Most dplyr verbs use tidy evaluation in some way. Tidy evaluation is a special type of non-standard evaluation used throughout the tidyverse. There are two basic forms found in dplyr:
arrange()
, count()
,
filter()
, group_by()
, mutate()
,
and summarise()
use data masking so that
you can use data variables as if they were variables in the environment
(i.e. you write my_variable
not
df$my_variable
).
across()
, relocate()
,
rename()
, select()
, and pull()
use tidy selection so you can easily choose variables
based on their position, name, or type
(e.g. starts_with("x")
or
is.numeric
).
To determine whether a function argument uses data masking or tidy
selection, look at the documentation: in the arguments list, you’ll see
<data-masking>
or
<tidy-select>
.
Data masking and tidy selection make interactive data exploration fast and fluid, but they add some new challenges when you attempt to use them indirectly such as in a for loop or a function. This vignette shows you how to overcome those challenges. We’ll first go over the basics of data masking and tidy selection, talk about how to use them indirectly, and then show you a number of recipes to solve common problems.
This vignette will give you the minimum knowledge you need to be an effective programmer with tidy evaluation. If you’d like to learn more about the underlying theory, or precisely how it’s different from non-standard evaluation, we recommend that you read the Metaprogramming chapters in Advanced R.
Data masking makes data manipulation faster because it requires less
typing. In most (but not all1) base R functions you need to refer to
variables with $
, leading to code that repeats the name of
the data frame many times:
The dplyr equivalent of this code is more concise because data
masking allows you to need to type starwars
once:
The key idea behind data masking is that it blurs the line between the two different meanings of the word “variable”:
env-variables are “programming” variables that
live in an environment. They are usually created with
<-
.
data-variables are “statistical” variables that
live in a data frame. They usually come from data files
(e.g. .csv
, .xls
), or are created manipulating
existing variables.
To make those definitions a little more concrete, take this piece of code:
It creates a env-variable, df
, that contains two
data-variables, x
and y
. Then it extracts the
data-variable x
out of the env-variable df
using $
.
I think this blurring of the meaning of “variable” is a really nice
feature for interactive data analysis because it allows you to refer to
data-vars as is, without any prefix. And this seems to be fairly
intuitive since many newer R users will attempt to write
diamonds[x == 0 | y == 0, ]
.
Unfortunately, this benefit does not come for free. When you start to program with these tools, you’re going to have to grapple with the distinction. This will be hard because you’ve never had to think about it before, so it’ll take a while for your brain to learn these new concepts and categories. However, once you’ve teased apart the idea of “variable” into data-variable and env-variable, I think you’ll find it fairly straightforward to use.
The main challenge of programming with functions that use data masking arises when you introduce some indirection, i.e. when you want to get the data-variable from an env-variable instead of directly typing the data-variable’s name. There are two main cases:
When you have the data-variable in a function argument (i.e. an
env-variable that holds a promise2), you need to embrace the
argument by surrounding it in doubled braces, like
filter(df, {{ var }})
.
The following function uses embracing to create a wrapper around
summarise()
that computes the minimum and maximum values of
a variable, as well as the number of observations that were
summarised:
When you have an env-variable that is a character vector, you
need to index into the .data
pronoun with [[
,
like summarise(df, mean = mean(.data[[var]]))
.
The following example uses .data
to count the number of
unique values in each variable of mtcars
:
Note that .data
is not a data frame; it’s a special
construct, a pronoun, that allows you to access the current variables
either directly, with .data$x
or indirectly with
.data[[var]]
. Don’t expect other functions to work with
it.
Many data masking functions also use dynamic dots, which gives you
another useful feature: generating names programmatically by using
:=
instead of =
. There are two basics forms,
as illustrated below with tibble()
:
If you have the name in an env-variable, you can use glue syntax to interpolate in:
If the name should be derived from a data-variable in an argument, you can use embracing syntax:
Learn more in ?rlang::`dyn-dots`
.
Data masking makes it easy to compute on values within a dataset. Tidy selection is a complementary tool that makes it easy to work with the columns of a dataset.
