Column-wise operations

It’s often useful to perform the same operation on multiple columns, but copying and pasting is both tedious and error prone:

df %>% 
  group_by(g1, g2) %>% 
  summarise(a = mean(a), b = mean(b), c = mean(c), d = mean(d))

(If you’re trying to compute mean(a, b, c, d) for each row, instead see vignette("rowwise"))

This vignette will introduce you to the across() function, which lets you rewrite the previous code more succinctly:

df %>% 
  group_by(g1, g2) %>% 
  summarise(across(a:d, mean))

We’ll start by discussing the basic usage of across(), particularly as it applies to summarise(), and show how to use it with multiple functions. We’ll then show a few uses with other verbs. We’ll finish off with a bit of history, showing why we prefer across() to our last approach (the _if(), _at() and _all() functions) and how to translate your old code to the new syntax.

library(dplyr, warn.conflicts = FALSE)

Basic usage

across() has two primary arguments:

Here are a couple of examples of across() in conjunction with its favourite verb, summarise(). But you can use across() with any dplyr verb, as you’ll see a little later.

starwars %>% 
  summarise(across(where(is.character), n_distinct))
#> # A tibble: 1 × 8
#>    name hair_color skin_color eye_color   sex gender homeworld species
#>   <int>      <int>      <int>     <int> <int>  <int>     <int>   <int>
#> 1    87         13         31        15     5      3        49      38

starwars %>% 
  group_by(species) %>% 
  filter(n() > 1) %>% 
  summarise(across(c(sex, gender, homeworld), n_distinct))
#> # A tibble: 9 × 4
#>   species    sex gender homeworld
#>   <chr>    <int>  <int>     <int>
#> 1 Droid        1      2         3
#> 2 Gungan       1      1         1
#> 3 Human        2      2        16
#> 4 Kaminoan     2      2         1
#> # … with 5 more rows

starwars %>% 
  group_by(homeworld) %>% 
  filter(n() > 1) %>% 
  summarise(across(where(is.numeric), ~ mean(.x, na.rm = TRUE)))
#> # A tibble: 10 × 4
#>   homeworld height  mass birth_year
#>   <chr>      <dbl> <dbl>      <dbl>
#> 1 Alderaan    176.  64         43  
#> 2 Corellia    175   78.5       25  
#> 3 Coruscant   174.  50         91  
#> 4 Kamino      208.  83.1       31.5
#> # … with 6 more rows

Because across() is usually used in combination with summarise() and mutate(), it doesn’t select grouping variables in order to avoid accidentally modifying them:

df <- data.frame(g = c(1, 1, 2), x = c(-1, 1, 3), y = c(-1, -4, -9))
df %>% 
  group_by(g) %>% 
  summarise(across(where(is.numeric), sum))
#> # A tibble: 2 × 3
#>       g     x     y
#>   <dbl> <dbl> <dbl>
#> 1     1     0    -5
#> 2     2     3    -9

Multiple functions

You can transform each variable with more than one function by supplying a named list of functions or lambda functions in the second argument:

min_max <- list(
  min = ~min(.x, na.rm = TRUE), 
  max = ~max(.x, na.rm = TRUE)
)
starwars %>% summarise(across(where(is.numeric), min_max))
#> # A tibble: 1 × 6
#>   height_min height_max mass_min mass_max birth_year_min birth_year_max
#>        <int>      <int>    <dbl>    <dbl>          <dbl>          <dbl>
#> 1         66        264       15     1358              8            896
starwars %>% summarise(across(c(height, mass, birth_year), min_max))
#> # A tibble: 1 × 6
#>   height_min height_max mass_min mass_max birth_year_min birth_year_max
#>        <int>      <int>    <dbl>    <dbl>          <dbl>          <dbl>
#> 1         66        264       15     1358              8            896

Control how the names are created with the .names argument which takes a glue spec:

starwars %>% summarise(across(where(is.numeric), min_max, .names = "{.fn}.{.col}"))
#> # A tibble: 1 × 6
#>   min.height max.height min.mass max.mass min.birth_year max.birth_year
#>        <int>      <int>    <dbl>    <dbl>          <dbl>          <dbl>
#> 1         66        264       15     1358              8            896
starwars %>% summarise(across(c(height, mass, birth_year), min_max, .names = "{.fn}.{.col}"))
#> # A tibble: 1 × 6
#>   min.height max.height min.mass max.mass min.birth_year max.birth_year
#>        <int>      <int>    <dbl>    <dbl>          <dbl>          <dbl>
#> 1         66        264       15     1358              8            896

