Why does R 'sample' some columns more than others? - r

Question

I am testing the impact of missing data on regression analysis. So, using a simulated dataset, I want to randomly remove a proportion of observations (not entire rows) from a designated set of columns. I am using 'sample' to do this. Unfortunately, this is making some columns have much more missing values than others. See an example below:
#Data frame with 5 columns, 10 rows
DF = data.frame(A = paste(letters[1:10]),B = rnorm(10, 1, 10), C = rnorm(10, 1, 10), D = rnorm(10, 1, 10), E = rnorm(10,1,10))
#Function to randomly delete a proportion (ProportionRemove) of records per column, for a designated set of columns (ColumnStart - ColumnEnd)
RandomSample = function(DataFrame,ColumnStart, ColumnEnd,ProportionRemove){
#ci is the opposite of the proportion
ci = 1-ProportionRemove
Missing = sapply(DataFrame[(ColumnStart:ColumnEnd)], function(x) x[sample(c(TRUE, NA), prob = c(ci,ProportionRemove), size = length(DataFrame), replace = TRUE)])}
#Randomly sample column 2 - 5 within DF, deleting 80% of the observation per column
Test = RandomSample(DF, 2, 5, 0.8)
I understand there is an element of randomness to this, but in 10 trials (10*4 = 40 columns), 17 of the columns had no data, and in one trial, one column still had 6 records (rather than the expected ~2) - see below.
B C D E
[1,] NA 24.004402 7.201558 NA
[2,] NA NA NA NA
[3,] NA 4.029659 NA NA
[4,] NA NA NA NA
[5,] NA 29.377632 NA NA
[6,] NA 3.340918 -2.131747 NA
[7,] NA NA NA NA
[8,] NA 15.967318 NA NA
[9,] NA NA NA NA
[10,] NA -8.078221 NA NA
In summary, I want to replace a propotion of observations with NAs in each column.
Any help is greatly appreciated!!!

Answer 1

This makes sense to me. As #Frank suggested (in a since-deleted comment ... *sigh*), "randomness" can give you really non-random-looking results (Dilbert: Tour of Accounting, 2001-10-25).
If you want random samples with guaranteed ratios, try this:
guaranteedSampling <- function(DataFrame, ProportionRemove) {
n <- max(1L, floor(nrow(DataFrame) * ProportionRemove))
inds <- replicate(ncol(DataFrame), sample(nrow(DataFrame), size=n), simplify=FALSE)
DataFrame[] <- mapply(`[<-`, DataFrame, inds, MoreArgs=list(NA), SIMPLIFY=FALSE)
DataFrame
}
set.seed(2)
guaranteedSampling(DF[2:5], 0.8)
# B C D E
# 1 NA NA NA NA
# 2 NA NA NA NA
# 3 NA NA NA NA
# 4 6.792463 10.582938 NA NA
# 5 NA NA -0.612816 NA
# 6 NA -2.278758 NA NA
# 7 NA NA NA 2.245884
# 8 NA NA NA 5.993387
# 9 7.863310 NA 9.042127 NA
# 10 NA NA NA NA

Answer 2

Further to #joran's comment, you either wanted nrow(DataFrame) or length(x)
The specific impact in your example is that you are producing a vector with 5 elements (because DF has 5 variables) each with 0.8 probability of being NA and 0.2 of being TRUE.
Then this statement (which is what the sapply is doing to each column you specify and in this case I'm applying to DF$B only):
DF$B[sample(c(TRUE, NA), prob=c(0.2, 0.8), size = 5, replace=TRUE)]
does something that isn't immediately obvious to the uninitiated*. This:
sample(c(TRUE, NA), prob=c(0.2, 0.8), size = 5, replace=TRUE)
gives a logical vector, which when used to extract elements of a vector is silently recycled. So lets say you end up with:
NA TRUE NA TRUE NA
When you subset DF$B you end up getting this:
DF$B[c(NA, TRUE, NA, TRUE, NA, NA, TRUE, NA, TRUE, NA)]
Notice in your example how the top 5 numbers always follow the same pattern as the bottom 5 numbers. This explains why so many columns ended up being all NA, because there is a 0.32768 probability of getting 5 out of 5 NA which gets recycled to the whole column.
The other issue with your code is that the function doesn't actually do anything useful because you didn't specify any return value. Here it is corrected and cleaned up and using http://adv-r.had.co.nz/Style.html:
random_sample <- function(x, col_start, col_end, p) {
sapply(x[col_start:col_end],
function(y) y[sample(c(TRUE, NA), prob = c(1-p, p), size = length(y), replace = TRUE)])
}
*The uninitiated in this case includes me! I had no idea that logical vectors were recycled when used to extract until having a look at this question.