Feature imbalance for categorical columns

Source: vignettes/pkgdown/inspect_imb_examples.Rmd

Illustrative data: starwars

The examples below make use of the starwars and storms data from the dplyr package

# some example data
data(starwars, package = "dplyr")
data(storms, package = "dplyr")

For illustrating comparisons of dataframes, use the starwars data and produce two new dataframes star_1 and star_2 that randomly sample the rows of the original and drop a couple of columns.

library(dplyr)
star_1 <- starwars %>% sample_n(50)
star_2 <- starwars %>% sample_n(50) %>% select(-1, -2)

inspect_imb() for a single dataframe

Understanding categorical columns that are dominated by a single level can be useful. inspect_imb() returns a tibble containing categorical column names (col_name); the most frequently occurring categorical level in each column (value) and pctn & cnt the percentage and count which the value occurs. The tibble is sorted in descending order of pcnt.

## # A tibble: 8 × 4
##   col_name   value      pcnt   cnt
##   <chr>      <chr>     <dbl> <int>
## 1 gender     masculine 75.9     66
## 2 sex        male      69.0     60
## 3 hair_color none      42.5     37
## 4 species    Human     40.2     35
## 5 eye_color  brown     24.1     21
## 6 skin_color fair      19.5     17
## 7 homeworld  Naboo     12.6     11
## 8 name       Ackbar     1.15     1

A barplot is printed by passing the result to the show_plot() function:

inspect_imb() for two dataframes

When a second dataframe is provided, inspect_imb() returns a tibble that compares the frequency of the most common categorical values of the first dataframe to those in the second. The p_value column contains a measure of evidence for whether the true frequencies are equal or not.

inspect_imb(star_1, star_2)
## # A tibble: 8 × 7
##   col_name   value     pcnt_1 cnt_1 pcnt_2 cnt_2 p_value
##   <chr>      <chr>      <dbl> <int>  <dbl> <int>   <dbl>
## 1 gender     masculine     84    42     80    40   0.795
## 2 sex        male          78    39     74    37   0.815
## 3 hair_color none          48    24     46    23   1    
## 4 species    Human         32    16     32    16   1    
## 5 eye_color  blue          24    12     NA    NA  NA    
## 6 homeworld  Tatooine      14     7     NA    NA  NA    
## 7 skin_color fair          12     6     14     7   1    
## 8 name       Ackbar         2     1     NA    NA  NA
inspect_imb(star_1, star_2) %>% show_plot()

  • Smaller p_value indicates stronger evidence against the null hypothesis that the true frequency of the most common values is the same.
  • The visualisation illustrates the significance of the difference using a coloured bar overlay. Orange bars indicate evidence of equality of the imbalance, while blue bars indicate inequality. If a p_value cannot be calculated, no coloured bar is shown.
  • The significance level can be specified using the alpha argument to inspect_imb(). The default is alpha = 0.05.