Focused Information Criterion

Select a candidate statistical model by the estimated risk of its estimator for a declared focus parameter, allowing the preferred model to change when the inferential target changes.

Version

v3 · 2026-09-06 · History

Domain-specific #

1861

Origin domain

statistics

Subdomain

statistical model selection

Aliases

Focussed Information Criterion, FIC

Core Idea

The Focused Information Criterion (FIC) is a statistical model-selection method that asks which candidate model estimates a declared focus parameter most accurately. The focus, written mu, is an estimand meaningful under every candidate model: a regression mean at a specified covariate value, a quantile, a treatment contrast, a survival probability at a chosen time and patient profile, or another scientifically chosen functional. Candidate model S supplies an estimator mu_hat_S. FIC estimates the risk of that estimator for this particular focus and selects a candidate with minimum estimated focus risk. Change the focus and the selected model may legitimately change.^[1]

That target dependence distinguishes FIC from a generic global-fit score. AIC, BIC, and related criteria attach one score to a fitted model without receiving the downstream estimand as an input. In its original likelihood formulation, FIC instead compares the precision of submodel-based estimators of mu. The model that best approximates the overall data-generating distribution need not minimize mean squared error for one functional of it. FIC therefore refuses the assumption that one selected model must be best for every intended inference.^[1]

The canonical theory makes the bias–variance comparison estimable through local misspecification. Write a wide model with always-included parameter theta and optional parameter vector gamma; the narrow model fixes gamma=gamma_0. The data-generating sequence is represented as.

f_true(y) = f(y; theta_0, gamma_0 + delta/sqrt(n)).

Candidate submodel S frees only selected components of gamma. For a scalar focus, the large-sample error has a model-dependent limit.

sqrt(n)(mu_hat_S - mu_true) => Lambda_S,

whose squared-error risk can be read as r(S)=E[Lambda_S^2]=b_S(delta)^2+v_S. Smaller models commonly reduce estimation variance but incur focus-relevant bias by fixing omitted directions; larger models reduce that bias but estimate more parameters. Because delta is not consistently estimable in this local experiment, the original FIC uses a bias-corrected estimate of the limiting risk and chooses its minimizer.^[1]

This setup is not a claim that nature literally changes with sample size. It is an asymptotic calibration regime in which omission bias and estimation variance remain comparable, rather than one automatically overwhelming the other. Extensions change the model family, loss, focus family, or robustness assumptions, but retain the operational identity: declare the inferential focus; derive candidate-specific focus risk under explicit assumptions; estimate those risks; select by the declared minimum-risk rule. A method that merely gives extra narrative attention to one coefficient, or minimizes an overall lack-of-fit measure and inspects the focus afterward, is not FIC.

Structural Signature

Sig role-phrases:

• the candidate model family — a declared set of models that give coherent estimators of the same scientific quantity
• the common focus parameter mu — the estimand chosen before scoring and defined meaningfully across all candidates
• the candidate focus estimators mu_hat_S — one estimator of that same focus obtained under each candidate model
• the misspecification regime — the assumptions under which omitted directions, model error, and sampling variation are compared, canonically gamma_0 + delta/sqrt(n)
• the focus-specific loss or precision target — canonically limiting mean squared error, or another explicitly derived risk
• the bias and variability contributions — candidate-dependent components whose weights and directions depend on the focus derivatives
• the estimable risk surrogate — the FIC value or extension-specific estimate that is justified for the stated framework
• the minimum estimated-risk selector — the rule returning an argmin candidate, including ties and near-ties
• the post-selection inferential boundary — the distinction between selecting an estimator and establishing uncertainty guarantees after data-dependent selection

The locked recognition test is:

declared common focus -> candidate focus estimators -> explicit misspecification/loss regime -> estimated focus-specific risks -> minimum-risk model choice.

Every arrow is load-bearing. If the focus is not common across models, the scores compare different targets. If risk is not derived for that focus, the procedure is a generic model score. If the misspecification regime is unstated, the bias estimate has no warranted interpretation. If the final model is chosen by fit, convention, or significance after calculating FIC, the selector has been displaced.

