Hannan–Quinn information criterion¶

In statistics, the Hannan–Quinn information criterion (HQC) is a criterion for model selection.

Version: v1 · 2026-09-28 · History
Domain-specific #: 9798
Domain group: Formal Sciences
Origin domain: Experimental Design & Statistics
Subdomains: Model Selection, Time Series Analysis → Experimental Design & Statistics

Core Idea¶

The Hannan–Quinn information criterion is a likelihood-based rule for selecting among fitted statistical models while penalizing unnecessary parameters. In a common form, HQC = -2 log L_max + 2k log log n, where L_max is the maximized likelihood, k the number of estimated parameters, and n the sample size; the candidate with the smallest value is preferred. Fit lowers the first term, while complexity raises the second, so a parameter is retained only when its improvement in likelihood outweighs the slowly increasing penalty.

The log log n penalty places Hannan–Quinn between familiar alternatives asymptotically. Akaike's criterion uses an approximately constant penalty per parameter and targets predictive or Kullback–Leibler efficiency, but can continue selecting overlarge models with nonvanishing probability. The Bayesian information criterion uses k log n and penalizes complexity more strongly. Under suitable regularity and finite true-model assumptions, the Hannan–Quinn rate is sufficient for strong consistency: the selected order eventually equals the true order almost surely. Its connection to the law of the iterated logarithm makes the very slowly diverging penalty near a lower boundary for such consistency in important time-series settings. Finite-sample variants differ in constants, likelihood conventions, and parameter counts.

HQC is not a hypothesis test, posterior model probability, or universal guarantee that the chosen model predicts best. Strong consistency depends on assumptions about the candidate family, dependence, sample size, and existence of a true finite-dimensional model; misspecification changes the interpretation. Its small asymptotic penalty difference may also be outweighed in practice by estimation and tuning details. The abstraction is minimally diverging complexity control: likelihood gains are discounted by a penalty that grows slowly enough to preserve fit yet strongly enough, under stated conditions, to eliminate persistent overfitting.

Structural Signature¶

Sig role-phrases:

the candidate model family — fitted statistical alternatives varying in parameter count or order
the maximized likelihood — best in-sample fit available to each candidate
the parameter count k — estimated complexity entering the penalty
the effective sample size n — data quantity driving penalty growth under the adopted convention
the HQ score — negative twice log likelihood plus twice k times log log n in the common form
the minimization rule — preference for the candidate with the smallest penalized criterion
the fit–complexity exchange — extra parameter retained only when its likelihood improvement exceeds its charge
the slowly diverging penalty — asymptotic rate stronger than AIC's constant penalty and weaker than BIC's log n penalty
the strong-consistency result — eventual almost-sure recovery of a finite true order under suitable regularity assumptions
the assumption boundary — misspecification, dependence, finite samples, parameter-count choices, and tuning preventing interpretation as a test, posterior probability, or universal predictive optimum

What It Is Not¶

Not a hypothesis test. The score ranks fitted candidates without producing a test statistic and calibrated rejection rule by itself.
Not a posterior model probability. Its likelihood penalty does not supply Bayesian probabilities absent a separate model and prior analysis.
Not a universal guarantee of best prediction. Strong consistency and predictive efficiency are different objectives.
Not assumption-free recovery of truth. The consistency result depends on a suitable candidate family, regularity, dependence structure, and finite true order.
Not necessarily meaningful under model misspecification in the same way. If no candidate is true, “eventually selects the true model” cannot apply.
Not invariant to every parameter count and likelihood convention. Constants, effective sample size, and treatment of variance or initial conditions can alter finite-sample rankings.
Not merely AIC or BIC under another name. Its slowly diverging log-log penalty occupies a distinct asymptotic position.

Scope of Application¶

The Hannan–Quinn information criterion is a statistical-selection instrument and applies when likelihood models are compared under a slowly diverging complexity penalty and the target is consistent finite-order selection under appropriate assumptions.

Autoregressive order selection. Candidate lag orders fitted to the same series receive comparable likelihood and parameter counts.
Time-series model comparison. The criterion balances improved fit against additional dynamic parameters.
Asymptotic theory. The log-log penalty is studied near the rate needed for strong consistency in relevant settings.
AIC comparison. HQC's diverging penalty is contrasted with predictive-efficiency objectives and persistent overfit risk.
BIC comparison. Its weaker-than-log-n penalty reveals a different fit–parsimony tradeoff.
Finite-sample analysis. Constants, effective observations, initial conditions, variance parameters, and tie rules affect rankings.
Sensitivity and diagnostics. Residuals, stability, and out-of-sample behavior complement the minimum score.
Applicability boundary. HQC is not a hypothesis test, posterior probability, universal prediction guarantee, or assumption-free truth selector; formula, constant, likelihood, n, k, candidate family, dependence, missingness, fitting method, finite-true-model premise, misspecification, and uncertainty must accompany any selected order.

