Pith. sign in

REVIEW 4 major objections 4 minor 15 references

Two or more Gaussian shocks break non-Gaussian SVAR identification invisibly to residual normality tests; the paper shows a normality test on bootstrap replications of the impact matrix catches the failure.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 18:28 UTC pith:Z3S2AYWC

load-bearing objection Useful, honest diagnostic paper for a real problem; the central asymptotic size claim rests on unproven bootstrap and Edgeworth conditions, so it needs a rigorous referee before I'd lean on it. the 4 major comments →

arxiv 2607.17275 v1 pith:Z3S2AYWC submitted 2026-07-19 econ.EM stat.ME

Two Gaussians, Too Many: A bootstrap-based approach to assess identifiability in non-Gaussian structural Vector Autoregressions

Paritosh Shankarrao Junare This is my paper
classification econ.EM stat.ME
keywords non-Gaussian SVARidentificationbootstrap diagnosticGaussian shockspre-testing biasnormal-inverse Gaussianstructural vector autoregressionmoving block bootstrap
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper takes aim at a blind spot in non-Gaussian structural VAR analysis: the usual pre-test for identification, testing reduced-form residuals for normality, cannot see the one failure mode that actually breaks identification — two or more Gaussian structural shocks. The author proposes to bootstrap the estimated impact matrix (the matrix linking structural shocks to observed innovations) and test the bootstrap replications for normality. Under valid identification (at most one Gaussian shock) the conditional bootstrap distribution of the impact matrix is asymptotically normal; when two or more shocks are Gaussian, the likelihood is flat along rotations inside the Gaussian subspace and the replications follow a non-normal distribution, so the diagnostic reduces to a standard normality test. The paper shows the test keeps correct size across the entire null — including the single-Gaussian boundary, where the full Fisher information is singular — and, by letting bootstrap replications and sample size diverge jointly (M/T → 0), does not induce pre-testing bias. A practitioner who wants to know whether non-Gaussian identification actually holds can check it directly, instead of relying on residual tests that cannot distinguish valid from invalid specifications.

Core claim

Central claim: identification validity in a non-Gaussian structural VAR can be read off the bootstrap distribution of the impact-matrix estimator. Under the null (at most one Gaussian shock), the non-Gaussian maximum-likelihood impact-matrix estimator is asymptotically normal, and so are its bootstrap replications. When two or more shocks are Gaussian, the likelihood is flat along rotations of the Gaussian subspace, the estimator fails to concentrate, and the replications are non-normal. This holds at the single-Gaussian boundary, where Fisher information is singular but the impact parameters remain identified. Jointly diverging replications and sample size (M/T→0) make the test pivotal.

What carries the argument

The key object is the studentized bootstrap statistic Q*_T = Σ̂^{-1/2} T^{1/2}(B̂*_T − B̂_T), the standardized gap between resampled and estimated impact matrices, asymptotically standard normal under valid identification. The diagnostic statistic d*_{T,M} splits into a resampling term (asymptotically N(0,1)) and a centering term M^{1/2}(Ĝ*_T − Φ); the joint-divergence condition M T^{-2ρ} → 0 (ρ = 1/2, so M/T → 0) kills the centering term and with it the pre-testing bias. The boundary case relies on profile information: at the Gaussian limit of the NIG family the full Fisher information is singular, yet the impact parameters retain non-singular profile information.

Load-bearing premise

The whole size-control and no-pre-testing-bias argument rests on an assumption the paper states but does not fully prove: that the bootstrap reproduces the sampling behavior of this two-step estimator with an error that shrinks at the rate the theory needs (halving when the sample size quadruples); if the bootstrap converges more slowly, the centering term the test depends on does not vanish and the guarantees fail.

