How do you specify of possible arguments states in functions using non-standard evaluation?

Question

I am learning about programming with tidy evaluation and non-standard evaluation and have been trying to work out how to constrain the possible states of an argument in a function.

For instance given a data set:

set.seed(123)
data <- data_frame(GROUP_ONE = sample(LETTERS[1:3], 10, replace = TRUE), 
                   GROUP_TWO = sample(letters[4:6], 10, replace = TRUE), 
                   result = rnorm(10))

I can create a function which has an argument I use to group the data using a quosure like so:

my_function <- function(data, group = GROUP_ONE){

  require(dplyr)
  require(magrittr)

  group <- enquo(group)

  result <- data %>% 
    group_by(!!group) %>% 
    summarise(mean=mean(result))

  return(result)
}

and this does what I want

my_function(data)

# A tibble: 3 x 2
  GROUP_ONE       mean
      <chr>      <dbl>
1         A  1.5054975
2         B  0.2817966
3         C -0.5129904

and I can supply a different group:

my_function(data, group = GROUP_TWO)

# A tibble: 3 x 2
  GROUP_TWO       mean
      <chr>      <dbl>
1         d -0.3308130
2         e  0.2352483
3         f  0.7347437

However, I cannot group by a column for which is not present in the data.

e.g.

 my_function(data, group = GROUP_THREE)

Error in grouped_df_impl(data, unname(vars), drop) : Column GROUP_THREE is unknown

I would like to add a step at the beginning of the function so that the function stops with a custom error message if the group argument is not GROUP_ONE or GROUP_TWO

something like:

if(!group %in% c(~GROUP_ONE, ~GROUP_TWO)) stop("CUSTOM ERROR MESSAGE")

except this does not work as you apparently you can't put quosures in a vector. It should be possible to convert the quosure to a string somehow and have a vector of strings but I can't figure out how.

How is this done?

Answer 1

I think you need quo_name (from dplyr or rlang), which transforms a quoted symbol to a string:

my_function <- function(data, group = GROUP_ONE){

    require(dplyr)
    require(magrittr)

    group <- enquo(group)

    if(!quo_name(group) %in% names(data)) stop("CUSTOM ERROR MESSAGE")

    result <- data %>% 
        group_by(!!group) %>% 
        summarise(mean=mean(result))

    return(result)
}

# > my_function(data, GROUP_THREE)
# Error in my_function(data, GROUP_THREE) : CUSTOM ERROR MESSAGE

---

Edit

As noted by lionel in comment: except for quo_name, there are many other alternatives including base R as.character and as_string from rlang.

Answer 2

quo_name() is for transforming arbitrary expressions to text so that isn't robust for checking symbols.

If you expect only symbols, and if those symbols should only represent data frames columns, you don't need quosures. In this case you can capture with enexpr() (and there will be ensym() in the next version of rlang):

group <- enexpr(group)
stopifnot(is_symbol(group))  # Or some custom error

Then turn it to a string for the check:

as_string(group) %in% names

You can then unquote the symbol just like you unquote the quosure.

df %>% group_by(!! group)

Alternatively if you need quosures you can check the contained expression:

expr <- get_expr(quo)
is_symbol(expr) && as_string(expr) %in% names

That should be the preferred UI because group_by() has mutate semantics, so you can do stuff like this: df %>% group_by(as.factor(col)). This also means that it's hopeless to try to provide custom error messages, unless you want to capture the error, parse it to make sure it's a "symbol not found" one, and rethrow another error.