marginal_effects(): report covariates of the NTN mu design

[davidhbernstein]

Bug

For sfm(model_name = "NTN") fitted with uhet/muhet, marginal_effects() built its covariate list from the σ_u design alone (colnames(Z)[keep], where Z is z_spec$Z). The mu_spec δs were then looked up by name for those covariates only. As a result:

• uhet = ~ z1, muhet = ~ z2 returned a z1 column and no z2 column, with no warning. Yet z2 moves E[u] through μ.
• muhet = ~ z alone, where σ_u is intercept-only, stopped with "every column of the variance-determinant design is constant".

Both contradict ?marginal_effects, which says a covariate in one design only "contributes only through that one".

Fix

The chain-rule loop already handled a covariate missing from one design, since .sfa_me_delta_by_name() zero-fills it. It was simply never given the μ-only names. The covariate list for the truncated-normal branch is now the union, by name, of the non-constant columns of both designs. The σ_u δ is zero-filled the same way.

Other paths are unaffected:

• the scaling path, half-normal and exponential (NHN_Z, NE_Z);
• psfm (TRE_Z, GTRE_Z), including component = "h".

They build the same nm and d as before. All 160 existing marginal-effects expectations pass unchanged.

Test

test-m4-truncnormal-marginal.R gains "a covariate in the mu design only is reported, not dropped".

• It fits uhet = ~ z1, muhet = ~ z2.
• It checks dE_u.dz1, dE_u.dz2 and dVar_u.dz2 at three rows against numDeriv derivatives of the closed-form truncated-normal moments. Each covariate is moved in its own design, independently of the package's helpers, with a 1e-5 relative tolerance.
• It also checks the μ-only fit.

On main this test fails: z2 is absent and the μ-only call errors. On this branch it passes.

Docs (issue #3)

?marginal_effects gains a table of which fits are supported, their link, and which covariates get a column. The links were checked against sfm(): the NTN het path honours z_link, and psfm uses the variance link. \value now lists the "component" attribute and the sigma_h/E_h/Var_h/dE_h.d<z> names used for component = "h". \usage is unchanged.

Local suite

5239 passed, 0 failed, 4 errors. All 4 errors are plm::pdata.frame(): already a pdata.frame in test-panel-boundary.R and test-gtre-draws.R. They come from the local plm 2.6.3, not this change, and main's CI is green on those files.

🤖 Generated with Claude Code

https://claude.ai/code/session_014oJL17ewiahA8pWbemnqNp

---

Generated by Claude Code