Sharpening Identification in Large Structural VARs Using Narrative Restrictions

Jan Pr\"user; Lukas Berend

arxiv: 2505.19244 · v3 · pith:TVMVDX2Enew · submitted 2025-05-25 · 💰 econ.EM

Sharpening Identification in Large Structural VARs Using Narrative Restrictions

Lukas Berend , Jan Pr\"user This is my paper

Pith reviewed 2026-05-22 00:57 UTC · model grok-4.3

classification 💰 econ.EM

keywords structural VARnarrative restrictionssign restrictionsimpulse responsesBayesian estimationhigh-dimensional modelsstructural shocksUS macroeconomy

0 comments

The pith

Narrative restrictions imposed as prior constraints on structural shocks sharpen identification and narrow uncertainty in large structural VARs.

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

The paper introduces a framework for high-dimensional structural vector autoregressions that incorporates narrative restrictions directly as linear inequality constraints on the structural shocks through the prior distribution. This extends sign-restricted methods by allowing many such restrictions while maintaining computational feasibility via a new sampling algorithm that avoids rejection sampling inefficiencies. The approach also includes a factor structure on the error terms to manage large numbers of variables. In an application identifying ten shocks across thirty-nine U.S. macroeconomic and financial series, the restrictions produce tighter posterior distributions around impulse responses and more distinct economic interpretations of the shocks. A sympathetic reader would see this as enabling more reliable policy analysis from high-dimensional time-series data.

Core claim

The authors develop a high-dimensional SVAR with a factor structure in the errors that admits a large number of linear inequality restrictions on impact impulse responses and on the structural shocks themselves. Narrative restrictions enter the model as prior constraints on the shocks, and an efficient sampling algorithm draws from the resulting posterior. Applied to U.S. data, the method identifies ten shocks and shows that the added restrictions reduce uncertainty in the impulse responses while clarifying the economic meaning of each shock.

What carries the argument

A prior distribution that encodes narrative restrictions as linear inequality constraints on the structural shocks, paired with a factor structure on the reduced-form errors and a scalable MCMC sampler.

If this is right

Posterior uncertainty around impulse responses decreases as more narrative restrictions are added.
The identified shocks acquire clearer economic labels and produce more distinct dynamic responses.
The sampler remains practical even when the number of variables and restrictions grows substantially.
Identification of a larger set of shocks becomes feasible compared with sign restrictions alone.

Where Pith is reading between the lines

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

The same prior-constraint approach could be tested on datasets from other economies to check whether narrative information consistently narrows bands.
Combining these restrictions with external instruments might reveal whether the two identification strategies reinforce or conflict with each other.
Allowing the factor loadings or the inequality bounds to vary over time could capture evolving interpretations of the same narrative events.

Load-bearing premise

The narrative restrictions chosen by the researcher correctly describe historical events and translate without distortion into linear inequalities on the unobserved structural shocks.

What would settle it

Re-estimating the same thirty-nine-variable U.S. model while dropping the narrative restrictions produces impulse-response bands of identical or greater width and shocks whose responses overlap as much as before.

read the original abstract

We propose a high-dimensional structural vector autoregression framework with a factor structure in the error terms that accommodates a large number of linear inequality restrictions on both impact impulse responses and structural shocks. Our framework extends recent advances in large sign-restricted VARs by allowing narrative restrictions to be imposed directly through constraints on structural shocks via prior distributions, thereby sharpening identification and enhancing the economic interpretability of the structural shocks. To estimate the model, we develop a computationally efficient sampling algorithm that scales well with both model dimension and the number of imposed restrictions, while avoiding the low acceptance-rate problems associated with existing rejection-based approaches. We apply our methodology to a large-scale structural VAR model of the U.S. economy, identifying ten structural shocks and tracing their dynamic effects across thirty-nine macroeconomic and financial variables. The empirical application demonstrates that the incorporation of narrative restrictions improves structural identification in high-dimensional settings by reducing the uncertainty surrounding impulse response functions and facilitating a clearer economic interpretation of the identified structural shocks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a rejection-free sampler for adding narrative restrictions on shocks in large factor-augmented SVARs, but the abstract supplies no numbers on how well it actually tightens identification or enforces the constraints.