Clarity¶

Hannan–Quinn information criterion balances maximized likelihood against a parameter penalty growing as \(2k\log\log n\) under a common convention. It is not a universal measure of model truth, and constants, effective sample size, likelihood comparability, and candidate family must be stated. Its asymptotic rationale differs from AIC's predictive target and BIC's stronger penalty. The sharper model-selection question is whether candidates are fit to the same data under valid likelihood assumptions and whether consistent order selection, rather than finite-sample prediction alone, is the goal.

Manages Complexity¶

Hannan–Quinn criterion compresses each fitted candidate model to maximized likelihood, parameter count, sample size, and one penalty formula. The analyst ranks models by a single comparable score while retaining the fit–complexity tradeoff. AIC, HQC, and BIC branches differ chiefly in penalty growth and asymptotic objective, so selection behavior can be anticipated without re-deriving each method. This compression is valid only among models fit to comparable data and likelihoods; residual diagnostics, specification, dependence, and finite-sample behavior remain outside the score. The selected minimum organizes comparison but does not certify truth.

Abstract Reasoning¶

Penalization move. Combine maximized log-likelihood with the Hannan–Quinn parameter penalty involving log log sample size to score candidate models. Selection move. Choose the lowest criterion among models fitted to the same observations under compatible likelihood conventions. Asymptotic move. Use the penalty's growth to reason about strong consistency under the theorem's regularity and candidate-model assumptions. Diagnostic move. Compare HQC with AIC, BIC, residual evidence, and stability rather than treating a small score difference as decisive. Boundary move. HQC is not a hypothesis-test p-value, cross-validation estimate, or guarantee of best finite-sample prediction or causal truth.

Knowledge Transfer¶

Within the home domain. The Hannan–Quinn information criterion transfers across time-series, regression, econometrics, and statistical model selection when maximized likelihood is balanced against its specific log-log sample-size penalty. Sample size, parameter count, likelihood convention, candidate set, and asymptotic consistency retain roles. Beyond the home domain (C — selection criterion). It applies literally to compatible fitted models in any discipline. Its boundary is inferential: scores compare only the supplied candidates, small differences may be unstable, regularity assumptions matter, and HQC is not a p-value, causal test, cross-validation estimate, or guarantee of best finite-sample prediction.

Examples¶

Canonical¶

Several autoregressive models are fit to the same series. For each, HQC=-2 log Lmax+2k log log n is computed with one sample-size and parameter-count convention; the smallest score is selected. Extra lags improve fit but pay a slowly diverging penalty stronger asymptotically than AIC's constant per-parameter charge and weaker than BIC's log n charge. Under suitable finite-true-order and regularity conditions, this rate can recover the true order almost surely eventually. The result is a selection rule, not a hypothesis-test p-value or posterior probability.

Mapped back: AR orders are the candidate model family, fit the maximized likelihood, complexity the parameter count k, data the effective sample size n, formula the HQ score, and minimum the minimization rule through the fit–complexity exchange.

Applied / In Practice¶

An analyst reports HQ, AIC, and BIC alongside residual checks and out-of-sample performance. Dependence, misspecification, small samples, and ambiguous effective n make criteria disagree, so subject-matter plausibility remains part of selection. Negative log-likelihood conventions and intercept/variance counting are aligned across software. Strong consistency is not advertised as universal predictive superiority in finite data.

Mapped back: Comparative penalties expose the slowly diverging penalty, while finite-sample and modeling limits define the assumption boundary around the strong-consistency result.

Structural Tensions¶

T1 — Identity versus admissible variation. Hannan–Quinn information criterion must remain recognizable across legitimate variants. Admissible variation is bounded by this condition: Candidate lag orders fitted to the same series receive comparable likelihood and parameter counts. The stable element is expressed by this invariant: In statistics, the Hannan–Quinn information criterion (HQC) is a criterion for model selection. Treating every surface change as a new abstraction fragments the identity, while allowing a change to the constitutive relation produces a false positive.

Diagnostic: After the proposed variation, can an analyst still establish this invariant: In statistics, the Hannan–Quinn information criterion (HQC) is a criterion for model selection?

T2 — Recognition versus proxy. The domain needs observable or inferential evidence for Hannan–Quinn information criterion, but the evidence is not automatically the identity. The working recognition rule is: the assumption boundary — misspecification, dependence, finite samples, parameter-count choices, and tuning preventing interpretation as a test, posterior probability, or universal predictive optimum. A familiar indicator can occur without the defining relation, and the relation can persist when a customary detector is unavailable.

