tests/testthat/test_boin.R

In escalation: Modular Approach to Dose Finding Clinical Trials

test_that('BOIN recommendations match published example.', {

  # p.25 of Han, Pan, Zhang, Liu & Yuan (2019)

  num_doses <- 5
  target <- 0.3
  boin_fitter <- get_boin(num_doses = num_doses, target = target)

  x <- fit(boin_fitter, '1NNN')
  expect_equal(recommended_dose(x), 2)
  expect_true(continue(x))

  x <- fit(boin_fitter, '1NNN 2NNN')
  expect_equal(recommended_dose(x), 3)
  expect_true(continue(x))

  x <- fit(boin_fitter, '1NNN 2NNN 3NTT')
  expect_equal(recommended_dose(x), 2)
  expect_true(continue(x))

  x <- fit(boin_fitter, '1NNN 2NNN 3NTT 2NTN')
  expect_equal(recommended_dose(x), 3)
  expect_true(continue(x))

  x <- fit(boin_fitter, '1NNN 2NNN 3NTT 2NTN 3NNN')
  expect_equal(recommended_dose(x), 3)
  expect_true(continue(x))

  x <- fit(boin_fitter, '1NNN 2NNN 3NTT 2NTN 3NNN 3NNN')
  expect_equal(recommended_dose(x), 4)
  expect_true(continue(x))

  x <- fit(boin_fitter, '1NNN 2NNN 3NTT 2NTN 3NNN 3NNN 4TTT')
  expect_equal(recommended_dose(x), 3)
  expect_true(continue(x))

  x <- fit(boin_fitter, '1NNN 2NNN 3NTT 2NTN 3NNN 3NNN 4TTT 3NTN')
  expect_equal(recommended_dose(x), 3)
  expect_true(continue(x))

  x <- fit(boin_fitter, '1NNN 2NNN 3NTT 2NTN 3NNN 3NNN 4TTT 3NTN 3NNT')
  expect_equal(recommended_dose(x), 3)
  expect_true(continue(x))

  x <- fit(boin_fitter, '1NNN 2NNN 3NTT 2NTN 3NNN 3NNN 4TTT 3NTN 3NNT 3TNN')
  expect_equal(recommended_dose(x), 3)
  expect_true(continue(x))

  expect_equal(n_at_dose(x), c(3, 6, 18, 3, 0))
  expect_equal(tox_at_dose(x), c(0, 1, 5, 3, 0))
})


test_that('boin_selector supports correct interface.', {

  num_doses <- 5
  target <- 0.3

  model_fitter <- get_boin(num_doses = num_doses, target = target)

  # Example 1, using outcome string
  x <- fit(model_fitter, '1NNN 2NTT')

  expect_equal(tox_target(x), 0.3)
  expect_true(is.numeric(tox_target(x)))

  expect_equal(num_patients(x), 6)
  expect_true(is.integer(num_patients(x)))

  expect_equal(cohort(x), c(1,1,1, 2,2,2))
  expect_true(is.integer(cohort(x)))
  expect_equal(length(cohort(x)), num_patients(x))

  expect_equal(doses_given(x), c(1,1,1, 2,2,2))
  expect_true(is.integer(doses_given(x)))
  expect_equal(length(doses_given(x)), num_patients(x))

  expect_equal(tox(x), c(0,0,0, 0,1,1))
  expect_true(is.integer(tox(x)))
  expect_equal(length(tox(x)), num_patients(x))

  expect_true(all((model_frame(x) - data.frame(patient = c(1,2,3,4,5,6),
                                               cohort = c(1,1,1,2,2,2),
                                               dose = c(1,1,1,2,2,2),
                                               tox = c(0,0,0,0,1,1))) == 0))
  expect_equal(nrow(model_frame(x)), num_patients(x))

  expect_equal(num_doses(x), 5)
  expect_true(is.integer(num_doses(x)))

