daniel1noble / BACE / 25330902065

91%

push

github

Fix two source bugs in sim_bace + add numerical correctness tests

While writing the numerical-correctness test fixtures (test-numerical
_correctness.R) I tripped over two real source bugs:

1. sim_bace produced a lowercase 'species' column. The rest of BACE
   (bace_imp examples, bundled .rda datasets, ran_phylo_form
   conventions) uses capital 'Species'. This forced anyone feeding
   sim_bace output into bace_imp to rename the column first --
   easy to miss, no helpful error message. Fixed by emitting
   'Species' directly from sim_bace's output assembly.

2. generate_default_beta_matrix(n_predictors = 1) crashed with
   "subscript out of bounds": R's `2:1` ⇒ c(2, 1) quirk fed
   beta_matrix[2, 1] on a 1x1 matrix. Fixed by guarding the inner
   loop behind n_predictors >= 2 and adding an n_predictors < 1
   error.

Both fixes have explicit regression tests in test-simulate.R
(test_that "sim_bace output uses 'Species'" / "generate_default_beta
_matrix handles n_predictors = 1 cleanly").

New file tests/testthat/test-numerical_correctness.R adds 5 tests
that go beyond the existing shape/structure smoke tests to verify
BACE actually recovers known structure under controlled simulation:

  - BACE imputed values correlate with truth on hidden cells when
    phylogenetic signal is high (40 spp, signal=0.8: r > 0.30).
  - BACE outperforms a column-mean baseline (80 spp, signal=0.8,
    averaged over the last 3 imputation runs: bace_rmse < 0.95 *
    baseline_rmse).
  - phylo_signal_summary recovers a known simulated lambda > 0.30
    when signal=0.9 (typical estimate ~ 0.5 at n=60).
  - pool_posteriors produces a well-formed pooled MCMCglmm with
    finite Sol and positive intercept variance.
  - assess_convergence runs successfully on both short (under-
    converged) and long (more-converged) chains -- tracks summary
    stability without insisting on flaky pass/fail thresholds.

All gated with skip_on_cran() so they execute in the GHA workflow
(NOT_CR... (continued)

9 of 9 new or added lines in 2 files covered. (100.0%)

2476 of 3670 relevant lines covered (67.47%)

53.59 hits per line