Making removal of redundant covariates reproducible

R/Normalization.R

@@ -159,7 +159,7 @@ tidyCovariateData <- function(covariateData,
             inner_join(covariateData$valueCounts, by = "covariateId") %>%
             select(.data$analysisId, .data$covariateId, n) %>%
             collect()
-          valueCounts <- valueCounts[order(valueCounts$analysisId, -valueCounts$n), ]
+          valueCounts <- valueCounts[order(valueCounts$analysisId, -valueCounts$n, valueCounts$covariateId), ]
           deleteCovariateIds <- c(deleteCovariateIds, valueCounts$covariateId[!duplicated(valueCounts$analysisId)])
           ignoreCovariateIds <- valueCounts$covariateId
           ParallelLogger::logInfo("Removing ", length(deleteCovariateIds), " redundant covariates")

tests/testthat/test-tidyCovariates.R

@@ -93,3 +93,33 @@ test_that("tidyCovariateData on Temporal Data", {
   tidy <- tidyCovariateData(covariateData)
   expect_equal(length(pull(tidy$analysisRef, analysisId)), length(pull(covariateData$analysisRef, analysisId)))
 })
+
+test_that("Removal of redundant covariates is reproducible", {
+  # Create data with two covariates from the same analysis, with equal prevalence:
+  covariates <- tibble(
+    covariateId = rep(c(1,2), 5),
+    rowId = seq_len(10),
+    covariateValue = 1
+  )
+  covariateRef <- tibble(
+    covariateId = c(1,2),
+    analysisId = c(1, 1)
+  )
+  metaData <- list(populationSize = 10)
+  covariateData <- Andromeda::andromeda(
+    covariates = covariates,
+    covariateRef = covariateRef
+  )
+  attr(covariateData, "metaData") <- metaData
+  class(covariateData) <- "CovariateData"
+
+  # Repeat removal of redundant covariates multiple times to evaluate consistency:
+  covariateIds <- c()
+  for (i in seq_len(10)) {
+    tidy <- tidyCovariateData(covariateData, minFraction = 0, normalize = FALSE, removeRedundancy = TRUE)
+    covariateIds[i] <- tidy$covariates |>
+      distinct(covariateId) |>
+      pull()
+  }
+  expect_equal(length(unique(covariateIds)), 1)
+})