Diagnostics follow those roles. Recompute the criterion for a materially different focus; identical values can be legitimate, but focus-invariant code when focus derivatives should differ indicates an implementation error. Compare estimated bias and variance components; a zero bias term obtained merely by assuming the widest model true should be labeled rather than treated as data-proved. Inspect near-ties instead of converting small score differences into categorical certainty. Stress-test the candidate family: FIC ranks the supplied models and does not certify that an adequate model was supplied. Finally, use post-selection or model-averaging methods when inferential coverage after selection matters; the minimum score alone is not a confidence guarantee.^[2]

What It Is Not

FIC is not a universal goodness-of-fit test. It does not ask whether a model is absolutely adequate, and a low FIC value has no context-free scale. It is a relative comparison among declared candidates for one declared target.

It is not AIC with a coefficient highlighted in the report. AIC estimates a global predictive or Kullback–Leibler objective under its assumptions; FIC changes the objective itself to the risk of a target estimator. The two can coincide in special cases—including the original paper's one-departure setting and certain averaged log-density focuses—without becoming identical definitions.^[1]

It is not “choose the smallest standard error.” Estimated precision must include focus-relevant modeling bias as well as sampling variance. Ignoring bias automatically favors restrictive models; ignoring variance automatically favors the widest candidate.

It is not a focus parameter, a loss function, or a candidate model. Those are inputs. Nor is it model averaging: FIC selects a candidate by estimated focus risk, whereas focused model averaging combines candidate estimators using weights. The two methods share theory and may be used together, but selection and averaging are different terminal operations.^[2]

It is not a proof that the locally misspecified wide model is true. The original likelihood development assumes a declared wide/narrow framework and studies local departures around the narrow model; model-robust, semiparametric, and other FIC variants change that validity package. Results cannot be transferred between them by acronym alone.

Scope of Application

FIC applies where several plausible statistical models estimate the same scientifically meaningful quantity and the analyst can derive or approximate each estimator's risk under a defensible comparison regime. Its original scope is large-sample likelihood inference over nested submodels. Subsequent formulations cover regression, generalized linear models, survival analysis, semiparametric and nonparametric settings, quantile targets, time series, high-dimensional procedures, and focused model averaging. The domain can change; the common-focus and risk-estimation obligations do not.^[2]

Good applications begin with the question rather than the model. “What is the two-year survival probability for this profile?” supplies a focus. “Which covariates make the model best?” does not yet say best for what. After defining the focus, the analyst declares protected and optional terms, candidate subsets, loss, reference or wide model, and the asymptotic or finite-sample approximation used for risk. The candidate set must be scientifically coherent: an estimator with a different target or support cannot be made comparable by putting its score in the same table.

FIC is less suitable when no stable estimand is shared across candidates, when the scientific aim is faithful whole-distribution approximation rather than target estimation, when the candidate set is generated adaptively without accounting for that search, or when sample size is too small for the chosen asymptotics and no validated finite-sample alternative exists. It also does not replace residual checks, sensitivity analysis, external validation, or substantive model criticism.

Clarity

FIC forces the phrase “best model” to be completed. The analyst must name the target, the loss, the candidate class, and the regime under which the score estimates risk. This turns an ambiguous global superlative into an auditable conditional claim: “among these candidates, under this local-misspecification approximation and squared-error loss, model S has the smallest estimated risk for mu.”

The focus also localizes disagreements. Two analysts can select different models without contradiction if they estimate different quantities or use different loss. A survival-probability focus for one patient profile and a median-survival focus for a population are not interchangeable. Conversely, analysts claiming the same focus and validity package should be able to reproduce the same derivatives, information matrices, risk components, and selection rule.

A clear report includes the full FIC table, not only the winner; labels the loss orientation; shows bias and variance contributions where available; states tie handling; identifies protected and optional parameters; and distinguishes the selected estimator's estimated risk from its post-selection sampling uncertainty.

