Nothing
# tests/testthat/test-wRandomforest.R
# Smoke tests for wRandomforest(). Uses make_test_data() from helper-wmodels.R.
#
# If RStudio's R session crashes (not a normal testthat failure) while
# running this file specifically, see helper-wmodels.R's comment on strata
# sparsity -- randomForest::randomForest()'s stratified sampling is
# implemented in compiled code and can segfault rather than error cleanly
# when a strata x cluster combination has too few observations.
test_data <- make_test_data()
col.x <- which(names(test_data) %in% c("x1", "x2", "x3", "x4", "x5"))
test_that("wRandomforest() runs on small synthetic classification data", {
skip_if_not_installed("randomForest")
skip_if_not_installed("survey")
fit <- wRandomforest(data = test_data, y = as.factor(test_data$HBP), col.x = col.x,
cluster = "SDMVPSU", strata = "SDMVSTRA", weights = "WTSAF2YR",
method = "dCV", k = 4, R = 1,
ntree = 50)
expect_s3_class(fit, "w.randomforest")
expect_true(is.numeric(fit$mtry$all))
expect_true(is.numeric(fit$mtry$optimal.mtry))
expect_true(is.numeric(fit$evaluation_log$weighted.test.error))
expect_s3_class(fit$model, "randomForest")
})
test_that("wRandomforest() respects a user-supplied design object", {
skip_if_not_installed("randomForest")
skip_if_not_installed("survey")
des <- survey::svydesign(ids = ~SDMVPSU, strata = ~SDMVSTRA,
weights = ~WTSAF2YR, nest = TRUE, data = test_data)
fit <- wRandomforest(y = as.factor(test_data$HBP), col.x = col.x, design = des,
method = "dCV", k = 4, R = 1, ntree = 50)
expect_s3_class(fit, "w.randomforest")
})
test_that("wRandomforest() runs with method = 'JKn'", {
skip_if_not_installed("randomForest")
skip_if_not_installed("survey")
fit <- wRandomforest(data = test_data, y = as.factor(test_data$HBP), col.x = col.x,
cluster = "SDMVPSU", strata = "SDMVSTRA", weights = "WTSAF2YR",
method = "JKn", ntree = 50)
expect_s3_class(fit, "w.randomforest")
expect_true(is.numeric(fit$mtry$optimal.mtry))
expect_true(is.numeric(fit$evaluation_log$weighted.test.error))
})
test_that("wRandomforest() rejects method = 'bootstrap' and 'subbootstrap' with an informative error", {
# wRandomforest() intentionally does not support top-level bootstrap/
# subbootstrap replicate-weight methods, since randomForest::randomForest()
# already performs bootstrap resampling internally for every tree. See
# wRandomforest()'s @details for the rationale.
skip_if_not_installed("randomForest")
skip_if_not_installed("survey")
expect_error(
wRandomforest(data = test_data, y = as.factor(test_data$HBP), col.x = col.x,
cluster = "SDMVPSU", strata = "SDMVSTRA", weights = "WTSAF2YR",
method = "bootstrap", ntree = 50),
"not supported by wRandomforest"
)
expect_error(
wRandomforest(data = test_data, y = as.factor(test_data$HBP), col.x = col.x,
cluster = "SDMVPSU", strata = "SDMVSTRA", weights = "WTSAF2YR",
method = "subbootstrap", ntree = 50),
"not supported by wRandomforest"
)
})
test_that("REGRESSION: eval.error() computes bootstrap error against each replicate's OOB holdout (not the full sample), even though wRandomforest() currently never calls it with this method", {
# This mirrors the equivalent error.f() regression test for wElnet(). The
# same weights.name bug existed here: for bootstrap/subbootstrap,
# weights.name was never reconstructed per-fold, so it stayed fixed at
# the plain full-sample weights column for all B replicates, silently
# computing error against the entire original sample instead of each
# replicate's actual out-of-bag (OOB) held-out units.
#
# NOTE: wRandomforest() itself explicitly rejects method = "bootstrap"/
# "subbootstrap" (see the test above) and therefore never reaches this
# code path today. This test calls eval.error() directly to verify the
# underlying utility function is correct regardless, since it is a
# standalone function that could be reused or called elsewhere.
skip_if_not_installed("survey")
set.seed(42)
rw_data <- replicate_weights(data = test_data, method = "subbootstrap",
cluster = "SDMVPSU", strata = "SDMVSTRA",
weights = "WTSAF2YR", B = 5, rw.test = TRUE)
# Build l.yhat the same way wRandomforest() would, for fold 1 only.
train_col <- "rw_r_1_train_1"
test_col <- "sw_r_1_test_1"
expect_true(test_col %in% colnames(rw_data)) # column must exist at all post-fix
idx_train <- which(rw_data[, train_col] > 0)
y_fac <- as.factor(rw_data$HBP)
rf_fold1 <- randomForest::randomForest(
x = rw_data[idx_train, col.x], y = y_fac[idx_train],
xtest = rw_data[-idx_train, col.x],
weights = as.numeric(rw_data[idx_train, train_col])
)
l.yhat <- list()
l.yhat[["yhat_rw_r_1_train_1"]] <- rf_fold1$test$predicted
error_vec <- eval.error(data = rw_data, l.yhat = l.yhat, method = "subbootstrap",
cv.error.ind = FALSE, R = 1, k = 5, B = 5,
y = y_fac, objective = "binary:logistic",
weights = "WTSAF2YR")
# Manually recompute fold 1's expected error using only the actual OOB
# units (positive sw_r_1_test_1 weight), independent of eval.error()'s
# internal logic, as ground truth.
idx_test <- which(rw_data[, test_col] > 0)
expected_loss <- as.numeric(y_fac[idx_test] != l.yhat[["yhat_rw_r_1_train_1"]])
expected_error_fold1 <- sum(expected_loss * rw_data[idx_test, test_col]) / sum(rw_data[idx_test, test_col])
expect_equal(unname(error_vec[1]), expected_error_fold1, tolerance = 1e-6)
})
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.