tests/testthat/test-estimate-liability.R

################################################################################

test_that("estimate_liability_multi errors clearly when target_phenotype is missing for PA", {
  # Regression test: with method = "PA" and target_phenotype left at its
  # default (NULL), cur_target_id was built as paste0("g_", NULL) == "g_",
  # which matches no real covariance-matrix column, so PA_algorithm()
  # failed later with an unrelated-looking "Target ID not found" error
  # instead of a clear message about the actual, upstream problem.
  tbl <- tibble::tibble(
    fid = "fam1", pid = c("proband", "mom"), role = c("o", "m"),
    lower_p1 = c(-Inf, -Inf), upper_p1 = c(0.5, 0.8),
    lower_p2 = c(-Inf, -Inf), upper_p2 = c(0.3, 0.6)
  )

  expect_error(
    estimate_liability_multi(.tbl = tbl, h2_vec = c(0.5, 0.5),
                             genetic_corrmat = diag(2), full_corrmat = diag(2),
                             phen_names = c("p1", "p2"),
                             fid = "fid", pid = "pid", role = "role",
                             method = "PA"),
    "target_phenotype"
  )
})

test_that("estimate_liability_multi does not validate target_phenotype for Gibbs", {
  # target_phenotype is never consulted by Gibbs_estimator(), so a mismatch
  # with phen_names should not block a Gibbs run.
  tbl <- tibble::tibble(
    fid = "fam1", pid = c("proband", "mom"), role = c("o", "m"),
    lower_p1 = c(-Inf, -Inf), upper_p1 = c(0.5, 0.8),
    lower_p2 = c(-Inf, -Inf), upper_p2 = c(0.3, 0.6)
  )

  expect_no_error(
    estimate_liability_multi(.tbl = tbl, h2_vec = c(0.5, 0.5),
                             genetic_corrmat = diag(2), full_corrmat = diag(2),
                             phen_names = c("p1", "p2"),
                             fid = "fid", pid = "pid", role = "role",
                             method = "Gibbs", target_phenotype = "not_a_real_phenotype",
                             out = 1, tol = 0.5)
  )
})

################################################################################
# .tbl and family_graphs are mutually exclusive inputs
################################################################################

test_that("estimate_liability_single errors when both .tbl and family_graphs are given", {
  tbl <- tibble::tibble(
    fid = "fam1", pid = c("proband", "mom"), role = c("o", "m"),
    lower = c(-Inf, 1.5), upper = c(Inf, 1.5)
  )
  fam <- data.frame(fam = 1, i = c("proband", "mom"), f = c("0", "0"), m = c("mom", "0"))
  thresholds <- data.frame(i = c("proband", "mom"), lower = -Inf, upper = 0.8)
  graph <- prepare_graph(fam, icol = "i", fcol = "f", mcol = "m", node_attributes = thresholds)
  fg <- get_family_graphs(graph, ndegree = 2, proband_vec = "proband", fid = "fid")

  expect_error(
    estimate_liability_single(.tbl = tbl, family_graphs = fg, h2 = 0.5,
                              pid = "pid", fid = "fid", role = "role", method = "PA"),
    "only one"
  )
})

test_that("estimate_liability_multi errors when both .tbl and family_graphs are given", {
  tbl <- tibble::tibble(
    fid = "fam1", pid = c("proband", "mom"), role = c("o", "m"),
    lower_p1 = c(-Inf, -Inf), upper_p1 = c(0.5, 0.8),
    lower_p2 = c(-Inf, -Inf), upper_p2 = c(0.3, 0.6)
  )
  fam <- data.frame(fam = 1, i = c("proband", "mom"), f = c("0", "0"), m = c("mom", "0"))
  thresholds <- data.frame(i = c("proband", "mom"), lower_p1 = -Inf, upper_p1 = 0.8,
                           lower_p2 = -Inf, upper_p2 = 0.8)
  graph <- prepare_graph(fam, icol = "i", fcol = "f", mcol = "m", node_attributes = thresholds)
  fg <- get_family_graphs(graph, ndegree = 2, proband_vec = "proband", fid = "fid")

