Question

I’m working with some code that has crucial similarities to the following trivial script:

scores <- matrix(rnorm(4*20), ncol=4,nrow=20)
result <- matrix(NA, ncol=2, nrow=20)
index <- as.logical(rbinom(20,1,.2))
result[index, 1:3] <- cbind(1, scores[index,3:4])

where index is a logical vector and the sum(index) is typically greater than 1but can occasionally be 1 or 0.

The script fails in the case where sum(index) == 1:

> scores <- matrix(rnorm(4*20), ncol=4,nrow=20)
> result <- matrix(NA, ncol=3, nrow=20)
> index <- c(rep(FALSE, 19),TRUE)
> result[index, 1:3] <- cbind(1, scores[index,3:4])
Error in result[index, 1:3] <- cbind(1, scores[index, 3:4]) : 
  number of items to replace is not a multiple of replacement length
> cbind(1, scores[index,3:4])
     [,1]       [,2]
[1,]    1 -0.1780255
[2,]    1 -0.6840048
> #should be:
> c(1, scores[index,3:4])
[1]  1.0000000 -0.1780255 -0.6840048

and where sum(index) ==0:

> scores <- matrix(rnorm(4*20), ncol=4,nrow=20)
> result <- matrix(NA, ncol=3, nrow=20)
> index <- rep(FALSE, 20)
> result[index, 1:3] <- cbind(1, scores[index,3:4])
Warning message:
In cbind(1, scores[index, 3:4]) :
  number of rows of result is not a multiple of vector length (arg 1)
> #cbinding to a zero-row matrix returns an error

The obvious solution to this problem is the following:

scores <- matrix(rnorm(4*20), ncol=4,nrow=20)
result <- matrix(NA, ncol=3, nrow=20)
index <- as.logical(rbinom(20,1,.1))
if(sum(index) > 1){
    result[index, 1:3] <- cbind(1, scores[index,3:4])
}else{
    if(sum(index) ==1){
        result[index, 1:3] <- c(1, scores[index,3:4])
    }
}

However, I’m interested in advice on how to code to avoid this error without having to write a bunch of if statements. Is there a trick to binding an atomic vector to an nx2 matrix OR 2-length vector (n=1) such that the result is always an nx3 matrix? Extra points if the script can do this without producing an error when n=0.

If it weren’t for like an hour of debugging, I wouldn’t have identified this issue–it was buried quite a few functions down in a batch processing script. Any general advice on coding in a way of avoiding such ‘gotchas’?

Answer 1

Generally adding drop=FALSE to mtx[1,] calls will avoid the difficulties that arise with single row extractions and subsequent operations that assume a matrix structure:

result[index, 1:2] <- cbind(1, scores[1, 3:4, drop=FALSE])  # no error
# Also adding a third column to avoid dimension mismatch
scores <- matrix(rnorm(4*20), ncol=4,nrow=20)
result <- matrix(NA, ncol=3, nrow=20)
index <- as.logical(rbinom(20,1,.2))
result[index, 1:3] <- cbind(1, scores[index,3:4, drop=FALSE])

I haven’t quite figured out how you want us to do to avoid errors with assignment of zero row objects to zero row objects. You should instead check for length(index)==0

(The real problem is that you were assigning a three column matrix to a two column target. Oh, I see you tried to fix that, except you were still trying to assign to a third column that was not there in the dimensions.)