read the letter

The main point is that Berend and Prüser combine a factor error structure with narrative restrictions imposed as linear inequalities on structural shocks through the prior, then supply a custom sampler meant to avoid the acceptance-rate collapse that comes with rejection sampling in high dimensions. That combination is the concrete advance over the sign-restricted large VAR papers they cite. The 39-variable US application with ten shocks is presented as evidence that the extra restrictions produce narrower impulse responses and more interpretable shocks, which is the sort of practical payoff people doing policy work care about. If the sampler scales and the restrictions bind as claimed, it could be a useful addition for anyone already estimating big VARs with mixed identification information. The soft spots are straightforward. The abstract asserts efficiency gains and sharper identification without showing acceptance rates, run-time comparisons, or posterior mass checks outside the narrative region. The stress-test worry about leftover mass on non-compliant draws is therefore still open; if the prior is implemented as a soft penalty rather than a hard truncation, the reported credible sets could be wider than a strict rejection procedure would deliver. The central modeling assumption—that the researcher’s narrative restrictions are correctly specified and introduce no systematic bias—also sits more heavily on the prior here than in a pure sign-restriction setup. This is for macroeconometricians who already work with large sign-restricted VARs and want to layer in narrative information without hitting computational walls. A reader who needs to handle thirty-plus variables and cares about feasible computation would find the sampler details worth examining. It deserves a serious referee to check whether the algorithm actually delivers the claimed tightening and whether the empirical gains survive alternative prior specifications.

Referee Report

3 major / 2 minor

Summary. The paper proposes a factor-augmented structural VAR framework for high-dimensional data that incorporates a large number of linear inequality restrictions on both impact impulse responses and structural shocks. Narrative restrictions are imposed directly via prior distributions on the structural shocks rather than rejection sampling. A new MCMC algorithm is developed that is claimed to scale with model dimension and number of restrictions while avoiding low acceptance rates. The method is applied to a 39-variable U.S. macroeconomic dataset to identify ten structural shocks, with the claim that the narrative restrictions materially reduce uncertainty in impulse responses and improve economic interpretability.

Significance. If the sampling algorithm enforces the narrative restrictions such that posterior mass outside the compliant region is negligible, the framework would offer a scalable way to combine sign and narrative information in large SVARs, addressing a practical bottleneck in the existing literature. The empirical illustration with ten shocks across thirty-nine series provides a concrete demonstration of potential gains in interpretability, though the absence of direct quantitative comparisons to rejection-based or sign-only benchmarks leaves the magnitude of improvement open.

major comments (3)

[§3.2] §3.2 (prior construction for narrative restrictions): the linear inequality constraints on structural shocks are incorporated through a prior, but the text does not specify whether the prior is a truncated density, an auxiliary-variable scheme with exact normalization, or a soft penalty. If the implementation is soft or the normalizing constant is approximated, the posterior can retain non-negligible mass on non-compliant regions, which would leave credible sets wider than a correctly enforced hard-constraint procedure and undermine the central sharpening claim.
[§4] §4 (sampling algorithm and efficiency claims): the paper asserts that the new sampler scales well and avoids the low acceptance-rate problems of rejection sampling, yet no explicit acceptance rates, effective sample sizes, or computational-complexity comparison versus standard rejection sampling is reported for the high-dimensional factor-augmented case. Without these diagnostics, the efficiency advantage remains an unverified assertion rather than a demonstrated property.
[Section 5] Empirical application (Section 5, impulse-response figures): the reported reduction in IRF uncertainty is presented visually but without tabulated metrics (e.g., average credible-set width or ratio of widths with versus without narrative restrictions) relative to a pure sign-restricted baseline estimated on the same dataset. This omission makes it impossible to quantify how much the narrative restrictions actually sharpen identification in the 39-variable setting.

minor comments (2)

[§2] The notation distinguishing the factor loadings, structural impact matrix, and narrative-constraint matrix could be made more explicit in §2 to avoid confusion when the same symbols appear in both the factor and restriction blocks.
[Figures] Figure captions should state the exact number of narrative restrictions imposed in each shock identification to allow readers to assess the information content added beyond sign restrictions.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment point by point below and indicate the revisions we will incorporate to improve clarity and strengthen the evidence for our claims.

read point-by-point responses

Referee: [§3.2] §3.2 (prior construction for narrative restrictions): the linear inequality constraints on structural shocks are incorporated through a prior, but the text does not specify whether the prior is a truncated density, an auxiliary-variable scheme with exact normalization, or a soft penalty. If the implementation is soft or the normalizing constant is approximated, the posterior can retain non-negligible mass on non-compliant regions, which would leave credible sets wider than a correctly enforced hard-constraint procedure and undermine the central sharpening claim.

