Home / Curated / Mathematical Methods / Model-robust nonlinearity-consistent sandwich variance and Ratio of Asymptotic Variances (RAV) specification test

Model-robust nonlinearity-consistent sandwich variance and Ratio of Asymptotic Variances (RAV) specification test

Mathematical Methods AIMech:ModelRobustNonlinearityConsistentSandwichAndRAVSpecificationTest DRAFT

Aliases: Models as Approximations I model-robust linear regression framework (Buja et al. 2019), Ratio of Asymptotic Variances (RAV_j) regressor-specific misspecification diagnostic and test, Canonical Noise and Approximation Estimation Offset (EO) decomposition, Case-wise and pairwise adjusted-regressor weighted slope representations

A model-robust theory and diagnostic methodology for ordinary least squares (OLS) linear regression under arbitrary joint regressor-response distributions P = P_{Y, X_vec} without assuming linearity or homoskedasticity (Buja, Berk, Brown, George, Pitkin, Traskin, Zhao, & Zhang 2019). Viewing the population OLS parameter beta(P) = E[X_vec X_vec']^{-1} E[X_vec Y] = E[X_vec X_vec']^{-1} E[X_vec mu(X_vec)] (Equations 3-4) as a statistical functional of the joint distribution P, Proposition 4.1 proves that regressor ancillarity fails whenever the conditional mean mu(x_vec) = E[Y | X_vec = x_vec] is nonlinear because beta(P) then depends on the marginal regressor distribution P_{X_vec}. Decomposing the response as Y = beta' X_vec + eta(X_vec) + eps with canonical nonlinearity eta(X_vec) = mu(X_vec) - beta' X_vec and noise eps = Y - mu(X_vec) satisfying E[X_vec eta(X_vec)] = 0, E[X_vec eps] = 0, E[eta(X_vec) eps] = 0, and conditional MSE m^2(X_vec) = sigma^2(X_vec) + eta^2(X_vec) (Section 5, Equations 6-9), the framework separates the Total Estimation Offset beta_hat - beta(P) into the uncorrelated Noise Estimation Offset beta_hat - beta(X) and Approximation Estimation Offset beta(X) - beta(P), where beta(X) = (X'X)^{-1} X' mu(X) is the conditional target of fixed-X regression (Equations 10-15), both of order O_P(N^{-1/2}) with asymptotic covariance matrices E[X_vec X_vec']^{-1} E[sigma^2(X_vec) X_vec X_vec'] E[X_vec X_vec']^{-1} and E[X_vec X_vec']^{-1} E[eta^2(X_vec) X_vec X_vec'] E[X_vec X_vec']^{-1} whose sum is consistently estimated by the Eicker-Huber-White sandwich estimator AV_hat_sand[beta] = lim_{M -> infty} M V_{P_hat_N}[beta^*_M] (Propositions 7, 8.1, 8.2, Equations 16-20). Using adjusted regressors X_{j.} = X_j - beta_{-j}' X_vec_{-j}, the j-th slope admits case-wise and pairwise weighted-average representations (Section 10, Equations 22-23), and the Ratio of Asymptotic Variances RAV[beta_j, m^2] = AV_total[beta_j] / AV_lin[beta_j] = E[m^2(X_vec) X_{j.}^2] / (E[m^2(X_vec)] E[X_{j.}^2]) = w_sigma RAV[beta_j, sigma^2] + w_eta RAV[beta_j, eta^2] (Definition 11.4, Lemma E.2, Equation 27) quantifies regressor-specific protective (RAV < 1) vs. anti-protective (RAV > 1) standard-error distortion within sharp range [P-min X_{j.}^2 / E[X_{j.}^2], P-max X_{j.}^2 / E[X_{j.}^2]] (Proposition 11.5) and yields the regressor-specific test statistic RAV_hat_j = N <r^2, X_{j,.hat}^2> / (||r||^2 ||X_{j,.hat}||^2) with asymptotic null variance (E[delta^4]/E[delta^2]^2)(E[X_{j.}^4]/E[X_{j.}^2]^2) - 1 (Proposition 12.2, Equation 30) and a permutation test (Section 12.3).

