INLA group node conflict diagnostic
Algorithms AIMech:INLAGroupNodeConflictDiagnostic DRAFT
A group-wise Bayesian model diagnostic for latent Gaussian models that uses two INLA fits per selected group to compare remaining-data and within-group information about its latent linear predictor vector. Fitted mean and covariance summaries yield a quadratic conflict score with a rank-based Gaussian chi-square reference. Nuisance-parameter updating, replication dependence and numerical rank choices determine the interpretation.
Identity Notes & Concept Scope
No normalized identity match for the INLA group-conflict candidate among all 442 pinned AIO and 560 maintained IDs, names, aliases and scoped synonyms. Fixed-string searches include ignored and hidden files in the recovery, older discovery and original Documents checkouts, returning 185, 0 and 0 lines. Hit paths were classified and canonical diagnostic records and the relevant pending INLA/source notes inspected; generated mirrors and long metadata/search-index lines are not separate method definitions and were not all read line by line. The semantic assessment covers 61 AIO rows and 214 native names/definitions: 61 and 213 prior rows were verified unchanged, and the one new multiple-testing definition was read. Ten full AIO nodes with incident edges and generated records and eight native context records were read or verified against prior full readings. This distinct procedure fixes the split to group-level latent predictors and computes its summaries with INLA. It is not generic INLA fitting, arbitrary separator diagnostics, Gaussian-message screening, simultaneous maxT inference or a software-package instance.
Claims & Grounded Evidence
7 Claims
Scope
#grouped-latent-predictors
Partition the observed data indices into declared groups. For each group j, split the corresponding vector of latent linear predictors eta_j into a copy informed by the remaining groups and a copy informed by the selected group. This specialization uses latent Gaussian models and INLA approximations. It assesses one group versus the rest on the latent predictor scale, rather than arbitrary separators, all possible partitions, or pointwise predictive residuals.
Mechanism
#two-fits-and-summaries
Perform a remaining-data INLA fit with the selected group responses omitted, retaining predictions at those indices, then a within-group fit. The fitted distributions must be proper and have the finite moments required for the summaries. Obtain the predictor-vector means and covariance matrices through linear combinations. Use remaining minus within for the mean difference. The inspected helper makes two fits per group, supplies the remaining fit through control.update to the within fit, and adds the two marginal covariance matrices. The within distribution therefore also depends on information supplied by the remaining fit; its notation as conditioning only on group data is abbreviated.
Mechanism
#replication-and-covariance
The sum S=S_remaining+S_within is the covariance of a difference under independent draws from the two fitted marginal laws, conditional on the observed data and chosen fits. If the intended replication instead shares a nuisance draw, its cross-covariance C contributes S=S_remaining+S_within-C-C^T. Cutting feedback does not by itself remove that dependence. Distinguish this posterior replication choice from dependence of fitted summaries over repeated data; separate fits alone do not prove sampling calibration.
Mechanism
#quadratic-reference
For mean difference m and a positive semidefinite covariance S, the reported discrepancy is Q=m^T*S^+*m. With a justified centered Gaussian reference on the retained r-dimensional subspace, use the upper chi-square_r tail. S^+ denotes the Moore–Penrose pseudoinverse; rank deficiency is allowed. A null outside the affine support requires separate treatment because the pseudoinverse drops nullspace displacement. A numerical cutoff must define the same retained subspace for both Q and r. Approximate posterior normality is not by itself a theorem about repeated-data p-value calibration.
Limitation
#posterior-as-prior-is-not-a-general-cut
The paper equates passing the remaining-data hyperparameter posterior as the within-fit prior with a multiple-imputation cut. Static tracing of the pinned 2026 ordinary integration path instead shows prior-density substitution followed by a new local-data marginal likelihood and new integration weights. Generically, if q_minus(theta) is the supplied prior approximation and Z_j(theta) the local evidence, an ordinary update weights theta proportionally to q_minus(theta)*Z_j(theta); a frozen-weight cut retains q_minus(theta) while normalizing the conditional local distribution at each theta. These coincide only under additional conditions, such as constant local evidence. The helper does not itself configure the USER_EXPERT supplied-weight exception. This is a bounded source-code inference, not a runtime result or a claim about the exact 2017 release.
Limitation
#implementation-rank-mismatch
In the pinned helper, degrees of freedom count covariance eigenvalues above the absolute threshold 0.001, whereas inla.ginv with tol=0.001 retains singular values above 0.001 times the largest singular value. These can select different ranks: for diagonal covariance (100, 0.01), the former counts two directions and the latter retains one. Rescaling predictor units can change the absolute-rank count. This discrepancy qualifies that implementation; it is not the mathematical rank convention of the diagnostic and was not tested by running R-INLA.
Limitation
#calibration-and-implementation-scope
The source leaves a systematic study of power and calibration for future work. A small score flags incompatibility under the chosen model, grouping, priors, approximation and reference; it does not identify which data source is wrong or establish a posterior probability of equality. Benjamini–Hochberg adjustments in the examples do not establish general FDR control for dependent approximate conflict scores. The inspected experimental helper requires data.frame input and a named response and rejects a predictor A matrix. Its supported software interface is narrower than the broad latent Gaussian model family.
⚠️ Curated Limitations & Negative Results
- This is the INLA linear-predictor specialization of node-conflict assessment, not the general INLA fitting algorithm, arbitrary separator splitting, or a simultaneous maxT adjustment.
- The paper’s posterior-as-prior description is not accepted as a general frozen nuisance cut. The inspected 2026 ordinary integration path reweights hyperparameters using local data; historical releases, user-selected exceptional paths and runtime behavior remain unverified.
- Independent marginal replication and paired shared-nuisance replication have different covariances. Neither a normal approximation nor a pseudoinverse alone establishes null support, calibration or error control.
- The pinned helper’s absolute rank and relative pseudoinverse cutoffs can disagree and depend differently on scale. No correction to upstream R-INLA code or claim of a released fix is made.
- All 22 pages of arXiv v4 were read, including Figures 1–5, Table 1 and references. Its margin date is 2017-11-01 and title-page date 2017-11-02. The one-dimensional introductory difference has inconsistent sign labels; the record consistently uses remaining minus within. A prose/Table 1 discrepancy is retained in the discovery notes without importing empirical values.
- Official source inspection is limited to the named revision and cited functions. No R-INLA installation, compilation, model fitting, MCMC, benchmark, biomedical inference or repeated-data calibration was performed. Constructed algebraic examples illustrate weighting, covariance and rank boundaries only.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:BayesianNetwork
Bayesian Network
|
related_match | The diagnostic splits information about latent quantities in Bayesian hierarchical graphical models. BayesianNetwork provides representation context, not the INLA approximation or group-wise diagnostic; this is a related mapping, not an asserted subclass or equivalence. |
Evidence Sources & Primary Literature
8 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
ferkingstad2017
|
Egil Ferkingstad, Leonhard Held, Håvard Rue
|
primary paper
full_text
|
ID: https://doi.org/10.1002/sta4.163
sha256:4f6e83ce3665…
|
cut2026
|
R-INLA contributors
|
official documentation
documentation
|
ID: https://github.com/hrue/r-inla/blob/5bc4c7f223c679bb4b99171aaca490d72e96494c/rinla/R/cut.R
sha256:f7a36f50ca94…
|
sections2026
|
R-INLA contributors
|
official documentation
documentation
|
ID: https://github.com/hrue/r-inla/blob/5bc4c7f223c679bb4b99171aaca490d72e96494c/rinla/R/sections.R
sha256:4e01dd8f3c94…
|
parse2026
|
R-INLA contributors
|
official documentation
documentation
|
ID: https://github.com/hrue/r-inla/blob/5bc4c7f223c679bb4b99171aaca490d72e96494c/inlaprog/src/inla-parse.c
sha256:f07f2defde06…
|
inlac2026
|
R-INLA contributors
|
official documentation
documentation
|
ID: https://github.com/hrue/r-inla/blob/5bc4c7f223c679bb4b99171aaca490d72e96494c/inlaprog/src/inla.c
sha256:6bd722684049…
|
domin2026
|
R-INLA contributors
|
official documentation
documentation
|
ID: https://github.com/hrue/r-inla/blob/5bc4c7f223c679bb4b99171aaca490d72e96494c/gmrflib/domin-interface.c
sha256:8f9afb51e92d…
|
integration2026
|
R-INLA contributors
|
official documentation
documentation
|
ID: https://github.com/hrue/r-inla/blob/5bc4c7f223c679bb4b99171aaca490d72e96494c/gmrflib/approx-inference.c
sha256:56e71bd2e5d9…
|
utils2026
|
R-INLA contributors
|
official documentation
documentation
|
ID: https://github.com/hrue/r-inla/blob/5bc4c7f223c679bb4b99171aaca490d72e96494c/rinla/R/utils.R
sha256:aba074dd4c42…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-13T01:35:28Z
Updated: 2026-10-13T01:40:17Z
Source File:
data/curated/INLAGroupNodeConflictDiagnostic.yaml