What would settle it

Two checks would settle the core claim. (1) At the single-Gaussian boundary, refit with a distribution family whose approach to Gaussianity is first-order (Student-t with diverging degrees of freedom rather than NIG, where the kurtosis term is O(1/α²)): the paper's Lemma B.1 parity cancellation depends on the second-order structure, so if bootstrap replications of the impact matrix are detectably non-normal at the boundary under a first-order family, the 'valid across the entire null' claim is specific to NIG. (2) Estimate the sup-norm distance between the conditional bootstrap distribution of

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners can replace residual-based normality pre-tests with a normality test on bootstrap replications of the impact matrix; the paper shows residual tests cannot separate valid from invalid identification, while the bootstrap diagnostic can.
  • A researcher who suspects one Gaussian shock need not determine which shock it is or switch specifications: estimating all shocks under the NIG family leaves the impact-matrix estimator consistent and asymptotically normal even at the Gaussian boundary, with valid standard errors.
  • Because the diagnostic is asymptotically pivotal and ancillary under the joint M, T divergence, subsequent impulse-response inference conditioned on it is not distorted — unlike inference conditioned on residual pre-tests, which the weak-identification simulations show can substantially over- or under-cover parameters.
  • A univariate version of the test localizes, in large samples, which shocks are responsible for the identification failure, extending the procedure to partially identified non-Gaussian SVARs.
  • Under the alternative of two or more Gaussian shocks the test is consistent, with power increasing in the sample size, while M governs the finite-sample size-power trade-off.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The boundary-validity proof is specific to the NIG family, whose approach to Gaussianity is second-order in the boundary parameter (excess kurtosis proportional to 1/α²); for families approaching Gaussianity at a first-order rate (e.g., Student-t with diverging degrees of freedom, excess kurtosis proportional to 1/ν), the parity cancellation that keeps the boundary nuisance out of the impact matri
  • Power is demonstrated against fixed alternatives (two exactly Gaussian shocks). The behavior against local alternatives — a shock with small but nonzero excess kurtosis — is the open regime the paper itself flags; the M/T → 0 rate that secures size control also caps finite-sample power.
  • The mechanism is generic — a bootstrap distribution departs from its nominal limit exactly when the criterion is flat along an unidentified direction — so the same template could diagnose identification failures in moment-based estimators that relax full independence, an extension the paper mentions but does not develop.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper develops a bootstrap diagnostic for checking whether a non-Gaussian SVAR satisfies the identifying condition that at most one structural shock is Gaussian. Under the null of valid identification, the studentized bootstrap distribution of the NGML impact-matrix estimator is claimed to be asymptotically standard normal, so the diagnostic reduces to a normality test on bootstrap replications. The paper further claims validity at the single-Gaussian boundary, where the full-parameter information matrix is singular, and that joint divergence of M and T (with M/T→0) makes the test asymptotically pivotal and free of pre-testing bias. Monte Carlo experiments with NIG shocks and an empirical uncertainty SVAR illustrate the procedure.

Significance. If the theoretical claims are fully established, the paper offers a useful and operationally simple diagnostic that directly targets the number of Gaussian structural shocks, and it provides a concrete argument for why conditioning on this diagnostic need not distort later inference. The simulation and empirical sections are informative and the idea of using the bootstrap distribution itself as a diagnostic is likely to be of independent interest. However, the central asymptotic results rely on high-level conditions that are asserted rather than proved for the two-step NGML estimator, so the contribution is currently conditional on unverifiable regularity assumptions.