Identity Notes & Concept Scope Distinct from AIMech:ValidPostSelectionInferencePoSIConstant (Berk, Brown, Buja, Zhang, & Zhao 2013), which studies fixed-design homoskedastic Gaussian simultaneous post-selection confidence intervals across submodels M in M_all. Distinct from AIMech:PostModelSelectionPredictorPoSIConfidenceIntervals (Bachoc, Leeb, & Pötscher 2019) and AIMech:UniformlyValidPostModelSelectionConfidenceIntervals (Bachoc, Preinerstorfer, & Steinberger 2020), which study post-model-selection simultaneous coverage constants. Buja, Berk, Brown, George, Pitkin, Traskin, Zhao, & Zhang (2019, Statistical Science 34(4):523-544, arXiv:1404.1578v4, DOI:10.1214/18-STS693) establish the foundational model-robust theory of linear OLS under nonlinear misspecification and random regressors: proving the failure of regressor ancillarity under nonlinearity (Proposition 4.1), the canonical orthogonal decomposition into Noise Estimation Offset beta_hat - beta(X) and Approximation Estimation Offset beta(X) - beta(P) (Sections 5-7, Proposition 7), the exact characterization of the plug-in sandwich estimator as the M -> infty limit of the M-of-N pairs bootstrap (Proposition 8.2), the case-wise and pairwise adjusted-regressor weighted slope representations (Section 10, Equations 22-23), and the Ratio of Asymptotic Variances (RAV) functional, sharp bounds, and regressor-specific specification test RAV_hat_j (Sections 11-12, Definition 11.4, Propositions 11.5 & 12.2, Appendix E).

Claims & Grounded Evidence

4 Claims

Mechanism #regressor-ancillarity-failure-and-canonical-noise-nonlinearity-decomposition

For a joint regressor-response distribution P with integrable second moments and full-rank second-moment matrix E[X_vec X_vec'], the population OLS functional beta(P) = argmin_beta E[(Y - beta' X_vec)^2] = E[X_vec X_vec']^{-1} E[X_vec Y] = E[X_vec X_vec']^{-1} E[X_vec mu(X_vec)] (Section 3, Equations 3-4) depends on P_{X_vec} whenever the conditional expectation mu(x_vec) = E[Y | X_vec = x_vec] is nonlinear: Proposition 4.1 (Section 4, page 10, proved in Appendix E.1, page 36) proves that beta(P) is independent of the marginal regressor distribution P_{X_vec} if and only if mu(X_vec) = beta_0' X_vec almost surely for some beta_0, invalidating conditionality/ancillarity of X_vec under model misspecification. Decomposing the response as Y = beta' X_vec + eta(X_vec) + eps with population residual delta = Y - beta' X_vec = eta(X_vec) + eps, nonlinearity eta(X_vec) = mu(X_vec) - beta' X_vec, and noise eps = Y - mu(X_vec) (Section 5, page 11, Equations 6-9) yields exact orthogonality E[X_vec delta] = E[X_vec eta(X_vec)] = E[X_vec eps] = 0, E[eta(X_vec) eps] = 0, and pointwise conditional MSE decomposition m^2(X_vec) = E[delta^2 | X_vec] = sigma^2(X_vec) + eta^2(X_vec).

supports Source: Models as Approximations I: Consequences Illustrated with Linear Regression · Locator: Sections 3-5, Pages 9-12, Equations (1)-(9), Proposition 4.1; Appendix E.1, Page 36
supports Source: Models as Approximations I: Consequences Illustrated with Linear Regression (Statistical Science Crossref Record) · Locator: Statistical Science 34(4):523-544 abstract

Mechanism #estimation-offset-pythagorean-clt-and-pairs-bootstrap-sandwich-limit

Conditioning on the random design matrix X = (X_vec_1, ..., X_vec_N)' defines the conditional target beta(X) = E[beta_hat | X] = (X'X)^{-1} X' mu(X) (Section 5, page 12, Equation 11) and decomposes the Total Estimation Offset beta_hat - beta(P) into the Noise Estimation Offset beta_hat - beta(X) = (X'X)^{-1} X' eps_vec and the Approximation Estimation Offset beta(X) - beta(P) = (X'X)^{-1} X' eta_vec (Equation 12), which are exactly uncorrelated for finite N (Cov(beta_hat - beta(X), beta(X) - beta(P)) = 0, Equation 13) and satisfy the Pythagorean covariance identity V[beta_hat] = E[V[beta_hat | X]] + V[beta(X)] (Equations 14-15) and the decomposed Central Limit Theorems N^{1/2}(beta_hat - beta(X)) ->^D N(0, E[X_vec X_vec']^{-1} E[sigma^2(X_vec) X_vec X_vec'] E[X_vec X_vec']^{-1}), N^{1/2}(beta(X) - beta(P)) ->^D N(0, E[X_vec X_vec']^{-1} E[eta^2(X_vec) X_vec X_vec'] E[X_vec X_vec']^{-1}), and N^{1/2}(beta_hat - beta(P)) ->^D N(0, E[X_vec X_vec']^{-1} E[m^2(X_vec) X_vec X_vec'] E[X_vec X_vec']^{-1}) (Section 7, page 16, Proposition 7, Equation 16). Furthermore, Proposition 8.1 proves that the Eicker-Huber-White plug-in sandwich estimator AV_hat_sand[beta] = (X'X/N)^{-1} (1/N sum_{i=1}^N r_i^2 x_vec_i x_vec_i') (X'X/N)^{-1} (Equation 17) is consistent under combined nonlinearity and heteroskedasticity, while Proposition 8.2 (Section 8.2, page 18, Equations 18-20) proves that for a fixed sample P_hat_N, the M-of-N x-y bootstrap estimator beta^*_M satisfies M^{1/2}(beta^*_M - beta_hat) ->^D N(0, AV_hat_sand[beta]) and AV_hat_sand[beta] = lim_{M -> infty} M V_{P_hat_N}[beta^*_M].