Underneath all functions that use tidy selection is the tidyselect package. It provides a miniature domain specific language that makes it easy to select columns by name, position, or type. For example:
select(df, 1)
selects the first column;
select(df, last_col())
selects the last column.
select(df, c(a, b, c))
selects columns
a
, b
, and c
.
select(df, starts_with("a"))
selects all columns
whose name starts with “a”; select(df, ends_with("z"))
selects all columns whose name ends with “z”.
select(df, where(is.numeric))
selects all numeric
columns.
You can see more details in ?dplyr_tidy_select
.
As with data masking, tidy selection makes a common task easier at the cost of making a less common task harder. When you want to use tidy select indirectly with the column specification stored in an intermediate variable, you’ll need to learn some new tools. Again, there are two forms of indirection:
When you have the data-variable in an env-variable that is a function argument, you use the same technique as data masking: you embrace the argument by surrounding it in doubled braces.
The following function summarises a data frame by computing the mean of all variables selected by the user:
When you have an env-variable that is a character vector, you
need to use all_of()
or any_of()
depending on
whether you want the function to error if a variable is not found.
The following code uses all_of()
to select all of the
variables found in a character vector; then !
plus
all_of()
to select all of the variables not found
in a character vector:
The following examples solve a grab bag of common problems. We show you the minimum amount of code so that you can get the basic idea; most real problems will require more code or combining multiple techniques.
If you check the documentation, you’ll see that .data
never uses data masking or tidy select. That means you don’t need to do
anything special in your function:
If you want the user to supply an expression that’s passed onto an argument which uses data masking or tidy select, embrace the argument:
my_summarise <- function(data, group_var) {
data %>%
group_by({{ group_var }}) %>%
summarise(mean = mean(mass))
}
This generalises in a straightforward way if you want to use one user-supplied expression in multiple places:
my_summarise2 <- function(data, expr) {
data %>% summarise(
mean = mean({{ expr }}),
sum = sum({{ expr }}),
n = n()
)
}
If you want the user to provide multiple expressions, embrace each of them:
my_summarise3 <- function(data, mean_var, sd_var) {
data %>%
summarise(mean = mean({{ mean_var }}), sd = sd({{ sd_var }}))
}
If you want to use the name of a variable in the output, you can
embrace the variable name on the left-hand side of :=
with
{{
:
my_summarise4 <- function(data, expr) {
data %>% summarise(
"mean_{{expr}}" := mean({{ expr }}),
"sum_{{expr}}" := sum({{ expr }}),
"n_{{expr}}" := n()
)
}
my_summarise5 <- function(data, mean_var, sd_var) {
data %>%
summarise(
"mean_{{mean_var}}" := mean({{ mean_var }}),
"sd_{{sd_var}}" := sd({{ sd_var }})
)
}
If you want to take an arbitrary number of user supplied expressions,
use ...
. This is most often useful when you want to give
the user full control over a single part of the pipeline, like a
group_by()
or a mutate()
.
my_summarise <- function(.data, ...) {
.data %>%
group_by(...) %>%
summarise(mass = mean(mass, na.rm = TRUE), height = mean(height, na.rm = TRUE))
}
starwars %>% my_summarise(homeworld)
#> # A tibble: 49 × 3
#> homeworld mass height
#> <chr> <dbl> <dbl>
#> 1 Alderaan 64 176.
#> 2 Aleen Minor 15 79
#> 3 Bespin 79 175
#> 4 Bestine IV 110 180
#> # ℹ 45 more rows
starwars %>% my_summarise(sex, gender)
#> `summarise()` has grouped output by 'sex'. You can override using the `.groups`
#> argument.
#> # A tibble: 6 × 4
#> # Groups: sex [5]
#> sex gender mass height
#> <chr> <chr> <dbl> <dbl>
#> 1 female feminine 54.7 172.
#> 2 hermaphroditic masculine 1358 175
#> 3 male masculine 80.2 179.
