菜单
THE core skill. Enforces component-labelled paths — never describe or compute a non-mu path as a mean effect. Use for ANY prose, docs, message, or code that states what a drmSEM path MEANS.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | drmsem-semantics |
| description | THE core skill. Enforces component-labelled paths — never describe or compute a non-mu path as a mean effect. Use for ANY prose, docs, message, or code that states what a drmSEM path MEANS. |
A drmSEM edge is (from, to, component, link, term). component is the
distributional parameter of the endogenous node to that from targets. The
recurring fatal mistake this skill prevents: treating EVERY path as a mean
effect. A path to sigma is not a path to the mean; a path to zi is not a
path to the conditional mean.
| component | targets | link (typical) | correct reading of an X→Y path |
|---|---|---|---|
mu | conditional mean/location | family link (identity/log/logit/...) | X shifts the expected value of Y |
sigma | residual scale / dispersion | log | X changes spread/dispersion of Y, not its mean |
nu | shape (e.g. Student df) | log | X changes tail/shape of Y's distribution |
zi | zero-inflation probability | logit | X changes the chance of a structural (extra) zero |
hu | hurdle probability | logit | X changes the probability of crossing the hurdle |
sd(group) (sd_*) | random-effect scale | log | X changes among-group heterogeneity, not the mean |
rho12 | bivariate residual correlation | tanh | X changes the residual correlation between two responses |
(Links match drm_nominal_link() in R/edges.R. mu's link follows the family.)
component from paths(object) / the edge table — never assume mu.sigma of Y", not "effect on Y".mu, sigma, nu, zi, hu, sd(group),
rho12, "component-labelled path", "distribution-mediated effect".sigma(abundance): a log-scale increase in residual dispersion."sigma.total_effects(mediation = "distribution") / indirect_effects(); this is the
distribution-mediated effect.mu-path coefficient by a sigma-path coefficient — there
is no coefficient-product mediation across components or for non-Gaussian /
cross-link paths. It is banned (see AGENTS.md, simulation-effects).plot.drm_sem() colours edges by component for this
reason; keep that distinction in text too.Say so and say what to try: a different family, mean-only mediation, or noting
that distribution-mediated effects fall back to mean propagation when no
realized-value sampler exists (check_sem() reports the sampler flag).
Keep drmSEM CRAN-clean — R CMD check 0 errors/0 warnings, examples that need drmTMB wrapped in \dontrun, Remotes for the GitHub engine, no compiled code in drmSEM itself, DESCRIPTION hygiene, ASCII, reasonable file sizes, reproducible RNG. Use before release or when touching DESCRIPTION/examples/CI.
Reader-facing docs for drmSEM — README, the vignette, and pkgdown reference sync. Pair equations + R syntax + interpretation, show component-labelled output and the distribution-mediated decomposition, and tell users what to try when a path/family is unsupported. Use when writing user-facing prose.
How drmSEM talks to drmTMB — ONLY through R/extractors.R. The drmTMB return-shape facts, the dpar:term coefficient convention, and the rule to confirm parameterizations against a live fit before trusting samplers.
Build basis sets and run any-component d-separation / Fisher's C correctly — adjacency = direct parent on ANY component, causal ordering, conditioning on parents, LRT df bookkeeping, honest framing as drmSEM's definition.
R package craft for drmSEM — roxygen2 on every export, S3 conventions, the document/test/check dev loop, NAMESPACE sync, cli messaging, tests shipped with implementation, small focused commits.
The simulation-based effect engine — do()-style topological propagation, mean vs distribution mediation, direct/indirect/total decomposition, Monte-Carlo coefficient uncertainty. Use whenever computing or describing an effect estimand. Coefficient-product mediation is banned.