Manages Complexity

Model-selection complexity grows combinatorially when q optional parameters produce up to 2^q submodels. FIC gives these candidates a common focus-specific currency. Once the wide-model information, focus derivatives, and subset projections are computed, many candidate risks can be compared without separately inventing a scientific objective for every model.

The method also prevents unnecessary commitment to a single explanatory model for all downstream tasks. One fitted-data context can maintain a focus-indexed selection map: model S_1 for estimating a central mean, S_2 for a tail probability, and S_3 for a profile-specific survival quantity. That is more complex than publishing one winner, but it is controlled complexity because each difference is traced to an explicit target and risk decomposition.

Complexity returns when many focus parameters, losses, or candidate families are explored. Then average or weighted FIC, focused model averaging, prespecified focus families, and sensitivity surfaces can summarize the landscape. These are extensions, not permission to search over focuses until a preferred model wins. Preserve the mapping from each scientific question to its score calculation.

Abstract Reasoning

Focus invariance check. Verify that mu(theta,gamma) denotes one quantity under every candidate. Parameter names alone do not establish this; transformations, links, and conditional-versus-marginal interpretations can change the estimand.

Local experiment. In the original theory, place the true parameter at gamma_0 + delta/sqrt(n). For each subset S, derive the limiting distribution of sqrt(n)(mu_hat_S-mu_true). The focus gradient converts omitted parameter directions into bias; the information matrix converts estimated directions into variance.

Risk decomposition. Under squared-error loss, write r(S)=b_S(delta)^2+v_S. A candidate is favored only through this sum. Bias and variance are not separately optimized objectives; they are components of focus risk under the declared scaling.

Original likelihood FIC. In Claeskens and Hjort's notation, the real-data criterion for subset S can be written

FIC_S = (psi_hat_full - psi_hat_S)^2 + 2 omega_hat_S^T K_hat_S omega_hat_S.

Here omega records the focus-relevant derivative direction after accounting for always-included parameters, K_S is the subset information quantity, and the two psi terms are the corresponding full- and subset-model estimates of the local focus departure. This expression estimates limiting squared-error risk up to an additive constant independent of S; therefore its argmin, not its absolute magnitude, is operative.^[1]

Selection uncertainty. Treat FIC_S as an estimated risk. The local parameter delta remains noisy in the limit experiment, so selection does not converge to an omniscient oracle for every local departure. Report close competitors and assess the stability of the minimizer.

Focus perturbation. Change the covariate profile, time horizon, quantile, contrast, or loss while holding data and candidates fixed. The resulting model map reveals which components matter to which target. An unexpected invariant map can be a substantive result or a sign that focus dependence was omitted from the code.

Assumption ladder. Separate claims proved under the original nested likelihood setup from those needing Cox, semiparametric, nonparametric, robust, or high-dimensional FIC theory. A shared minimum-risk pattern does not make formulas portable without re-derivation.

Knowledge Transfer

Within statistics, FIC transfers by preserving roles rather than formulas. A generalized linear model changes the likelihood and focus derivative; a Cox model adds a baseline hazard and censoring structure; a semiparametric model changes nuisance estimation; a quantile target changes the estimator and risk. In every case, the analyst can ask: What is the common focus? What candidate estimators exist? What is their focus risk? Which terms estimate bias and variance? Which model minimizes the justified estimate?

The method transfers especially well between scientific questions posed on the same data. In medicine, models for average treatment effect, subgroup risk, and a tail probability can differ. In ecology, a model chosen for a population mean may not be best for a rare-event probability. In forecasting, different horizons or regions of covariate space can define different focuses. The transfer is literal only when each application supplies statistical estimators and a valid risk calculation.

Outside statistics, “choose the representation that best serves the downstream question” is a useful analogy, but it is not automatically FIC. Generic target-dependent evaluation is covered by Selection and Evaluation. Literal FIC requires the sampling-theoretic focus estimator, risk approximation, and misspecification assumptions that make the score an information criterion rather than a metaphor.