  expect_equal(dose_indices(x), 1:5)
  expect_true(is.integer(dose_indices(x)))
  expect_equal(length(dose_indices(x)), num_doses(x))

  expect_equal(recommended_dose(x), 1)
  expect_true(is.integer(recommended_dose(x)))
  expect_equal(length(recommended_dose(x)), 1)

  expect_equal(continue(x), TRUE)
  expect_true(is.logical(continue(x)))

  expect_equal(n_at_dose(x), c(3,3,0,0,0))
  expect_true(is.integer(n_at_dose(x)))
  expect_equal(length(n_at_dose(x)), num_doses(x))

  expect_equal(n_at_dose(x, dose = 0), 0)
  expect_true(is.integer(n_at_dose(x, dose = 0)))
  expect_equal(length(n_at_dose(x, dose = 0)), 1)

  expect_equal(n_at_dose(x, dose = 1), 3)
  expect_true(is.integer(n_at_dose(x, dose = 1)))
  expect_equal(length(n_at_dose(x, dose = 1)), 1)

  expect_equal(n_at_dose(x, dose = 'recommended'), 3)
  expect_true(is.integer(n_at_dose(x, dose = 'recommended')))
  expect_equal(length(n_at_dose(x, dose = 'recommended')), 1)

  expect_equal(n_at_recommended_dose(x), 3)
  expect_true(is.integer(n_at_recommended_dose(x)))
  expect_equal(length(n_at_recommended_dose(x)), 1)

  expect_equal(is_randomising(x), FALSE)
  expect_true(is.logical(is_randomising(x)))
  expect_equal(length(is_randomising(x)), 1)

  expect_equal(unname(prob_administer(x)), c(0.5,0.5,0,0,0))
  expect_true(is.numeric(prob_administer(x)))
  expect_equal(length(prob_administer(x)), num_doses(x))

  expect_equal(tox_at_dose(x), c(0,2,0,0,0))
  expect_true(is.integer(tox_at_dose(x)))
  expect_equal(length(tox_at_dose(x)), num_doses(x))

  expect_true(is.numeric(empiric_tox_rate(x)))
  expect_equal(length(empiric_tox_rate(x)), num_doses(x))

  expect_true(is.numeric(mean_prob_tox(x)))
  expect_equal(length(mean_prob_tox(x)), num_doses(x))

  expect_true(is.numeric(median_prob_tox(x)))
  expect_equal(length(median_prob_tox(x)), num_doses(x))

  expect_true(is.logical(dose_admissible(x)))
  expect_equal(length(dose_admissible(x)), num_doses(x))

  expect_true(is.numeric(prob_tox_quantile(x, p = 0.9)))
  expect_equal(length(prob_tox_quantile(x, p = 0.9)), num_doses(x))

  expect_true(is.numeric(prob_tox_exceeds(x, 0.5)))
  expect_equal(length(prob_tox_exceeds(x, 0.5)), num_doses(x))

  expect_true(is.logical(supports_sampling(x)))

  expect_error(prob_tox_samples(x))
  expect_error(prob_tox_samples(x, tall = TRUE))

  # Expect summary to not error. This is how that is tested, apparently:
  expect_error(summary(x), NA)
  expect_output(print(x))
  expect_true(tibble::is_tibble(as_tibble(x)))
  expect_true(nrow(as_tibble(x)) >= num_doses(x))