If you’d prefer all summaries with the same function to be grouped together, you’ll have to expand the calls yourself:

starwars %>% summarise(
  across(c(height, mass, birth_year), ~min(.x, na.rm = TRUE), .names = "min_{.col}"),
  across(c(height, mass, birth_year), ~max(.x, na.rm = TRUE), .names = "max_{.col}")
)
#> # A tibble: 1 × 6
#>   min_height min_mass min_birth_year max_height max_mass max_birth_year
#>        <int>    <dbl>          <dbl>      <int>    <dbl>          <dbl>
#> 1         66       15              8        264     1358            896

(One day this might become an argument to across() but we’re not yet sure how it would work.)

We cannot however use where(is.numeric) in that last case because the second across() would pick up the variables that were newly created (“min_height”, “min_mass” and “min_birth_year”).

We can work around this by combining both calls to across() into a single expression that returns a tibble:

starwars %>% summarise(
  tibble(
    across(where(is.numeric), ~min(.x, na.rm = TRUE), .names = "min_{.col}"),
    across(where(is.numeric), ~max(.x, na.rm = TRUE), .names = "max_{.col}")  
  )
)
#> # A tibble: 1 × 6
#>   min_height min_mass min_birth_year max_height max_mass max_birth_year
#>        <int>    <dbl>          <dbl>      <int>    <dbl>          <dbl>
#> 1         66       15              8        264     1358            896

Alternatively we could reorganize results with relocate():

starwars %>% 
  summarise(across(where(is.numeric), min_max, .names = "{.fn}.{.col}")) %>% 
  relocate(starts_with("min"))
#> # A tibble: 1 × 6
#>   min.height min.mass min.birth_year max.height max.mass max.birth_year
#>        <int>    <dbl>          <dbl>      <int>    <dbl>          <dbl>
#> 1         66       15              8        264     1358            896

Current column

If you need to, you can access the name of the “current” column inside by calling cur_column(). This can be useful if you want to perform some sort of context dependent transformation that’s already encoded in a vector:

df <- tibble(x = 1:3, y = 3:5, z = 5:7)
mult <- list(x = 1, y = 10, z = 100)

df %>% mutate(across(all_of(names(mult)), ~ .x * mult[[cur_column()]]))
#> # A tibble: 3 × 3
#>       x     y     z
#>   <dbl> <dbl> <dbl>
#> 1     1    30   500
#> 2     2    40   600
#> 3     3    50   700

Gotchas

Be careful when combining numeric summaries with where(is.numeric):

df <- data.frame(x = c(1, 2, 3), y = c(1, 4, 9))

df %>% 
  summarise(n = n(), across(where(is.numeric), sd))
#>    n x        y
#> 1 NA 1 4.041452

Here n becomes NA because n is numeric, so the across() computes its standard deviation, and the standard deviation of 3 (a constant) is NA. You probably want to compute n() last to avoid this problem:

df %>% 
  summarise(across(where(is.numeric), sd), n = n())
#>   x        y n
#> 1 1 4.041452 3

Alternatively, you could explicitly exclude n from the columns to operate on:

df %>% 
  summarise(n = n(), across(where(is.numeric) & !n, sd))
#>   n x        y
#> 1 3 1 4.041452

Another approach is to combine both the call to n() and across() in a single expression that returns a tibble:

df %>% 
  summarise(
    tibble(n = n(), across(where(is.numeric), sd))
  )
#>   n x        y
#> 1 3 1 4.041452

Other verbs

So far we’ve focused on the use of across() with summarise(), but it works with any other dplyr verb that uses data masking:

For some verbs, like group_by(), count() and distinct(), you can omit the summary functions:

across() doesn’t work with select() or rename() because they already use tidy select syntax; if you want to transform column names with a function, you can use rename_with().

filter()

