Nothing
test_that("all supported model structures agree with dense Gaussian equations", {
d <- vam_fixture(nstudent = 24L, nyear = 3L)
configurations <- expand.grid(persistence = c("GP", "rGP", "CP", "VP", "ZP"),
REML = c(FALSE, TRUE), stringsAsFactors = FALSE)
configurations$student.side <- "R"
configurations <- rbind(configurations,
data.frame(persistence = "GP", REML = c(FALSE, TRUE),
student.side = "G"))
for (i in seq_len(nrow(configurations))) {
cfg <- configurations[i, ]
fit <- suppressWarnings(GPvam(d, ~as.factor(year)+x,
persistence = cfg$persistence,
student.side = cfg$student.side, REML = cfg$REML,
max.iter.EM = 3, verbose = FALSE))
ref <- dense_reference(fit)
context <- paste(cfg$student.side, cfg$persistence, cfg$REML)
expect_equal(as.numeric(fit$ybetas), ref$beta, tolerance = 1e-7, info = context)
expect_equal(as.numeric(fit$eta.hat), ref$eta, tolerance = 1e-7, info = context)
expect_equal(as.numeric(fit$loglik), ref$loglik, tolerance = 1e-7, info = context)
expect_equal(unname(as.matrix(fit$betacov)), unname(ref$betacov), tolerance = 1e-7,
info = context)
expect_equal(unname(as.matrix(fit$vinv)), unname(ref$Vinv), tolerance = 1e-7,
info = context)
expect_equal(as.numeric(fit$yhat), as.vector(fit$X %*% ref$beta + fit$Z %*% ref$eta),
tolerance = 1e-7, info = context)
p <- ncol(fit$X)
q <- ncol(fit$Z)
teacher_index <- if (cfg$student.side == "G") seq.int(fit$num.student+1L, q) else seq_len(q)
sd <- sqrt(diag(ref$C)[p+teacher_index])
expect_equal(fit$teach.effects$std_error, unname(sd), tolerance = 1e-4, info = context)
expect_false(fit$converged, info = context)
expect_true(all(is.finite(fit$loglik_history)), info = context)
set.seed(1741L+i)
bias <- bias.test.custom(fit, k_vectors = rep(1, ncol(fit$X)),
n_perms = 19, plot = FALSE, verbose = FALSE)
expect_true(all(bias$permutation_results$Permutation_P_Value >= 1/20 &
bias$permutation_results$Permutation_P_Value <= 1), info = context)
# Independently identify exchangeable blocks from actual teacher labels,
# rather than reproducing the diagnostic's block-width calculation.
permutation <- seq_len(ncol(fit$G))
offset <- if (cfg$student.side == "G") fit$num.student else 0L
for (year in unique(fit$teach.effects$teacher_year)) {
effects <- fit$teach.effects
ids <- unique(effects$teacher[effects$teacher_year == year])
first <- which(effects$teacher_year == year & effects$teacher == ids[1])
last <- which(effects$teacher_year == year & effects$teacher == tail(ids, 1))
permutation[offset+first] <- offset+last
permutation[offset+last] <- offset+first
}
if (offset > 1L) permutation[c(1L, offset)] <- c(offset, 1L)
expect_equal(unname(as.matrix(fit$G[permutation, permutation])),
unname(as.matrix(fit$G)), tolerance = 1e-10, info = context)
}
})
test_that("a one-year random-intercept fit matches an independent ANOVA solution", {
# Orthogonal balanced design: ML/REML variance estimates have closed forms.
# Between/within moments below are exact finite-sample likelihood solutions.
groups <- 8L
per_group <- 6L
group_means <- seq(-2.1, 2.1, length.out = groups)
within <- c(-0.7, -0.4, -0.1, 0.1, 0.4, 0.7)
d <- data.frame(student = seq_len(groups*per_group),
teacher = rep(seq_len(groups), each = per_group), year = 1,
y = 3 + rep(group_means, each = per_group) + rep(within, groups))
ss_within <- groups * sum(within^2)
residual_var <- ss_within/(groups*(per_group-1L))
for (reml in c(FALSE, TRUE)) {
teacher_var <- sum(group_means^2)/(groups-as.integer(reml)) - residual_var/per_group
fit <- GPvam(d, ~1, persistence = "CP", REML = reml,
max.iter.EM = 1000, tol1 = 1e-11, verbose = FALSE)
expect_true(fit$converged)
expect_equal(unname(coef(fit)), 3, tolerance = 1e-8)
expect_equal(as.numeric(fit$G[1, 1]), teacher_var, tolerance = 1e-4)
expect_equal(as.numeric(solve(fit$R.inv)[1, 1]), residual_var, tolerance = 1e-4)
expect_equal(fit$loglik, dense_reference(fit)$loglik, tolerance = 1e-8)
# CP, ZP, VP, GP and rGP coincide when there are no future periods.
for (persistence in c("GP", "rGP", "VP", "ZP")) {
same <- GPvam(d, ~1, persistence = persistence, REML = reml,
max.iter.EM = 1000, tol1 = 1e-11, verbose = FALSE)
expect_equal(same$loglik, fit$loglik, tolerance = 1e-7)
expect_equal(as.numeric(same$G[1, 1]), teacher_var, tolerance = 1e-4)
}
}
default <- GPvam(d, persistence = "CP", max.iter.EM = 1000,
tol1 = 1e-11, verbose = FALSE)
expect_equal(unname(coef(default)), 3, tolerance = 1e-8)
expect_error(GPvam(d, student.side = "G"), "repeated observed scores")
})
test_that("ML and REML distinguish fixed-effect uncertainty", {
d <- vam_fixture(nstudent = 24L, nyear = 2L)
ml <- suppressWarnings(GPvam(d, ~as.factor(year)+x, persistence = "CP",
REML = FALSE, max.iter.EM = 4, verbose = FALSE))
reml <- suppressWarnings(GPvam(d, ~as.factor(year)+x, persistence = "CP", REML = TRUE,
max.iter.EM = 4, verbose = FALSE))
expect_false(isTRUE(all.equal(as.matrix(ml$G), as.matrix(reml$G))))
expect_equal(ml$loglik, dense_reference(ml)$loglik, tolerance = 1e-8)
expect_equal(reml$loglik, dense_reference(reml)$loglik, tolerance = 1e-8)
})
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.