Using dplyr window functions to calculate percentiles

Question

I have a working solution but am looking for a cleaner, more readable solution that perhaps takes advantage of some of the newer dplyr window functions.

Using the mtcars dataset, if I want to look at the 25th, 50th, 75th percentiles and the mean and count of miles per gallon ("mpg") by the number of cylinders ("cyl"), I use the following code:

library(dplyr)
library(tidyr)

# load data
data("mtcars")

# Percentiles used in calculation
p <- c(.25,.5,.75)

# old dplyr solution 
mtcars %>% group_by(cyl) %>% 
  do(data.frame(p=p, stats=quantile(.$mpg, probs=p), 
                n = length(.$mpg), avg = mean(.$mpg))) %>%
  spread(p, stats) %>%
  select(1, 4:6, 3, 2)

# note: the select and spread statements are just to get the data into
#       the format in which I'd like to see it, but are not critical

Is there a way I can do this more cleanly with dplyr using some of the summary functions (n_tiles, percent_rank, etc.)? By cleanly, I mean without the "do" statement.

Thank you

Answer 1

In dplyr 1.0, summarise can return multiple values, allowing the following:

library(tidyverse)

mtcars %>% 
  group_by(cyl) %>%  
  summarise(quantile = scales::percent(c(0.25, 0.5, 0.75)),
            mpg = quantile(mpg, c(0.25, 0.5, 0.75)))

Or, you can avoid a separate line to name the quantiles by going with enframe:

mtcars %>% 
  group_by(cyl) %>%  
  summarise(enframe(quantile(mpg, c(0.25, 0.5, 0.75)), "quantile", "mpg"))

    cyl quantile   mpg
  <dbl> <chr>    <dbl>
1     4 25%       22.8
2     4 50%       26  
3     4 75%       30.4
4     6 25%       18.6
5     6 50%       19.7
6     6 75%       21  
7     8 25%       14.4
8     8 50%       15.2
9     8 75%       16.2

NOTE: As of dplyr 1.1.0, returning multiple rows per group with summarise is deprecated. Instead, use reframe, as in:

mtcars %>% 
 group_by(cyl) %>%  
 reframe(enframe(quantile(mpg, c(0.25, 0.5, 0.75)), "quantile", "mpg"))

---

Answer for previous versions of dplyr

library(tidyverse)

mtcars %>% 
  group_by(cyl) %>% 
  summarise(x=list(enframe(quantile(mpg, probs=c(0.25,0.5,0.75)), "quantiles", "mpg"))) %>% 
  unnest(x)

    cyl quantiles   mpg
1     4       25% 22.80
2     4       50% 26.00
3     4       75% 30.40
4     6       25% 18.65
5     6       50% 19.70
6     6       75% 21.00
7     8       25% 14.40
8     8       50% 15.20
9     8       75% 16.25

This can be turned into a more general function using tidyeval:

q_by_group = function(data, value.col, ..., probs=seq(0,1,0.25)) {

  groups=enquos(...)
  
  data %>% 
    group_by(!!!groups) %>% 
    summarise(x = list(enframe(quantile({{value.col}}, probs=probs), "quantiles", "mpg"))) %>% 
    unnest(x)
}

q_by_group(mtcars, mpg)
q_by_group(mtcars, mpg, cyl)
q_by_group(mtcars, mpg, cyl, vs, probs=c(0.5,0.75))
q_by_group(iris, Petal.Width, Species)

Answer 2

If you're up for using purrr::map, you can do it like this!

library(tidyverse)

mtcars %>%
  tbl_df() %>%
  nest(-cyl) %>%
  mutate(Quantiles = map(data, ~ quantile(.$mpg)),
         Quantiles = map(Quantiles, ~ bind_rows(.) %>% gather())) %>% 
  unnest(Quantiles)

#> # A tibble: 15 x 3
#>      cyl key   value
#>    <dbl> <chr> <dbl>
#>  1     6 0%     17.8
#>  2     6 25%    18.6
#>  3     6 50%    19.7
#>  4     6 75%    21  
#>  5     6 100%   21.4
#>  6     4 0%     21.4
#>  7     4 25%    22.8
#>  8     4 50%    26  
#>  9     4 75%    30.4
#> 10     4 100%   33.9
#> 11     8 0%     10.4
#> 12     8 25%    14.4
#> 13     8 50%    15.2
#> 14     8 75%    16.2
#> 15     8 100%   19.2

^{Created on 2018-11-10 by the reprex package (v0.2.1)}

One nice thing about this approach is the output is tidy, one observation per row.

