This is a technical description of the tidyselect syntax.

library(tidyselect)
library(magrittr)

# For better printing
mtcars <- tibble::as_tibble(mtcars)
iris <- tibble::as_tibble(iris)

To illustrate the semantics of tidyselect, we’ll use variants of dplyr::select() and dplyr::rename() that return the named vector of locations for the selected or renamed elements:

select_loc <- function(data, ...) {
  eval_select(rlang::expr(c(...)), data)
}

rename_loc <- function(data, ...) {
  eval_rename(rlang::expr(c(...)), data)
}

Sets of variables

The tidyselect syntax is all about sets of variables, internally represented by integer vectors of locations. For example, c(1L, 2L) represents the set of the first and second variables, as does c(1L, 2L, 1L).

If a vector of locations contains duplicates, they are normally treated as the same element, since they represent sets. An exception to this occurs with named elements whose names differ. If the names don’t match, they are treated as different elements in order to allow renaming a variable to multiple names (see section on Renaming variables).

The syntax of tidyselect is generally designed for set combination. For instance, c(foo(), bar()) represents the union of the variables in foo() and those in bar().

Bare names

Within data-expressions (see Evaluation section), bare names represent their own locations, i.e. a set of size 1. The following expressions are equivalent:

mtcars %>% select_loc(mpg:hp, -cyl, vs)
#>  mpg disp   hp   vs 
#>    1    3    4    8

mtcars %>% select_loc(1:4, -2, 8)
#>  mpg disp   hp   vs 
#>    1    3    4    8

The : operator

: can be used to select consecutive variables between two locations. It returns the corresponding sequence of locations.

mtcars %>% select_loc(2:4)
#>  cyl disp   hp 
#>    2    3    4

Because bare names represent their own locations, it is easy to select a range of variables:

mtcars %>% select_loc(cyl:hp)
#>  cyl disp   hp 
#>    2    3    4

Boolean operators

Boolean operators provide a more intuitive approach to set combination. Though sets are internally represented with vectors of locations, they could equally be represented with a full logical vector of inclusion indicators (taking the which() of this vector would then recover the locations). The boolean operators should be considered in terms of the logical representation of sets.

The | operator takes the union of two sets:

iris %>% select_loc(starts_with("Sepal") | ends_with("Width"))
#> Sepal.Length  Sepal.Width  Petal.Width 
#>            1            2            4

The & operator takes the intersection of two sets:

iris %>% select_loc(starts_with("Sepal") & ends_with("Width"))
#> Sepal.Width 
#>           2

The ! operator takes the complement of a set:

iris %>% select_loc(!ends_with("Width"))
#> Sepal.Length Petal.Length      Species 
#>            1            3            5

Taking the intersection with a complement produces a set difference:

iris %>% select_loc(starts_with("Sepal") & !ends_with("Width"))
#> Sepal.Length 
#>            1

Dots, c(), and unary -

tidyselect functions can take dots like dplyr::select(), or a named argument like tidyr::pivot_longer(). In the latter case, the dots syntax is accessible via c(). In fact ... syntax is implemented through c(...) and is thus completely equivalent.

mtcars %>% select_loc(mpg, disp:hp)
#>  mpg disp   hp 
#>    1    3    4

mtcars %>% select_loc(c(mpg, disp:hp))
#>  mpg disp   hp 
#>    1    3    4

Dots and c() are syntax for:

  • Set union or set difference
  • Renaming variables

Non-negative inputs are recursively joined with union(). The precedence is left-associative, just like with boolean operators. These expressions are all syntax for set union:

iris %>% select_loc(starts_with("Sepal"), ends_with("Width"), Species)
#> Sepal.Length  Sepal.Width  Petal.Width      Species 
#>            1            2            4            5

iris %>% select_loc(starts_with("Sepal") | ends_with("Width") | Species)
#> Sepal.Length  Sepal.Width  Petal.Width      Species 
#>            1            2            4            5

iris %>% select_loc(union(union(starts_with("Sepal"), ends_with("Width")), 5L))
#> Sepal.Length  Sepal.Width  Petal.Width      Species 
#>            1            2            4            5

