larmarange commented 7 months ago

Expected Behavior

Labelled version 2.13.0 is about to be submitted to CRAN.

In this new version, a bug has been fixed in set_value_labels(). The function now calls systematically haven::labelled() and report an error when trying to assign not accepted values.

Current Behavior

With the dev version of labelled, some tests of REDCapTidier are failing. cf. https://win-builder.r-project.org/incoming_pretest/labelled_2.13.0_20240405_134651/reverseDependencies/summary.txt

After exploring, it seems to occur here in test-utils.R:

out <- multi_choice_to_labels(
    db_data = db_data_classic,
    db_metadata = db_metadata_classic,
    raw_or_label = "haven"
  ) %>%
    suppressWarnings(classes = c(
      "empty_parse_warning",
      "field_missing_categories",
      "duplicate_labels"
    ))

Trying to understand when it was coming from, it seems related to the variable "radio_dtxt_error" in db_metadata_classic.

Variable db_data$radio_dtxt_error contains only NA and is of class logical. And it has no value labels in metadata. Therefore the identified labels is only NA and of class logical.

If you try to apply value labels to a logical vector, you will get an error from haven because value labels could be applied only to numeric or to character vectors.

In apply_labs_haven(), I see these lines of code:

  if (!is.null(attr(labels_cast, "problems"))) {
    # If there was parsing problem fall back to character
    x <- as.character(x)
    labels_cast <- force_cast(labels, character())
  }

In this specific case, this condition is not triggerred. In this specific case, it would be important to convert x to character or numeric before calling set_value_labels(). (in current CRAN version of labelled, it doesn't produce an error while it should have done).

It seems that changing the code to

  if (!is.null(attr(labels_cast, "problems")) || is.logical(x)) {
    # If there was parsing problem fall back to character
    x <- as.character(x)
    labels_cast <- force_cast(labels, character())
  }

is enough to solve the bug.

Would you like to consider such change? Could you check if all tests pass?

If yes, would you consider to submit a new version on CRAN before I resubmit the dev version of labelled on CRAN?

How to Reproduce the Bug:

Install dev version of labelled and run tests of REDCapTidier.

Failure Logs

If applicable, submit any failure logs or error messages here.

Checklist

Before submitting this issue, please check and verify below that the submission meets the below criteria:

[ ] The issue is atomic
[ ] The issue description is documented
[ ] The issue title describes the problem succinctly
[ ] Developers are assigned to the issue
[ ] Labels are assigned to the issue