Harmonize Value Labels

This vignette follows up on the Working With A Crosswalk Table. In the that vignette, you learned how to remove variables that cannot be harmonized with subset_surveys() and harmonize variable names with harmonize_survey_variables().

As a result of these steps, you have a list of surveys, or surveys saved in files that are harmonization candidates. They now need a consistent numerical coding, labelling with special attention given to missing values and other special values.

Harmonize value codes and labels

The function harmonize_values() solves problems in the following situations:

When the data are read from an SPSS file, in one dataset the variable survey1$trust has no user-defined missing values, but in another dataset the variable survey2$trust does have missing values defined. The two variables cannot be combined. We add harmonized missing values to the missing value range, even if they are not present among the observations.
The labels are not matching in survey1$trust and survey2$trust. We harmonize the labels, and record their initial values for reproducibility.
The missing value ranges in survey1$trust and survey2$trust do not match. We harmonize the missing values, and record their initial values for reproducibility.
There are unexpected labels present in the range of substantive or missing values. They are taken out from the value range with a special code and marked with a special label.

Scenario 1

All values are present, and only the missing values are recoded.

v1 <- labelled_spss_survey(
  c(1, 0, 1, 9),
  labels = c(
    "yes" = 1,
    "no" = 0,
    "inap" = 9
  ),
  na_values = 9
)

h1 <- harmonize_values(
  x = v1,
  harmonize_labels = list(
    from = c("^yes", "^no", "^inap"),
    to = c("trust", "not_trust", "inap"),
    numeric_values = c(1, 0, 99999)
  ),
  id = "survey1"
)

str(h1)
#>  'retroharmonize_labelled_spss_survey' num [1:4]     1     0     1 99999
#>  - attr(*, "labels")= Named num [1:5] 0 1 99997 99998 99999
#>   ..- attr(*, "names")= chr [1:5] "not_trust" "trust" "do_not_know" "declined" ...
#>  - attr(*, "label")= chr "v1"
#>  - attr(*, "na_values")= num [1:3] 99997 99998 99999
#>  - attr(*, "survey1_name")= chr "v1"
#>  - attr(*, "survey1_values")= Named num [1:3] 0 1 99999
#>   ..- attr(*, "names")= chr [1:3] "0" "1" "9"
#>  - attr(*, "survey1_label")= chr "v1"
#>  - attr(*, "survey1_labels")= Named num [1:3] 1 0 9
#>   ..- attr(*, "names")= chr [1:3] "yes" "no" "inap"
#>  - attr(*, "survey1_na_values")= num 9
#>  - attr(*, "id")= chr "survey1"

the attribute survey1_values may be used to restore the original coding.
the attribute survey1_labels may be used to restore the original labelling.
the attribute na_values can re-define if a category should be treated as missing.

The to_numeric() method converts the missing value range to NA_real_.

Scenario 2

The original variable is of class haven::labelled_spss(). It has an invalid missing value.

v2 <- haven::labelled_spss(
  c(1, 1, 0, 8),
  labels = c(
    "yes" = 1,
    "no" = 0,
    "declined" = 8
  ),
  na_values = 8
)

h2 <- harmonize_values(
  v2,
  harmonize_labels = list(
    from = c("^yes", "^no", "^inap"),
    to = c("trust", "not_trust", "inap"),
    numeric_values = c(1, 0, 99999)
  ),
  id = "survey2"
)
str(h2)
#>  'retroharmonize_labelled_spss_survey' num [1:4] 1 1 0 8
#>  - attr(*, "labels")= Named num [1:5] 0 1 99997 99998 99999
#>   ..- attr(*, "names")= chr [1:5] "not_trust" "trust" "do_not_know" "declined" ...
#>  - attr(*, "label")= chr "v2"
#>  - attr(*, "na_values")= num [1:3] 99997 99998 99999
#>  - attr(*, "survey2_name")= chr "v2"
#>  - attr(*, "survey2_values")= Named num [1:3] 0 1 8
#>   ..- attr(*, "names")= chr [1:3] "0" "1" "8"
#>  - attr(*, "survey2_label")= chr "v2"
#>  - attr(*, "survey2_labels")= Named num [1:3] 1 0 8
#>   ..- attr(*, "names")= chr [1:3] "yes" "no" "declined"
#>  - attr(*, "survey2_na_values")= num 8
#>  - attr(*, "id")= chr "survey2"

We apply the code 99901 for this value and label it as invalid_label.