  expect_error(
    estimate_liability_multi(.tbl = tbl, family_graphs = fg, h2_vec = c(0.5, 0.5),
                             genetic_corrmat = diag(2), full_corrmat = diag(2),
                             phen_names = c("p1", "p2"),
                             fid = "fid", pid = "pid", role = "role", method = "PA",
                             target_phenotype = "p1"),
    "only one"
  )
})

################################################################################
# PA correctness, against a hand-built PA_algorithm() ground truth
################################################################################

test_that("estimate_liability_single (PA) matches a direct PA_algorithm call", {
  tbl <- tibble::tibble(
    fid = "fam1", pid = c("proband", "mom"), role = c("o", "m"),
    lower = c(-Inf, 1.5), upper = c(Inf, 1.5)
  )
  res <- estimate_liability_single(.tbl = tbl, h2 = 0.5, pid = "pid", fid = "fid", role = "role",
                                   method = "PA", out = c(1))

  gt_cov <- construct_covmat(h2 = 0.5, fam_vec = c("m"))
  gt <- PA_algorithm(mu = rep(0, 3), covmat = gt_cov, target_id = "g",
                     lower = c(-Inf, -Inf, 1.5), upper = c(Inf, Inf, 1.5))

  expect_equal(res$est, gt$est)
  expect_equal(res$var, gt$var)
})

test_that("estimate_liability_single (PA, useMixture) matches a direct PA_algorithm call", {
  tbl <- tibble::tibble(
    fid = "fam1", pid = c("proband", "mom"), role = c("o", "m"),
    lower = c(-Inf, -Inf), upper = c(Inf, 1.2),
    K_i = c(NA, 0.05), K_pop = c(NA, 0.10)
  )
  res <- estimate_liability_single(.tbl = tbl, h2 = 0.5, pid = "pid", fid = "fid", role = "role",
                                   method = "PA", out = c(1), useMixture = TRUE)

  gt_cov <- construct_covmat(h2 = 0.5, fam_vec = c("m"))
  gt <- PA_algorithm(mu = rep(0, 3), covmat = gt_cov, target_id = "g",
                     lower = c(-Inf, -Inf, -Inf), upper = c(Inf, Inf, 1.2),
                     K_i = c(NA, NA, 0.05), K_pop = c(NA, NA, 0.10))

  expect_equal(res$est, gt$est)
  expect_equal(res$var, gt$var)
})

test_that("estimate_liability_multi (PA, useMixture) matches a direct PA_algorithm call", {
  tbl <- tibble::tibble(
    fid = "fam1", pid = c("proband", "mom"), role = c("o", "m"),
    lower_p1 = c(-Inf, -Inf), upper_p1 = c(Inf, 1.2), K_i_p1 = c(NA, 0.05), K_pop_p1 = c(NA, 0.10),
    lower_p2 = c(-Inf, -Inf), upper_p2 = c(Inf, 1.3), K_i_p2 = c(NA, 0.06), K_pop_p2 = c(NA, 0.12)
  )
  res <- estimate_liability_multi(.tbl = tbl, h2_vec = c(0.5, 0.5),
                                  genetic_corrmat = diag(2), full_corrmat = diag(2),
                                  phen_names = c("p1", "p2"), pid = "pid", fid = "fid", role = "role",
                                  method = "PA", out = c(1), useMixture = TRUE, target_phenotype = "p1")

  gt_cov <- construct_covmat(h2 = c(0.5, 0.5), genetic_corrmat = diag(2), full_corrmat = diag(2),
                             fam_vec = c("m"), phen_names = c("p1", "p2"))
  gt <- PA_algorithm(mu = rep(0, nrow(gt_cov)), covmat = gt_cov, target_id = "g_p1",
                     lower = rep(-Inf, 6), upper = c(Inf, Inf, 1.2, Inf, Inf, 1.3),
                     K_i = c(NA, NA, 0.05, NA, NA, 0.06), K_pop = c(NA, NA, 0.10, NA, NA, 0.12))

  expect_equal(res$est, gt$est)
  expect_equal(res$var, gt$var)
})