Examples

Canonical: One local nuisance direction

For i=1,...,n, suppose independent observations satisfy

Y_i ~ N(mu + gamma, 1) and Z_i ~ N(gamma, 1),

with the two streams independent. The focus is mu. The narrow model fixes gamma=0, giving mu_hat_N=Y_bar. The wide model estimates gamma, giving gamma_hat=Z_bar and mu_hat_W=Y_bar-Z_bar. Under the local sequence gamma=delta/sqrt(n),

sqrt(n)(mu_hat_N-mu) ~ N(delta,1) and sqrt(n)(mu_hat_W-mu) ~ N(0,2).

Their true scaled mean squared errors are therefore r_N=delta^2+1 and r_W=2: the narrow estimator saves variance but can be biased; the wide estimator removes that bias and pays an extra variance unit. Let D=sqrt(n) Z_bar, so D ~ N(delta,1). In the original one-direction FIC, the criterion values reduce to

FIC_N=D^2 and FIC_W=2.

Thus FIC chooses the narrow model when |D|<sqrt(2) and the wide model otherwise, exactly the one-dimensional rule derived by Claeskens and Hjort. The example also exposes the method's limit: D remains noisy under local misspecification, so estimated-risk selection need not equal the oracle comparison based on unknown delta in every sample.^[1]

Mapped back: The two nested candidates share focus mu; their estimators are Y_bar and Y_bar-Z_bar; gamma=delta/sqrt(n) supplies the misspecification regime; squared-error risks split into local bias and variance; D supplies the risk-estimation statistic; and the minimum of D^2 and 2 determines the selected model while leaving selection uncertainty visible.

Applied / In Practice: Focused covariate selection in a Cox model

For censored survival data, Hjort and Claeskens consider a Cox hazard

h_i(u)=h_0(u) exp(x_i^T beta + z_i^T gamma).

The x covariates are protected in every model; subsets of the z covariates may be included. A focus mu(beta,gamma,H_0) might be a patient's survival probability at a specified time or a median survival time. Each subset produces its own estimate of that same focus. The Cox-specific FIC estimates each estimator's mean squared error and chooses the subset with the smallest value. Because the focus derivatives differ, a subset preferred for median survival can differ from one preferred for a survival probability, and a patient profile can change which covariates matter most to the target.^[3]

The 2006 paper develops this method rather than merely importing the parametric formula, and illustrates it with simulations and a Danish skin-cancer survival study. The application keeps model criticism and censoring assumptions separate from selection: FIC ranks declared subsets for a declared quantity; it does not certify the Cox proportional-hazards form or turn the chosen subset into a causal model.

Mapped back: Protected and optional covariates define the candidate family; the profile- and time-indexed survival functional is the common focus; partial-likelihood and baseline-hazard machinery produce candidate estimators; Cox-specific asymptotics supply focus MSE estimates; their minimum selects a subset; and changing survival quantity or profile can change the selected model without contradiction.

Structural Tensions

Target relevance ↔ global adequacy. Optimizing one focus can select a model that is poor for other summaries of the data-generating process. Diagnostic: evaluate residuals and global adequacy separately, and recompute risk for every downstream estimand that will actually be used.

Omission bias ↔ estimation variance. Restrictive models can stabilize estimates while distorting the focus; wide models reduce omission bias while estimating more nuisance directions. Diagnostic: display the focus-specific bias and variance contributions instead of reading model size as a proxy for either.

Local comparability ↔ large misspecification. The delta/sqrt(n) regime keeps bias and variance on one scale, but gross misspecification may lie outside that approximation. Diagnostic: perturb the wide model, inspect residuals, and compare with robust or simulation-based risk estimates.

Question specificity ↔ researcher flexibility. A declared focus makes selection scientifically relevant; choosing the focus after seeing which model wins creates a new multiplicity channel. Diagnostic: timestamp or preregister the focus-to-analysis mapping and report all explored focuses.