We cannot directly use across() in filter() because we need an extra step to combine the results. To that end, filter() has two special purpose companion functions:

starwars %>% 
  filter(if_any(everything(), ~ !is.na(.x)))
#> # A tibble: 87 × 14
#>   name         height  mass hair_…¹ skin_…² eye_c…³ birth…⁴ sex   gender homew…⁵
#>   <chr>         <int> <dbl> <chr>   <chr>   <chr>     <dbl> <chr> <chr>  <chr>  
#> 1 Luke Skywal…    172    77 blond   fair    blue       19   male  mascu… Tatooi…
#> 2 C-3PO           167    75 <NA>    gold    yellow    112   none  mascu… Tatooi…
#> 3 R2-D2            96    32 <NA>    white,… red        33   none  mascu… Naboo  
#> 4 Darth Vader     202   136 none    white   yellow     41.9 male  mascu… Tatooi…
#> # … with 83 more rows, 4 more variables: species <chr>, films <list>,
#> #   vehicles <list>, starships <list>, and abbreviated variable names
#> #   ¹​hair_color, ²​skin_color, ³​eye_color, ⁴​birth_year, ⁵​homeworld
starwars %>% 
  filter(if_all(everything(), ~ !is.na(.x)))
#> # A tibble: 29 × 14
#>   name         height  mass hair_…¹ skin_…² eye_c…³ birth…⁴ sex   gender homew…⁵
#>   <chr>         <int> <dbl> <chr>   <chr>   <chr>     <dbl> <chr> <chr>  <chr>  
#> 1 Luke Skywal…    172    77 blond   fair    blue       19   male  mascu… Tatooi…
#> 2 Darth Vader     202   136 none    white   yellow     41.9 male  mascu… Tatooi…
#> 3 Leia Organa     150    49 brown   light   brown      19   fema… femin… Aldera…
#> 4 Owen Lars       178   120 brown,… light   blue       52   male  mascu… Tatooi…
#> # … with 25 more rows, 4 more variables: species <chr>, films <list>,
#> #   vehicles <list>, starships <list>, and abbreviated variable names
#> #   ¹​hair_color, ²​skin_color, ³​eye_color, ⁴​birth_year, ⁵​homeworld

_if, _at, _all

Prior versions of dplyr allowed you to apply a function to multiple columns in a different way: using functions with _if, _at, and _all() suffixes. These functions solved a pressing need and are used by many people, but are now superseded. That means that they’ll stay around, but won’t receive any new features and will only get critical bug fixes.

Why do we like across()?

Why did we decide to move away from these functions in favour of across()?

  1. across() makes it possible to express useful summaries that were previously impossible:

    df %>%
      group_by(g1, g2) %>% 
      summarise(
        across(where(is.numeric), mean), 
        across(where(is.factor), nlevels),
        n = n(), 
      )
  2. across() reduces the number of functions that dplyr needs to provide. This makes dplyr easier for you to use (because there are fewer functions to remember) and easier for us to implement new verbs (since we only need to implement one function, not four).

  3. across() unifies _if and _at semantics so that you can select by position, name, and type, and you can now create compound selections that were previously impossible. For example, you can now transform all numeric columns whose name begins with “x”: across(where(is.numeric) & starts_with("x")).

  4. across() doesn’t need to use vars(). The _at() functions are the only place in dplyr where you have to manually quote variable names, which makes them a little weird and hence harder to remember.

Why did it take so long to discover across()?

It’s disappointing that we didn’t discover across() earlier, and instead worked through several false starts (first not realising that it was a common problem, then with the _each() functions, and most recently with the _if()/_at()/_all() functions). But across() couldn’t work without three recent discoveries:

How do you convert existing code?

Fortunately, it’s generally straightforward to translate your existing code to use across():

For example:

df %>% mutate_if(is.numeric, mean, na.rm = TRUE)
# ->
df %>% mutate(across(where(is.numeric), mean, na.rm = TRUE))

df %>% mutate_at(vars(c(x, starts_with("y"))), mean)
# ->
df %>% mutate(across(c(x, starts_with("y")), mean, na.rm = TRUE))

df %>% mutate_all(mean)
# ->
df %>% mutate(across(everything(), mean))

There are a few exceptions to this rule: