Inference Method Glossary

This page is the reference for the raw enum values that appear in inference_options() and in the inference tables returned by summary(), inference_table(), contrast(), test_effect(), anova(), compare(), and confint().

Printed output may use reader-facing display columns such as display_status, display_reason, and what_to_do_next. The values below are the stable machine-readable contract for tests, reports, and downstream tooling.

expected_status

expected_status is used by inference_options() to say whether a route is usable on the current fit.

Value	Definition
available	The method can run on this fit and produce the requested quantity.
not_assessed	The method is known but refused for this fit, with the reason named in expected_reliability_reason or reason_code.
not_yet_wired	The method is part of the inference contract but is not exposed through the current R bridge.

expected_reliability_reason and reliability_reason

expected_reliability_reason is the reason inference_options() attaches to a route on its map. reliability_reason is the reason the engine attaches to an actual inference row when a verb runs. The vocabularies overlap but are not identical, and where the engine has recorded an observed reason for a route, inference_options() reports the observed value in preference to its own prediction — so on a fitted LMM the Wald route usually shows the engine’s asymptotic_wald_z_fallback rather than the two route-map entries marked (route map) below.

Value	Definition	Example
not_available	No additional reliability reason is attached to the row; inspect status and reason_code for the decision.	A refused cluster-bootstrap p-value row uses a specific reason_code while reliability itself is not available.
asymptotic_wald_z_fallback	A t reference distribution was the target, degrees of freedom could not be computed, and a standard normal was substituted; the row is graded low.	inference_table() and summary() on a boundary fit report their Wald rows with this reason. This is the reason code most users meet first.
asymptotic_wald_z_at_boundary (route map)	Wald z can run, but boundary or reduced-rank fits can make asymptotic uncertainty too optimistic.	Defined for the inference_options() route map; in practice the engine’s observed asymptotic_wald_z_fallback takes precedence.
interior_converged (route map)	The fit is interior and the fast Wald route has no boundary caveat from the route map. An interior optimum says nothing about model specification, so the value deliberately does not claim it (the pre-0.2.0 spelling interior_converged_well_specified overclaimed and is retired).	Defined for the route map; superseded by the observed reason in current output.
satterthwaite_unavailable_at_boundary	Satterthwaite degrees of freedom are not certified when variance parameters are on the boundary.	test_effect(fit, term, method = "satterthwaite") refuses on a singular fit.
kenward_roger_unavailable_at_boundary	Kenward-Roger degrees of freedom and covariance adjustment are not certified when variance parameters are on the boundary.	test_effect(fit, term, method = "kenward_roger") refuses on a singular fit.
satterthwaite_finite_difference_approximation	The route uses a finite-difference Satterthwaite degrees-of-freedom approximation.	The standard reason on Satterthwaite rows for interior fits.
kenward_roger_approximation	The route uses Kenward-Roger finite-sample degrees of freedom with covariance adjustment.	The standard reason on Kenward-Roger rows for interior fits.
parametric_bootstrap_monte_carlo	The p-value comes from parametric-bootstrap replicates; reliability is governed by replicate count and Monte-Carlo standard error.	contrast(fit, L, method = "bootstrap") rows carry this reason.
glmm_joint_laplace_active_hessian_wald	GLMM Wald inference from the joint-Laplace fit’s certified Hessian.	The summary()$inference rows of a method = "joint_laplace" GLMM fit (the printed coefficient block shows the statistics, not the reason).
bootstrap_monte_carlo_replicates (route map)	Reliability is governed by replicate count and Monte-Carlo standard error.	The inference_options() prediction for bootstrap routes; the observed row value is parametric_bootstrap_monte_carlo.
bootstrap_insufficient_replicates	The bootstrap route ran, but the requested replicate count is too small for a moderate reliability grade.	A demonstration call with nsim = 10 uses this reason.
bootstrap_not_certified	A bootstrap payload was returned without a certified p-value.	A bootstrap LRT row with no engine-certified p-value uses this reason.
inference_unavailable_by_policy	The method is refused by an explicit policy for this fit and requested target.	A bootstrap LRT on a REML fit returns this reliability reason with a more specific reason_code.
bootstrap_lrt_requires_ml	A bootstrap likelihood-ratio test requires ML-fitted models, not REML-fitted models.	test_effect(fit, term, method = "bootstrap_lrt") refuses on a REML fit.
bootstrap_cluster_resample_p_value_unavailable	Cluster resampling describes estimator distributions and does not certify fixed-effect p-values in the current schema.	test_effect(fit, term, method = "cluster_bootstrap", group = "subj") refuses with this reason.
profile_likelihood_ci	Profile-likelihood confidence intervals are available for this requested parameter set.	inference_options() marks profile_ci available on an ML fit away from the boundary.
profile_beta_unavailable_under_reml	Fixed-effect beta profile intervals require an ML fit and are not certified from REML profile payloads.	confint(fit, method = "profile") on a REML fit records beta rows with this reason.
profile_ci_unavailable_at_boundary	Profile intervals are not certified for boundary or reduced-rank fits.	inference_options() refuses profile_ci on a singular fit.

The method column has one value worth listing here as well: not_computed, which appears on GLMM rows fitted by the default pirls_profiled estimator, whose standard errors and p-values are withheld (see vignette("glmm")).

reliability — the grade rubric

reliability is a closed four-value scale: "high", "moderate", "low", or "not_available". This section is the rubric (rubric version 1, mixeff 0.2.0): the exact mapping the code implements from observable facts to a grade. The grade is a summary of which method produced the number and what is known about that method’s approximation — nothing more.

Grade	Earned by	Where you see it
high	A model-based covariance route with no approximation beyond the model itself (the LMM model_based fixed-effect covariance).	LMM Wald/Satterthwaite covariance status.
moderate	A named, documented approximation within the validated envelope: finite-difference Satterthwaite df; the joint-Laplace active-Hessian GLMM covariance; the working-Hessian approximation under its explicit opt-in; a certified bootstrap with at least 999 successful replicates and a finite Monte Carlo SE.	Most certified inference rows.
low	The route ran but with a known weakening: a t target that fell back to standard-normal z (degenerate df), or a bootstrap below the 999-replicate floor.	Boundary-fit Wald rows; small-nsim bootstrap results.
not_available	No graded quantity: the row is a refusal, or reliability is not defined for the target.	Withheld GLMM rows; refusal rows.

The 999-replicate bootstrap floor is a grading convention, documented rather than asserted: at B = 999 bootstrap p-values resolve on an exact 1/(B+1) = 0.001 grid, and the Monte Carlo SE of a rate near 0.05 is about 0.0069. B = 998 does not differ materially from B = 999 — which is why the replicate accounting (successful count, Monte Carlo SE) ships alongside the grade, so you can judge the evidence directly instead of trusting the label.

What a grade does not mean: no grade asserts that the model is correctly specified, that the estimate is unbiased, or that a p-value is exactly calibrated. Those are properties of the model and the data; mixeff’s measured calibration evidence for its routes lives in the benchmarking article’s simulation study, with Wilson intervals and replicate accounting.

What “certified” means

Several surfaces use the word certified. It has one definition:

The engine’s inference contract marks this estimator/target combination available: the required numerical invariants passed, the value was produced by the named method, and that method is inside the package’s validated support envelope.

“Certified” does not mean statistically guaranteed correct, and it does not mean the model is well specified. An absent certificate is never treated as a certificate: surfaces that cannot see a certified inference table withhold rather than assume. The uncertified alternatives are always explicitly labelled (for example wald_z_working_hessian rows under the inference = "working_hessian" opt-in).

reason_code

reason_code is attached to actual unavailable or not-assessed rows and to typed errors from inference verbs. It may match expected_reliability_reason, but the two fields answer different questions: expected_reliability_reason predicts or grades a route, while reason_code names the concrete refusal that actually occurred.

Value	Definition	Example
satterthwaite_unavailable_at_boundary	Satterthwaite inference is refused because the fit is boundary or reduced-rank.	contrast(fit, L, method = "satterthwaite") on a singular fit returns a not-assessed row.
kenward_roger_unavailable_at_boundary	Kenward-Roger inference is refused because the fit is boundary or reduced-rank.	test_effect(fit, term, method = "kenward_roger") on a singular fit returns a not-assessed row.
bootstrap_lrt_requires_ml	Bootstrap likelihood-ratio testing is refused because at least one fit was REML-fitted.	compare(reduced, full, method = "bootstrap") refuses unless the models are ML fits.
bootstrap_lrt_reduced_formula_failed	The reduced formula for a term bootstrap LRT could not be constructed.	A term name that cannot be dropped from the model formula produces this code.
bootstrap_lrt_engine_refused	R validation passed, but the Rust bootstrap-LRT bridge refused the payload.	A malformed or unsupported engine-side bootstrap LRT request is returned with this code.
bootstrap_lrt_requires_nested_models	Compared models are not ordered as a nested reduced and larger alternative pair.	A bootstrap LRT comparison with no degrees-of-freedom increase in the alternative fit refuses.
bootstrap_lrt_requires_nested_model_frames	The reduced model uses variables that are absent from the alternative model frame.	A comparison where the reduced formula references a variable not carried by the full fit refuses.
bootstrap_lrt_requires_same_observations	The compared fits do not share identical model-frame values for shared columns.	A bootstrap LRT comparison after different row filtering refuses.
bootstrap_lrt_requires_same_weights	The compared fits use different case weights.	A weighted reduced fit compared to an unweighted full fit refuses.
bootstrap_cluster_resample_p_value_unavailable	Cluster bootstrap is available as an estimator-distribution target but not as a fixed-effect p-value certificate.	test_effect(..., method = "cluster_bootstrap", group = "subj") returns a not-assessed p-value row.
cluster_bootstrap_multifactor_ambiguous	A cluster-bootstrap term test needs an explicit grouping factor when the model has multiple random-effect groups.	test_effect(..., method = "cluster_bootstrap") on a crossed-random-effect model refuses until group is supplied.
profile_beta_unavailable_under_reml	Fixed-effect beta profile intervals are absent by contract under REML.	confint(fit, method = "profile") on a REML fit adds beta rows with NA limits and this code.
fixed_effect_inference_table_unavailable_legacy_object	A legacy or revived object lacks the Rust fixed-effect inference artifact.	inference_table() on an older saved fit returns not-assessed rows with this code.