  # Example 2, using trivial outcome string
  x <- fit(model_fitter, '')

  expect_equal(tox_target(x), 0.3)
  expect_true(is.numeric(tox_target(x)))

  expect_equal(num_patients(x), 0)
  expect_true(is.integer(num_patients(x)))

  expect_equal(cohort(x), integer(0))
  expect_true(is.integer(cohort(x)))
  expect_equal(length(cohort(x)), num_patients(x))

  expect_equal(doses_given(x), integer(0))
  expect_true(is.integer(doses_given(x)))
  expect_equal(length(doses_given(x)), num_patients(x))

  expect_equal(tox(x), integer(0))
  expect_true(is.integer(tox(x)))
  expect_equal(length(tox(x)), num_patients(x))

  expect_equal(num_tox(x), 0)
  expect_true(is.integer(num_tox(x)))

  expect_equal(dose_indices(x), 1:5)
  expect_true(is.integer(dose_indices(x)))
  expect_equal(length(dose_indices(x)), num_doses(x))

  mf <- model_frame(x)
  expect_equal(nrow(mf), 0)
  expect_equal(ncol(mf), 4)

  expect_equal(num_doses(x), 5)
  expect_true(is.integer(num_doses(x)))

  expect_equal(recommended_dose(x), 1)
  expect_true(is.integer(recommended_dose(x)))
  expect_equal(length(recommended_dose(x)), 1)

  expect_equal(continue(x), TRUE)
  expect_true(is.logical(continue(x)))

  expect_equal(n_at_dose(x), c(0,0,0,0,0))
  expect_true(is.integer(n_at_dose(x)))
  expect_equal(length(n_at_dose(x)), num_doses(x))

  expect_equal(n_at_dose(x, dose = 0), 0)
  expect_true(is.integer(n_at_dose(x, dose = 0)))
  expect_equal(length(n_at_dose(x, dose = 0)), 1)

  expect_equal(n_at_dose(x, dose = 1), 0)
  expect_true(is.integer(n_at_dose(x, dose = 1)))
  expect_equal(length(n_at_dose(x, dose = 1)), 1)

  expect_equal(n_at_dose(x, dose = 'recommended'), 0)
  expect_true(is.integer(n_at_dose(x, dose = 'recommended')))
  expect_equal(length(n_at_dose(x, dose = 'recommended')), 1)

  expect_equal(n_at_recommended_dose(x), 0)
  expect_true(is.integer(n_at_recommended_dose(x)))
  expect_equal(length(n_at_recommended_dose(x)), 1)

  expect_equal(is_randomising(x), FALSE)
  expect_true(is.logical(is_randomising(x)))
  expect_equal(length(is_randomising(x)), 1)

  expect_true(is.numeric(prob_administer(x)))
  expect_equal(length(prob_administer(x)), num_doses(x))

  expect_equal(tox_at_dose(x), c(0,0,0,0,0))
  expect_true(is.integer(tox_at_dose(x)))
  expect_equal(length(tox_at_dose(x)), num_doses(x))

  expect_true(is.numeric(empiric_tox_rate(x)))
  expect_equal(length(empiric_tox_rate(x)), num_doses(x))

  expect_true(is.numeric(mean_prob_tox(x)))
  expect_equal(length(mean_prob_tox(x)), num_doses(x))

  expect_true(is.numeric(median_prob_tox(x)))
  expect_equal(length(median_prob_tox(x)), num_doses(x))

  expect_true(is.logical(dose_admissible(x)))
  expect_equal(length(dose_admissible(x)), num_doses(x))

  expect_true(is.numeric(prob_tox_quantile(x, p = 0.9)))
  expect_equal(length(prob_tox_quantile(x, p = 0.9)), num_doses(x))

  expect_true(is.numeric(prob_tox_exceeds(x, 0.5)))
  expect_equal(length(prob_tox_exceeds(x, 0.5)), num_doses(x))

  expect_true(is.logical(supports_sampling(x)))

  expect_error(prob_tox_samples(x))
  expect_error(prob_tox_samples(x, tall = TRUE))

  # Expect summary to not error. This is how that is tested, apparently:
  expect_error(summary(x), NA)
  expect_output(print(x))
  expect_true(tibble::is_tibble(as_tibble(x)))
  expect_true(nrow(as_tibble(x)) >= num_doses(x))


