使用 dplyr window 函数计算百分位数

Using dplyr window functions to calculate percentiles

我有一个可行的解决方案,但我正在寻找一个更清晰、更具可读性的解决方案,它可能利用了一些较新的 dplyr window 函数。

使用 mtcars 数据集,如果我想查看第 25、50、75 个百分位数以及每加仑英里数的平均值和计数 ("mpg") 按气缸数 ("cyl" ), 我使用以下代码:

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

有没有一种方法可以使用 dplyr 使用一些汇总函数(n_tiles、percent_rank 等)更干净地完成此操作?干净地,我的意思是没有 "do" 语句。

谢谢

不确定如何在 dplyr 中避免 do(),但是您可以使用 c()as.list() 以及 data.table 以非常简单的方式做到这一点:

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

如果您希望它们按 cyl 列排序,请将 by 替换为 keyby

dplyr 1.0中,summarise可以return多个值,允许如下:

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)))

或者,您可以使用 enframe:

来避免使用单独的行来命名分位数
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

以前版本的回答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

可以使用 tidyeval 将其转换为更通用的函数:

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)

这是一个dplyr方法,使用了broom包的tidy()功能,不幸的是它仍然需要do(),但它更简单。

library(dplyr)
library(broom)

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

给出:

    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

注意 t() 的使用,因为 broom 包没有命名数字的方法。

这是基于我的

如果您准备好使用 purrr::map,您可以这样做!

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

reprex package (v0.2.1)

创建于 2018-11-10

这种方法的一个好处是输出很整洁,每行一个观察值。

这是一个相当易读的解决方案,它以整洁的格式使用 dplyrpurrr 到 return 分位数:

代码

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)))
       })

结果

# 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

此解决方案仅使用 dplyrtidyr,允许您在 dplyr 链中指定分位数,并利用 tidyr::crossing() 到 "stack"分组和汇总之前的数据集的多个副本。

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

结果:

# 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

unique() 是让 dplyr::summarise() 知道您每组只需要一个值的必要条件。

这是使用 dplyrpurrrrlang 组合的解决方案:

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

reprex package (v0.2.0) 创建于 2018-10-01。

编辑(2019-04-17):

dplyr 0.8.0 起,funs 函数已被弃用,取而代之的是使用 list 将所需函数传递到作用域 dplyr 函数中。因此,上面的实现变得更加直接。我们不再需要担心使用 !!! 取消引用函数。请看下面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

reprex package (v0.2.0) 创建于 2019-04-17。

do() 实际上是正确的用法,因为它是为分组转换而设计的。将其视为映射到数据框组的 lapply()。 (对于这样一个专门的功能,像“do”这样的通用名称并不理想。但可能为时已晚。)

从道德上讲,在每个 cyl 组中,您想将 quantile() 应用于 mpg 列:

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

除非那不起作用,因为 quantile() 不是 return 数据框;您必须明确地转换其输出。由于此更改相当于用数据框包装 quantile(),因此您可以使用 gestalt 函数组合运算符 %>>>%:

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

回答了很多不同的方式。 dplyr distinct 改变了我想做的事情..

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)

渲染

# 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

另一种实现方式,unnest_wider/longer

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

如果你想为多个变量做这件事,你可以在分组之前收集:

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)