#> 4 none feminine NaN 96
#> # ℹ 2 more rows
When you use ...
in this way, make sure that any other
arguments start with .
to reduce the chances of argument
clashes; see https://design.tidyverse.org/dots-prefix.html for more
details.
Sometimes it can be useful for a single expression to return multiple columns. You can do this by returning an unnamed data frame:
quantile_df <- function(x, probs = c(0.25, 0.5, 0.75)) {
tibble(
val = quantile(x, probs),
quant = probs
)
}
x <- 1:5
quantile_df(x)
#> # A tibble: 3 × 2
#> val quant
#> <dbl> <dbl>
#> 1 2 0.25
#> 2 3 0.5
#> 3 4 0.75
This sort of function is useful inside summarise()
and
mutate()
which allow you to add multiple columns by
returning a data frame:
df <- tibble(
grp = rep(1:3, each = 10),
x = runif(30),
y = rnorm(30)
)
df %>%
group_by(grp) %>%
summarise(quantile_df(x, probs = .5))
#> # A tibble: 3 × 3
#> grp val quant
#> <int> <dbl> <dbl>
#> 1 1 0.361 0.5
#> 2 2 0.541 0.5
#> 3 3 0.456 0.5
df %>%
group_by(grp) %>%
summarise(across(x:y, ~ quantile_df(.x, probs = .5), .unpack = TRUE))
#> # A tibble: 3 × 5
#> grp x_val x_quant y_val y_quant
#> <int> <dbl> <dbl> <dbl> <dbl>
#> 1 1 0.361 0.5 0.174 0.5
#> 2 2 0.541 0.5 -0.0110 0.5
#> 3 3 0.456 0.5 0.0583 0.5
Notice that we set .unpack = TRUE
inside
across()
. This tells across()
to
unpack the data frame returned by quantile_df()
into its respective columns, combining the column names of the original
columns (x
and y
) with the column names
returned from the function (val
and
quant
).
If your function returns multiple rows per group, then
you’ll need to switch from summarise()
to
reframe()
. summarise()
is restricted to
returning 1 row summaries per group, but reframe()
lifts
this restriction:
If you want the user to provide a set of data-variables that are then
transformed, use across()
and pick()
:
my_summarise <- function(data, summary_vars) {
data %>%
summarise(across({{ summary_vars }}, ~ mean(., na.rm = TRUE)))
}
starwars %>%
group_by(species) %>%
my_summarise(c(mass, height))
#> # A tibble: 38 × 3
#> species mass height
#> <chr> <dbl> <dbl>
#> 1 Aleena 15 79
#> 2 Besalisk 102 198
#> 3 Cerean 82 198
#> 4 Chagrian NaN 196
#> # ℹ 34 more rows
You can use this same idea for multiple sets of input data-variables:
my_summarise <- function(data, group_var, summarise_var) {
data %>%
group_by(pick({{ group_var }})) %>%
summarise(across({{ summarise_var }}, mean))
}
Use the .names
argument to across()
to
control the names of the output.
If you have a character vector of variable names, and want to operate
on them with a for loop, index into the special .data
pronoun:
This same technique works with for loop alternatives like the base R
apply()
family and the purrr map()
family:
(Note that the x
in .data[[x]]
is always
treated as an env-variable; it will never come from the data.)
Many Shiny input controls return character vectors, so you can use
the same approach as above: .data[[input$var]]
.
library(shiny)
ui <- fluidPage(
selectInput("var", "Variable", choices = names(diamonds)),
tableOutput("output")
)
server <- function(input, output, session) {
data <- reactive(filter(diamonds, .data[[input$var]] > 0))
output$output <- renderTable(head(data()))
}
See https://mastering-shiny.org/action-tidy.html for more details and case studies.
dplyr’s filter()
is inspired by base R’s
subset()
. subset()
provides data masking, but
not with tidy evaluation, so the techniques described in this chapter
don’t apply to it.↩︎
In R, arguments are lazily evaluated which means that until you attempt to use, they don’t hold a value, just a promise that describes how to compute the value. You can learn more at https://adv-r.hadley.nz/functions.html#lazy-evaluation↩︎