Nothing
# EX1: a p-value tie used to add or remove several terms in one step while the
# history recorded a single step. Every step must move exactly one term and the
# AIC history must describe the model after that step.
selection_aic_for <- function(response, terms, data, fit) {
form <- if (length(terms) == 0) {
stats::as.formula(paste(response, "~ 1"))
} else {
stats::as.formula(paste(response, "~", paste(terms, collapse = " + ")))
}
stats::AIC(stats::glm(form, data = data, family = fit$family))
}
test_that("p-both history records one term and one AIC per step (EX1)", {
m <- glm(y ~ ., data = stepwise, family = binomial(link = "logit"))
invisible(capture.output(k <- blr_step_p_both(m)))
expect_equal(k$steps, length(k$orders))
expect_equal(k$steps, length(k$method))
expect_equal(k$steps, length(k$aic))
expect_equal(k$steps, length(k$bic))
expect_equal(k$steps, length(k$dev))
terms <- character(0)
for (i in seq_len(k$steps)) {
if (k$method[i] == "addition") {
terms <- c(terms, k$orders[i])
} else {
terms <- setdiff(terms, k$orders[i])
}
expect_equal(k$aic[i], selection_aic_for("y", terms, stepwise, m),
tolerance = 1e-8)
}
ref <- glm(y ~ ., data = stepwise[, c("y", k$predictors)],
family = binomial(link = "logit"))
expect_equal(coef(k$model), coef(ref))
})
test_that("p-forward history records one term and one AIC per step (EX1)", {
m <- glm(y ~ ., data = stepwise, family = binomial(link = "logit"))
invisible(capture.output(k <- blr_step_p_forward(m)))
expect_equal(k$steps, length(k$predictors))
expect_equal(k$steps, length(k$aic))
expect_equal(k$steps, length(k$bic))
expect_equal(k$steps, length(k$dev))
for (i in seq_len(k$steps)) {
expect_equal(k$aic[i],
selection_aic_for("y", k$predictors[seq_len(i)], stepwise, m),
tolerance = 1e-8)
}
ref <- glm(y ~ ., data = stepwise[, c("y", k$predictors)],
family = binomial(link = "logit"))
expect_equal(coef(k$model), coef(ref))
})
test_that("AIC selectors break ties by candidate order and still move one term", {
set.seed(14)
d <- data.frame(x_a = rnorm(200), noise = rnorm(200))
d$x_b <- d$x_a
d$y <- rbinom(200, 1, plogis(2 * d$x_a))
fwd_model <- glm(y ~ x_a + x_b + noise, data = d, family = binomial)
invisible(capture.output(fwd <- blr_step_aic_forward(fwd_model)))
expect_equal(fwd$predictors[1], "x_a")
expect_equal(fwd$steps, length(fwd$predictors))
expect_equal(fwd$steps, length(fwd$aics))
# backward: each x_a/x_b pair is repeated with the two columns swapped and
# the same response, so dropping either one gives exactly the same AIC while
# dropping signal carries noise worsens it. Candidate order must pick x_a.
set.seed(15)
n <- 100
db <- data.frame(x_a = rnorm(n), x_b = rnorm(n), noise = rnorm(n))
db$y <- rbinom(n, 1, plogis(2 * db$noise))
swapped <- db
swapped$x_a <- db$x_b
swapped$x_b <- db$x_a
db <- rbind(db, swapped)
bwd_model <- glm(y ~ x_a + x_b + noise, data = db, family = binomial)
invisible(capture.output(bwd <- blr_step_aic_backward(bwd_model)))
# The two drops tie exactly only if floating point agrees bit for bit, which
# varies by platform/BLAS; assert the candidate-order rule only for an exact tie.
aic_drop_a <- AIC(glm(y ~ x_b + noise, data = db, family = binomial))
aic_drop_b <- AIC(glm(y ~ x_a + noise, data = db, family = binomial))
if (identical(aic_drop_a, aic_drop_b)) {
expect_equal(bwd$predictors[1], "x_a")
} else {
expect_equal(bwd$predictors[1], if (aic_drop_a < aic_drop_b) "x_a" else "x_b")
}
expect_equal(bwd$steps, length(bwd$predictors))
expect_equal(bwd$steps, nrow(bwd$result) - 1L)
})
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.