################################################################################
# pid reported for .tbl input
################################################################################

test_that("estimate_liability_single reports the proband's own pid for .tbl input, not fid", {
  # Regression test: the .tbl-input branch always set pid := cur_fid
  # (copying estimate_liability_multi()'s graph-input branch, where pid
  # and fid are the same by convention), instead of looking up the
  # proband's actual pid (role == "o") the way
  # estimate_liability_multi()'s own .tbl-input branch already did. Any
  # dataset where fid and the proband's pid genuinely differ - the normal
  # case - got the family id back in the pid column instead of the
  # individual's id.
  tbl <- tibble::tibble(
    fid = "family_42", pid = c("person_107", "person_108"), role = c("o", "m"),
    lower = c(-Inf, 1.5), upper = c(Inf, 1.5)
  )
  res <- estimate_liability_single(.tbl = tbl, h2 = 0.5, pid = "pid", fid = "fid", role = "role",
                                   method = "PA", out = c(1))
  expect_equal(res$pid, "person_107")
})

################################################################################
# Gibbs_estimator's out handling for multiple phenotypes
################################################################################

test_that("Gibbs_estimator respects out for multiple phenotypes instead of always returning genetic", {
  # Regression test: updated_out used to be
  # which(str_detect(colnames(cov), "g_")), which always selected every
  # phenotype's genetic-liability column regardless of `out`. Requesting
  # out = 2 ("full") silently returned genetic liability mislabeled as
  # "full", and out = c(1, 2) crashed outright, since updated_out (length
  # = n phenotypes) and return_names (length = n phenotypes * length(out))
  # disagreed.
  tbl <- tibble::tibble(
    fid = "fam1", pid = c("proband", "mom"), role = c("o", "m"),
    lower_p1 = c(-Inf, -Inf), upper_p1 = c(Inf, 1.2),
    lower_p2 = c(-Inf, -Inf), upper_p2 = c(Inf, 1.3)
  )

  res_genetic <- estimate_liability_multi(.tbl = tbl, h2_vec = c(0.5, 0.5),
                                          genetic_corrmat = diag(2), full_corrmat = diag(2),
                                          phen_names = c("p1", "p2"), pid = "pid", fid = "fid", role = "role",
                                          method = "Gibbs", out = c(1), tol = 1)
  res_full <- estimate_liability_multi(.tbl = tbl, h2_vec = c(0.5, 0.5),
                                       genetic_corrmat = diag(2), full_corrmat = diag(2),
                                       phen_names = c("p1", "p2"), pid = "pid", fid = "fid", role = "role",
                                       method = "Gibbs", out = c(2), tol = 1)
  res_both <- estimate_liability_multi(.tbl = tbl, h2_vec = c(0.5, 0.5),
                                       genetic_corrmat = diag(2), full_corrmat = diag(2),
                                       phen_names = c("p1", "p2"), pid = "pid", fid = "fid", role = "role",
                                       method = "Gibbs", out = c(1, 2), tol = 1)

  expect_setequal(colnames(res_genetic), c("fid", "pid", "genetic_p1_est", "genetic_p1_se", "genetic_p2_est", "genetic_p2_se"))
  expect_setequal(colnames(res_full), c("fid", "pid", "full_p1_est", "full_p1_se", "full_p2_est", "full_p2_se"))
  expect_setequal(colnames(res_both), c("fid", "pid",
                                        "genetic_p1_est", "genetic_p1_se", "full_p1_est", "full_p1_se",
                                        "genetic_p2_est", "genetic_p2_se", "full_p2_est", "full_p2_se"))