  # Example 3, using tibble
  outcomes <- tibble(
    cohort = c(1,1,1, 2,2,2),
    dose = c(1,1,1, 2,2,2),
    tox = c(0,0, 0,0, 1,1)
  )
  x <- fit(model_fitter, outcomes)

  expect_equal(tox_target(x), 0.3)
  expect_true(is.numeric(tox_target(x)))

  expect_equal(num_patients(x), 6)
  expect_true(is.integer(num_patients(x)))

  expect_equal(cohort(x), c(1,1,1, 2,2,2))
  expect_true(is.integer(cohort(x)))
  expect_equal(length(cohort(x)), num_patients(x))

  expect_equal(doses_given(x), c(1,1,1, 2,2,2))
  expect_true(is.integer(doses_given(x)))
  expect_equal(length(doses_given(x)), num_patients(x))

  expect_equal(tox(x), c(0,0,0, 0,1,1))
  expect_true(is.integer(tox(x)))
  expect_equal(length(tox(x)), num_patients(x))

  expect_true(all((model_frame(x) - data.frame(patient = c(1,2,3,4,5,6),
                                               cohort = c(1,1,1,2,2,2),
                                               dose = c(1,1,1,2,2,2),
                                               tox = c(0,0,0,0,1,1))) == 0))
  expect_equal(nrow(model_frame(x)), num_patients(x))

  expect_equal(num_doses(x), 5)
  expect_true(is.integer(num_doses(x)))

  expect_equal(dose_indices(x), 1:5)
  expect_true(is.integer(dose_indices(x)))
  expect_equal(length(dose_indices(x)), num_doses(x))

  expect_equal(recommended_dose(x), 1)
  expect_true(is.integer(recommended_dose(x)))
  expect_equal(length(recommended_dose(x)), 1)

  expect_equal(continue(x), TRUE)
  expect_true(is.logical(continue(x)))

  expect_equal(n_at_dose(x), c(3,3,0,0,0))
  expect_true(is.integer(n_at_dose(x)))
  expect_equal(length(n_at_dose(x)), num_doses(x))

  expect_equal(n_at_dose(x, dose = 0), 0)
  expect_true(is.integer(n_at_dose(x, dose = 0)))
  expect_equal(length(n_at_dose(x, dose = 0)), 1)

  expect_equal(n_at_dose(x, dose = 1), 3)
  expect_true(is.integer(n_at_dose(x, dose = 1)))
  expect_equal(length(n_at_dose(x, dose = 1)), 1)

  expect_equal(n_at_dose(x, dose = 'recommended'), 3)
  expect_true(is.integer(n_at_dose(x, dose = 'recommended')))
  expect_equal(length(n_at_dose(x, dose = 'recommended')), 1)

  expect_equal(n_at_recommended_dose(x), 3)
  expect_true(is.integer(n_at_recommended_dose(x)))
  expect_equal(length(n_at_recommended_dose(x)), 1)

  expect_equal(is_randomising(x), FALSE)
  expect_true(is.logical(is_randomising(x)))
  expect_equal(length(is_randomising(x)), 1)

  expect_equal(unname(prob_administer(x)), c(0.5,0.5,0,0,0))
  expect_true(is.numeric(prob_administer(x)))
  expect_equal(length(prob_administer(x)), num_doses(x))

  expect_equal(tox_at_dose(x), c(0,2,0,0,0))
  expect_true(is.integer(tox_at_dose(x)))
  expect_equal(length(tox_at_dose(x)), num_doses(x))

  expect_true(is.numeric(empiric_tox_rate(x)))
  expect_equal(length(empiric_tox_rate(x)), num_doses(x))

  expect_true(is.numeric(mean_prob_tox(x)))
  expect_equal(length(mean_prob_tox(x)), num_doses(x))

  expect_true(is.numeric(median_prob_tox(x)))
  expect_equal(length(median_prob_tox(x)), num_doses(x))

  expect_true(is.logical(dose_admissible(x)))
  expect_equal(length(dose_admissible(x)), num_doses(x))

  expect_true(is.numeric(prob_tox_quantile(x, p = 0.9)))
  expect_equal(length(prob_tox_quantile(x, p = 0.9)), num_doses(x))

  expect_true(is.numeric(prob_tox_exceeds(x, 0.5)))
  expect_equal(length(prob_tox_exceeds(x, 0.5)), num_doses(x))

  expect_true(is.logical(supports_sampling(x)))

  expect_error(prob_tox_samples(x))
  expect_error(prob_tox_samples(x, tall = TRUE))

