tests/testthat/test_point_laplace.R

In ebnm: Solve the Empirical Bayes Normal Means Problem

context("Point Laplace")

n <- 1000
set.seed(1)
s <- rnorm(n, 1, 0.1)
x <- c(rexp(n / 2, rate = 0.1), rep(0, n / 2)) + rnorm(n, sd = s)

true_pi0 <- 0.5
true_scale <- 10
true_mean <- 0

true_g <- laplacemix(pi = c(true_pi0, 1 - true_pi0),
                     mean = rep(true_mean, 2),
                     scale = c(0, true_scale))

pl.res <- ebnm(x, s, prior_family = "point_laplace")

test_that("Basic functionality works", {
  pl.res2 <- ebnm_point_laplace(x, s)
  pl.res$call <- pl.res2$call <- NULL
  expect_identical(pl.res, pl.res2)
  expect_equal(pl.res[[g_ret_str()]], true_g, tolerance = 0.1)
})

test_that("Mode estimation works", {
  pl.res2 <- ebnm_point_laplace(x, s, mode = "est")
  expect_equal(pl.res2[[g_ret_str()]], true_g, tolerance = 0.5)
  expect_false(identical(pl.res2[[g_ret_str()]]$mean[1], true_mean))
})

test_that("Fixing the scale works", {
  pl.res2 <- ebnm_point_laplace(x, s, scale = true_scale)
  expect_equal(pl.res2[[g_ret_str()]], true_g, tolerance = 0.1)
  expect_equal(pl.res2[[g_ret_str()]]$scale[2], true_scale)
})

test_that("Fixing g works", {
  pl.res2 <- ebnm_point_laplace(x, s, g_init = pl.res[[g_ret_str()]], fix_g = TRUE)
  expect_identical(pl.res[[g_ret_str()]], pl.res2[[g_ret_str()]])
  expect_equal(as.numeric(pl.res[[llik_ret_str()]]), as.numeric(pl.res2[[llik_ret_str()]]))
})

test_that("Initializing g works", {
  pl.res2 <- ebnm_point_laplace(x, s, g_init = true_g)
  expect_equal(pl.res[[llik_ret_str()]], pl.res2[[llik_ret_str()]])
})

test_that("Output parameter works", {
  pl.res <- ebnm_point_laplace(x, s, output = c("fitted_g"))
  pl.res$call <- NULL
  expect_identical(names(pl.res), "fitted_g")
})

# test_that("Infinite and zero SEs give expected results", {
#   x <- c(rep(0, 5), rep(5, 5))
#   s <- rep(1, 10)
#   # s[6] <- 0
#   s[10] <- Inf
#
#   pl.res <- ebnm_point_laplace(x, s, output = ebnm_output_all())
#
#   # expect_equal(pl.res[[df_ret_str()]][[pm_ret_str()]][6], x[6])
#   # expect_equal(pl.res[[df_ret_str()]][[pm_ret_str()]]2[6], x[6]^2)
#   expect_equal(pl.res[[df_ret_str()]][[pm_ret_str()]][10], 0)
#   expect_equal(pl.res[[df_ret_str()]][[pm2_ret_str()]][10],
#                2 * pl.res[[g_ret_str()]]$scale[2]^2 * pl.res[[g_ret_str()]]$pi[2])
#   expect_equal(pl.res[[df_ret_str()]][[lfsr_ret_str()]][10],
#                pl.res[[g_ret_str()]]$pi[1] + pl.res[[g_ret_str()]]$pi[2] / 2)
# })

test_that("Can fix g with one component", {
  g_init <- laplacemix(pi = 1,
                       scale = true_scale,
                       mean = true_mean)
  pl.res <- ebnm_point_laplace(x, s, g_init = g_init, fix_g = TRUE)

  g_init2 <- laplacemix(pi = c(0, 1),
                        scale = c(0, true_scale),
                        mean = rep(true_mean, 2))
  pl.res2 <- ebnm_point_laplace(x, s, g_init = g_init2, fix_g = TRUE)

  expect_equal(pl.res$log_likelihood, pl.res2$log_likelihood)
})

test_that("Null case estimates pi0 = 1", {
  x <- rnorm(n, s = 0.5)
  pl.res <- ebnm_point_laplace(x, s = 1)
  expect_equal(pl.res[[g_ret_str()]]$pi[1], 1)
})

test_that("Very large observations give reasonable results", {
  scl <- 1e8
  pl.res <- ebnm_point_laplace(x, s, mode = "estimate")
  pl.res.lg <- ebnm_point_laplace(scl*x, scl*s, mode = "estimate")

  expect_equal(pl.res[[g_ret_str()]]$pi[1], pl.res.lg[[g_ret_str()]]$pi[1])
  expect_equal(scl * pl.res[[g_ret_str()]]$scale[2], pl.res.lg[[g_ret_str()]]$scale[2])
  expect_equal(scl * pl.res[[g_ret_str()]]$mean[1], pl.res.lg[[g_ret_str()]]$mean[1])
})

test_that("Very small observations give reasonable results", {
  scl <- 1e-8
  pl.res <- ebnm_point_laplace(x, s, mode = "estimate")
  pl.res.sm <- ebnm_point_laplace(scl*x, scl*s, mode = "estimate")

  expect_equal(pl.res[[g_ret_str()]]$pi[1], pl.res.sm[[g_ret_str()]]$pi[1])
  expect_equal(scl * pl.res[[g_ret_str()]]$scale[2], pl.res.sm[[g_ret_str()]]$scale[2])
  expect_equal(scl * pl.res[[g_ret_str()]]$mean[1], pl.res.sm[[g_ret_str()]]$mean[1])
})

# test_that("predict method works as expected", {
#   pl.res <- ebnm_point_laplace(x, s)
#   pl.res2 <- predict(pl.res, list(x = 1:10, s = 1))
#   expect_equal(pl.res$fitted_g, pl.res2$fitted_g)
# })