major comments (4)
  1. [Appendix B.3, Condition B.1 and Eq. (B.33)] The proof of Proposition 3.1(i) rests on the assertion that supx |Ĝ*_T(x)−Φ(x)| = O_p(T^{−ρ}) with ρ = 1/2. This Edgeworth rate is stated in Remark B.4, not proved, and standard Edgeworth theorems cited there do not cover a two-step estimator whose second stage uses generated residuals and whose boundary nuisance is estimated at rate T^{−1/4}. If the actual rate is slower, the centering term M^{1/2}Ω̂_T^{−1/2}(Ĝ*_T−Φ) need not vanish under M/T→0, so the claimed asymptotic size and the no-pre-testing-bias property are not established. The condition should either be proved or explicitly maintained as an assumption, with the joint rate condition adjusted accordingly.
  2. [Appendix B.2, Condition (BV) and Lemma B.3] Condition (BV) — first-order validity of the residual-based moving block bootstrap composed with the two-step NGML map — is introduced as 'maintained' and then used as a premise in Lemma B.3 and Proposition 2.2. The cited results of Brüggemann et al. (2016) and Gonçalves and White (2004) concern reduced-form VARs and standard extremum estimators, not the present two-step procedure with generated residuals. Without a proof or a clearly stated assumption covering this composition, the bootstrap normality result underlying the diagnostic is unsupported.
  3. [Appendix B.1, Proposition 2.1 and Lemma 2.1] The boundary case is argued through expansion (B.1) and the parity-cancellation Lemma B.1. The proof is heuristic: the second-order approach of the NIG density to Gaussianity and the cancellation ¯c=0 are asserted rather than derived from the NIG density, and the slow rate ρ̂_{k,T}=O_p(T^{−1/4}) is not shown to be harmless for the studentized bootstrap distribution. Since Proposition 2.1 is what extends the diagnostic to the full null, a rigorous treatment is needed before claiming validity across the boundary.
  4. [Appendix B.3, Lemma B.4] The divergence under H1 is not proved in the manuscript. Lemma B.4 asserts that affected bootstrap replications converge to a non-normal rotation-induced distribution and that any consistent normality test diverges, but the proof is a sketch and the general n≥3 case is left to a 'maintained' statement. The power claim of Proposition 3.1(ii) rests on this lemma, so it should either be proved for the stated alternatives or the consistency claim should be narrowed accordingly.
minor comments (4)
  1. [Abstract and Remark 3.2] The paper repeatedly states the joint rate as M/T→0, but the formal condition is M T^{−2ρ}=o_p(1), which reduces to M/T→0 only when ρ=1/2. The text should consistently distinguish the general condition from the Edgeworth special case.
  2. [Section 4.5] The uniformity statement for p-values under H0, π^*_{M,T,S}(x)→x, is stated without reference to the joint M,T,S asymptotics. It should be formulated as a precise limit theorem under the same joint regime used elsewhere.
  3. [Table 1 caption] The caption should clarify which columns correspond to the Doornik–Hansen and Jarque–Bera tests respectively; currently the table layout is ambiguous.
  4. [Algorithm A.2, step 4] The centering formula uses û^*_{jl+s} both before and after centering; the notation should be cleaned up so that the definition of the centered residuals is unambiguous.

Circularity Check

1 steps flagged

Central bootstrap-normality result is assumed via maintained Condition (BV); the diagnostic's size guarantee is conditional on an input equivalent to its conclusion.

specific steps
  1. self definitional [Appendix B.2, Condition (BV); Lemma B.3; Proposition 2.2 proof (B.2.4)]
    "Condition (BV) [Bootstrap validity of the two-step estimator]. The residual-based MBB of Algorithm A.2, composed with the NGML second step, is first-order valid: conditional on the data and under H0, the bootstrap estimator ˆλ∗ T admits the same first-order expansion as ˆλ T ... This two-step composition is maintained; the β-mixing and (2+δ)-moment conditions it requires are verified in Lemma B.2."

    Proposition 2.2 — and hence Corollary 2.1, the basis for the diagnostic's claim that the bootstrap replications are asymptotically normal under H0 — is proved by invoking Condition (BV) in Lemma B.3. But Condition (BV) is defined as exactly the first-order bootstrap validity of the composed residual-based MBB + NGML estimator, and it is labeled 'maintained,' not derived from the proposition's stated assumptions (A.1, Assumptions 2.1–2.3). The proof of Proposition 2.2 therefore assumes the target result as an input; the 'asymptotic normality of the bootstrap distribution' is a restatement of Condition (BV), not an independent derivation.

