Bernoulli probit

HOBBS form: y(i) ~ bernoulli_probit(eta);

The package test file performs:

Same model in HOBBS, Stan, and JAGS

param eta(1);
block eta(1) {
eta(1) ~ dnorm(0, 2);
for (i = 1:n) y(i) ~ bernoulli_probit(eta(1));
}
data {
  int<lower=1> n;
  array[n] int<lower=0, upper=1> y;
}
parameters {
  vector[1] eta;
}
model {
  eta[1] ~ normal(0, 2);
  for (i in 1:n) y[i] ~ bernoulli(Phi(eta[1]));
}
model {
  eta[1] ~ dnorm(0, 0.25)
  p <- phi(eta[1])
  for (i in 1:n) { y[i] ~ dbern(p) }
}

Run this test

From the documentation project itself:

library(hobbs)
library(testthat)
source("validation/source/tests/testthat/helper-distributions.R")
testthat::test_file("validation/source/tests/testthat/test-dist-bernoulli-probit.R")

From the HOBBS package source tree, the normal package test workflow is simpler:

devtools::test(filter = "dist-bernoulli-probit")

Download the test file · helper-distributions.R

Complete test file

test_that("bernoulli_probit recovers eta", {
  set.seed(121); truth <- 0.6; y <- rbinom(120, 1, pnorm(truth))
  model <- 'param eta(1);
block eta(1) {
eta(1) ~ dnorm(0, 2);
for (i = 1:n) y(i) ~ bernoulli_probit(eta(1));
}'
  d <- hobbs_test_draws(model, list(y = y)); expect_posterior_near(d, "eta[1]", truth, 0.35)
})

test_that("bernoulli_probit gives the correct posterior", {
  set.seed(121)
  truth <- 0.6
  y <- rbinom(120, 1, pnorm(truth))

  model <- 'param eta(1);
block eta(1) {
eta(1) ~ dnorm(0, 2);
for (i = 1:n) y(i) ~ bernoulli_probit(eta(1));
}'

  d <- hobbs_test_draws(model, list(y = y))

  log_posterior <- function(eta) {
    dnorm(eta, 0, 2, log = TRUE) +
      sum(dbinom(y, 1, pnorm(eta), log = TRUE))
  }

  expect_numerical_posterior(
    d, "eta[1]", log_posterior,
    lower = -2.5, upper = 2.5,
    mean_tolerance = 0.04, sd_tolerance = 0.03
  )
})

test_that("bernoulli_probit matches Stan and JAGS", {
  skip_if_reference_samplers_missing()
  set.seed(121)
  truth <- 0.6
  y <- rbinom(120, 1, pnorm(truth))
  n <- length(y)

  hobbs_model <- 'param eta(1);
block eta(1) {
eta(1) ~ dnorm(0, 2);
for (i = 1:n) y(i) ~ bernoulli_probit(eta(1));
}'

  stan_model <- '
data {
  int<lower=1> n;
  array[n] int<lower=0, upper=1> y;
}
parameters {
  vector[1] eta;
}
model {
  eta[1] ~ normal(0, 2);
  for (i in 1:n) y[i] ~ bernoulli(Phi(eta[1]));
}'

  jags_model <- '
model {
  eta[1] ~ dnorm(0, 0.25)
  p <- phi(eta[1])
  for (i in 1:n) { y[i] ~ dbern(p) }
}'

  d_hobbs <- hobbs_test_draws(hobbs_model, list(y = y))
  d_stan <- stan_test_draws(stan_model, list(n = n, y = y), "eta")
  d_jags <- jags_test_draws(jags_model, list(n = n, y = y), "eta")

  expect_posterior_matches_reference(d_hobbs, d_stan, d_jags, "eta[1]")
})