  # Expect summary to not error. This is how that is tested, apparently:
  expect_error(summary(x), NA)
  expect_output(print(x))
  expect_true(tibble::is_tibble(as_tibble(x)))
  expect_true(nrow(as_tibble(x)) >= num_doses(x))

})


test_that('BOIN advises stopping when indicated', {

  num_doses <- 5
  target <- 0.3

  # Under these parameters, 0/1 tox will see escalation, 1/2 tox will see
  # descalation, and 3/3 or 3/4 will see elimination.
  # BOIN::get.boundary(target = target, ncohort = 1, cohortsize = 5)

  boin_fitter <- get_boin(num_doses = num_doses, target = target,
                          use_stopping_rule = TRUE)

  # Design should continue
  x <- fit(boin_fitter, '1T')
  expect_equal(recommended_dose(x), 1)
  expect_true(continue(x))
  expect_equal(dose_admissible(x), rep(TRUE, num_doses(x)))

  # Design should continue
  x <- fit(boin_fitter, '1TT')
  expect_equal(recommended_dose(x), 1)
  expect_true(continue(x))
  expect_equal(dose_admissible(x), rep(TRUE, num_doses(x)))

  # Design should stop
  x <- fit(boin_fitter, '1TTT')
  expect_true(is.na(recommended_dose(x)))
  expect_false(continue(x))
  expect_equal(dose_admissible(x), rep(FALSE, num_doses(x)))

  # Design should escalate
  x <- fit(boin_fitter, '1N')
  expect_equal(recommended_dose(x), 2)
  expect_true(continue(x))
  expect_equal(dose_admissible(x), rep(TRUE, num_doses(x)))

  # Design should de-escalate
  x <- fit(boin_fitter, '1N 2TN')
  expect_equal(recommended_dose(x), 1)
  expect_true(continue(x))
  expect_equal(dose_admissible(x), rep(TRUE, num_doses(x)))

  # Design should not yet stop
  x <- fit(boin_fitter, '1N 2TN 1TT')
  expect_equal(recommended_dose(x), 1)
  expect_true(continue(x))
  expect_equal(dose_admissible(x), rep(TRUE, num_doses(x)))

  # Now design should stop
  x <- fit(boin_fitter, '1N 2TN 1TTT')
  expect_true(is.na(recommended_dose(x)))
  expect_false(continue(x))
  expect_equal(dose_admissible(x), rep(FALSE, num_doses(x)))

  # If those 3 in 4 DLTs occurred at a higher dose, trial should continue but
  # toxic dose and those doses above should be inadmissible
  x <- fit(boin_fitter, '1N 4TTTT')
  expect_equal(recommended_dose(x), 3)
  expect_true(continue(x))
  expect_equal(dose_admissible(x), c(TRUE, TRUE, TRUE, FALSE, FALSE))

})


test_that('BOIN stopping rule can be turned off.', {

  num_doses <- 5
  target <- 0.3

  # Under these parameters, 0/1 tox will see escalation, 1/2 tox will see
  # descalation, and 3/3 or 3/4 will see elimination.
  # BOIN::get.boundary(target = target, ncohort = 1, cohortsize = 5)

  boin_fitter <- get_boin(num_doses = num_doses, target = target,
                          use_stopping_rule = FALSE)


