Can dplyr package be used for conditional mutating

Question

Can the mutate be used when the mutation is conditional (depending on the values of certain column values)?

This example helps showing what I mean.

structure(list(a = c(1, 3, 4, 6, 3, 2, 5, 1), b = c(1, 3, 4, 
2, 6, 7, 2, 6), c = c(6, 3, 6, 5, 3, 6, 5, 3), d = c(6, 2, 4, 
5, 3, 7, 2, 6), e = c(1, 2, 4, 5, 6, 7, 6, 3), f = c(2, 3, 4, 
2, 2, 7, 5, 2)), .Names = c("a", "b", "c", "d", "e", "f"), row.names = c(NA, 
8L), class = "data.frame")

  a b c d e f
1 1 1 6 6 1 2
2 3 3 3 2 2 3
3 4 4 6 4 4 4
4 6 2 5 5 5 2
5 3 6 3 3 6 2
6 2 7 6 7 7 7
7 5 2 5 2 6 5
8 1 6 3 6 3 2

I was hoping to find a solution to my problem using the dplyr package (and yes I know this not code that should work, but I guess it makes the purpose clear) for creating a new column g:

 library(dplyr)
 df <- mutate(df,
         if (a == 2 | a == 5 | a == 7 | (a == 1 & b == 4)){g = 2},
         if (a == 0 | a == 1 | a == 4 | a == 3 |  c == 4) {g = 3})

The result of the code I am looking for should have this result in this particular example:

  a b c d e f  g
1 1 1 6 6 1 2  3
2 3 3 3 2 2 3  3
3 4 4 6 4 4 4  3
4 6 2 5 5 5 2 NA
5 3 6 3 3 6 2 NA
6 2 7 6 7 7 7  2
7 5 2 5 2 6 5  2
8 1 6 3 6 3 2  3

Does anyone have an idea about how to do this in dplyr? This data frame is just an example, the data frames I am dealing with are much larger. Because of its speed I tried to use dplyr, but perhaps there are other, better ways to handle this problem?

Answer 1

Since you ask for other better ways to handle the problem, here's another way using data.table:

require(data.table) ## 1.9.2+
setDT(df)
df[a %in% c(0,1,3,4) | c == 4, g := 3L]
df[a %in% c(2,5,7) | (a==1 & b==4), g := 2L]

Note the order of conditional statements is reversed to get g correctly. There's no copy of g made, even during the second assignment - it's replaced in-place.

On larger data this would have better performance than using nested if-else, as it can evaluate both 'yes' and 'no' cases, and nesting can get harder to read/maintain IMHO.

---

Here's a benchmark on relatively bigger data:

# R version 3.1.0
require(data.table) ## 1.9.2
require(dplyr)
DT <- setDT(lapply(1:6, function(x) sample(7, 1e7, TRUE)))
setnames(DT, letters[1:6])
# > dim(DT) 
# [1] 10000000        6
DF <- as.data.frame(DT)

DT_fun <- function(DT) {
    DT[(a %in% c(0,1,3,4) | c == 4), g := 3L]
    DT[a %in% c(2,5,7) | (a==1 & b==4), g := 2L]
}

DPLYR_fun <- function(DF) {
    mutate(DF, g = ifelse(a %in% c(2,5,7) | (a==1 & b==4), 2L, 
            ifelse(a %in% c(0,1,3,4) | c==4, 3L, NA_integer_)))
}

BASE_fun <- function(DF) { # R v3.1.0
    transform(DF, g = ifelse(a %in% c(2,5,7) | (a==1 & b==4), 2L, 
            ifelse(a %in% c(0,1,3,4) | c==4, 3L, NA_integer_)))
}

system.time(ans1 <- DT_fun(DT))
#   user  system elapsed 
#  2.659   0.420   3.107 

system.time(ans2 <- DPLYR_fun(DF))
#   user  system elapsed 
# 11.822   1.075  12.976 

system.time(ans3 <- BASE_fun(DF))
#   user  system elapsed 
# 11.676   1.530  13.319 

identical(as.data.frame(ans1), as.data.frame(ans2))
# [1] TRUE

identical(as.data.frame(ans1), as.data.frame(ans3))
# [1] TRUE

Not sure if this is an alternative you'd asked for, but I hope it helps.