Diagnostic: Does the evidence establish the defining claim—In statistics, the Hannan–Quinn information criterion (HQC) is a criterion for model selection—or only a correlated sign?

T3 — Definition versus operational judgment. A compact definition aids reuse, whereas actual classification in cross-domain formal modeling can require expert decisions about boundary conditions, measurements, conventions, or exceptions. The log log n penalty places Hannan–Quinn between familiar alternatives asymptotically. The definition must constrain those judgments without pretending that every admissible case can be recognized from a label alone.

Diagnostic: Which observation would make a competent practitioner reject the classification under the stated definition?

T4 — Scope versus overextension. Hannan–Quinn information criterion has a genuine habitat in which candidate lag orders fitted to the same series receive comparable likelihood and parameter counts. Yet HQC is not a hypothesis test, posterior probability, universal prediction guarantee, or assumption-free truth selector; formula, constant, likelihood, n, k, candidate family, dependence, missingness, fitting method, finite-true-model premise, misspecification, and uncertainty must accompany any selected order. A useful application map therefore has to be broad enough to cover recurring practice and narrow enough to exclude merely topical or metaphorical occurrences.

Diagnostic: Can the claimed application fill the same carrier and relation roles, or has only the name traveled?

T5 — Transfer versus domain accent. Knowledge about Hannan–Quinn information criterion can travel within its home domain, and some structural lessons may travel farther. The Hannan–Quinn information criterion transfers across time-series, regression, econometrics, and statistical model selection when maximized likelihood is balanced against its specific log-log sample-size penalty. What transfers must be separated from the specialist vocabulary, warrant, and closure conditions that remain anchored in cross-domain formal modeling.

Diagnostic: Is the receiving case a literal instance of Hannan–Quinn information criterion, a co-instance of Theory, or only an analogy?

T6 — Autonomy versus reduction. Hannan–Quinn information criterion is a strict specialization of Evaluation, but the edge does not erase the domain differentia. The broader node supplies only the necessary structural relation; cross-domain formal modeling supplies the carrier, warrant, boundary, and exception conditions expressed by this identity: In statistics, the Hannan–Quinn information criterion (HQC) is a criterion for model selection. The entry is over-split if those conditions add no discriminating work and under-specified if the parent alone is used for cases that require them.

Diagnostic: Can a domain expert use the added conditions to distinguish Hannan–Quinn information criterion from another case that equally instantiates Evaluation?

Structural–Framed Character¶

Hannan–Quinn information criterion is mixed: structurally specifiable but materially dependent on its disciplinary frame. Its structural side consists of the carrier the candidate model family — fitted statistical alternatives varying in parameter count or order and the constitutive relation In statistics, the Hannan–Quinn information criterion (HQC) is a criterion for model selection. Its framed side comes from cross-domain formal modeling, which fixes what the terms denote, what counts as evidence, and when a qualification or exception defeats the classification.

Across the principal tests, the entry is not merely a free-floating pattern. Evaluative weight: the identity can be stated descriptively even when its use has practical or normative consequences. Practice dependence: the assumption boundary — misspecification, dependence, finite samples, parameter-count choices, and tuning preventing interpretation as a test, posterior probability, or universal predictive optimum. Institutional stabilization: disciplinary conventions may stabilize the name and test without necessarily creating every underlying event or relation. Vocabulary portability: the invariant is In statistics, the Hannan–Quinn information criterion (HQC) is a criterion for model selection. Import versus recognition: an outside case qualifies literally only if the same typed roles and collapse condition are available; otherwise the comparison is analogical.

The reusable remainder is Evaluation under a reviewed subsumption relation. That node preserves the necessary cross-domain organization after the cross-domain formal modeling-specific carrier, evidence, and exceptions are removed. Hannan–Quinn information criterion remains autonomous because its recognition and collapse conditions distinguish cases that the parent alone leaves together.

Structural Core vs. Domain Accent¶

What is skeletal. The portable skeleton is a typed carrier organized by a constitutive relation, an invariant, a recognition test, and a collapse condition. Here the carrier is the candidate model family — fitted statistical alternatives varying in parameter count or order. The decisive relation is In statistics, the Hannan–Quinn information criterion (HQC) is a criterion for model selection, which also states the controlling invariant at this level. Stripped of specialist nouns, this organization is represented by Theory.

What is domain-bound. cross-domain formal modeling supplies the actual objects or agents, admissible transformations, units or conventions, standards of warrant, and named exceptions. In this case, recognition requires evidence for the assumption boundary — misspecification, dependence, finite samples, parameter-count choices, and tuning preventing interpretation as a test, posterior probability, or universal predictive optimum. Admissible variation is bounded by the condition that candidate lag orders fitted to the same series receive comparable likelihood and parameter counts, and the classification collapses when the score ranks fitted candidates without producing a test statistic and calibrated rejection rule by itself. These are constitutive differentia, not illustrative decoration.

