WISH: deprecate data.frame(..., row.names = "name of column to use as rownames") and add explicit alternative

[jan-glx]

Currently, data.frame provides and argument row.names that not only allows for supplying row names to be used for the new data frame but also allows for specifying a column whose values to use as row names:

row.names	NULL or a single integer or character string specifying a column to be used as row names, or a character or integer vector giving the row names for the data frame.

However, in the case of a single row data frame, there is no way to distinguish these two meanings, leading to inconsistent behavior:

data.frame(id=paste0("row", 1:2), val=1:2) |> data.frame(row.names="id")
#      val
# row1   1
# row2   2
data.frame(id=paste0("row", 1:1), val=1:1) |> data.frame(row.names="id")
#      id val
# id row1   1

To prevent the resulting edge-case bugs, developers have to avoid this feature and instead use code such as rownames(df) <- df$id; df$id <- NULL. or :

 col_as_rownames <- function(data, col.as.row.names) {
    stopifnot(length(col.as.row.names)==1)
    idx <- if(is.numeric(col.as.row.names)) col.as.row.names else which(colnames(data) == col.as.row.names)
    data.frame(data[, -idx, drop = FALSE], row.names=data[, idx])
}
data.frame(id=paste0("row", 1:2), val=1:2) |> col_as_rownames("id")
data.frame(id=paste0("row", 1:1), val=1:1)  |> col_as_rownames("id")

• ?base::data.frame should be updated to caution against using row.names with a column name or index and point to an safe alternative
• base::data.frame should provide an argument col.as.row.names or so to achieve the same in a simple and safe way.

[J-Moravec]

data.frame() here is consistent with other functions, such as read.table() (and derived read.csv(), etc.).

Specifying which column to be interpreted as rownames is IMHO much more useful than specifying rownames directly.
You can always setup rownames with rownames(x) = y. Transforming a column of a data.frame into rownames, while removing it, takes a bit more.

rownames(x) = x[[i]]
x[[i]] = NULL

In data analysis, when taking data from foreign source, I ended up interpreting pre-existing column as names of DF much more often than specifying the names explicitly.

[HenrikBengtsson]

However, in the case of a single row data frame, there is no way to distinguish these two meanings, leading to inconsistent behavior:

A random thought: R could warn or even produce an error when it's ambiguous what row.names is meant to be. One option for distinguishing the two case could be to use the AsIs class, e.g. row.names = I(one_col_name) to mean "column name" rather than a vector of row names. Alternatively, one could use I(.) for declaring "row names".

[jan-glx]

@J-Moravec

Specifying which column to be interpreted as rownames is IMHO much more useful than specifying rownames directly.

Precisely. Which is why it is very annoying that it stops working when the data.frame has only one row.

data.frame() here is consistent with other functions, such as read.table() (and derived read.csv(), etc.).
Then any change should be implemented there as well.

In data analysis, when taking data from foreign source, I ended up interpreting pre-existing column as names of DF much more often than specifying the names explicitly.
So do I (using data.table almost exclusively), but sometimes when using other packages one is forced to go back to rownames which is where consistent behavior would be very useful.

@HenrikBengtsson

Throwing an error could be breaking existing code, a warning in the one row case might be helpful when debugging such an issue but would often come to late.
I kind of like your idea of using e.g. I - looks more R-y than adding another overly long parameter name. I(.) for "row names" would feel more natural for me, but would imply changing the default for the one-row case. Are there other resonable options? E.g. expressions to force interpretation as column name?