Answer 2

Use ifelse

df %>%
  mutate(g = ifelse(a == 2 | a == 5 | a == 7 | (a == 1 & b == 4), 2,
               ifelse(a == 0 | a == 1 | a == 4 | a == 3 |  c == 4, 3, NA)))

Added - if_else: Note that in dplyr 0.5 there is an if_else function defined so an alternative would be to replace ifelse with if_else; however, note that since if_else is stricter than ifelse (both legs of the condition must have the same type) so the NA in that case would have to be replaced with NA_real_ .

df %>%
  mutate(g = if_else(a == 2 | a == 5 | a == 7 | (a == 1 & b == 4), 2,
               if_else(a == 0 | a == 1 | a == 4 | a == 3 |  c == 4, 3, NA_real_)))

Added - case_when Since this question was posted dplyr has added case_when so another alternative would be:

df %>% mutate(g = case_when(a == 2 | a == 5 | a == 7 | (a == 1 & b == 4) ~ 2,
                            a == 0 | a == 1 | a == 4 | a == 3 |  c == 4 ~ 3,
                            TRUE ~ NA_real_))

Answer 3

case_when is now a pretty clean implementation of the SQL-style case when:

structure(list(a = c(1, 3, 4, 6, 3, 2, 5, 1), b = c(1, 3, 4, 
2, 6, 7, 2, 6), c = c(6, 3, 6, 5, 3, 6, 5, 3), d = c(6, 2, 4, 
5, 3, 7, 2, 6), e = c(1, 2, 4, 5, 6, 7, 6, 3), f = c(2, 3, 4, 
2, 2, 7, 5, 2)), .Names = c("a", "b", "c", "d", "e", "f"), row.names = c(NA, 
8L), class = "data.frame") -> df


df %>% 
    mutate( g = case_when(
                a == 2 | a == 5 | a == 7 | (a == 1 & b == 4 )     ~   2,
                a == 0 | a == 1 | a == 4 |  a == 3 | c == 4       ~   3
))

Using dplyr 0.7.4

The manual: http://dplyr.tidyverse.org/reference/case_when.html

Answer 4

dplyr now has a function case_when that offers a vectorised if. The syntax is a little strange compared to mosaic:::derivedFactor as you cannot access variables in the standard dplyr way, and need to declare the mode of NA, but it is considerably faster than mosaic:::derivedFactor.

df %>%
mutate(g = case_when(a %in% c(2,5,7) | (a==1 & b==4) ~ 2L, 
                     a %in% c(0,1,3,4) | c == 4 ~ 3L, 
                     TRUE~as.integer(NA)))

EDIT: If you're using dplyr::case_when() from before version 0.7.0 of the package, then you need to precede variable names with '.$' (e.g. write .$a == 1 inside case_when).

Benchmark: For the benchmark (reusing functions from Arun 's post) and reducing sample size:

require(data.table) 
require(mosaic) 
require(dplyr)
require(microbenchmark)

DT <- setDT(lapply(1:6, function(x) sample(7, 10000, TRUE)))
setnames(DT, letters[1:6])
DF <- as.data.frame(DT)

DPLYR_case_when <- function(DF) {
  DF %>%
  mutate(g = case_when(a %in% c(2,5,7) | (a==1 & b==4) ~ 2L, 
                       a %in% c(0,1,3,4) | c==4 ~ 3L, 
                       TRUE~as.integer(NA)))
}

DT_fun <- function(DT) {
  DT[(a %in% c(0,1,3,4) | c == 4), g := 3L]
  DT[a %in% c(2,5,7) | (a==1 & b==4), g := 2L]
}

DPLYR_fun <- function(DF) {
  mutate(DF, g = ifelse(a %in% c(2,5,7) | (a==1 & b==4), 2L, 
                    ifelse(a %in% c(0,1,3,4) | c==4, 3L, NA_integer_)))
}

mosa_fun <- function(DF) {
  mutate(DF, g = derivedFactor(
    "2" = (a == 2 | a == 5 | a == 7 | (a == 1 & b == 4)),
    "3" = (a == 0 | a == 1 | a == 4 | a == 3 |  c == 4),
    .method = "first",
    .default = NA
  ))
}

microbenchmark(
  DT_fun(DT),
  DPLYR_fun(DF),
  DPLYR_case_when(DF),
  mosa_fun(DF),
  times=20
)

This gives:

            expr        min         lq       mean     median         uq        max neval
         DT_fun(DT)   1.503589   1.626971   2.054825   1.755860   2.292157   3.426192    20
      DPLYR_fun(DF)   2.420798   2.596476   3.617092   3.484567   4.184260   6.235367    20
DPLYR_case_when(DF)   2.153481   2.252134   6.124249   2.365763   3.119575  72.344114    20
       mosa_fun(DF) 396.344113 407.649356 413.743179 412.412634 416.515742 459.974969    20

Answer 5

The derivedFactor function from mosaic package seems to be designed to handle this. Using this example, it would look like:

library(dplyr)
library(mosaic)
df <- mutate(df, g = derivedFactor(
     "2" = (a == 2 | a == 5 | a == 7 | (a == 1 & b == 4)),
     "3" = (a == 0 | a == 1 | a == 4 | a == 3 |  c == 4),
     .method = "first",
     .default = NA
     ))

(If you want the result to be numeric instead of a factor, you can wrap derivedFactor in an as.numeric call.)

derivedFactor can be used for an arbitrary number of conditionals, too.