  # genetic and full liability must actually differ (h2 = 0.5, not 0 or 1) -
  # catches the bug where "full" silently returned the genetic values
  expect_false(isTRUE(all.equal(res_full$full_p1_est, res_genetic$genetic_p1_est, tolerance = 1e-3)))

  # out = c(1,2) must reproduce the separate single-out runs (within Monte
  # Carlo noise, same seed-free Gibbs run - loose tolerance)
  expect_equal(res_both$genetic_p1_est, res_genetic$genetic_p1_est, tolerance = 0.2)
  expect_equal(res_both$full_p1_est, res_full$full_p1_est, tolerance = 0.2)
})

test_that("Gibbs_estimator's out handling works for the graph path too, not just .tbl", {
  # Regression test: the first fix for the above used
  # match(paste0(c("g","o")[out], "_", phen_names), colnames(cov)), which
  # assumes .tbl-style column names ("g_p1", "o_p1"). The graph path names
  # its columns from the proband's own id instead (e.g. "proband_g_p1" for
  # genetic, just "proband_p1" - no "o" marker at all - for full liability
  # via graph_based_covariance_construction_multi()), so every name-based
  # match failed (all NA), surfacing as "out is not of the required
  # format" warnings and a dimnames-length error. Fixed by position
  # instead of name: both paths guarantee genetic liability is the 1st and
  # full liability the 2nd entry of each phenotype's block in cov.
  fam <- data.frame(fam = 1, i = c("proband", "mom"), f = c("0", "0"), m = c("mom", "0"))
  thresholds <- data.frame(i = c("proband", "mom"), lower_p1 = -Inf, upper_p1 = 0.8,
                           lower_p2 = -Inf, upper_p2 = 0.8)
  graph <- prepare_graph(fam, icol = "i", fcol = "f", mcol = "m", node_attributes = thresholds)
  fg <- get_family_graphs(graph, ndegree = 2, proband_vec = "proband", fid = "fid")

  res_genetic <- estimate_liability_multi(family_graphs = fg, h2_vec = c(0.5, 0.5),
                                          genetic_corrmat = diag(2), full_corrmat = diag(2),
                                          phen_names = c("p1", "p2"), fid = "fid", pid = "pid",
                                          method = "Gibbs", out = c(1), tol = 1)
  res_full <- estimate_liability_multi(family_graphs = fg, h2_vec = c(0.5, 0.5),
                                       genetic_corrmat = diag(2), full_corrmat = diag(2),
                                       phen_names = c("p1", "p2"), fid = "fid", pid = "pid",
                                       method = "Gibbs", out = c(2), tol = 1)
  res_both <- estimate_liability_multi(family_graphs = fg, h2_vec = c(0.5, 0.5),
                                       genetic_corrmat = diag(2), full_corrmat = diag(2),
                                       phen_names = c("p1", "p2"), fid = "fid", pid = "pid",
                                       method = "Gibbs", out = c(1, 2), tol = 1)

  expect_setequal(colnames(res_genetic), c("fid", "pid", "genetic_p1_est", "genetic_p1_se", "genetic_p2_est", "genetic_p2_se"))
  expect_setequal(colnames(res_full), c("fid", "pid", "full_p1_est", "full_p1_se", "full_p2_est", "full_p2_se"))
  expect_setequal(colnames(res_both), c("fid", "pid",
                                        "genetic_p1_est", "genetic_p1_se", "full_p1_est", "full_p1_se",
                                        "genetic_p2_est", "genetic_p2_se", "full_p2_est", "full_p2_se"))
  expect_false(isTRUE(all.equal(res_full$full_p1_est, res_genetic$genetic_p1_est, tolerance = 1e-3)))
})

Try the LTFGRS package in your browser

Any scripts or data that you put into this service are public.

LTFGRS documentation built on Nov. 20, 2026, 5:10 p.m.