Intercoder reliability tests

Tidycomm provides the test_icr() function to conveniently compute intercoder reliability tests for several variables and reliability estimates at the same time.

Test data has to be structured in a long format, with one column indicating the unit (e.g., article 1, article 2, etc.), one column indicating the coder (either by a string or a numeric ID), and one column each per coded variable to test.

For demonstration purposes, we will use the fbposts data included in Tidycomm that consists of 45 political Facebook posts (identified by post_id) coded by six coders (identified by coder_id) for various formal (post type, number of pictures used in post) and populism-related (attacks on elites, references to ‘the people’, othering) features:

fbposts
#> # A tibble: 270 x 7
#>    post_id coder_id type  n_pictures pop_elite pop_people pop_othering
#>      <int>    <int> <chr>      <int>     <int>      <int>        <int>
#>  1       1        1 photo          1         0          0            0
#>  2       1        2 photo          1         0          0            0
#>  3       1        3 photo          1         0          0            0
#>  4       1        4 photo          1         0          0            0
#>  5       1        5 photo          1         0          0            0
#>  6       1        6 photo          1         0          0            0
#>  7       2        1 photo          1         0          0            0
#>  8       2        2 photo          1         0          0            0
#>  9       2        3 photo          1         0          0            0
#> 10       2        4 photo          1         0          0            0
#> # ... with 260 more rows

Basic use

test_icr() computes various intercoder reliability estimates for all specified variables. The first two arguments (in a pipe) are the unit-identifying variable and the coder-identifying variable, followed by the test variables:

fbposts %>% 
  test_icr(post_id, coder_id, pop_elite, pop_people, pop_othering)
#> # A tibble: 3 x 8
#>   Variable     n_Units n_Coders n_Categories Level   Agreement Holstis_CR
#>   <chr>          <int>    <int>        <int> <chr>       <dbl>      <dbl>
#> 1 pop_elite         45        6            6 nominal     0.733      0.861
#> 2 pop_people        45        6            2 nominal     0.778      0.916
#> 3 pop_othering      45        6            4 nominal     0.867      0.945
#> # ... with 1 more variable: Krippendorffs_Alpha <dbl>

If no test variables are specified, all variables in the dataset (excluding the unit and coder variables) will be tested:

fbposts %>% 
  test_icr(post_id, coder_id)
#> # A tibble: 5 x 8
#>   Variable     n_Units n_Coders n_Categories Level   Agreement Holstis_CR
#>   <chr>          <int>    <int>        <int> <chr>       <dbl>      <dbl>
#> 1 type              45        6            4 nominal     1          1    
#> 2 n_pictures        45        6            7 nominal     0.822      0.930
#> 3 pop_elite         45        6            6 nominal     0.733      0.861
#> 4 pop_people        45        6            2 nominal     0.778      0.916
#> 5 pop_othering      45        6            4 nominal     0.867      0.945
#> # ... with 1 more variable: Krippendorffs_Alpha <dbl>

Reliability estimates

Currently, test_icr() supports the following reliability estimates:

agreement: Simple percent agreement.
holsti: Holsti’s \(CR\) (mean pairwise percent agreement).
kripp_alpha: Krippendorff’s \(\alpha\).
cohens_kappa: Cohen’s \(\kappa\) (only available for two coders).
fleiss_kappa: Fleiss’ \(\kappa\).
brennan_prediger: Brennan & Prediger’s \(\kappa\) (for more than two coders, von Eye’s (2006) proposed extension to multiple coders is computed).

By default, test_icr() will output simple percent agreement, Holsti’s \(CR\), and Krippendorff’s \(\alpha\) as reliability estimates. You can add other estimates by setting their name to TRUE in the function call (and remove the default ones by setting them to FALSE):

fbposts %>% 
  test_icr(post_id, coder_id, fleiss_kappa = TRUE, agreement = FALSE)
#> # A tibble: 5 x 8
#>   Variable    n_Units n_Coders n_Categories Level  Holstis_CR Krippendorffs_Alp~
#>   <chr>         <int>    <int>        <int> <chr>       <dbl>              <dbl>
#> 1 type             45        6            4 nomin~      1                  1    
#> 2 n_pictures       45        6            7 nomin~      0.930              0.880
#> 3 pop_elite        45        6            6 nomin~      0.861              0.339
#> 4 pop_people       45        6            2 nomin~      0.916              0.287
#> 5 pop_otheri~      45        6            4 nomin~      0.945              0.566
#> # ... with 1 more variable: Fleiss_Kappa <dbl>

Variable levels

By default, test_icr() assumes all test variables to be nominal. You can set other variable levels by passing a named vector of the form c(variable_name = "variable_level") to the levels argument.

fbposts %>% 
  test_icr(post_id, coder_id, levels = c(n_pictures = "ordinal"))
#> # A tibble: 5 x 8
#>   Variable     n_Units n_Coders n_Categories Level   Agreement Holstis_CR
#>   <chr>          <int>    <int>        <int> <chr>       <dbl>      <dbl>
#> 1 type              45        6            4 nominal     1          1    
#> 2 n_pictures        45        6            7 ordinal     0.822      0.930
#> 3 pop_elite         45        6            6 nominal     0.733      0.861
#> 4 pop_people        45        6            2 nominal     0.778      0.916
#> 5 pop_othering      45        6            4 nominal     0.867      0.945
#> # ... with 1 more variable: Krippendorffs_Alpha <dbl>

Nominal test variables can be represented by either integer codes or string labels, whereas ordinal variables must be represented by integer codes, and interval/ratio variables must be numeric (integer or float).