full rationale

The paper's main theoretical guarantee — that, under the null, the conditional bootstrap distribution of the NGML impact-matrix estimator is asymptotically normal — is not fully derived. Proposition 2.2's proof relies on Condition (BV), which is labeled 'Bootstrap validity of the two-step estimator' and explicitly 'maintained.' That condition states essentially the same first-order validity that Proposition 2.2 claims. Corollary 2.1 and Proposition 3.1 inherit this circularity, so the diagnostic's claimed correct asymptotic size under the full null is conditional on an input equivalent to its conclusion. Separately, the O_p(T^{-1/2}) Edgeworth rate of Condition B.1/Remark B.4 is asserted rather than established for the two-step estimator at the single-Gaussian boundary; this is an omitted proof rather than a circular step. The paper does contain non-circular components: the diagnostic is not fitted to its target, the Monte Carlo evidence is independent, and the comparison with residual-based pre-tests is meaningful. But the central first-principles claim of bootstrap asymptotic normality reduces, by the paper's own proof structure, to a maintained condition that is the result itself.

Axiom & Free-Parameter Ledger

2 free parameters · 7 axioms · 0 invented entities

The central claim rests on the null hypothesis that at most one shock is Gaussian, the independence/i.i.d. assumption on shocks, the stability of the VAR, and a set of regularity conditions. Critically, the proof of the diagnostic's size control depends on an Edgeworth expansion of the bootstrap distribution with rate T^{-1/2} (Condition B.1) and on the bootstrap validity of the two-step NGML estimator (Condition BV), both of which are asserted, not proven. No new entities are introduced.

free parameters (2)
  • M (bootstrap replications per test sequence) = M = (1/2)T^{3/5} (also (1/3)T^{3/5} in simulations)
    User-selected number of bootstrap replications for each normality test sequence; controls the size-power trade-off. The theory only requires M/T → 0.
  • block length l = l = floor(5.03 T^{1/4})
    Standard rule from Hall et al. (1995) for moving block bootstrap; not fitted to data, but a practical choice.
axioms (7)
  • domain assumption Assumption 2.1: det(I_np − Π z) ≠ 0 for |z| ≤ 1 (VAR stability)
    Required for stationarity and the bootstrap CLT for the score (Lemma B.2).
  • domain assumption Assumption 2.2: structural shocks are i.i.d., mean-zero, mutually independent, with diagonal covariance
    The identification result relies on independence; the i.i.d. design is used for the bootstrap score CLT.
  • domain assumption Assumption 2.3: at most one structural shock is Gaussian (the null hypothesis)
    This is the condition being tested; the diagnostic's validity under H0 is the claim.
  • domain assumption Regularity conditions A.1 (1)-(7): compact parameter space, smooth densities, integrability, positive definite Fisher information
    Standard ML regularity needed for asymptotic normality of the NGML estimator.
  • ad hoc to paper Condition B.1: the conditional bootstrap distribution converges to the normal at rate O_p(T^{-ρ}) with ρ=1/2
    Asserted for the studentized bootstrap statistic; the size control of the diagnostic (Proposition 3.1) depends on this rate. Not proven for the two-step NGML estimator.
  • ad hoc to paper Condition (BV): residual-based moving block bootstrap is first-order valid for the two-step NGML estimator
    Needed to justify the bootstrap CLT for the score; the proof refers to Gonçalves and White (2004) and Brüggemann et al. (2016) but does not verify their conditions for the NGML two-step estimator.
  • domain assumption Joint divergence M T^{-2ρ} = o_p(1) with ρ=1/2, i.e., M/T → 0
    This is part of the procedure, not an axiom; included because the size control depends on it.

pith-pipeline@v1.3.0-alltime-deepseek · 47822 in / 10803 out tokens · 93105 ms · 2026-08-01T18:28:08.776429+00:00 · methodology