  # Design should escalate
  x <- fit(boin_fitter, '1N')
  expect_equal(recommended_dose(x), 2)
  expect_true(continue(x))
  expect_equal(dose_admissible(x), rep(TRUE, num_doses(x)))

  # Design should de-escalate
  x <- fit(boin_fitter, '1N 2TN')
  expect_equal(recommended_dose(x), 1)
  expect_true(continue(x))
  expect_equal(dose_admissible(x), rep(TRUE, num_doses(x)))

  # Design should not stop here
  x <- fit(boin_fitter, '1N 2TN 1TT')
  expect_equal(recommended_dose(x), 1)
  expect_true(continue(x))
  expect_equal(dose_admissible(x), rep(TRUE, num_doses(x)))

  # Design should not stop here either
  x <- fit(boin_fitter, '1N 2TN 1TTT')
  expect_equal(recommended_dose(x), 1)
  expect_true(continue(x))
  expect_equal(dose_admissible(x), rep(TRUE, num_doses(x)))

  # If those 3 in 4 DLTs occurred at a higher dose, the trial should still
  # continue and all doses should be admissible
  x <- fit(boin_fitter, '1N 4TTTT')
  expect_equal(recommended_dose(x), 3)
  expect_true(continue(x))
  expect_equal(dose_admissible(x), rep(TRUE, num_doses(x)))

  # Compare to tests above, this shows that the stopping rule has been disabled.
})


test_that('BOIN prob_tox_exceeds matches boin package', {

  num_doses <- 5
  target <- 0.3
  boin_fitter <- get_boin(num_doses = num_doses, target = target)

  outcomes <- data.frame(
    cohort = c(1,1, 2,2, 3,3, 5,5),
    dose = c(1,1, 2,2, 3,3, 5,5),
    tox = c(0,0, 0,0, 1,1, 1,0)
  )
  x <- fit(boin_fitter, outcomes)

  prob_tox_1 <- prob_tox_exceeds(x, threshold = target)

  prob_tox_2 <- as.character(x$boin_fit$p_overdose)
  prob_tox_2[prob_tox_2 == '----'] <- NA
  prob_tox_2 <- as.numeric(prob_tox_2)

  # Expect NAs in same place:
  expect_true(all(is.na(prob_tox_1) == is.na(prob_tox_2)))

  # And similar values where not NA
  expect_true(all(abs(prob_tox_1[!is.na(prob_tox_1)] -
                        prob_tox_2[!is.na(prob_tox_2)]) < 0.01))

})


test_that('boin_selector respects eliminated doses', {

  model <- get_boin(num_doses = 3, target = 0.25, use_stopping_rule = TRUE)

  # After 3/3 tox, it should descalate but never re-escalate:
  fit <- model %>% fit('2TTT')
  expect_equal(fit %>% recommended_dose(), 1)
  expect_true(fit %>% continue())
  expect_equal(fit %>% dose_admissible(), c(TRUE, FALSE, FALSE))

  fit <- model %>% fit('2TTT 1N')
  expect_equal(fit %>% recommended_dose(), 1)
  expect_true(fit %>% continue())
  expect_equal(fit %>% dose_admissible(), c(TRUE, FALSE, FALSE))

  fit <- model %>% fit('2TTT 1NN')
  expect_equal(fit %>% recommended_dose(), 1)
  expect_true(fit %>% continue())
  expect_equal(fit %>% dose_admissible(), c(TRUE, FALSE, FALSE))

  fit <- model %>% fit('2TTT 1NNN')
  expect_equal(fit %>% recommended_dose(), 1)
  expect_true(fit %>% continue())
  expect_equal(fit %>% dose_admissible(), c(TRUE, FALSE, FALSE))

  fit <- model %>% fit('2TTT 1NNNNNNNNNNN')
  expect_equal(fit %>% recommended_dose(), 1)
  expect_true(fit %>% continue())
  expect_equal(fit %>% dose_admissible(), c(TRUE, FALSE, FALSE))

})