two questions for R beginners
John Sorkin wrote:
Please take what follows not as an ad hominem statement, but rather as an attempt to improve what is already an excellent program, that has been built as a result of many, many hours of dedicated work by many, many unpaid, unsung volunteers. It troubles me a bit that when a confusing aspect of R is pointed out the response is not to try to improve the language so as to avoid the confusion, but rather to state that the confusion is inherent in the language. I understand that to make changes that would avoid the confusing aspect of the language that has been discussed in this thread would take time and effort by an R wizard (which I am not), time and effort that would not be compensated in the traditional sense. This does not mean that we should not acknowledge the confusion. If we what R to be the de facto lingua franca of statistical analysis doesn't it make sense to strive for syntax that is as straight forward and consistent as possible?
I think you've misunderstood the argument. It would not be hard to make the suggested change. I don't object to it because it would be too much work, I object to it because I think it is not an improvement. Dataframes and matrices are different, and there is no way to avoid that fact. The arguments in favour of the change seem to be these: - Dataframes and matrices are similar in some respects, so they should be similar in more. In fact, I believe that the source of confusion is the fact that the are similar, so this would not improve things. People would still be confused by the differences, which are unavoidable. - Using $ to extract a column of a matrix would be convenient. I agree, it saves 4 keystrokes to type X$column instead of X[,"column"]. But I think it increases confusion, so the savings are not worthwhile. For example, the col2rgb function returns a matrix with rows named red, green and blue. But under your proposal, I'd still need to use X["red",] to extract the red component, because columns are components, but rows are not. You are complaining that the lack of $ for matrices is an unnecessary asymmetry, and unnecessary asymmetries are confusing. But your proposal introduces a new one! - Some functions return matrices when I expect a dataframe, or vice versa. That will continue to be true regardless of whether the proposed change is made. You need to read the documentation. If it is unclear, it should be improved, the language shouldn't be changed so that sloppy documentation is accurate. - You suggested this so anyone who disagrees must be lazy. Which really is an ad hominem argument, despite your disclaimer. I think you should respect the fact that there are people who disagree with the value of your suggestion. (Which is also an ad hominem attack, but isn't central to my argument.) Duncan Murdoch
Again, please understand that my comment is made with deepest respect for the many people who have unselfishly contributed to the R project. Many thanks to each and every one of you. John
Karl Ove Hufthammer <karl at huftis.org> 3/2/2010 4:00 AM >>>
On Mon, 01 Mar 2010 10:00:07 -0500 Duncan Murdoch <murdoch at stats.uwo.ca> wrote:
Suppose X is a dataframe or a matrix. What would you expect to get from
X[1]? What about as.vector(X), or as.numeric(X)?
All this of course depends on type of object one is speaking of. There are plenty of surprises available, and it's best to use the most logical way of extracting. E.g., to extract the top-left element of a 2D structure (data frame or matrix), use 'X[1,1]'. Luckily, R provides some shortcuts. For example, you can write 'X[2,3]' on a data frame, just as if it was a matrix, even though the underlying structure is completely different. (This doesn't work on a normal list; there you have to type the whole 'X[[2]][3]'.) The behaviour of the 'as.' functions may sometimes be surprising, at least for me. For example, 'as.data.frame' on a named vector gives a single-column data frame, instead of a single-row data frame. (I'm not sure what's the recommended way of converting a named vector to row data frame, but 'as.data.frame(t(X))' works, even though both 'X' and 't(X)' looks like a row of numbers.)
The point is that a dataframe is a list, and a matrix isn't. If users
don't understand that, then they'll be confused somewhere. Making
matrices more list-like in one respect will just move the confusion
elsewhere. The solution is to understand the difference.
My main problem is not understanding the difference, which is easy, but knowing which type of I have when I get the output a function in a package. If I know the object is a named vector or a matrix with column names, it's easy enough to type 'X[,"colname"]', and if it's a data frame one may use the shortcut 'X$colname'. Usually, it *is* documented what the return value of a function is, but just looking at the output is much faster, and *usually* gives the correct answer. For example, 'mean' applied on a data frame gives a named vector, not a data frame, which is somewhat surprising (given that the columns of a data frame may be of different types, while the elements of a vector may not). (And yes, I know that it's *documented* that it returns a named vector.) On the other hand, perhaps it is surprising that 'mean' works on data frames at all. :-)