Answer 3

This is a dplyr approach that uses the tidy() function of the broom package, unfortunately it still requires do(), but it is a lot simpler.

library(dplyr)
library(broom)

mtcars %>%
    group_by(cyl) %>%
    do( tidy(t(quantile(.$mpg))) )

which gives:

    cyl   X0.  X25.  X50.  X75. X100.
  (dbl) (dbl) (dbl) (dbl) (dbl) (dbl)
1     4  21.4 22.80  26.0 30.40  33.9
2     6  17.8 18.65  19.7 21.00  21.4
3     8  10.4 14.40  15.2 16.25  19.2

Note the use of t() since the broom package does not have a method for named numerics.

This is based on my earlier answer for summary() here.

Answer 4

Not sure how to avoid do() in dplyr, but you can do this with c() and as.list() with data.table in a pretty straightforward manner:

require(data.table) 
as.data.table(mtcars)[, c(as.list(quantile(mpg, probs=p)), 
                        avg=mean(mpg), n=.N), by=cyl]
#    cyl   25%  50%   75%      avg  n
# 1:   6 18.65 19.7 21.00 19.74286  7
# 2:   4 22.80 26.0 30.40 26.66364 11
# 3:   8 14.40 15.2 16.25 15.10000 14

Replace by with keyby if you want them ordered by cyl column.

Answer 5

Answered many diffrent ways. dplyr distinct made the difference for what I wanted to do..

mtcars %>%
   select(cyl, mpg) %>%
   group_by(cyl) %>%
   mutate( qnt_0   = quantile(mpg, probs= 0),
           qnt_25  = quantile(mpg, probs= 0.25),
           qnt_50  = quantile(mpg, probs= 0.5),
           qnt_75  = quantile(mpg, probs= 0.75),
           qnt_100 = quantile(mpg, probs= 1),
              mean = mean(mpg),
                sd = sd(mpg)
          ) %>%
   distinct(qnt_0 ,qnt_25 ,qnt_50 ,qnt_75 ,qnt_100 ,mean ,sd)

renders

# A tibble: 3 x 8
# Groups:   cyl [3]
  qnt_0 qnt_25 qnt_50 qnt_75 qnt_100  mean    sd   cyl
  <dbl>  <dbl>  <dbl>  <dbl>   <dbl> <dbl> <dbl> <dbl>
1  17.8   18.6   19.7   21      21.4  19.7  1.45     6
2  21.4   22.8   26     30.4    33.9  26.7  4.51     4
3  10.4   14.4   15.2   16.2    19.2  15.1  2.56     8

Answer 6

This solution uses dplyr and tidyr only, lets you specify your quantiles in the dplyr chain, and takes advantage of tidyr::crossing() to "stack" multiple copies of the dataset prior to grouping and summarising.

diamonds %>%  # Initial data
  tidyr::crossing(pctile = 0:4/4) %>%  # Specify quantiles; crossing() is like expand.grid()
  dplyr::group_by(cut, pctile) %>%  # Indicate your grouping var, plus your quantile var
  dplyr::summarise(quantile_value = quantile(price, unique(pctile))) %>%  # unique() is needed
  dplyr::mutate(pctile = sprintf("%1.0f%%", pctile*100))  # Optional prettification

Result:

# A tibble: 25 x 3
# Groups:   cut [5]
         cut pctile quantile_value
       <ord>  <chr>          <dbl>
 1      Fair     0%         337.00
 2      Fair    25%        2050.25
 3      Fair    50%        3282.00
 4      Fair    75%        5205.50
 5      Fair   100%       18574.00
 6      Good     0%         327.00
 7      Good    25%        1145.00
 8      Good    50%        3050.50
 9      Good    75%        5028.00
10      Good   100%       18788.00
11 Very Good     0%         336.00
12 Very Good    25%         912.00
13 Very Good    50%        2648.00
14 Very Good    75%        5372.75
15 Very Good   100%       18818.00
16   Premium     0%         326.00
17   Premium    25%        1046.00
18   Premium    50%        3185.00
19   Premium    75%        6296.00
20   Premium   100%       18823.00
21     Ideal     0%         326.00
22     Ideal    25%         878.00
23     Ideal    50%        1810.00
24     Ideal    75%        4678.50
25     Ideal   100%       18806.00

The unique() is necessary to let dplyr::summarise() know that you only want one value per group.

