REDCapCAST/tests/testthat/test-csv-exports.R

74 lines
2.4 KiB
R

# Set up the path and data -------------------------------------------------
metadata <- read.csv(
get_data_location("ExampleProject_DataDictionary_2018-06-07.csv"),
stringsAsFactors = TRUE
)
records <- read.csv(get_data_location("ExampleProject_DATA_2018-06-07_1129.csv"),
stringsAsFactors = TRUE)
redcap_output_csv1 <- REDCap_split(records, metadata)
# Test that basic CSV export matches reference ------------------------------
test_that("CSV export matches reference", {
expect_known_hash(redcap_output_csv1, "f74558d1939c17d9ff0e08a19b956e26")
})
# Test that REDCap_split can handle a focused dataset
records_red <- records[!records$redcap_repeat_instrument == "sale",
!names(records) %in% metadata$field_name[metadata$form_name == "sale"] &
!names(records) == "sale_complete"]
records_red$redcap_repeat_instrument <- as.character(records_red$redcap_repeat_instrument)
redcap_output_red <- REDCap_split(records_red, metadata)
test_that("REDCap_split handles subset dataset",
{
testthat::expect_length(redcap_output_red,1)
})
# Test that R code enhanced CSV export matches reference --------------------
if (requireNamespace("Hmisc", quietly = TRUE)) {
test_that("R code enhanced export matches reference", {
redcap_output_csv2 <-
REDCap_split(REDCap_process_csv(records), metadata)
expect_known_hash(redcap_output_csv2, "34f82cab35bf8aae47d08cd96f743e6b")
})
}
if (requireNamespace("readr", quietly = TRUE)) {
context("Compatibility with readr")
metadata <- readr::read_csv(get_data_location("ExampleProject_DataDictionary_2018-06-07.csv"))
records <- readr::read_csv(get_data_location("ExampleProject_DATA_2018-06-07_1129.csv"))
redcap_output_readr <- REDCap_split(records, metadata)
expect_matching_elements <- function(FUN) {
FUN <- match.fun(FUN)
expect_identical(lapply(redcap_output_readr, FUN),
lapply(redcap_output_csv1, FUN))
}
test_that("Result of data read in with `readr` will match result with `read.csv`",
{
# The list itself
expect_identical(length(redcap_output_readr), length(redcap_output_csv1))
expect_identical(names(redcap_output_readr), names(redcap_output_csv1))
# Each element of the list
expect_matching_elements(names)
expect_matching_elements(dim)
})
}