Minimum score ↔ selection uncertainty. One candidate must attain the smallest estimated score, but close values may reflect estimation noise rather than a stable performance gap. Diagnostic: report the full score profile, bootstrap or simulate stability where justified, and consider focused model averaging.

Candidate closure ↔ model discovery. FIC compares supplied models efficiently, but a missing structural feature cannot win. Diagnostic: run specification checks and broaden the candidate family on scientific grounds before treating the ranking as exhaustive.

Method autonomy ↔ reduction to generic selection. FIC is a selection rule, but reducing it to “score options and take the minimum” erases the common estimand, sampling distribution, local bias, focus derivative, and risk-validity package. Diagnostic: replace the statistical models with arbitrary options; if the entire account survives without reconstructing focus-estimator risk, it describes Selection, not FIC.

Structural–Framed Character

Focused Information Criterion is mixed-structural with an aggregate framedness score of 0.25. Its mathematical selection operator is highly formal, while literal recognition remains bounded to statistical model selection.

1. Vocabulary travels (0.5). “Focus,” “risk,” “candidate,” and “selection” travel widely, but focus parameter, limiting MSE, information matrix, and local misspecification have fixed statistical meanings.
2. Evaluative weight (0.0). “Best” means minimum declared statistical risk, not moral, social, or aesthetic approval.
3. Institutional origin (0.0). Publication and software conventions affect implementations but do not constitute the criterion.
4. Human-practice bound (0.0). Once focus, candidates, and validity regime are encoded, the calculation and selection can be automated.
5. Import versus recognize (0.75). A system must be represented as competing statistical estimators under sampling risk before FIC is literally recognizable; generic target-sensitive choice is only analogy.

The result is a formally structural operation inside a sharply delimited inferential vocabulary, not a socially framed category and not a substrate-independent prime.

Structural Core vs. Domain Accent

Structural core. Several representations support one declared downstream quantity. Each induces a different error profile. Estimate the target-relevant cost of each representation under explicit uncertainty, then choose the minimum rather than optimizing representation-wide fit.

Domain accent. The representations are statistical models; the target is a common estimand; uncertainty comes from sampling and misspecification; error is focus-specific risk; bias and variance follow from focus derivatives and information; and the canonical comparison uses a 1/sqrt(n) local departure from a narrow model.

Three-part test. Remove names such as Cox, likelihood, and regression and the focus-risk selector remains. Change the statistical substrate and the same roles can be re-derived. Remove the common estimand, sampling-risk calculation, or validity regime and only generic Selection, Evaluation, Optimization, and Trade-offs remain. The retained domain accent is therefore indispensable, making FIC a reusable domain-specific abstraction rather than a prime or a mere paper title.

Instantiates / Related Primes

• prime:selection — proposed strict subsumption parent. FIC takes a candidate population and a focus-risk criterion, then gives one or more models greater retention. It adds common-estimand, risk-estimation, and misspecification commitments absent from generic Selection.
• prime:statistical_inference — proposed strict presupposition parent. Focus estimators, limiting distributions, information matrices, and model-selection risk are meaningful only inside sample-to-process reasoning with quantified uncertainty. FIC is not all Statistical Inference, and inference can proceed without selecting a model.
• prime:trade_offs — related mechanism. The canonical squared-error criterion balances omission bias against estimation variance. The live Trade-offs prime also requires a general feasible-frontier and substitution structure not needed to recognize every FIC extension, so it remains a close explanatory neighbor rather than a proposed direct parent.
• prime:evaluation — related operation. FIC evaluates models under a declared criterion, but model scoring is intermediate; candidate retention is the terminal identity.
• domain_specific:underfitting — related failure state. A narrow model may have high focus bias, but FIC is the two-sided selector, not the restrictive-model failure it can diagnose or avoid.
• prime:nonparametric_methods — related scope. FIC has nonparametric extensions, but nonparametricity is one model-family choice rather than a genus of FIC.
• domain_specific:regression — related application family. Many FIC constructions compare regression submodels; FIC also applies beyond regression and does not define the response–predictor model itself.