supports Source: Models as Approximations I: Consequences Illustrated with Linear Regression · Locator: Sections 5-8, Pages 12-19, Equations (10)-(20), Proposition 7, Propositions 8.1-8.2; Appendix C, Pages 32-34
supports Source: Models as Approximations I: Consequences Illustrated with Linear Regression (arXiv:1404.1578 Abstract) · Locator: arXiv:1404.1578v4 abstract

Mechanism #adjusted-regressor-weighted-slopes-and-rav-convex-bounds

Using the population adjusted regressor X_{j.} = X_j - argmin_{L} E[(X_j - L(X_vec_{-j}))^2] and empirical adjusted regressor X_{j,.hat} (Section 9, pages 19-20, Proposition 9.1), Section 10 (pages 20-21, Equations 22-23) proves that beta_hat_j = <y, X_{j,.hat}> / ||X_{j,.hat}||^2 admits both the case-wise weighted slope representation beta_hat_j = sum_{i=1}^N v_i (y_i / X_{i,j.hat}) with weights v_i = X_{i,j.hat}^2 / ||X_{j,.hat}||^2 and the pairwise weighted slope representation beta_hat_j = sum_{i,k} w_{i,k} (y_i - y_k) / (X_{i,j.hat} - X_{k,j.hat}) with weights w_{i,k} = (X_{i,j.hat} - X_{k,j.hat})^2 / (2 N ||X_{j,.hat}||^2) summing to 1. Section 11 and Appendix E (pages 21-25, 36-40, Definition 11.4, Lemma E.2, Proposition 11.5, Proposition E.3, Equations 24-28) define the Ratio of Asymptotic Variances RAV[beta_j, m^2] = AV_total[beta_j] / AV_lin[beta_j] = E[m^2(X_vec) X_{j.}^2] / (E[m^2(X_vec)] E[X_{j.}^2]), prove the convex combination identity RAV[beta_j, m^2] = w_sigma RAV[beta_j, sigma^2] + w_eta RAV[beta_j, eta^2] with w_sigma = E[sigma^2(X_vec)] / E[m^2(X_vec)] and w_eta = E[eta^2(X_vec)] / E[m^2(X_vec)] (Equation 27), and establish the sharp range [P-min X_{j.}^2 / E[X_{j.}^2], P-max X_{j.}^2 / E[X_{j.}^2]] classifying misspecification as protective (RAV < 1, when m^2 concentrates near X_{j.} = 0) or anti-protective (RAV > 1, when m^2 concentrates in the tails of X_{j.}^2).

supports Source: Models as Approximations I: Consequences Illustrated with Linear Regression · Locator: Sections 9-11, Pages 19-25, Equations (21)-(28), Proposition 9.1, Definition 11.4, Lemma 11.4, Proposition 11.5; Appendix E.2-E.4, Pages 36-40, Lemma E.2, Proposition E.3

Mechanism #regressor-specific-rav-specification-test-and-null-asymptotics

Section 12 (pages 25-27, Definition 12.1, Proposition 12.2, Equations 29-30) defines the plug-in test statistic RAV_hat_j = SE_sand[beta_hat_j]^2 / SE_lin[beta_hat_j]^2 = N <r^2, X_{j,.hat}^2> / (||r||^2 ||X_{j,.hat}||^2) where r = Y - X beta_hat, and proves in Proposition 12.2 (proved in Appendix F, page 41) that under the null hypothesis H_0 that delta^2 and X_{j.}^2 are independent (implied by well-specification and homoskedasticity) and fourth moments exist, N^{1/2}(RAV_hat_j - 1) ->^D N(0, (E[delta^4]/E[delta^2]^2)(E[X_{j.}^4]/E[X_{j.}^2]^2) - 1), which equals N(0, 8) when delta and X_{j.} are Gaussian, and pairs this normal test with a heavy-tail-robust permutation test (Section 12.3, page 27) that randomly permutes the squared residuals r_i^2 against X_{i,j.hat}^2.

