Question

I’m currently trying to build an LDA model on a dataset which contains some missing (NA) values. I want to, for example, impute the mean for NA values. From what I understand, I can set na.action=na.omit in the lda and predict functions which will remove the observations when building the model, and force return of NA when making predictions.

my.dat <- as.data.frame(cbind(
    c(0, 1, 0, 1, 1, 0),
    c(5, 8, 9, 1, -1, NA),
    c(-2.4, -4.0, -4.4, -0.5, 0.7, -0.3)
))
mod <- lda(my.dat[,-1], my.dat[,1], na.action=na.omit)
predict(mod, my.dat[,-1], na.action=na.omit)

But I want now to impute the means where I have an NA value. So, I can define my own na.impute function. But, I cannot understand what is passed to this function, and what I need to return.

na.impute <- function (object) {
    print(object)
    object
}

which gives me output:

[1] g x
<0 rows> (or 0-length row.names)

which doesn’t make much sense to me. I cannot find any guidance in the documentation. What exactly is object, and how am I supposed to manipulate it to overwrite NA values?

Answer 1

Here is the first way how to find out what is object:

na.impute <- function (object) {
  browser()
  print(object)
  object
}

lda(my.dat[,-1], my.dat[,1], na.action=na.impute)
# Called from: na.action(structure(list(g = grouping, x = x), class = "data.frame"))
Browse[1]> str(object)
# 'data.frame': 0 obs. of  2 variables:
#  $ g: num  0 1 0 1 1 0
#  $ x: matrix [1:6, 1:2] 5 8 9 1 -1 NA -2.4 -4 -4.4 -0.5 ...
#   ..- attr(*, "dimnames")=List of 2
#   .. ..$ : NULL
#   .. ..$ : chr  "V2" "V3"
Browse[1]> object$g
# [1] 0 1 0 1 1 0
Browse[1]> object$x
#      V2   V3
# [1,]  5 -2.4
# [2,]  8 -4.0
# [3,]  9 -4.4
# [4,]  1 -0.5
# [5,] -1  0.7
# [6,] NA -0.3
# attr(,"class")
# [1] "matrix"

So it really is an unusual object: structure(list(g = grouping, x = x), class = "data.frame"). Another way to see this, let us inspect function lda:

lda
# function (x, ...) 
# UseMethod("lda")
# <bytecode: 0x0e3583fc>
# <environment: namespace:MASS>
methods(lda)
# [1] lda.collapsed.gibbs.sampler lda.data.frame*             lda.default*               
# [4] lda.formula*                lda.matrix*                
# 
#    Non-visible functions are asterisked

In this case we are interested in lda.data.frame. Since it is asterisked we have to use either MASS:::lda.data.frame or getAnywhere("lda.data.frame") to see the source code:

function (x, ...) 
{
    res <- lda(structure(data.matrix(x), class = "matrix"), ...)
    cl <- match.call()
    cl[[1L]] <- as.name("lda")
    res$call <- cl
    res
}
<bytecode: 0x067c3248>
<environment: namespace:MASS>

Now we can see that lda.matrix is needed, so again using one of two functions:

function (x, grouping, ..., subset, na.action) 
{
    if (!missing(subset)) {
        x <- x[subset, , drop = FALSE]
        grouping <- grouping[subset]
    }
    if (!missing(na.action)) {
        dfr <- na.action(structure(list(g = grouping, x = x), 
            class = "data.frame"))
        grouping <- dfr$g
        x <- dfr$x
    }
    res <- lda.default(x, grouping, ...)
    cl <- match.call()
    cl[[1L]] <- as.name("lda")
    res$call <- cl
    res
}
<bytecode: 0x067bf7b8>
<environment: namespace:MASS>

And finally here we find a call of na.action which is what we expected. Now this is a function which replaces NA values with column means:

na.impute <- function (object) {
  temp <- object$x
  k <- which(is.na(temp), arr.ind = TRUE)
  temp[k] <- colMeans(temp, na.rm = TRUE)[k[, 2]]
  structure(list(g = object$g, x = as.matrix(temp)), class = "data.frame")
}
lda(my.dat[,-1], my.dat[,1], na.action=na.impute)
# Call:
# lda(my.dat[, -1], my.dat[, 1], na.action = na.impute)
#
# Prior probabilities of groups:
#   0   1 
# 0.5 0.5 
#
# Group means:
#         V2        V3
# 0 6.133333 -2.366667
# 1 2.666667 -1.266667
#
# Coefficients of linear discriminants:
#           LD1
# V2 -0.8155124
# V3 -1.1614265

Now considering predict and na.action it is unavailable option: see getAnywhere("predict.lda"), there is no usage of this argument.