The proposed minimal direct DAG set is prime:selection by strict subsumption and prime:statistical_inference by strict presupposition. The two roles are independent: one supplies the terminal survivor-producing operation, the other the probabilistic inferential substrate.

Relationships to Other Abstractions

Current abstraction Focused Information Criterion Domain-specific

Parents (2) — more general patterns this builds on

• Focused Information Criterion is a kind of Selection Prime

  prime:selection — proposed strict subsumption parent. FIC takes a candidate population and a focus-risk criterion, then gives one or more models greater retention.

• Focused Information Criterion presupposes Statistical Inference Prime

  prime:statistical_inference — proposed strict presupposition parent. Focus estimators, limiting distributions, information matrices, and model-selection risk are meaningful only inside sample-to-process reasoning with quantified.

Hierarchy paths (5) — routes to 5 parentless roots

• Focused Information Criterion → Selection

Show alternative paths (4)

Neighborhood in Abstraction Space

Focused Information Criterion sits in a sparse region of the domain-specific corpus (75^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

• Variational Bayesian Methods — 0.84
• Learnable Function Class — 0.83
• Regression — 0.83
• Optimality criterion — 0.83
• Statistical Contrast — 0.83

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With

• Akaike information criterion (AIC). AIC estimates a model-wide predictive/Kullback–Leibler objective; FIC estimates risk for a declared focus estimator. Tell: Does changing mu enter the score calculation?
• Bayesian information criterion (BIC). BIC applies a sample-size-dependent dimension penalty for a different asymptotic selection objective. Tell: Is the criterion derived from target-estimator risk or from likelihood plus a generic parameter penalty?
• Goodness of fit. Fit diagnostics ask whether modeled distributions reproduce relevant data features. Tell: Can every candidate fit inadequately even though one has minimum FIC? Yes.
• Minimum standard error selection. Variance alone ignores focus bias from restrictions. Tell: Where is the omitted-direction bias term estimated?
• Cross-validation. Cross-validation estimates a prespecified out-of-sample loss by resampling. It becomes focus-like only when the held-out loss is explicitly tied to the target; it is not FIC merely because it selects a model. Tell: Is the comparison based on a derived focus-risk approximation or a resampled predictive loss?
• Focused model averaging. Averaging combines estimators; FIC selection returns a model or tied set. Tell: Is the terminal output one candidate or a weighted estimator?
• prime:trade_offs. Bias–variance tension explains the method, but the prime does not supply the common estimand, local experiment, information calculation, or selection rule. Tell: Can the account calculate which statistical submodel minimizes focus risk?
• domain_specific:underfitting. Underfitting is a failure caused by excessive restriction. FIC can select a narrow model when its variance saving outweighs its focus bias. Tell: Is restriction being diagnosed as failure or evaluated conditionally for one target?
• Post-selection inference. Selecting by estimated risk does not by itself calibrate confidence intervals conditional on selection. Tell: Is uncertainty coverage after the data-dependent choice actually derived?
• Focus parameter. mu names the target; FIC names the procedure comparing its candidate estimators. Tell: Is the object being estimated, or the rule choosing how to estimate it?

References

[1] Gerda Claeskens and Nils Lid Hjort, “The Focused Information Criterion,” Journal of the American Statistical Association 98, no. 464 (2003): 900–916. Publisher record and DOI. The publisher metadata corrects a page-range swap in the frozen candidate article's bibliography. registry ↩a ↩b ↩c ↩d ↩e ↩f

[2] Gerda Claeskens and Nils Lid Hjort, Model Selection and Model Averaging, Cambridge Series in Statistical and Probabilistic Mathematics 27, Cambridge University Press, 2008, especially Chapter 6, “The Focussed Information Criterion.” Publisher chapter record and DOI. registry ↩a ↩b ↩c

[3] Nils Lid Hjort and Gerda Claeskens, “Focused Information Criteria and Model Averaging for the Cox Hazard Regression Model,” Journal of the American Statistical Association 101, no. 476 (2006): 1449–1464. Publisher record and DOI. registry ↩