topic-multiple-columns: Taking multiple columns without '...'

In hadley/rlang: Functions for Base Types and Core R and 'Tidyverse' Features

Taking multiple columns without ...

Description

In this guide we compare ways of taking multiple columns in a single function argument.

As a refresher (see the programming patterns article), there are two common ways of passing arguments to data-masking functions. For single arguments, embrace with {{:

my_group_by <- function(data, var) {
  data %>% dplyr::group_by({{ var }})
}

my_pivot_longer <- function(data, var) {
  data %>% tidyr::pivot_longer({{ var }})
}

For multiple arguments in ..., pass them on to functions that also take ... like group_by(), or pass them within c() for functions taking tidy selection in a single argument like pivot_longer():

# Pass dots through
my_group_by <- function(.data, ...) {
  .data %>% dplyr::group_by(...)
}

my_pivot_longer <- function(.data, ...) {
  .data %>% tidyr::pivot_longer(c(...))
}

But what if you want to take multiple columns in a single named argument rather than in ...?

Using tidy selections

The idiomatic tidyverse way of taking multiple columns in a single argument is to take a tidy selection (see the Argument behaviours section). In tidy selections, the syntax for passing multiple columns in a single argument is c():

mtcars %>% tidyr::pivot_longer(c(am, cyl, vs))

Since ⁠{{⁠ inherits behaviour, this implementation of my_pivot_longer() automatically allows multiple columns passing:

my_pivot_longer <- function(data, var) {
  data %>% tidyr::pivot_longer({{ var }})
}

mtcars %>% my_pivot_longer(c(am, cyl, vs))

For group_by(), which takes data-masked arguments, we'll use across() as a bridge (see Bridge patterns).

my_group_by <- function(data, var) {
  data %>% dplyr::group_by(across({{ var }}))
}

mtcars %>% my_group_by(c(am, cyl, vs))

When embracing in tidyselect context or using across() is not possible, you might have to implement tidyselect behaviour manually with tidyselect::eval_select().

Using external defusal

To implement an argument with tidyselect behaviour, it is necessary to defuse the argument. However defusing an argument which had historically behaved like a regular argument is a rather disruptive breaking change. This is why we could not implement tidy selections in ggplot2 facetting functions like facet_grid() and facet_wrap().

An alternative is to use external defusal of arguments. This is what formula interfaces do for instance. A modelling function takes a formula in a regular argument and the formula defuses the user code:

my_lm <- function(data, f, ...) {
  lm(f, data, ...)
}

mtcars %>% my_lm(disp ~ drat)

Once created, the defused expressions contained in the formula are passed around like a normal argument. A similar approach was taken to update facet_ functions to tidy eval. The vars() function (a simple alias to quos()) is provided so that users can defuse their arguments externally.

ggplot2::facet_grid(
  ggplot2::vars(cyl),
  ggplot2::vars(am, vs)
)

You can implement this approach by simply taking a list of defused expressions as argument. This list can be passed the usual way to other functions taking such lists:

my_facet_grid <- function(rows, cols, ...) {
  ggplot2::facet_grid(rows, cols, ...)
}

Or it can be spliced with !!!:

my_group_by <- function(data, vars) {
  stopifnot(is_quosures(vars))
  data %>% dplyr::group_by(!!!vars)
}

mtcars %>% my_group_by(dplyr::vars(cyl, am))

A non-approach: Parsing lists

Intuitively, many programmers who want to take a list of expressions in a single argument try to defuse an argument and parse it. The user is expected to supply multiple arguments within a list() expression. When such a call is detected, the arguments are retrieved and spliced with ⁠!!!⁠. Otherwise, the user is assumed to have supplied a single argument which is injected with ⁠!!⁠. An implementation along these lines might look like this:

my_group_by <- function(data, vars) {
  vars <- enquo(vars)

  if (quo_is_call(vars, "list")) {
    expr <- quo_get_expr(vars)
    env <- quo_get_env(vars)
    args <- as_quosures(call_args(expr), env = env)
    data %>% dplyr::group_by(!!!args)
  } else {
    data %>% dplyr::group_by(!!vars)
  }
}

This does work in simple cases:

mtcars %>% my_group_by(cyl) %>% dplyr::group_vars()
#> [1] "cyl"

mtcars %>% my_group_by(list(cyl, am)) %>% dplyr::group_vars()
#> [1] "cyl" "am"

However this parsing approach quickly shows limits:

mtcars %>% my_group_by(list2(cyl, am))
#> Error in `group_by()`: Can't add columns.
#> i `..1 = list2(cyl, am)`.
#> i `..1` must be size 32 or 1, not 2.

Also, it would be better for overall consistency of interfaces to use the tidyselect syntax c() for passing multiple columns. In general, we recommend to use either the tidyselect or the external defusal approaches.

hadley/rlang documentation built on Dec. 17, 2024, 3:26 a.m.