Pass arguments to dplyr functions

Question

I want to parameterise the following computation using dplyr that finds which values of Sepal.Length are associated with more than one value of Sepal.Width:

library(dplyr)

iris %>%
    group_by(Sepal.Length) %>%
    summarise(n.uniq=n_distinct(Sepal.Width)) %>%
    filter(n.uniq > 1)

Normally I would write something like this:

not.uniq.per.group <- function(data, group.var, uniq.var) {
    iris %>%
        group_by(group.var) %>%
        summarise(n.uniq=n_distinct(uniq.var)) %>%
        filter(n.uniq > 1)
}

However, this approach throws errors because dplyr uses non-standard evaluation. How should this function be written?

Answer 1

Here's the way to do it from rlang 0.4 using curly curly {{ pseudo operator :

library(dplyr)

not.uniq.per.group <- function(data, group.var, uniq.var) {
  data %>%
    group_by({{ group.var }}) %>%
    summarise(n.uniq = n_distinct({{ uniq.var }})) %>%
    filter(n.uniq > 1)
}

iris %>% not.uniq.per.group(Sepal.Length, Sepal.Width)
#> # A tibble: 25 x 2
#>    Sepal.Length n.uniq
#>           <dbl>  <int>
#>  1          4.4      3
#>  2          4.6      4
#>  3          4.8      3
#>  4          4.9      5
#>  5          5        8
#>  6          5.1      6
#>  7          5.2      4
#>  8          5.4      4
#>  9          5.5      6
#> 10          5.6      5
#> # ... with 15 more rows

Answer 2

Like the old dplyr versions up to 0.5, the new dplyr has facilities for both standard evaluation (SE) and nonstandard evaluation (NSE). But they are expressed differently than before.

If you want an NSE function, you pass bare expressions and use enquo to capture them as quosures. If you want an SE function, just pass quosures (or symbols) directly, then unquote them in the dplyr calls. Here is the SE solution to the question:

library(tidyverse)
library(rlang)

f1 <- function(df, grp.var, uniq.var) {
   df %>%
       group_by(!!grp.var) %>%
       summarise(n_uniq = n_distinct(!!uniq.var)) %>%
       filter(n_uniq > 1)  
}

a <- f1(iris, quo(Sepal.Length), quo(Sepal.Width))
b <- f1(iris, sym("Sepal.Length"), sym("Sepal.Width"))
identical(a, b)
#> [1] TRUE

Note how the SE version enables you to work with string arguments - just turn them into symbols first using sym(). For more information, see the programming with dplyr vignette.

Answer 3

You need to use the standard evaluation versions of the dplyr functions (just append '_' to the function names, ie. group_by_ & summarise_) and pass strings to your function, which you then need to turn into symbols. To parameterise the argument of summarise_, you will need to use interp(), which is defined in the lazyeval package. Concretely:

library(dplyr)
library(lazyeval)

not.uniq.per.group <- function(df, grp.var, uniq.var) {
    df %>%
        group_by_(grp.var) %>%
        summarise_( n_uniq=interp(~n_distinct(v), v=as.name(uniq.var)) ) %>%
        filter(n_uniq > 1)
}

not.uniq.per.group(iris, "Sepal.Length", "Sepal.Width")

---

Note that in recent versions of dplyr the standard evaluation versions of the dplyr functions have been "soft deprecated" in favor of non-standard evaluation.

See the Programming with dplyr vignette for more information on working with non-standard evaluation.

Answer 4

In the current version of dplyr (0.7.4) the use of the standard evaluation function versions (appended '_' to the function name, e.g. group_by_) is deprecated. Instead you should rely on tidyeval when writing functions.

Here's an example of how your function would look then:

# definition of your function
not.uniq.per.group <- function(data, group.var, uniq.var) {
  # enquotes variables to be used with dplyr-functions
  group.var <- enquo(group.var)
  uniq.var <- enquo(uniq.var)

  # use '!!' before parameter names in dplyr-functions
  data %>%
    group_by(!!group.var) %>%
    summarise(n.uniq=n_distinct(!!uniq.var)) %>%
    filter(n.uniq > 1)
}

# call of your function
not.uniq.per.group(iris, Sepal.Length, Sepal.Width)

If you want to learn all about the details, there's an excellent vignette by the dplyr-team on how this works.

Answer 5

In the devel version of dplyr (soon to be released 0.6.0), we can also make use of slightly different syntax for passing the variables.

f1 <- function(df, grp.var, uniq.var) {
   grp.var <- enquo(grp.var)
   uniq.var <- enquo(uniq.var)

   df %>%
       group_by(!!grp.var) %>%
       summarise(n_uniq = n_distinct(!!uniq.var)) %>%
       filter(n_uniq >1)  


}

res2 <- f1(iris, Sepal.Length, Sepal.Width) 
res1 <- not.uniq.per.group(iris, "Sepal.Length", "Sepal.Width")
identical(res1, res2)
#[1] TRUE

Here enquo takes the arguments and returns the value as a quosure (similar to substitute in base R) by evaluating the function arguments lazily and inside the summarise, we ask it to unquote (!! or UQ) so that it gets evaluated.

Answer 6

You can avoid lazyeval by using do to call an anonymous function and then using get. This solution can be used more generally to employ multiple aggregations. I usually write the function separately.

library(dplyr)

not.uniq.per.group <- function(df, grp.var, uniq.var) {
  df %>%
    group_by_(grp.var) %>%
    do((function(., uniq.var) {
      with(., data.frame(n_uniq = n_distinct(get(uniq.var))))
    }      
  )(., uniq.var)) %>%
  filter(n_uniq > 1)
}

not.uniq.per.group(iris, "Sepal.Length", "Sepal.Width")

Answer 7

I've written a function in the past that does something similar to what you're doing, except that it explores all the columns outside the primary key and looks for multiple unique values per group.

find_dups = function(.table, ...) {
  require(dplyr)
  require(tidyr)
  # get column names of primary key
  pk <- .table %>% select(...) %>% names
  other <- names(.table)[!(names(.table) %in% pk)]
  # group by primary key,
  # get number of rows per unique combo,
  # filter for duplicates,
  # get number of distinct values in each column,
  # gather to get df of 1 row per primary key, other column,
  # filter for where a columns have more than 1 unique value,
  # order table by primary key
  .table %>%
    group_by(...) %>%
    mutate(cnt = n()) %>%
    filter(cnt > 1) %>%
    select(-cnt) %>%
    summarise_each(funs(n_distinct)) %>%
    gather_('column', 'unique_vals', other) %>%
    filter(unique_vals > 1) %>%
    arrange(...) %>%
    return
  # Final dataframe:
  ## One row per primary key and column that creates duplicates.
  ## Last column indicates how many unique values of
  ## the given column exist for each primary key.
}

This function also works with the piping operator:

dat %>% find_dups(key1, key2)