Unary - is normally syntax for set difference:

iris %>% select_loc(starts_with("Sepal"), -ends_with("Width"), -Sepal.Length)
#> named integer(0)

iris %>% select_loc(setdiff(setdiff(starts_with("Sepal"), ends_with("Width")), 1L))
#> named integer(0)

If the first ... or c() input is negative, an implicit everything() is appended.

iris %>% select_loc(-starts_with("Sepal"))
#> Petal.Length  Petal.Width      Species 
#>            3            4            5

iris %>% select_loc(everything(), -starts_with("Sepal"))
#> Petal.Length  Petal.Width      Species 
#>            3            4            5

iris %>% select_loc(setdiff(everything(), starts_with("Sepal")))
#> Petal.Length  Petal.Width      Species 
#>            3            4            5

In this case, unary - is syntax for set complement. Unary - and ! are equivalent:

iris %>% select_loc(-starts_with("Sepal"))
#> Petal.Length  Petal.Width      Species 
#>            3            4            5

iris %>% select_loc(!starts_with("Sepal"))
#> Petal.Length  Petal.Width      Species 
#>            3            4            5

Each level of c() is independent. In particular, a nested c() starting with - always stands for set complement:

iris %>% select_loc(c(starts_with("Sepal"), -Sepal.Length))
#> Sepal.Width 
#>           2

iris %>% select_loc(c(starts_with("Sepal"), c(-Sepal.Length)))
#> Sepal.Length  Sepal.Width Petal.Length  Petal.Width      Species 
#>            1            2            3            4            5

In boolean terms, these expressions are equivalent to:

iris %>% select_loc(starts_with("Sepal") & !Sepal.Length)
#> Sepal.Width 
#>           2

iris %>% select_loc(starts_with("Sepal") | !Sepal.Length)
#> Sepal.Length  Sepal.Width Petal.Length  Petal.Width      Species 
#>            1            2            3            4            5

In general, when unary - is used alone outside ... or c(), it stands for set complement.

Renaming variables

Name combination and propagation

When named inputs are provided in ... or c(), the selection is renamed. If the inputs are already named, the outer and inner names are combined with a ... separator:

mtcars %>% select_loc(foo = c(bar = mpg, baz = cyl))
#> foo...bar foo...baz 
#>         1         2

Otherwise the outer names is propagated to the selected elements according to the following rules:

  • With data frames, a numeric suffix is appended because columns must be uniquely named.

    mtcars %>% select_loc(foo = c(mpg, cyl))
    #> foo1 foo2 
    #>    1    2
  • With normal vectors, the name is simply assigned to all selected inputs.

    as.list(mtcars) %>% select_loc(foo = c(mpg, cyl))
    #> foo foo 
    #>   1   2

Combination and propagation can be composed by using nested c():

mtcars %>% select_loc(foo = c(bar = c(mpg, cyl)))
#> foo...bar1 foo...bar2 
#>          1          2

Set combination with named variables

Named elements have special rules to determine their identities in a set. Unnamed elements match any names:

  • a | c(foo = a) is equivalent to c(foo = a).
  • a & c(foo = a) is equivalent to c(foo = a).

Named elements with different names are distinct:

  • c(foo = a) & c(bar = a) is equivalent to c().
  • c(foo = a) | c(bar = a) is equivalent to c(foo = a, bar = a).

Because unnamed elements match any named ones, it is possible to select multiple elements and rename one of them:

iris %>% select_loc(!Species, foo = Sepal.Width)
#> Sepal.Length          foo Petal.Length  Petal.Width 
#>            1            2            3            4

Predicate functions

Predicate function objects can be supplied as input in an env-expression, typically with the selection helper where(). They are applied to all elements of the data, and should return TRUE or FALSE to indicate inclusion. Predicates in env-expressions are effectively expanded to the set of variables that they represent:

iris %>% select_loc(where(is.numeric))
#> Sepal.Length  Sepal.Width Petal.Length  Petal.Width 
#>            1            2            3            4

iris %>% select_loc(where(is.factor))
#> Species 
#>       5

iris %>% select_loc(where(is.numeric) | where(is.factor))
#> Sepal.Length  Sepal.Width Petal.Length  Petal.Width      Species 
#>            1            2            3            4            5

iris %>% select_loc(where(is.numeric) & where(is.factor))
#> named integer(0)

Selection helpers

We call selection helpers any function that inspects the currently active variables with peek_vars() and returns a selection.