Why it remains a domain-specific node. The reviewed DAG relation is subsumption to Evaluation. Outside cross-domain formal modeling, the parent captures only the reusable structural remainder. The specialist name remains literal only where the assumption boundary — misspecification, dependence, finite samples, parameter-count choices, and tuning preventing interpretation as a test, posterior probability, or universal predictive optimum can be established under the domain's standards of warrant.

This entry is a kind of Evaluation.

Immediate parent — Evaluation (subsumption). Hannan–Quinn information criterion is a domain-specific kind of Evaluation: In statistics, the Hannan–Quinn information criterion (HQC) is a criterion for model selection. The parent supplies the necessary broader identity—Apply a criterion-bearing frame to a bounded object, interpret its relevant features against that frame, and produce a verdict, score, rank, or action-guiding judgment.—while the candidate adds the source-domain carrier, recognition rule, and failure conditions. The defining source account begins: The Hannan–Quinn information criterion is a likelihood-based rule for selecting among fitted statistical models while penalizing unnecessary parameters.
Nearest catalog surface declined — Focused Information Criterion. Its rematch score was 0.256484. Retrieval proximity did not establish synonymy or parentage; the carrier, invariant, and collapse condition remain different.
Related reasoning operations. Evidence, comparison, boundary testing, and representation can support a case without becoming additional DAG parents.

Relationships to Other Abstractions¶

Current abstraction Hannan–Quinn information criterion Domain-specific

Parents (1) — more general patterns this builds on

Hannan–Quinn information criterion is a kind of Evaluation Prime

Hannan–Quinn information criterion is a domain-specific kind of Evaluation: In statistics, the Hannan–Quinn information criterion (HQC) is a criterion for model selection.

Hierarchy path (1) — routes to 1 parentless root

Hannan–Quinn information criterion → Evaluation → Comparison → Self Checking

Neighborhood in Abstraction Space¶

Hannan–Quinn information criterion sits in a sparse region of the domain-specific corpus (68^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2551 abstractions)

Nearest neighbors

Optimality criterion — 0.85
Bayes Factor — 0.85
Machine-Learning Learning Curve — 0.85
Focused Information Criterion — 0.84
Expectation–Maximization Algorithm — 0.83

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With¶

Evaluation. This is the reviewed immediate parent or structural prerequisite, not a synonym. Tell: retain Hannan–Quinn information criterion only when the domain-specific relation In statistics, the Hannan–Quinn information criterion (HQC) is a criterion for model selection. and its source-domain warrant are established; otherwise route the case to Evaluation.
Focused Information Criterion. This is the closest catalog retrieval surface, not an accepted synonym or parent. Tell: Ask which entry's carrier, invariant, and collapse test the case actually satisfies; shared vocabulary or a score of 0.730067 is insufficient.
Not a hypothesis test. The score ranks fitted candidates without producing a test statistic and calibrated rejection rule by itself. Tell: Require the positive recognition condition that the assumption boundary — misspecification, dependence, finite samples, parameter-count choices, and tuning preventing interpretation as a test, posterior probability, or universal predictive optimum.
Not a posterior model probability. Its likelihood penalty does not supply Bayesian probabilities absent a separate model and prior analysis. Tell: Replace the familiar surface feature and test whether in statistics, the Hannan–Quinn information criterion (HQC) is a criterion for model selection.
A detector, representation, or consequence. A method may reveal Hannan–Quinn information criterion, a notation may describe it, and an outcome may follow from it without any of those being identical to the abstraction. Tell: Would the defining relation remain if the present detector, notation, or downstream effect changed?
A metaphorical transfer. A case outside the home domain may resemble the structure while lacking its native role types and standards of warrant. Tell: If only the general organization survives, route the comparison to Theory rather than treating it as another Hannan–Quinn information criterion instance.

References¶

Frozen Wikipedia revision: https://en.wikipedia.org/wiki/Hannan%E2%80%93Quinn_information_criterion (revision 1305816318).
DOI: https://doi.org/10.1111/j.2517-6161.1979.tb01072.x
DOI: https://doi.org/10.5705/ss.202016.0011
Supporting reference preserved in the packet: https://academic.oup.com/jrsssb/article/41/2/190/7027596
Supporting reference preserved in the packet: http://www3.stat.sinica.edu.tw/statistica/J28N1/j28n111/j28n111.html
Supporting reference preserved in the packet: http://www3.stat.sinica.edu.tw/statistica/oldpdf/A3n214.pdf

The frozen Wikipedia revision is discovery provenance. The cited source set was reviewed for identity, formal or operational relation, and scope. The encyclopedia's structural synthesis is bounded to those claims; URL transport failure alone was not treated as substantive contradiction.