Error: Mutate can't convert haven labelled to characater

[mbcann01]

Overview

While working on the L2C codebook, I got this error. I went into the codebookr code to try to fix it. Nothing I was doing was working or really even making sense. Then I closed Rstudio, reopened it, and tried again. It magically started working. However, I want to keep a record of what I already tried in case it comes back.

Test file

I created a test qmd file for debugging. Here is what it contained.

library(dplyr, warn.conflicts = FALSE)
library(codebookr)

Load L2C data

path <- "/Users/bradcannell/Library/CloudStorage/OneDrive-TheUniversityofTexasHealthScienceCenteratHouston/01_research/L2C Teams/Participant Data/R data/combined_participant_data.rds"

combined_participant_data <- readr::read_rds(path)

Keep only the variables from the L2C data that seemed to be causing trouble.

df <- combined_participant_data %>% 
  # select(835:840)
  select(835)

print(codebook(df), "test.docx")

The print code above is where I was getting the error. But, here is the really interesting part. When I run devtools::load_all(), the code chunk above works, but when I run devtools::install() I get the same error as when I just load the package with library(codebookr).

I found a post on R Community that seemed relevant (https://community.rstudio.com/t/devtools-load-all-vs-install/51787/3). It said,

One possible reason is that with devtools::load_all() you make all the functions in your package available to the current R session, even those which are not exported via NAMESPACE. With library(package_name) you only get those functions that are exported via NAMESPACE.

It also said,

Keep in mind that library(package_name) is The Truth which devtools::load_all() "roughly simulates".

The person who posted the comment said they could recreate their error by running devtools::load_all(export_all = FALSE) instead of just devtools::load_all(). So, I tried to do the same. My thought was that the Roxygen code I added to the files I didn't want to appear on the pkgdown function list (@keywords internal) might be the issue. However, this didn't fix the problem. The code worked with devtools::load_all(export_all = FALSE) and not with devtools::install(). I went so far as to add @export to every R file, run devtools::document() again, and then run devtools::install() again. I was still getting the same error.

What cb_summary_stats_few_cats is being used?

Next, I tried to understand if there were any differences in the code for cb_summary_stats_few_cats() when I was using devtools::load_all(export_all = FALSE) vs. devtools::install(). So I:

1. Change cb_summary_stats_few_cats to return summary before anything except n is calculated.
2. Restart R.
3. devtools::load_all(export_all = FALSE)
4. Check running cb_summary_stats_few_cats only
5. Check running codebook

# Here is the change I made to cb_summary_stats_few_cats only
summary <- df %>%
    dplyr::count(.data[[.x]]) %>%
    # Rename the first column from the name of the variable being analyzed to
    # "cat"
    dplyr::rename(cat = 1)

    return(summary)

    # I commented out all the code below.

    # Change the category label for missing values from NA to "Missing"
    # If .x is a factor, then replace_na() won't work. Have to change to
    # character first.
    dplyr::mutate(
      cat = as.character(cat),
      cat = tidyr::replace_na(cat, "Missing")
    ) %>%
    # Calculate the cumulative total and percentage
    dplyr::mutate(
      cum_freq = cumsum(n),
      prop     = n / max(cum_freq),
      percent  = prop * 100
    ) %>%
    # Keep columns of interest
    dplyr::select(cat, n, cum_freq, percent) %>%
    # Format numeric results
    dplyr::mutate(
      dplyr::across(
        .cols = c(n, cum_freq, percent),
        .fns  = ~ format(.x, nsmall = digits, big.mark = ",")
      )
    )

Then ran just cb_summary_stats_few_cats.

cb_summary_stats_few_cats(df, "any_health_insurance", digits = 2)

This code chunk returned the expected dplyr summary table with cat and n only. Cat is a <S3: haven_labelled>. So, the change I made to the code was obviously recognized when I loaded the package files with devtools::load_all(export_all = FALSE). Then, I ran the chunk below.

print(codebook(df), "test.docx")

However, this code returned a codebook document with a summary table that also included cumulative frequency and percent (the code I commented out above). Therefore, the codebook() function is clearly using some other version of cb_summary_stats_few_cats(), but why and which one?

Before moving on, I checked really quickly to see if I could reproduce the error outside of my codebookr code.

df %>%
    dplyr::count(any_health_insurance) %>%
    # Rename the first column from the name of the variable being analyzed to
    # "cat"
    dplyr::rename(cat = 1) %>%
    # Change the category label for missing values from NA to "Missing"
    # If .x is a factor, then replace_na() won't work. Have to change to
    # character first.
    dplyr::mutate(
      cat = as.character(cat),
      cat = tidyr::replace_na(cat, "Missing")
    ) %>%
    # Calculate the cumulative total and percentage
    dplyr::mutate(
      cum_freq = cumsum(n),
      prop     = n / max(cum_freq),
      percent  = prop * 100
    ) %>%
    # Keep columns of interest
    dplyr::select(cat, n, cum_freq, percent) %>%
    # Format numeric results
    dplyr::mutate(
      dplyr::across(
        .cols = c(n, cum_freq, percent),
        .fns  = ~ format(.x, nsmall = 2, big.mark = ",")
      )
    )

I could not. The code above works fine. So, what version of the code is devtools::install() using and why isn't it working? To try to answer this, I decided to load the package with library() and then view the internals of the function.

1. Restart R
2. library(codebookr)
3. View the internals of cb_summary_stats_few_cats

library(dplyr, warn.conflicts = FALSE)
library(codebookr)

cb_summary_stats_few_cats

However, the code returned in the code chunk above looks correct (i.e., like the code that runs when using devtools::load_all()). So, I'm still not sure what the issue is.

For Stack overflow

At that point, I decided to create a post on Stack Overflow. I wanted to create a reproducable example, so my plan was to ask responders to clone my codebookr repo and use the haven labeled study data.

Question: devtools::load_all() works as expected, but devtools::install() does not

library(dplyr, warn.conflicts = FALSE)
library(codebookr)

study <- haven::read_dta("/Users/bradcannell/Dropbox/R/Packages/codebookr/inst/extdata/study.dta")

print(codebook(df), "test.docx")

However, this code ran just fine. Then I closed out RStudio entirely, went back to my L2C project, and then the codebook started running there too. I have no idea what changed. However, I'm saving this issue in case it comes up again.