鼠标r包中执行随机森林时出错

Question

Here is just example data:

# generation of correlated data   
matrixCR <- matrix(NA, nrow = 100, ncol = 100)
diag(matrixCR) <- 1
matrixCR[upper.tri (matrixCR, diag = FALSE)] <- 0.5
matrixCR[lower.tri (matrixCR, diag = FALSE)] <- 0.5
matrixCR[1:10,1:10]
L = chol(matrixCR)# Cholesky decomposition
nvars = dim(L)[1]
nobs = 200
set.seed(123)
rM = t(L) %*% matrix(rnorm(nvars*nobs), nrow=nvars, ncol=nobs)
rM1 <- t(rM)
rownames(rM1) <- paste("S", 1:200, sep = "") 
colnames(rM1) <- paste("M", 1:100, sep = "")
# introducing missing value to the dataset 
N <- 2000*0.05 # 5% random missing values 
inds <- round ( runif(N, 1, length(rM1)) )
rM1[inds] <- NA


# using random forest implemented in mice package 
require(mice)
out.imp <- mice(rM1, m = 5, method ="rf")
imp.data <- complete(out.imp)

I am getting following error:

 iter imp variable
  1   1  M1  M2Error in apply(forest, MARGIN = 1, FUN = function(s) sample(unlist(s),  : 
  dim(X) must have a positive length

I am not sure what is causing this problem ?

Answer 1

As I mentioned in my comment, when the method is set to randomforest ( rf ), the mice function is throwing an error whenever it gets to a column with only a single NA value, but runs fine with any other number of NA values.

I checked with the package author and this appears to be a bug. Until it's fixed, you can choose a different imputation method for those columns with a single NA value. For example:

# Count number of NA in each column
NAcount = apply(rM1, 2, function(x) sum(is.na(x)))

# Create a vector giving the imputation method to use for each column. 
# Set it to "rf" unless that column has exactly one NA value.
method = rep("rf", ncol(rM1))
method[which(NAcount==1)] = "norm"

# Run the imputation with the new "method" selections
out.imp <- mice(rM1, m = 5, method = method)

I realize that for consistency you may want to use the same imputation method for all the columns, but the above gives you an option if you're set on using the randomforest method.