[Rd] A suggestion for an amendment to tapply

[Andrew Robinson]

Dear R-developers,

when tapply() is invoked on factors that have empty levels, it returns
NA.  This behaviour is in accord with the tapply documentation, and is
reasonable in many cases.  However, when FUN is sum, it would also
seem reasonable to return 0 instead of NA, because "the sum of an
empty set is zero, by definition."

I'd like to raise a discussion of the possibility of an amendment to
tapply.

The attached patch changes the function so that it checks if there are
any empty levels, and if there are, replaces the corresponding NA
values with the result of applying FUN to the empty set.  Eg in the
case of sum, it replaces the NA with 0, whereas with mean, it replaces
the NA with NA, and issues a warning.

This change has the following advantage: tapply and sum work better
together.  Arguably, tapply and any other function that has a non-NA
response to the empty set will also work better together.
Furthermore, tapply shows a warning if FUN would normally show a
warning upon being evaluated on an empty set.  That deviates from
current behaviour, which might be bad, but also provides information
that might be useful to the user, so that would be good.

The attached script provides the new function in full, and
demonstrates its application in some simple test cases.

Best wishes,

Andrew
-- 
Andrew Robinson  
Department of Mathematics and Statistics            Tel: +61-3-8344-9763
University of Melbourne, VIC 3010 Australia         Fax: +61-3-8344-4599
http://www.ms.unimelb.edu.au/~andrewpr
http://blogs.mbs.edu/fishing-in-the-bay/ 
-------------- next part --------------
## The new function

my.tapply <- function (X, INDEX, FUN=NULL, ..., simplify=TRUE)
{
    FUN <- if (!is.null(FUN)) match.fun(FUN)
    if (!is.list(INDEX)) INDEX <- list(INDEX)
    nI <- length(INDEX)
    namelist <- vector("list", nI)
    names(namelist) <- names(INDEX)
    extent <- integer(nI)
    nx <- length(X)
    one <- as.integer(1)
    group <- rep.int(one, nx)#- to contain the splitting vector
    ngroup <- one
    for (i in seq.int(INDEX)) {
	index <- as.factor(INDEX[[i]])
	if (length(index) != nx)
	    stop("arguments must have same length")
	namelist[[i]] <- levels(index)#- all of them, yes !
	extent[i] <- nlevels(index)
	group <- group + ngroup * (as.integer(index) - one)
	ngroup <- ngroup * nlevels(index)
    }
    if (is.null(FUN)) return(group)
    ans <- lapply(split(X, group), FUN, ...)
    index <- as.numeric(names(ans))
    if (simplify && all(unlist(lapply(ans, length)) == 1)) {
	ansmat <- array(dim=extent, dimnames=namelist)
	ans <- unlist(ans, recursive = FALSE)
    }
    else  {
	ansmat <- array(vector("list", prod(extent)),
			dim=extent, dimnames=namelist)
    }
    ## old : ansmat[as.numeric(names(ans))] <- ans
    names(ans) <- NULL
    ansmat[index] <- ans
    if (sum(table(INDEX) < 1) > 0)
        ansmat[table(INDEX) < 1] <- do.call(FUN, list(c(NULL), ...)) 
    ansmat
}

## Check its utility

group <- factor(c(1,1,3,3), levels=c("1","2","3"))
x <- c(1,2,3,4)

## Ok with mean?

tapply(x, group, mean)
my.tapply(x, group, mean)

## Ok with sum?

tapply(x, group, sum)
my.tapply(x, group, sum)

## Check that other arguments are carried through

x <- c(NA,2,3,10)

tapply(x, group, sum, na.rm=TRUE)
tapply(x, group, mean, na.rm=TRUE)

my.tapply(x, group, sum, na.rm=TRUE)
my.tapply(x, group, mean, na.rm=TRUE)

## Check that listed groups work ok also

group.2 <- factor(c(1,2,3,3), levels=c("1","2","3"))

tapply(x, list(group, group.2), sum, na.rm=TRUE)
tapply(x, list(group, group.2), mean, na.rm=TRUE)

my.tapply(x, list(group, group.2), sum, na.rm=TRUE)
my.tapply(x, list(group, group.2), mean, na.rm=TRUE)