0 comments
read the original abstract

Standard pre-tests of normality on reduced-form innovations are insufficient to detect two or more Gaussian shocks and hence, the failure of identification in non-Gaussian SVARs. We instead propose a bootstrap-based approach to evaluate the asymptotic validity of this condition by measuring the divergence between the conditional bootstrap distribution of a maximum likelihood estimator and its limiting distribution under valid identification. We show that, under valid identification and certain regularity conditions, the conditional bootstrap distribution of the impact matrix is asymptotically normal, so the diagnostic reduces to a test of normality of the bootstrap replications. The diagnostic remains valid in the single-Gaussian case, where the shape parameter of the Gaussian shock lies on the boundary, and the full-parameter information is singular; this establishes its validity across the entire null. Under the null of valid identification, the diagnostic induces no pre-testing bias as bootstrap replications and sample size diverge jointly at an appropriate rate. The joint divergence ensures that the test statistic, conditional on the data, is asymptotically pivotal, so conditioning on the diagnostic does not distort subsequent inference. Monte Carlo simulations with Normal-Inverse Gaussian shocks show that the diagnostic attains near-exact nominal size under valid identification and detects the failure due to multiple Gaussian shocks with power increasing in the sample size. Under weak identification with a near-Gaussian shock, conditioning on the bootstrap diagnostic, unlike on residual-based normality pre-tests, preserves the probability coverage of the estimates. Based on estimates of a SVAR model in the macroeconomic and financial uncertainty literature, we demonstrate its potential as a practical, robust tool for validating non-Gaussian identification without pre-testing bias.

Figures

Figures reproduced from arXiv: 2607.17275 by Paritosh Shankarrao Junare.

Figure 1
Figure 1. Figure 1: Empirical distribution functions of the p-values, p ∗ M,T ,s(x), s = 1, 2, · · · , S of the multivariate Doornik-Hansen normality test for the estimates of BT for the three specifications - ICA 0, ICA 1 and No ICA - with sample size T = 500 and M = (1/2)T 3/5 , for the number of bootstrap sequences, S = 1000, across Monte Carlo simulations, NS = 500 The darker and lighter shades indicate the 50 and 90- per… view at source ↗
Figure 2
Figure 2. Figure 2: Univariate ICA 1: Empirical distribution function of p ∗ M,T ,S(x) from the univariate Doornik￾Hansen test for normality of each element of (bootstrapped) estimated on-impact matrix Bˆ∗ T for the (valid) specification ICA 1 (see Section 4.2), with sample size T = 500, the number of bootstrap replicates in a sequence, M = (1/2)T 3/5 , and the number of sequences/tests, S = 1000, across Monte Carlo Simulatio… view at source ↗
Figure 3
Figure 3. Figure 3: Univariate No ICA: Empirical distribution function of p ∗ M,T ,S(x) from the univariate Doornik￾Hansen test for normality of each element of (bootstrapped) estimated on-impact matrix Bˆ∗ T for the (invalid) specification No ICA (see Section 4.2), with sample size T = 500, the number of bootstrap replicates in a sequence, M = (1/2)T 3/5 , and the number of sequences/tests, S = 1000, across Monte Carlo Simul… view at source ↗
Figure 4
Figure 4. Figure 4: Estimated IRFs of the trivariate SVAR system Xt = (UMt, Yt, UF t) ′ identified through ICA with NGML estimation with NIG-distributed structural shocks, along with the bootstrap-based confidence intervals (CIs) at 68% level (dashed red lines). The solid lines report the impulse responses of Xt to one-standard deviation positive shock. The bootstrap-based CIs are constructed using residual-based MB bootstrap… view at source ↗
Figure 5
Figure 5. Figure 5: Empirical distribution functions of the p-values, p ∗ M,T ,s(x), s = 1, 2, · · · , S of the multivariate Doornik-Hansen normality test for the estimates of Bˆ T for the three specifications - ICA 0, ICA 1 and No ICA - with sample size T = 300 and M = (1/2)T 3/5 , for the number of bootstrap sequences, S = 1000, across Monte Carlo simulations, NS = 500. The darker and lighter shades represent the 50 and 90-… view at source ↗
Figure 6
Figure 6. Figure 6: Univariate ICA 1: Empirical distribution function of p ∗ M,T ,S(x) from the univariate Jarque￾Bera test for normality of each element of (bootstrapped) estimated on-impact matrix Bˆ∗ T for the (valid) specification ICA 1 (see Section 4.2), with sample size T = 500, the number of bootstrap replicates in a sequence, M = (5/3)T 3/5 , and the number of sequences/tests, S = 1000, across Monte Carlo Simulations … view at source ↗
Figure 7
Figure 7. Figure 7: Univariate No ICA: Empirical distribution function of p ∗ M,T ,S(x) from the univariate Jarque￾Bera test for normality of each element of (bootstrapped) estimated on-impact matrix Bˆ∗ T for the (valid) specification No ICA (see Section 4.2), with sample size T = 500, the number of bootstrap replicates in a sequence, M = (5/3)T 3/5 , and the number of sequences/tests, S = 1000, across Monte Carlo Simulation… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references