Answer 7

Here is a solution using a combination of dplyr, purrr, and rlang:

library(dplyr)
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
library(tidyr)
library(purrr)

# load data
data("mtcars")

# Percentiles used in calculation
p <- c(.25,.5,.75)

p_names <- paste0(p*100, "%")
p_funs <- map(p, ~partial(quantile, probs = .x, na.rm = TRUE)) %>% 
  set_names(nm = p_names)

# dplyr/purrr/rlang solution 
mtcars %>% 
  group_by(cyl) %>% 
  summarize_at(vars(mpg), funs(!!!p_funs))
#> # A tibble: 3 x 4
#>     cyl `25%` `50%` `75%`
#>   <dbl> <dbl> <dbl> <dbl>
#> 1     4  22.8  26    30.4
#> 2     6  18.6  19.7  21  
#> 3     8  14.4  15.2  16.2


#Especially useful if you want to summarize more variables
mtcars %>% 
  group_by(cyl) %>% 
  summarize_at(vars(mpg, drat), funs(!!!p_funs))
#> # A tibble: 3 x 7
#>     cyl `mpg_25%` `drat_25%` `mpg_50%` `drat_50%` `mpg_75%` `drat_75%`
#>   <dbl>     <dbl>      <dbl>     <dbl>      <dbl>     <dbl>      <dbl>
#> 1     4      22.8       3.81      26         4.08      30.4       4.16
#> 2     6      18.6       3.35      19.7       3.9       21         3.91
#> 3     8      14.4       3.07      15.2       3.12      16.2       3.22

Created on 2018-10-01 by the reprex package (v0.2.0).

Edit (2019-04-17):

As of dplyr 0.8.0, the funs function has been deprecated in favor of using list to pass the desired functions into scoped dplyr functions. As a result of this, the implementation above gets slightly more straightfoward. We no longer need to worry about unquoting the functions with the !!!. Please see the below reprex:

library(dplyr)
#> Warning: package 'dplyr' was built under R version 3.5.2
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
library(tidyr)
library(purrr)

# load data
data("mtcars")

# Percentiles used in calculation
p <- c(.25,.5,.75)

p_names <- paste0(p*100, "%")
p_funs <- map(p, ~partial(quantile, probs = .x, na.rm = TRUE)) %>% 
  set_names(nm = p_names)

# dplyr/purrr/rlang solution 
mtcars %>% 
  group_by(cyl) %>% 
  summarize_at(vars(mpg), p_funs)
#> # A tibble: 3 x 4
#>     cyl `25%` `50%` `75%`
#>   <dbl> <dbl> <dbl> <dbl>
#> 1     4  22.8  26    30.4
#> 2     6  18.6  19.7  21  
#> 3     8  14.4  15.2  16.2


#Especially useful if you want to summarize more variables
mtcars %>% 
  group_by(cyl) %>% 
  summarize_at(vars(mpg, drat), p_funs)
#> # A tibble: 3 x 7
#>     cyl `mpg_25%` `drat_25%` `mpg_50%` `drat_50%` `mpg_75%` `drat_75%`
#>   <dbl>     <dbl>      <dbl>     <dbl>      <dbl>     <dbl>      <dbl>
#> 1     4      22.8       3.81      26         4.08      30.4       4.16
#> 2     6      18.6       3.35      19.7       3.9       21         3.91
#> 3     8      14.4       3.07      15.2       3.12      16.2       3.22

Created on 2019-04-17 by the reprex package (v0.2.0).

