rbind(mids, not-mids) and complete() about mice HOT 9 CLOSED

I think revising the assignment to where in rbind.mids above to:

where <- rbind(x$where, is.na(y))

would suffice to fix this issue

from mice.

I had a similar issue.
I think the issue is that:

print(dim(D_rbind$data))
[1] 10 2
print(dim(D_rbind$where))
[1] 5 2

But I disagree with the solution you are proposing. I would instead suggest something like

where <- rbind(x$where, matrix(FALSE, ncol=ncol(y), nrow=nrow(y)) )

The reason being that when we use rbind.mids, I think the object 'y' should be left untouched (also the missing in y)

from mice.

What I wanted to say is that the dimensions of D_rbind$data and D_rbind$where are not consistent. I think this is what is causing issues then with the mice:::complete() function.

from mice.

I don't mind either way, I assumed that the where variable was supposed to indicate where missing data is located in the dataset, hence my proposal. Note that the assignment to imp in the rbind.mids function already ensures that no imputation is performed on the y dataset regardless of the value of where.

And yep, I agree regarding your reasoning for the misbehaviour in complete().

from mice.

I think the the 'where' determine how 'complete' will fills the missing.

If you run the code below, then the issue you describe is solved.

x <- rnorm(10)
D <- data.frame(x=x, y=2*x+rnorm(10))
D[2:4, 1] <- NA

D_mids <- mice(D[1:5,])

y <- D[6:10,]
D_rbind <- mice:::rbind.mids(x=D_mids, y=y)
D_rbind$where <- rbind(D_rbind$where, matrix(FALSE, ncol=ncol(y), nrow=nrow(y)) )

mice:::complete(D_rbind, 1, include=T)
print(D)

One issue which I still have is that if I introduce some NAs in the 'y' data, it fails:

x <- rnorm(10)
D <- data.frame(x=x, y=2*x+rnorm(10))
D[2:4, 1] <- NA

D_mids <- mice(D[1:5,])

y <- D[6:10,]
y[4:5,1] <- NA
D_rbind <- mice:::rbind.mids(x=D_mids, y=y)
D_rbind$where <- rbind(D_rbind$where, matrix(FALSE, ncol=ncol(y), nrow=nrow(y)) )

mice:::complete(D_rbind, 1, include=T)

with the error below:
Error in [<-.data.frame(*tmp*, where[, j], j, value = c(0.700183624923754, :
replacement has 5 rows, data has 3

from mice.

Using is.na(y) in assignment to where will work in both cases you describe, without error.

from mice.

Thanks. Just pushed a fix for this.

from mice.

All good, I get why where=F assignment is the correct approach. Just FYI - this means that in rbind.mids() the following code

   # The original data of y will be copied into the multiple imputed dataset, including the missing values of y.
    r <- !wy
    imp <- vector("list", ncol(y))
    for (j in visitSequence) {
      imp[[j]] <- 
        rbind(x$imp[[j]], 
              as.data.frame(matrix(NA, nrow = sum(!r[, j]), ncol = x$m, 
                                   dimnames = list(row.names(y)[r[, j] == FALSE], 1:m))))
    }
    names(imp) <- varnames

can be replaced with

imp <- x$imp

from mice.

Thanks. Great simplification.

from mice.