After modifying the user-defined missing value labels:

h2b <- harmonize_values(
  v2,
  harmonize_labels = list(
    from = c("^yes", "^no", "^decline"),
    to = c("trust", "not_trust", "inap"),
    numeric_values = c(1, 0, 99999)
  ),
  id = "survey2"
)

str(h2b)
#>  'retroharmonize_labelled_spss_survey' num [1:4]     1     1     0 99999
#>  - attr(*, "labels")= Named num [1:5] 0 1 99997 99998 99999
#>   ..- attr(*, "names")= chr [1:5] "not_trust" "trust" "do_not_know" "declined" ...
#>  - attr(*, "label")= chr "v2"
#>  - attr(*, "na_values")= num [1:3] 99997 99998 99999
#>  - attr(*, "survey2_name")= chr "v2"
#>  - attr(*, "survey2_values")= Named num [1:3] 0 1 99999
#>   ..- attr(*, "names")= chr [1:3] "0" "1" "8"
#>  - attr(*, "survey2_label")= chr "v2"
#>  - attr(*, "survey2_labels")= Named num [1:3] 1 0 8
#>   ..- attr(*, "names")= chr [1:3] "yes" "no" "declined"
#>  - attr(*, "survey2_na_values")= num 8
#>  - attr(*, "id")= chr "survey2"

Scenario 3

The original vector is of class haven_labelled, therefore it has no defined missing value range. We want to remove DK from the value range to the missing range as do_not_know. The original vector also has an unlabelled value (9). Because we believe that in this vector all values should have a value label, we treat it as an invalid observation.

var3 <- labelled::labelled(
  x = c(1, 6, 2, 9, 1, 1, 2),
  labels = c(
    "Tend to trust" = 1,
    "Tend not to trust" = 2,
    "DK" = 6
  )
)

h3 <- harmonize_values(
  x = var3,
  harmonize_labels = list(
    from = c(
      "^tend\\sto|^trust",
      "^tend\\snot|not\\strust", "^dk",
      "^inap"
    ),
    to = c(
      "trust",
      "not_trust", "do_not_know",
      "inap"
    ),
    numeric_values = c(1, 0, 99997, 99999)
  ),
  id = "S3_"
)

str(h3)
#>  'retroharmonize_labelled_spss_survey' num [1:7]     1 99997     0     9     1     1     0
#>  - attr(*, "labels")= Named num [1:5] 0 1 99997 99998 99999
#>   ..- attr(*, "names")= chr [1:5] "not_trust" "trust" "do_not_know" "declined" ...
#>  - attr(*, "label")= chr "var3"
#>  - attr(*, "S3__name")= chr "var3"
#>  - attr(*, "S3__values")= Named num [1:4] 0 1 9 99997
#>   ..- attr(*, "names")= chr [1:4] "2" "1" "9" "6"
#>  - attr(*, "S3__label")= chr "var3"
#>  - attr(*, "S3__labels")= Named num [1:3] 1 2 6
#>   ..- attr(*, "names")= chr [1:3] "Tend to trust" "Tend not to trust" "DK"
#>  - attr(*, "id")= chr "S3_"
#>  - attr(*, "na_values")= num [1:3] 99997 99998 99999

Base Types & Summary

as factor:

summary(as_factor(h3))
#>   not_trust       trust           9 do_not_know    declined        inap 
#>           2           3           1           1           0           0
levels(as_factor(h3))
#> [1] "not_trust"   "trust"       "9"           "do_not_know" "declined"   
#> [6] "inap"
unique(as_factor(h3))
#> [1] trust       do_not_know not_trust   9          
#> Levels: not_trust trust 9 do_not_know declined inap

as numeric:

summary(as_numeric(h3))
#>    Min. 1st Qu.  Median    Mean 3rd Qu.    Max.    NA's 
#>    0.00    0.25    1.00    2.00    1.00    9.00       1
unique(as_numeric(h3))
#> [1]  1 NA  0  9

as character:

summary(as_character(h3))
#>    Length     Class      Mode 
#>         7 character character
unique(as_character(h3))
#> [1] "trust"       "do_not_know" "not_trust"   "9"

Combination of harmonized values

You can combine labelled_spss_survey vectors if the metadata describing their current state is an exact match. This means that the labels, missing values and missing range are defined the same way, and the base type of the vector is matching numeric or character — though labelling character vectors makes little sense.

