impute.rf function.

...

other arguments to pass to randomForest.

Details

This Random Forest imputation algorithm has been developed as an alternative to logistic or polytomous regression, and can accommodate non-linear relations and interactions among the predictor variables without requiring them to be specified in the model. The algorithm takes a bootstrap sample of the data to simulate sampling variability, fits a set of classification trees, and chooses each imputed value as the prediction of a randomly chosen tree.

Value

A vector of imputed values of y.

Note

This algorithm has been tested on simulated data and in survival analysis of real data with artificially introduced missingness completely at random. There was slight bias in hazard ratios compared to polytomous regression, but coverage of confidence intervals was correct.

Author(s)

Anoop Shah

References

Shah AD, Bartlett JW, Carpenter J, Nicholas O, Hemingway H. Comparison of Random Forest and parametric imputation models for imputing missing data using MICE: a CALIBER study. American Journal of Epidemiology 2014; 179(6): 764–774. doi:10.1093/aje/kwt312 https://academic.oup.com/aje/article/179/6/764/107562

See Also

setRFoptions, mice.impute.rfcont, mice, mice.impute.rf, mice.impute.cart, randomForest

Examples

set.seed(1)

# A small sample dataset 
mydata <- data.frame(
    x1 = as.factor(c('this', 'this', NA, 'that', 'this')),
    x2 = 1:5,
    x3 = c(TRUE, FALSE, TRUE, NA, FALSE))
mice(mydata, method = c('logreg', 'norm', 'logreg'), m = 2, maxit = 2)
mice(mydata[, 1:2], method = c('rfcat', 'rfcont'), m = 2, maxit = 2)
mice(mydata, method = c('rfcat', 'rfcont', 'rfcat'), m = 2, maxit = 2)

# A larger simulated dataset
mydata <- simdata(100, x2binary = TRUE)
mymardata <- makemar(mydata)

cat('\nNumber of missing values:\n')
print(sapply(mymardata, function(x){sum(is.na(x))}))

# Test imputation of a single column in a two-column dataset
cat('\nTest imputation of a simple dataset')
print(mice(mymardata[, c('y', 'x2')], method = 'rfcat', m = 2, maxit = 2))

# Analyse data
cat('\nFull data analysis:\n')
print(summary(lm(y ~ x1 + x2 + x3, data = mydata)))

cat('\nMICE normal and logistic:\n')
print(summary(pool(with(mice(mymardata,
    method = c('', 'norm', 'logreg', '', ''), m = 2, maxit = 2),
    lm(y ~ x1 + x2 + x3)))))

# Set options for Random Forest
setRFoptions(ntree_cat = 10)

cat('\nMICE using Random Forest:\n')
print(summary(pool(with(mice(mymardata,
    method = c('', 'rfcont', 'rfcat', '', ''), m = 2, maxit = 2),
    lm(y ~ x1 + x2 + x3)))))

cat('\nDataset with unobserved levels of a factor\n')
data3 <- data.frame(x1 = 1:100, x2 = factor(c(rep('A', 25),
    rep('B', 25), rep('C', 25), rep('D', 25))))
data3$x2[data3$x2 == 'D'] <- NA
mice(data3, method = c('', 'rfcat'), m = 2, maxit = 2)