supports Source: Models as Approximations I: Consequences Illustrated with Linear Regression · Locator: Section 12, Pages 25-27, Equations (29)-(30), Definition 12.1, Proposition 12.2, Section 12.3; Appendix F, Page 41
supports Source: Models as Approximations I: Consequences Illustrated with Linear Regression (Statistical Science Crossref Record) · Locator: Statistical Science 34(4):523-544 abstract

⚠️ Curated Limitations & Negative Results

Model-robustness via the Eicker-Huber-White sandwich estimator AV_hat_sand[beta] or the x-y pairs bootstrap requires finite fourth moments of the regressors X_vec and residuals delta; as emphasized in Section 12.2 (page 26) and Section 13 (pages 27-28), model robustness is distinct from heavy-tail/outlier robustness and can suffer from high variance under heavy-tailed regressor or error distributions.
A population value RAV[beta_j, m^2] = 1 is necessary but not sufficient for linearity and homoskedasticity, because protective (RAV < 1) and anti-protective (RAV > 1) deviations of m^2(X_vec) across the domain of X_{j.}^2 can cancel in expectation E[m^2(X_vec) X_{j.}^2] = E[m^2(X_vec)] E[X_{j.}^2] (Section 12.1, page 25).
Four minor typographical notes in arXiv:1404.1578v4 were identified and verified during full-text inspection: (1) the Section 5 heading (page 11) prints 'DEPOMPOSITION' for 'DECOMPOSITION'; (2) Section 13 (page 28, line 4) prints 'boostrap' for 'bootstrap' and 'R (2008))' with an extra closing parenthesis; (3) Section 14 (page 29, line 9) prints 'dissatified' for 'dissatisfied'; and (4) Appendix E.1 (page 36) duplicates a 4-line paragraph in the proof of Proposition 4.1 (lines 20-29) and writes '\mu_o(\vec{X})' with a letter 'o' instead of '0'.
All 43 pages of the full-text PDF of Buja, Berk, Brown, George, Pitkin, Traskin, Zhao, & Zhang (2019, Statistical Science 34(4):523-544, arXiv:1404.1578v4, DOI:10.1214/18-STS693) were read. Constructed numerical checks in research/discovery/model-robust-nonlinearity-sandwich-and-rav-test-boundaries.py verify (1) Proposition 4.1 regressor-distribution dependence under nonlinearity vs. invariance under linearity and Equations 6-9 canonical orthogonality, (2) Equations 12-15 Estimation Offset Pythagorean decomposition and Proposition 8.2 convergence of M * V_{P_hat_N}[beta^*_M] to AV_hat_sand[beta], (3) Section 10 case-wise and pairwise adjusted-regressor slope identities and Lemma E.2 / Proposition 11.5 RAV convex decomposition and sharp bounds, and (4) Proposition 12.2 Gaussian null variance (= 8) and Section 12.3 permutation test detection of regressor-specific nonlinearity.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:LinearRegression` Linear Regression	related_match	Reconstructs ordinary least squares linear regression model-freely as a population projection functional beta(P) under nonlinearity and heteroskedasticity, with sandwich/bootstrap standard errors and the regressor-specific RAV_hat_j specification test (AIO:LinearRegression).

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`buja2019p1-pdf`	Models as Approximations I: Consequences Illustrated with Linear Regression Andreas Buja, Richard Berk, Lawrence Brown, Edward George, Emil Pitkin, Mikhail Traskin, Linda Zhao, Kai Zhang	primary paper full_text	ID: https://doi.org/10.1214/18-STS693 sha256:11d501663ce1…
`buja2019p1-crossref`	Models as Approximations I: Consequences Illustrated with Linear Regression (Statistical Science Crossref Record) Andreas Buja, Richard Berk, Lawrence Brown, Edward George, Emil Pitkin, Mikhail Traskin, Linda Zhao, Kai Zhang	primary paper abstract_only	ID: https://doi.org/10.1214/18-STS693 sha256:72ef24d40564…
`buja2019p1-abs`	Models as Approximations I: Consequences Illustrated with Linear Regression (arXiv:1404.1578 Abstract) Andreas Buja, Richard Berk, Lawrence Brown, Edward George, Emil Pitkin, Mikhail Traskin, Linda Zhao, Kai Zhang	primary paper abstract_only	ID: https://doi.org/10.48550/arXiv.1404.1578 sha256:38a840a331cd…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini 3

Created: 2026-10-17T17:55:00Z

Updated: 2026-10-17T17:55:00Z

Source File: data/curated/ModelRobustNonlinearityConsistentSandwichAndRAVSpecificationTest.yaml · Schema: curated_concept.yaml

Explore across X-Mech

Related computational & biological Mechs: Embedding Map Network Graph ProteinTraitsMech (ESM-2 pLM) DUFMech (Domain pLM) TraitMech CultureMech All 14 X-Mech Repositories