Stack per-study check results into one list of tibbles

Reads every per-study check file in dir and binds them element-wise into a single named list of tibbles. Each file is expected to hold a named list of tibbles written by a per-study checking script, typically the output of check_y_bounds, check_missingness_nona, check_balance, and check_attrition with a study_id supplied.

Usage

stack_checks(
  dir = "checks",
  pattern = "_checks\\.rds$",
  exclude = "^all_checks\\.rds$",
  elements = NULL,
  warn_schema = TRUE
)

Arguments

dir: Path to the directory holding the per-study check files (default "checks").
pattern: Regular expression identifying the per-study files (default "_checks\\.rds$").
exclude: Regular expression for files to skip, matched against the base name. Defaults to "^all_checks\\.rds$" so that re-running over a directory that already holds a stacked file is safe.
elements: Character vector naming the list elements to stack, or NULL (default) to stack the union of element names found across all files.
warn_schema: Logical. Warn when the files contributing one element disagree about their columns (default TRUE). Set FALSE in a project where that is expected: a corpus whose studies are stratified on different variables produces different grouping columns by design, and the warning cannot tell that apart from staleness.

Value

A named list of tibbles, one per element name.

Details

Elements that are NULL or have zero rows in a given file are dropped before binding, so a study that contributes nothing to one check does not break the stack for the others. An element present in some files but not others is still returned, built from whichever files have it.

Stale files hide in a successful stack

dplyr::bind_rows unions column names and fills the gaps with NA. That tolerance is what lets a study which skipped one check stack alongside studies that ran it, and it is also how a stale file hides: when one study's results predate a change in what a check returns, its rows arrive missing the new columns while still carrying whatever the old version put in the shared ones. The stack succeeds, the corpus looks fully re-run, and it is not. A real instance: two per-study files a day older than the rest left estimable as NA on eight rows and shifted the count of informative attrition tests. Hence the warning, and hence warn_schema.

Examples

d <- tempfile()
dir.create(d)
saveRDS(
  list(ybounds = data.frame(study_id = "a", variable = "Y_x", in_bounds = TRUE)),
  file.path(d, "a_checks.rds")
)
saveRDS(
  list(ybounds = data.frame(study_id = "b", variable = "Y_x", in_bounds = FALSE)),
  file.path(d, "b_checks.rds")
)
stack_checks(d)
#> $ybounds
#>   study_id variable in_bounds
#> 1        a      Y_x      TRUE
#> 2        b      Y_x     FALSE
#>