  1. [1]

    and Stegun, I

    Abramowitz, M. and Stegun, I. A. (1974).Handbook of Mathematical Functions, With For- mulas, Graphs, and Mathematical Tables,. Dover Publications, Inc., USA. Amengual, D., Fiorentini, G., and Sentana, E. (2022). Moment tests of independent compo- nents.SERIEs, 13(1):429–474. Amengual, D., Fiorentini, G., and Sentana, E. (2024). Specification tests for non...

    arXiv 1974

  2. [2]

    Condition (MBB)

    The required moment condition R |x|2fi(x;θ i,0)dx=E[ε 2 i,t]<∞holds byV ar(ε i,t)<∞. Condition (MBB). The block lengthl=l T in Algorithm A.2 satisfiesl→ ∞andl/T→0 asT→ ∞(e.g.l=⌊5.03T 1/4⌋), so that the number of blocksL=⌊T /l⌋ → ∞, and the block start indicesi 1, . . . , iL are drawn independently and uniformly from{0,1, . . . , T−l}, independently of the...

    2004

  3. [3]

    24We fix the block length in the residual-based MBB algorithm to the largest integer smaller than 5.03T 1/4, see Hall et al

    a normality test is applied toMreplications per test sequence, andSsuch sequences are used to summarize the empirical distribution of thep-values. 24We fix the block length in the residual-based MBB algorithm to the largest integer smaller than 5.03T 1/4, see Hall et al. (1995), Jentsch and Lunsford (2016), Jentsch and Lunsford (2019), Mertens and Ravn (2...

    1995

  4. [4]

    Since the true parameter lies at the boundary αk =∞, there is noO(α −1 k ) remainder (the leading departure from invariance isO(α −2 k ) and vanishes at first order)

    The annihilation IψθN=J ψθN={0}is thereforeexact. Since the true parameter lies at the boundary αk =∞, there is noO(α −1 k ) remainder (the leading departure from invariance isO(α −2 k ) and vanishes at first order). This proves (ii). B.1.3.Schur-complement information.The information matrixI θθ is symmetric and positive semidefinite, so range(Iθθ) =N ⊥, ...

    2017

  5. [5]

    (1) Atρ k = 0, the NIG density departs from the Gaussian limit case at orderρ 2 k (its excess kurtosis is 3ρ 2 k/σ2 k)

    Hence, ¯c= 0.□ B.1.4.Proof of Proposition 2.1.Under the representationρ k = 1/α k the Gaussian limit is the finite boundary pointρ k = 0 and the criterion is smooth inψforθin a neighborhood of this point. (1) Atρ k = 0, the NIG density departs from the Gaussian limit case at orderρ 2 k (its excess kurtosis is 3ρ 2 k/σ2 k). Hence, theρ k-score vanishes and...

    1999

  6. [6]

    When all shocks are non- Gaussian the model is a correctly specified, the information identity givesI ψ =J ψ, andV B collapses to (2.15)

    d − → N 0,I −1 ψ JψI −1 ψ , where the profile-likelihood construction follows Murphy and van der Vaart (2000), and the delta method throughB(β) diag(σ) yields (2.13). When all shocks are non- Gaussian the model is a correctly specified, the information identity givesI ψ =J ψ, andV B collapses to (2.15). B.2.Proof of Proposition 2.2.We collect all the mode...

