daniel1noble / BACE / 25340715961

91%

push

github

sim_bace: scale default sigmas down for non-gaussian variables

When response_type is gaussian, default sigma_species = sigma_residual
= 1 is innocuous: the response is on a linear scale. For variables
passed through a non-linear link (poisson via exp, binary/threshold/
categorical via probit/logit), the same defaults combined with
phylo_signal = 0.9 give sigma2_phylo = 18, which produces linear-
predictor SDs around 4-5. After exp() for poisson that is rates
spanning 10^-3 to 10^20 — a few species end up with counts in the
millions while most are zero, swamping the phylogenetic signal we
tried to encode.

Verified via Pagel's lambda recovery on simulated y:
  Before fix (sigma=1):
    poisson:    lambda_est ~ 0.07 (asked for 0.9)
    threshold:  lambda_est ~ 0.79
  After fix (sigma=sqrt(0.1) for non-gaussian):
    poisson:    lambda_est ~ 0.71
    threshold:  lambda_est ~ 0.79  (unchanged — was already OK)

The fix sets per-variable defaults based on each variable's type:
gaussian variables keep sigma = 1; non-gaussian variables get
sigma_species = sigma_residual = sqrt(0.1), which gives
sigma2_phylo = 1.8 + sigma2_other = 0.2 = total latent variance ~2
at phylo_signal = 0.9.

Effect on the imputation accuracy benchmark (multirep_v1_n20,
n_final=50, 20 reps per type):
                          v2 (no fix)            v3 (with fix)
  poisson NRMSE_pe        0.66                   0.81
  poisson Spearman        0.40                   0.59  (signal recovered)
  poisson paired_p        0.32                   0.046 (now significant)
  poisson BACE_better%    70%                    75%
  ordinal_K4 NRMSE        0.30                   0.22
  ordinal_K4 better%      95%                    100%
  binary better%          90%                    100%

Coverage stays near-nominal across types (0.85 - 0.92), confirming
BACE's intervals remain well-calibrated.

R CMD check: 0E/0W/1N. All 1251 tests pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

6 of 6 new or added lines in 1 file covered. (100.0%)

3356 of 3681 relevant lines covered (91.17%)

87.73 hits per line