Answer 8

Yet another way to accomplish this, with unnest_wider/longer

    mtcars %>%
       group_by(cyl) %>%
       summarise(quants = list(quantile(mpg, probs = c(.01, .1, .25, .5, .75, .90,.99)))) %>%
       unnest_wider(quants)

And if you wanted to do it for multiple variables, you could gather before the grouping:

mtcars %>%
   gather(key = 'metric', value = 'value', -cyl) %>%
   group_by(cyl, metric) %>%
   summarise(quants = list(quantile(value, probs = c(.01, .1, .25, .5, .75, .90,.99)))) %>%
  unnest_wider(quants)

Answer 9

Here's a fairly readable solution that uses dplyr and purrr to return quantiles in a tidy format:

Code

library(dplyr)
library(purrr)

mtcars %>% 
    group_by(cyl) %>% 
    do({x <- .$mpg
        map_dfr(.x = c(.25, .5, .75),
                .f = ~ data_frame(Quantile = .x,
                                  Value = quantile(x, probs = .x)))
       })

Result

# A tibble: 9 x 3
# Groups:   cyl [3]
    cyl Quantile Value
  <dbl>    <dbl> <dbl>
1     4     0.25 22.80
2     4     0.50 26.00
3     4     0.75 30.40
4     6     0.25 18.65
5     6     0.50 19.70
6     6     0.75 21.00
7     8     0.25 14.40
8     8     0.50 15.20
9     8     0.75 16.25

Answer 10

do() is in fact the correct idiom, since it’s designed for group-wise transformations. Think of it as an lapply() that maps over groups of a data frame. (For such a specialized function, a generic name like “do” is not ideal. But it’s probably too late to change it.)

Morally, within each cyl group, you want to apply quantile() to the mpg column:

library(dplyr)

p <- c(.2, .5, .75)

mtcars %>% 
  group_by(cyl) %>%
  do(quantile(.$mpg, p))

#> Error: Results 1, 2, 3 must be data frames, not numeric

Except that doesn’t work because quantile() doesn’t return a data frame; you must convert its output, explicitly. Since this alteration amounts to wrapping quantile() with a data frame, you can use the gestalt function composition operator %>>>%:

library(gestalt)
library(tibble)

quantile_tbl <- quantile %>>>% enframe("quantile")

mtcars %>% 
  group_by(cyl) %>%
  do(quantile_tbl(.$mpg, p))

#> # A tibble: 9 x 3
#> # Groups:   cyl [3]
#>     cyl quantile value
#>   <dbl> <chr>    <dbl>
#> 1     4 20%       22.8
#> 2     4 50%       26  
#> 3     4 75%       30.4
#> 4     6 20%       18.3
#> 5     6 50%       19.7
#> 6     6 75%       21  
#> 7     8 20%       13.9
#> 8     8 50%       15.2
#> 9     8 75%       16.2

Answer 11

You can use q_summarise() from my package timeplyr.

It's both tidy-based (using data-masking rules) and very fast as it uses collapse and data.table under the hood.

# To install, uncomment the below line
# remotes::install_github("NicChr/timeplyr")

library(tidyverse)
library(timeplyr)

mtcars %>%
  q_summarise(mpg, .by = cyl, probs = p)
#>    cyl   p25  p50   p75
#> 1:   4 22.80 26.0 30.40
#> 2:   6 18.65 19.7 21.00
#> 3:   8 14.40 15.2 16.25

mtcars %>%
  q_summarise(mpg, .by = cyl, probs = p, pivot = "long")
#>    cyl .quantile   mpg
#> 1:   4       p25 22.80
#> 2:   4       p50 26.00
#> 3:   4       p75 30.40
#> 4:   6       p25 18.65
#> 5:   6       p50 19.70
#> 6:   6       p75 21.00
#> 7:   8       p25 14.40
#> 8:   8       p50 15.20
#> 9:   8       p75 16.25

# Comparison when there are lots of groups

df <- tibble(g = sample.int(10^4, replace = TRUE),
                             x = rnorm(10^4))

bench::mark(timeplyr = q_summarise(df, x, .by = g,
                        pivot = "long", probs = seq(0, 1, 0.25)),
            dplyr = q_by_group(df, x, g, probs = seq(0, 1, 0.25)),
            check = FALSE)
#> Warning: Some expressions had a GC in every iteration; so filtering is disabled.
#> # A tibble: 2 x 6
#>   expression      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr> <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 timeplyr     27.7ms   31.4ms    29.7      2.06MB     5.95
#> 2 dplyr          1.5s     1.5s     0.665    5.33MB     5.99

^{Created on 2023-07-10 with reprex v2.0.2}