Please note that currently only the computation of Krippendorff’s \(\alpha\) is influenced by the variable level.

Missing values

Missing values in intercoder reliability tests can be ambiguous (did the coder forget to code this variable for this unit, or does the missing value indicate that none of the categories was deemed fitting?) and present an obstacle to several reliability estimates (of the currently implemented estimates, only Krippendorff’s \(\alpha\) can deal with missing values).

Thus, test_icr() will by default respond with a warning when NA values are present in the test variables and output NA for all reliability estimates but Krippendorff’s \(\alpha\):

# Introduce some missing values
fbposts$type[1] <- NA
fbposts$type[2] <- NA
fbposts$pop_elite[5] <- NA

fbposts %>% 
  test_icr(post_id, coder_id)
#> Warning: Variable 'type' contains missing values. Consider setting na.omit =
#> TRUE or recoding missing values
#> Warning: Variable 'pop_elite' contains missing values. Consider setting na.omit
#> = TRUE or recoding missing values
#> # A tibble: 5 x 8
#>   Variable     n_Units n_Coders n_Categories Level   Agreement Holstis_CR
#>   <chr>          <int>    <int>        <int> <chr>       <dbl>      <dbl>
#> 1 type              45        6            4 nominal    NA         NA    
#> 2 n_pictures        45        6            7 nominal     0.822      0.930
#> 3 pop_elite         45        6            6 nominal    NA         NA    
#> 4 pop_people        45        6            2 nominal     0.778      0.916
#> 5 pop_othering      45        6            4 nominal     0.867      0.945
#> # ... with 1 more variable: Krippendorffs_Alpha <dbl>

You can set na.omit = TRUE to exclude all units with NA values for a specific test variable from the computation for this variable:

fbposts %>% 
  test_icr(post_id, coder_id, na.omit = TRUE)
#> # A tibble: 5 x 8
#>   Variable     n_Units n_Coders n_Categories Level   Agreement Holstis_CR
#>   <chr>          <int>    <int>        <int> <chr>       <dbl>      <dbl>
#> 1 type              44        6            4 nominal     1          1    
#> 2 n_pictures        45        6            7 nominal     0.822      0.930
#> 3 pop_elite         44        6            6 nominal     0.727      0.858
#> 4 pop_people        45        6            2 nominal     0.778      0.916
#> 5 pop_othering      45        6            4 nominal     0.867      0.945
#> # ... with 1 more variable: Krippendorffs_Alpha <dbl>

‘Hidden’ missing values

There are situations in which all units are coded multiple times, but not by same coders (e.g., in research seminars). Consider the following sample data:

data <- tibble::tibble(
  unit = c(1, 1, 2, 2, 3, 3),
  coder = c('a', 'b', 'a', 'c', 'b', 'c'),
  code = c(1, 0, 1, 1, 0, 0)
)

data
#> # A tibble: 6 x 3
#>    unit coder  code
#>   <dbl> <chr> <dbl>
#> 1     1 a         1
#> 2     1 b         0
#> 3     2 a         1
#> 4     2 c         1
#> 5     3 b         0
#> 6     3 c         0

Each unit was coded two times, but no coder coded all units. Thus, the units-coders matrix will contain one NA value per (unit) row, indicating that one coder did not code the respective unit. Setting na.omit = TRUE in test_icr() will thus result in an empty units-coders matrix:

data %>%
  test_icr(unit, coder, code, na.omit = TRUE)
#> Error: Empty units-coders matrix detected. This is most likely due to none of the units having been coded by all coders. See vignette('v04_icr') for details.

If it is not relevant that all units were coded by same coders, consider setting a variable indicating each coding per unit as the coder_id:

data %>% 
  dplyr::group_by(unit) %>% 
  dplyr::mutate(coding = 1:dplyr::n()) %>% 
  dplyr::ungroup() %>% 
  test_icr(unit, coding, code)
#> # A tibble: 1 x 8
#>   Variable n_Units n_Coders n_Categories Level   Agreement Holstis_CR
#>   <chr>      <int>    <int>        <int> <chr>       <dbl>      <dbl>
#> 1 code           3        2            2 nominal     0.667      0.667
#> # ... with 1 more variable: Krippendorffs_Alpha <dbl>