summarise wtd.quantile by group

Question

I want to create a new df using Hmisc::wtd.quantile for a dataframe with many repeating dates. I am grouping by date, using summarize() to aggregate on date, and attempting to use wtd.quantile() on each date (with weights). This is also a pretty large dataset. Below is some sample code:

# sample data
# grouping_var = dt_time
require(Hmisc)
require(plyr)
require(dplyr)
df <- data.frame(type = sample(letters[1:2], 10e6, replace = TRUE), 
             score = sample(500:899, 10e6, replace = TRUE),
             dt_time = sample(seq(as.Date('2010/01/01'), 
                                  as.Date('2018/01/01'), 
                                  by="day"), 10e6, replace = TRUE),
             weight = sample(1.0:2.0, 10e6, replace = TRUE))
# my attempt:
ptiles <- df %>%
group_by(dt_time) %>%
plyr::ddply(~dt_time, dplyr::summarize,
            ptile10 = Hmisc::wtd.quantile(., .$score, weights = .$weight, 
probs = .1, na.rm = TRUE),
            ptile50 = Hmisc::wtd.quantile(., .$score, weights = .$weight, 
probs = .5, na.rm = TRUE),
            ptile90 = Hmisc::wtd.quantile(., .$score, weights = .$weight, 
probs = .9, na.rm = TRUE))

# desired df,
# where each new variable would be created using the
# wtd.quantile function:
desired_ptiles <- data.frame(dt_time = seq(as.Date('2010/01/01'),
                                       as.Date('2010/01/06'),
                                       by = "day"),
                         # only 6 because lol 10e6
                         ptile10 = sample(500:899, 6, replace = TRUE),
                         ptile50 = sample(500:899, 6, replace = TRUE),
                         ptile90 = sample(500:899, 6, replace = TRUE))

So far my efforts have resulted in this error:

Error in summarise_impl(.data, dots) :
Evaluation error: 'arg' must be NULL or a character vector.

And when using formula notation:

ptiles <- df %>%
    group_by(dt_time) %>%
plyr::ddply(~dt_time, dplyr::summarize,
ptile10 = Hmisc::wtd.quantile(., .$score, weights = .$weight,
                                     probs = .1, na.rm = TRUE),
ptile50 = Hmisc::wtd.quantile(., .$score, weights = .$weight, 
                                        probs = .5, na.rm = TRUE),
          ptile90 = Hmisc::wtd.quantile(., .$score, weights = .$weight,
          probs = .9, na.rm = TRUE))
# error message:
Error in summarise_impl(.data, dots) : 
  Evaluation error: 'arg' must be NULL or a character vector.

Am I approaching this in the wrong way? I've seen methods that use split() but that seems annoying. Is there a data.table method that will allow wtd.quantile() to be summarized in this way?

Thanks!

score 3 · Accepted Answer · answered Sep 20 '18 at 07:08

You don't need ddply when using group_by, as the data is already split by the grouping. Also, you don't need to define the data within summarize after grouping.

This works:

ptiles <- df %>%
  group_by(dt_time) %>%
  summarize(ptile10 = wtd.quantile(score, weights = weight, 
                                            probs = .1, na.rm = TRUE),
              ptile50 = wtd.quantile(score, weights = weight, 
                                            probs = .5, na.rm = TRUE),
              ptile90 = wtd.quantile(score, weights = weight, 
                                            probs = .9, na.rm = TRUE))

> ptiles
# A tibble: 2,923 x 4
      dt_time ptile10 ptile50 ptile90
       <date>   <dbl>   <dbl>   <dbl>
 1 2010-01-01   539.0     697   859.0
 2 2010-01-02   538.0     704   861.7
 3 2010-01-03   541.0     706   862.0
 4 2010-01-04   541.0     702   859.0
 5 2010-01-05   540.0     706   860.0
 6 2010-01-06   537.0     695   859.0
 7 2010-01-07   539.0     696   859.0
 8 2010-01-08   536.0     700   857.0
 9 2010-01-09   538.0     694   861.0
10 2010-01-10   538.4     701   859.0
# ... with 2,913 more rows

score 0 · Answer 2 · answered Sep 20 '18 at 11:02

Here's an alternative solution that uses a function (based on @LAP's solution) to go over a set of quantile values you provide (i.e. c(.1, .5, .9)) and automatically create the appropriate columns:

require(Hmisc)
require(plyr)
require(tidyverse)

# for reproducibility
set.seed(5)

# example dataset
df <- data.frame(type = sample(letters[1:2], 50000, replace = TRUE), 
                 score = sample(500:899, 50000, replace = TRUE),
                 dt_time = sample(seq(as.Date('2010/01/01'), 
                                      as.Date('2018/01/01'), 
                                      by="day"), 50000, replace = TRUE),
                 weight = sample(1.0:2.0, 50000, replace = TRUE))

# function to get quantiles per date
f = function(x) {
    df %>%
      group_by(dt_time) %>%
      dplyr::summarize(value = wtd.quantile(score, weights = weight, probs = x, na.rm = TRUE)) %>%
      mutate(ptile = x) }

# input quantiles of interest
q = c(.1, .5, .9)

# apply function to each quantile of interest
# and reshape
map_df(q, f) %>%
  spread(ptile, value, sep = "_")

# # A tibble: 2,923 x 4
#   dt_time    ptile_0.1 ptile_0.5 ptile_0.9
#   <date>         <dbl>     <dbl>     <dbl>
# 1 2010-01-01      543       670.      857 
# 2 2010-01-02      549       800       868.
# 3 2010-01-03      529.      705       829.
# 4 2010-01-04      543       690       850.
# 5 2010-01-05      528.      642.      870.
# 6 2010-01-06      534.      649       810.
# 7 2010-01-07      512.      668       801.
# 8 2010-01-08      554       664       869.
# 9 2010-01-09      533.      650       765 
#10 2010-01-10      535.      696       843.
# # ... with 2,913 more rows

summarise wtd.quantile by group

2 Answers2