Authors: We thank the referee for this observation. Section 3.2 specifies the prior on the structural shocks as a truncated multivariate normal that assigns zero density to any realizations violating the narrative restrictions. The truncation is handled via an auxiliary-variable representation that permits exact computation of the normalizing constant without approximation or softening. Consequently, the posterior places no mass outside the admissible region. We will revise the section to include the explicit density formula and a brief description of the auxiliary-variable construction to eliminate any ambiguity. revision: yes
Referee: [§4] §4 (sampling algorithm and efficiency claims): the paper asserts that the new sampler scales well and avoids the low acceptance-rate problems of rejection sampling, yet no explicit acceptance rates, effective sample sizes, or computational-complexity comparison versus standard rejection sampling is reported for the high-dimensional factor-augmented case. Without these diagnostics, the efficiency advantage remains an unverified assertion rather than a demonstrated property.

Authors: We agree that quantitative efficiency diagnostics are needed to substantiate the scalability claims. In the revised manuscript we will add a dedicated subsection (or appendix table) reporting acceptance rates, effective sample sizes, and wall-clock times for the proposed sampler on the 39-variable model, together with the corresponding figures obtained from a standard rejection-sampling implementation on the same specification. These additions will allow direct verification of the efficiency gains. revision: yes
Referee: [Section 5] Empirical application (Section 5, impulse-response figures): the reported reduction in IRF uncertainty is presented visually but without tabulated metrics (e.g., average credible-set width or ratio of widths with versus without narrative restrictions) relative to a pure sign-restricted baseline estimated on the same dataset. This omission makes it impossible to quantify how much the narrative restrictions actually sharpen identification in the 39-variable setting.

Authors: We concur that numerical summaries would complement the visual evidence and facilitate precise assessment of the sharpening effect. We will insert a table in Section 5 (or an appendix) that reports, for a representative set of impulse responses, the average credible-set widths under sign restrictions alone and under sign plus narrative restrictions, both estimated on the identical 39-variable dataset. Ratios of these widths will also be provided to quantify the reduction in uncertainty. revision: yes

Circularity Check

0 steps flagged

No significant circularity; new algorithmic framework is self-contained

full rationale

The paper proposes an original high-dimensional SVAR model with factor structure in errors and a new sampling algorithm to incorporate narrative restrictions as linear inequality constraints on shocks via priors. The central claims concern the computational efficiency of this algorithm and its ability to sharpen identification relative to rejection sampling, which are presented as properties of the newly developed procedure rather than re-derivations of quantities already fitted inside the paper. No equations reduce by construction to prior fits, no self-citation is invoked as a load-bearing uniqueness theorem, and the empirical application traces effects across variables without renaming known patterns as novel derivations. The method therefore stands as an independent contribution whose validity rests on the correctness of the proposed priors and sampler, not on internal circular reduction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Framework assumes a factor structure adequately captures cross-variable error dependence and that narrative information can be encoded as valid linear inequalities on shocks without misspecification.

free parameters (1)

prior hyperparameters governing narrative restrictions
Chosen to enforce the inequality constraints on structural shocks.

axioms (1)

domain assumption Factor structure in the reduced-form error covariance is sufficient to handle high-dimensional systems.
Invoked to keep the model tractable when the number of variables reaches 39.

pith-pipeline@v0.9.0 · 5690 in / 1244 out tokens · 34863 ms · 2026-05-22T00:57:20.652753+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We implement our inequality restrictions using truncated Gaussian prior distributions... ft ∼ N(0,I)1(ℓf_t < Rf_t ft < υf_t)
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We use the efficient truncated Normal generator provided by Botev (2017) in order to sample the factors.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.