The historic metadata, i.e. the earlier naming and coding of the variable do not have to match, they are added to all “inherited vectors”.

var1 <- labelled::labelled_spss(
  x = c(1, 0, 1, 1, 0, 8, 9),
  labels = c(
    "TRUST" = 1,
    "NOT TRUST" = 0,
    "DON'T KNOW" = 8,
    "INAP. HERE" = 9
  ),
  na_values = c(8, 9)
)

var2 <- labelled::labelled_spss(
  x = c(2, 2, 8, 9, 1, 1),
  labels = c(
    "Tend to trust" = 1,
    "Tend not to trust" = 2,
    "DK" = 8,
    "Inap" = 9
  ),
  na_values = c(8, 9)
)


h1 <- harmonize_values(
  x = var1,
  harmonize_label = "Do you trust the European Union?",
  harmonize_labels = list(
    from = c("^tend\\sto|^trust", "^tend\\snot|not\\strust", "^dk|^don", "^inap"),
    to = c("trust", "not_trust", "do_not_know", "inap"),
    numeric_values = c(1, 0, 99997, 99999)
  ),
  na_values = c(
    "do_not_know" = 99997,
    "inap" = 99999
  ),
  id = "survey1"
)

h2 <- harmonize_values(
  x = var2,
  harmonize_label = "Do you trust the European Union?",
  harmonize_labels = list(
    from = c("^tend\\sto|^trust", "^tend\\snot|not\\strust", "^dk|^don", "^inap"),
    to = c("trust", "not_trust", "do_not_know", "inap"),
    numeric_values = c(1, 0, 99997, 99999)
  ),
  na_values = c(
    "do_not_know" = 99997,
    "inap" = 99999
  ),
  id = "survey2"
)

For a single vector, you can use the concatenate() function, which, under the hood, calls the vctrs::vec_c method with some additional validation.

vctrs::vec_c(h1, h2)
#> Survey ID: multi-wave
#> <[13]>: Do you trust the European Union?
#> 1 0 1 1 0 99997 99999 0 0 99997 99999 1 1
#> Missing values: 99997, 99999
#> See all attributes multi-wave_name, multi-wave_values, multi-wave_label [...], survey2_na_values with attributes(x)

Binding surveys together

As soon as you have only compatible variables with matching names in two data frames, you can bind them together in a way that their history is preserved. You can do this with vctrs::vec_rbind or dplyr::bind_rows(). The generic rbind() will lose the labelling information.

a <- tibble::tibble(
  rowid = paste0("survey1", 1:length(h1)),
  hvar = h1,
  w = runif(n = length(h1), 0, 1)
)
b <- tibble::tibble(
  rowid = paste0("survey2", 1:length(h2)),
  hvar = h2,
  w = runif(n = length(h2), 0, 1)
)

c <- dplyr::bind_rows(a, b)

summary(c)
#> Do you trust the European Union?
#> Numeric values without coding:
#>    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
#>       0       0       1   30769   99997   99999 
#> Numeric representation:
#>    Min. 1st Qu.  Median    Mean 3rd Qu.    Max.    NA's 
#>  0.0000  0.0000  1.0000  0.5556  1.0000  1.0000       4 
#> Factor representation:
#>     rowid                    hvar         w          
#>  Length:13          not_trust  :4   Min.   :0.09785  
#>  Class :character   trust      :5   1st Qu.:0.21837  
#>  Mode  :character   do_not_know:2   Median :0.43933  
#>                     inap       :2   Mean   :0.40657  
#>                                     3rd Qu.:0.54071  
#>                                     Max.   :0.78784

print(c)
#> # A tibble: 13 × 3
#>    rowid    hvar                          w
#>    <chr>    <retroh_dbl>              <dbl>
#>  1 survey11     1 [trust]            0.265 
#>  2 survey12     0 [not_trust]        0.439 
#>  3 survey13     1 [trust]            0.458 
#>  4 survey14     1 [trust]            0.541 
#>  5 survey15     0 [not_trust]        0.666 
#>  6 survey16 99997 (NA) [do_not_know] 0.113 
#>  7 survey17 99999 (NA) [inap]        0.218 
#>  8 survey21     0 [not_trust]        0.788 
#>  9 survey22     0 [not_trust]        0.0979
#> 10 survey23 99997 (NA) [do_not_know] 0.710 
#> 11 survey24 99999 (NA) [inap]        0.218 
#> 12 survey25     1 [trust]            0.268 
#> 13 survey26     1 [trust]            0.505

While dplyr’s join functions may result in correct values, the metadata get lost. A new join method will be developed.