  • peek_vars() returns a character vector of names.
  • The returned selection can be any output conforming to the types described in the Data types section.

Examples of selection helpers are all_of(), contains(), or last_col(). These selection helpers are evaluated as env-expressions (see Evaluation section).

Supported data types

The following data types can be returned from selection helpers or forced via !! or force() (the latter works in tidyselect because it is treated as an env-expression, see Evaluation section):

  • Vectors of locations:

    iris %>% select_loc(force(c(1, 3)))
    #> Sepal.Length Petal.Length 
    #>            1            3
  • Vectors of names. These are matched and transformed to locations.

    iris %>% select_loc(force(c("Sepal.Length", "Petal.Length")))
    #> Sepal.Length Petal.Length 
    #>            1            3
  • Predicate functions. These are applied to all elements to determine inclusion.

    iris %>% select_loc(force(is.numeric))
    #> Sepal.Length  Sepal.Width Petal.Length  Petal.Width 
    #>            1            2            3            4

Evaluation

Data-expressions and env-expressions

tidyselect is not a typical tidy evaluation UI. The main difference is that there is no data masking. In a typical tidy eval function, expressions are evaluated with data-vars first in scope, followed by env-vars:

mask <- function(data, expr) {
  rlang::eval_tidy(rlang::enquo(expr), data)
}

foo <- 10
cyl <- 200

# `cyl` represents the data frame column here:
mtcars %>% mask(cyl * foo)
#>  [1] 60 60 40 60 80 60 80 40 40 60 60 80 80 80 80 80 80 40 40 40 40 80 80 80 80
#> [26] 40 40 40 80 60 80 40

It is possible to bypass the data frame variables by forcing symbols to be looked up in the environment with !! or .env:

mtcars %>% mask(!!cyl * foo)
#> [1] 2000
mtcars %>% mask(.env$cyl * foo)
#> [1] 2000

With tidyselect, there is no such hierarchical data masking. Instead, expressions are evaluated either in the context of the data frame or in the user environment, without overlap. The scope of lookup depends on the kind of expression:

  1. data-expressions are evaluated in the data frame only. This includes bare symbols, the boolean operators, -, :, and c(). You can’t refer to environment-variables in a data-expression:

    cyl_pos <- 2
    mtcars %>% select_loc(mpg | cyl_pos)
    #> Error: Can't subset columns that don't exist.
    #> ✖ Column `cyl_pos` doesn't exist.
  2. env-expressions are evaluated in the environment. This includes all calls other than those mentioned above, as well as symbols that are part of those calls. You can’t refer to data-variables in a data-expression:

    mtcars %>% select_loc(all_of(mpg))
    #> Error: object 'mpg' not found

Because the scoping is unambiguous, you can safely refer to env-vars in an env-expression, without having to worry about potential naming clashes with data-vars:

x <- data.frame(x = 1:3, y = 4:6, z = 7:9)

# `ncol(x)` is an env-expression, so `x` represents the data frame in
# the environment rather than the column in the data frame
x %>% select_loc(2:ncol(x))
#> y z 
#> 2 3

If you have variable names in a character vector, it is safe to refer to the env-var containing the names with all_of() because it is an env-expression:

y <- c("y", "z")
x %>% select_loc(all_of(y))
#> y z 
#> 2 3

Note that currently, env-vars are still allowed in some data-expressions, for compatibility. However this is in the process of being deprecated and you should see a note recommending to use all_of() instead. This note will become a deprecation warning in the future, and then an error.

