a. Motivation for the declared package
In declared: Functions for Declared Missing Values

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(declared)

The R ecosystem already has some very good packages that deal with labelled objects. In particular, the inter-connected packages haven and labelled provide all the functionality most users would ever need.

As nice and useful as these packages are, it has become apparent they have some fundamental design features that run, in some situations, against users' expectations. This often relates to the treatment of declared missing values that are instrumental for the social sciences.

The following minimal example (adapted from the vignette in package haven) illustrates the situation:

library(haven)
x1 <- labelled_spss(
  x = c(1:5, -91),
  labels = c("Missing" = -91),
  na_value = -91
)

The printed objects from this package nicely display some properties:

x1

There are 5 normal (non-missing) values (supposedly, they represent the number of children), and one declared missing value coded -91. This value acts as a missing value, but it is different from a regular missing value in R, coded NA. The latter stands for any missing information (something like an empty cell) regardless of the reason.

On the other hand, the cell is not empty, but the value -91 is not a valid value either. It cannot possibly represent -91 children in the household, but for instance, it could have meant the respondent did not want to respond. It is properly identified as missing with:

is.na(x1)

But when calculating a mean, for instance, the normal expectation is that value 99 would not play any role in the calculations (since it should be missing). However:

mean(x1)

This means the value 99 did play an active role despite being identified as "missing". In an ideal world, the expected mean would be 3, or at best, employ the argument na.rm = TRUE{.R} if the result is NA because of the declared missing value.

A solution to this problem is offered by package labelled, which has a function called user_na_to_na(){.R}:

library(labelled)
mean(
  user_na_to_na(x1),
  na.rm = TRUE
)

The declared solution

While solving the problem, this above solution forces two additional operations:

converting the (already) declared user missing values, and
employing the na.rm argument.

This should not be necessary, especially if (and it is extremely likely that) users may forget the declared missing values are not actually missing values. This scenario is quite possible, as many users previously using other software like SPSS or Stata where nothing else should be done after declaring the missing values may not realize more is needed.

To solve this problem, the declared package creates a similar object, where declared missing values are stored (hence interpreted as) regular NA missing values in R.

library(declared)
x2 <- declared(
  x = c(1:5, -91),
  labels = c("Missing" = -91),
  na_value = -91
)
x2

The print method makes it obvious the value -91 is not a regular number but an actual missing value. More importantly, this type of storage circumvents the need to convert user-defined missing values to regular NA values since they are already stored as regular NA values. The average value is calculated simply as follows:

mean(x2)

Notice that neither user_na_to_na(){.R}, nor employing na.rm = TRUE{.R} are necessary, and, despite being stored as an NA value, the value 99 is not equivalent to an empty cell. The information still exists, but it is simply ignored in the calculations.

At first glance, providing a class method for this function seems unnecessary because activating the argument na.rm will return the correct result. Explaining the importance of the class method requires a discussion about the base R decision to have this argument deactivated by default. This is most likely to alert users about possible problems in the data since a default value of TRUE would obscure such problems; the mean is calculated irrespective of potentially problematic NA values.

This is where differentiating between empty and declared missing values proves valuable. The declared missing values are neither problematic nor signal potential problems in the data, given that once a reason is declared, it is already known why a particular value is missing.

The genuinely problematic values are the empty NA values, and the custom class method still allows identifying such values if they exist:

mean(c(x2, NA))
mean(c(x2, NA), na.rm = TRUE)

Since all declared values are stored as regular NA values, the base function is.na(){.R}, as well as all related functions such as anyNA(){.R} etc., are unaware and can not differentiate between empty and declared missing values:

is.na(c(x2, NA))

To overcome this situation, package declared complementary provides an additional function to account for the difference:

is.empty(c(x2, NA))

All missing values, empty and declared, play well with the NA oriented, base functions such as na.omit(){.R} or na.exclude(){.R}:

na.omit(x2)

It should be made obvious the excellent packages haven and labelled are not inherently doing a bad thing: the very same result is obtained, just via a different route. Package declared was created as an alternative to the design philosophy of these packages, with a fundamental difference: instead of interpreting existing values as missing, the missing values are interpreted as existing.

It does so by storing an additional attribute containing the positions (indexes) of the regular NA values in the object, which should be treated as missing, and, even more so, to be interpreted as a particular missing response category, as specified in the value labels attribute.