Skip to content

v1.7.0 — pprodnormal3() correctness fix

Pre-release
Pre-release

Choose a tag to compare

View all tags
@Data-Wise Data-Wise released this 17 Aug 05:55
· 9 commits to dev since this release
v1.7.0
d56e27f
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Fixes a correctness bug in pprodnormal3() that returned silently wrong probabilities on ill-conditioned covariance. If you use three-variable products (serial indirect effects a1 * a2 * b), install this now.

remotes::install_github("Data-Wise/RMediation", ref = "dev")

Pre-release. main remains at 1.6.1, which is what CRAN serves. CRAN's update cadence window opens 2026-08-21; this release exists so the fix is available before then.

What was wrong

pprodnormal3() (and its alias p_prod3()) returned materially wrong values whenever the (X, Y) covariance block was ill-conditioned — with no warning, no error, and no diagnostic. Tightening tol did not help: the integrator converged to a stable but incorrect value.

cdf(), dist_quantile(), confint() and ci() invert this CDF by root-finding, so all of them inherited the error.

At rho = 0.999, across the support:

q 1.6.1 1.7.0 reference
-2.0 0.0044691 (92% rel. error) 0.05711903 0.05711903
0.0 0.3445185 0.49727060 0.49727060
0.5 0.6109671 (24%) 0.80671502 0.80671502
2.0 0.6506248 0.92056384 0.92056384

Maximum error is now 5.9e-09, down from 2.7e-01.

A second, previously unreported failure is also fixed: at large standardized means (mean / sd beyond roughly 8) the old integrator returned exactly 0 where the true probability is around 0.46. That one was independent of conditioning.

What changed

The default integration method is now "gauss" — tensor-product Gauss-Legendre on a domain partitioned at the coordinate axes, with the node count escalated automatically until successive rules agree to tol.

  • Results change. On well-conditioned input the new default agrees with the old to about 1e-7. On ill-conditioned input it differs substantially; that difference is the correction.
  • Faster, by roughly 8-20x at the covariances tested.
  • method = "hcubature" still selects the previous integrator, for cross-checking. Not recommended otherwise.
  • New nodes argument forces a fixed rule; new diagnostics argument exposes the convergence estimate. diagnostics defaults to FALSE, so the return value stays a bare numeric.
  • pprodnormal3() now warns rather than returning a wrong answer silently when quadrature hits the node cap without meeting tol.

Verification

  • Test suite: 339 passed, 0 failed, 0 warnings, 0 skipped
  • R CMD check --as-cran: 0 errors, 0 warnings, 0 notes
  • CI: 6/6 green across Ubuntu (release + devel), macOS, and Windows

Reference values are Gauss-Legendre results self-consistent to 1e-14 across n = 512 / 1024 / 2048, each independently within 3 SE of Monte Carlo at 1e7 draws.

Full detail: SPEC-27-prod3-integrator.md (design and evidence) and SPEC-28-review-triage.md (review triage).

Assets 2
Loading