mtcars %>% select_loc(cyl_pos)
#> Note: Using an external vector in selections is ambiguous.
#> ℹ Use `all_of(cyl_pos)` instead of `cyl_pos` to silence this message.
#> ℹ See <https://tidyselect.r-lib.org/reference/faq-external-vector.html>.
#> This message is displayed once per session.
#> cyl 
#>   2

Arithmetic operators

Within data-expressions (see Evaluation section), +, * and / are overridden to cause an error. This is to prevent confusion stemming from normal data masking usage where variables can be transformed on the fly:

mtcars %>% select_loc(cyl^2)
#> Error: Can't use arithmetic operator `^` in selection context.

mtcars %>% select_loc(mpg * wt)
#> Error: Can't use arithmetic operator `*` in selection context.

Selecting versus renaming

The select and rename variants take the same types of inputs and have the same type of return value. They have a few important differences.

All renaming inputs must be named

Unlike eval_select() which can select without renaming, eval_rename() expects a fully named selection. If one or several names are missing, an error is thrown.

mtcars %>% select_loc(mpg)
#> mpg 
#>   1

mtcars %>% rename_loc(mpg)
#> Error: All renaming inputs must be named.

Renaming to an existing variable name

If the input data is a data frame, tidyselect generally throws an error when duplicate column names are selected, in order to respect the invariant of unique column names.

# Lists can have duplicates
as.list(mtcars) %>% select_loc(foo = mpg, foo = cyl)
#> foo foo 
#>   1   2

# Data frames cannot
mtcars %>% select_loc(foo = mpg, foo = cyl)
#> Error: Names must be unique.
#> ✖ These names are duplicated:
#>   * "foo" at locations 1 and 2.

A selection can rename a variable to an existing name if the latter is not part of the selection:

mtcars %>% select_loc(cyl, cyl = mpg)
#> Error: Names must be unique.
#> ✖ These names are duplicated:
#>   * "cyl" at locations 1 and 2.

mtcars %>% select_loc(disp, cyl = mpg)
#> disp  cyl 
#>    3    1

This is not possible when renaming.

mtcars %>% rename_loc(cyl, cyl = mpg)
#> Error: All renaming inputs must be named.

mtcars %>% rename_loc(disp, cyl = mpg)
#> Error: All renaming inputs must be named.

However, the name conflict can be solved by renaming the existing variable to another name:

mtcars %>% select_loc(foo = cyl, cyl = mpg)
#> foo cyl 
#>   2   1

mtcars %>% rename_loc(foo = cyl, cyl = mpg)
#> foo cyl 
#>   2   1

Duplicate columns in data frames

Normally a data frame shouldn’t have duplicate names. However, the real world is messy and duplicates do happen in the wild. tidyselect tries to be as permissive as it can with these duplicates so that users can restore unique names with select() or rename().

First let’s create a data frame with duplicate names:

dups <- vctrs::new_data_frame(list(x = 1, y = 2, x = 3))

If the duplicates are not part of the selection, they are simply ignored:

dups %>% select_loc(y)
#> y 
#> 2

If the duplicates are selected, this is an error:

dups %>% select_loc(x)
#> Error: Names must be unique.
#> ✖ These names are duplicated:
#>   * "x" at locations 1 and 2.

The duplicate names can be repaired by renaming chosen locations:

dups %>% select_loc(x, foo = 3)
#>   x foo 
#>   1   3

dups %>% rename_loc(foo = 3)
#> foo 
#>   3

Acknowledgements

The tidyselect syntax was inspired by the base::subset() function written by Peter Dalgaard. The select parameter of subset.data.frame() is evaluated in a data mask where the column names are bound to their locations in the data frame. This allows : to create sequences of variable locations. The locations can be combined with c(). This selection interface set the tone for the development of the tidyselect syntax.

mtcars %>% subset(select = c(cyl, hp:wt))
#> # A tibble: 32 x 4
#>     cyl    hp  drat    wt
#>   <dbl> <dbl> <dbl> <dbl>
#> 1     6   110  3.9   2.62
#> 2     6   110  3.9   2.88
#> 3     4    93  3.85  2.32
#> 4     6   110  3.08  3.22
#> # … with 28 more rows