    2000

  7. [7]

    (2016)(Theorem 4.1), a standard Taylor expansion of the bootstrap score completes the proof

    and of the bootstrap estimatorˆλ∗ T Br¨ uggemann et al. (2016)(Theorem 4.1), a standard Taylor expansion of the bootstrap score completes the proof. Lemma B.2(Score Conditions under NIG shocks).For the likelihood function defined through Equations (2.4)-(2.6) under the parameter spaceΘ θi of the NIG distributed shocks, Θθi ={(α i, γi, δi, µi)∈(0,∞)×R×(0,∞...

    2016

  8. [8]

    By Assumption 2.2,ε t isi.i.d., so it is independent ofF t−1. Since∇ λℓt(λ0) is a measurable 42TWO GAUSSIANS, TOO MANY function ofε t, and independent ofF t−1, it follows from Billingsley (1995) (Section 35.2) that ∇λℓt(λ0) is a martingale difference sequence with respect toF t−1, since: E[∇λℓt(λ0)| Ft−1] =E[∇ λℓt(λ0)]. The latter expectation equals zero ...

    1995

  9. [11]

    by condition (7) of A.1. Under Assumption 2.1 the companion process is stable and, the NIG density being strictly positive onR, geometrically ergodic; hence∇ λℓt is geometricallyβ- mixing, as a measurable function of the geometrically ergodic companion process. These are the conditions under which the residual-based moving-block bootstrap of Br¨ uggemann ...

    2016

  10. [12]

    By the consistency ofˆλT through Lanne et al

    by a uniform law of large numbers under Condition (MBB) and the moment bounds of Lemma B.2 (Gon¸ calves and White, 2004), hence the limit is nonsingular. By the consistency ofˆλT through Lanne et al. (2017)(Theorem

    2004

  11. [13]

    and of ˆλ∗ T through Br¨ uggemann et al. (2016)(Theorem 4.1), 48TWO GAUSSIANS, TOO MANY combining with Lemma B.3 yields T 1/2ˆλ∗ T − ˆλT d∗ − →p N 0,I( ˆλT )−1 , and pre-multiplying by ˆΣ−1/2 λT , a consistent estimator of the asymptotic covariance of ˆλT , gives the studentized statement of Proposition 2.2. The argument presumes the regularity conditions...

    2016

  12. [14]

    = E[∇λℓt∇λℓ′ t] =I(λ 0); when one shock is Gaussian the same steps apply to the structural sub-vectorψ= (vec(Π) ′, β′, σ′)′ and yield the robust (sandwich) covariance of Corollary 2.1 (Proposition 2.1). Remark B.3(Why a moving-block bootstrap).Because the score is a martingale difference (Lemma B.2), it is serially uncorrelated, and its long-run variance ...

    2016

  13. [15]

    Proof. (i)Withk≥2 Gaussian shocks the Gaussian sub-block of the density is rotation- invariant Comon (1994), so the NGML criterion is flat alongO(k) and that information block is singular (Gouri´ eroux et al., 2017; Maxand, 2020). The bootstrap residuals are asymptoti- cally Gaussian in the subspace, so the bootstrap criterion is asymptotically flat there...

    1994

  14. [172]

    and Lunsford, K

    Jentsch, C. and Lunsford, K. G. (2016). Proxy SV ARs: Asymptotic Theory, Bootstrap Inference, and the Effects of Income Tax Changes in the United States.Working Paper, –(16-19). Institution: Federal Reserve Bank of Cleveland. Jentsch, C. and Lunsford, K. G. (2019). The Dynamic Effects of Personal and Corpo- rate Income Tax Changes in the United States: Co...

    arXiv 2016

  15. [1974]

    (henceforth,AM-SI), Section 9.6. (2) Behavior asr i,t → ∞: By the uniform asymptotic expansion of modified Bessel func- tionsAM-SI(Equation 9.7.2): Kν(z) = r π 2z e−z 1 + 4ν2 −1 8z +O(z −2) asz→ ∞ Taking ratios forν∈ {0,1,2}and simplifying: K0(r) K1(r) = 1− 1 8r +O(r −2) 1 + 3 8r +O(r −2) − →1 asr→ ∞(B.4) TWO GAUSSIANS, TOO MANY43 K2(r) K1(r) = 1 + 15 8r ...

    1997