tl;dr

90% of my usecases when putting a function that uses tidyverse with non-standard evaluation (NSE) into an R package can be resolved by importFrom rlang .data. In these cases, I’m only looking for a way to be able to use tidyverse internally in the package, and have it pass the R package check, rather than allowing the user to supply function arguments in the NSE form.

Why bother?

I don’t think there is any reason to specifically use tidyverse (dplyr, tidyr, ggplot2 and so on) in R packages, and I’d prefer base R solutions. However, if one already uses tidyverse on a daily basis and has created a bunch of functions that she wants to make into a package, this seems a reasonable usecase.

However, I can’t use tidyverse functions in a package the same way as I use them interactively. Otherwise when I run package check (devtools::check() or R CMD check), it will complain because I have used NSE in my functions. For example, this function:

#' Using non-standard evaluation in a function
#' Results in NOTE: no visible binding for global variable "state"
#'
#' @param df A tibble
#' @return A tibble
#' @importFrom dplyr filter
#' @export
filter_nse <- function(df) {
  filter(df, state == "treated")
}

when included in a dummy package pkgsandbox, leads to the following NOTE during package check.

checking R code for possible problems ... NOTE
filter_nse: no visible binding for global variable ‘state’
Undefined global functions or variables:
  state

.data

To suppress it, to my knowledge the way to go as of now (end of 2018) is to use .data from rlang. The function looks like this:

#' Using .data in a function, existing column
#'
#' @param df A tibble
#' @return A tibble
#' @importFrom dplyr filter
#' @importFrom rlang .data
#' @export
filter_dotdata <- function(df) {
  filter(df, .data$state == "treated")
}

Note that our filter_* functions are only intended to work with data Puromycin because of the specific column state:

head(Puromycin, n=4)

##   conc rate   state
## 1 0.02   76 treated
## 2 0.02   47 treated
## 3 0.06   97 treated
## 4 0.06  107 treated

table(Puromycin$state)

## 
##   treated untreated 
##        12        11

As expected, this works:

res <- pkgsandbox::filter_dotdata(Puromycin)
table(res$state)

## 
##   treated untreated 
##        12         0

Alternatives and problems

#' Using !! sym in a function, existing column
#'
#' @param df A tibble
#' @return A tibble
#' @importFrom dplyr filter
#' @importFrom rlang !! sym
#' @export
filter_bangbangsym <- function(df) {
  filter(df, !!sym("state") == "treated")
}

res <- pkgsandbox::filter_bangbangsym(Puromycin)
table(res$state)

## 
##   treated untreated 
##        12         0

state <- "North Carolina"
pkgsandbox::filter_dotdata(iris)

## Column `state` not found in `.data`

state <- "North Carolina"
res_wrong_iris1 <- pkgsandbox::filter_bangbangsym(iris)
nrow(res_wrong_iris1) # 0 because "North Carolina" != "treated"

## [1] 0

state <- "treated"
res_wrong_iris2 <- pkgsandbox::filter_bangbangsym(iris)
nrow(res_wrong_iris2) # this didn't apply any filtering  because "treated" == "treated"

## [1] 150

state <- "treated"
res_wrong_iris3 <- pkgsandbox::filter_nse(iris)
nrow(res_wrong_iris3)

## [1] 150