Partially identified heteroskedastic SVARs

This paper addresses the failure of point identification in heteroskedastic Structural Vector Autoregressions (SVARs) caused by eigenvalue multiplicity during variance breaks by deriving identified sets for impulse responses using zero or sign restrictions and applying robust Bayesian inference, illustrated through an empirical analysis of the global crude oil market.

The Gist

Imagine you are a detective trying to solve a crime in a chaotic city. You have a security camera (your data) that recorded a series of events, but the footage is blurry and mixed up. You know there were three different culprits (structural shocks) acting at the same time, but you can't tell who did what just by looking at the blurry video.

In the world of economics, this is the problem of identifying Structural Vector Autoregressions (SVARs). Economists want to know: "Did a supply shock cause the oil price to rise, or was it a demand shock?"

The Old Tool: The "Volatility Flashlight"

For a while, economists had a clever trick called Heteroskedasticity. Think of this as a "Volatility Flashlight."

Imagine the city has two distinct eras:

1. Era 1 (Calm): The criminals are sneaky and quiet. Their movements are small and consistent.
2. Era 2 (Chaos): Suddenly, the criminals get wild. One of them starts running around wildly (high variance), while the others stay calm.

If you look at the security footage, you can see who got wild. If the "Oil Supply" criminal suddenly started running around in Era 2, but the "Demand" criminal stayed calm, you can finally point at the camera and say, "That's the supply shock!" This works great if the criminals change their behavior differently.

The Problem: The "Twin" Criminals

But what if, in Era 2, two of the criminals start running around with the exact same intensity?

Maybe the "Supply" criminal and the "Aggregate Demand" criminal both start sprinting at the exact same speed. Now, your flashlight sees two blurry streaks moving together. You can't tell which streak belongs to whom. The math says: "We can't identify these two."

In the past, if this happened, economists had to throw up their hands and say, "We can't solve this case using volatility." They would have to use a completely different, often weaker, method.

The New Strategy: The "Detective's Notebook"

This paper, by Bacchiocchi and colleagues, says: "Don't give up! Just because the flashlight is blurry doesn't mean you can't solve the case. You just need to open your notebook."

The authors propose a new strategy: Combine the Volatility Flashlight with a few specific clues (Zero Restrictions).

Here is the analogy:

• The Flashlight (Heteroskedasticity): Tells you that two criminals are moving together in a specific way. It narrows down the possibilities but doesn't give a final answer.
• The Notebook (Zero Restrictions): You write down a rule based on economic theory. For example: "The Supply Shock cannot affect the Economy in the very first second."

When you combine the blurry flashlight image with this specific rule from your notebook, the math suddenly snaps into focus. Even though the two criminals were running at the same speed, the rule tells you exactly which one is which.

The "Set" vs. "Point" Identification

Sometimes, even with the notebook, you can't get a single, perfect answer. You might only be able to say, "The criminal is somewhere in this specific neighborhood."

• Point Identification: You catch the criminal and know their exact name. (The paper shows how to do this if you have just the right number of rules).
• Set Identification: You know the criminal is in a specific house, but you aren't sure which room. (The paper provides a way to calculate the "bounds" of this house).

The paper introduces a Robust Bayesian method. Think of this as a super-smart AI that runs millions of simulations. It doesn't just guess one answer; it maps out the entire "neighborhood" where the answer could be, giving you a range of possibilities that is statistically sound, even when the data is messy.

The Real-World Test: The Oil Market

To prove their theory, the authors applied this to the Global Crude Oil Market.

• They looked at oil production, economic activity, and oil prices.
• They found that the "volatility flashlight" failed because two of the shocks (likely supply and demand) changed their behavior in a way that looked identical to the data.
• Standard methods said, "Game over, we can't tell them apart."
• The New Method: By adding a few logical rules (like "Oil supply shocks don't instantly change global economic activity"), they were able to separate the shocks again. They successfully identified the oil supply shock and the demand shock, even though the data was "confused."

The Takeaway

This paper is a toolkit for economists who are stuck. When the data is too noisy to give a single clear answer, don't throw the data away. Instead, combine the statistical patterns of the data with a few well-chosen economic rules. This allows you to either find the exact answer or, at the very least, draw a very tight circle around where the answer must be.

In short: When the data is ambiguous, use a little bit of logic to clear up the fog.

Technical Summary

Here is a detailed technical summary of the paper "Partially identified heteroskedastic SVARs" by Bacchiocchi, Bastianin, Kitagawa, and Mirto.

1. Problem Statement

Structural Vector Autoregressive (SVAR) models often rely on heteroskedasticity (changes in the variance of structural shocks across different regimes) to identify structural parameters. Standard identification through heteroskedasticity requires that the shifts in the variances of the structural shocks be statistically distinct (i.e., the eigenvalues of the ratio of covariance matrices between regimes must be distinct).

However, in empirical applications, researchers frequently encounter situations where:

1. Volatility shifts exist but are proportional or statistically indistinguishable for some shocks (leading to eigenvalue multiplicity).
2. Standard statistical tests (e.g., Lütkepohl et al., 2020) fail to reject the null hypothesis that certain variance shifts are equal.

In these cases, the standard heteroskedastic SVAR (HSVAR) approach fails to provide point identification (a unique solution), leaving the model only set-identified or unidentified. Existing literature often suggests abandoning the HSVAR approach in favor of other identification schemes (like narrative restrictions or external instruments) or combining them in complex ways. This paper addresses the gap of how to proceed when heteroskedasticity alone is insufficient but still present.

2. Methodology

The authors propose a unified framework that combines heteroskedasticity with zero and sign restrictions to recover identification when variance shifts are not fully distinct.

A. Theoretical Framework

• Eigen-decomposition Approach: The identification of the structural matrix $C = A_0^{-1}$ and the variance shift matrix $\Lambda$ is cast as an eigen-decomposition problem: $\Omega_1^{-1,tr} \Omega_2 \Omega_1^{-1,tr'} = Q \Lambda Q'$, where $\Omega_1, \Omega_2$ are reduced-form covariance matrices, and $Q$ is an orthogonal matrix of eigenvectors.
• Handling Multiplicity: When eigenvalues are distinct, $Q$ is unique (up to permutation and sign). When eigenvalues are equal (multiplicity $m_i > 1$), the corresponding eigenvectors span a subspace of dimension $m_i$, and $Q$ is not unique.
• Hybrid Identification Strategy: The paper derives conditions under which zero restrictions (exclusion restrictions) and sign restrictions can pin down the specific eigenvectors within the ambiguous subspace.
  • Theorem 1 (Point Identification): Even with eigenvalue multiplicity, point identification is achievable if the number of zero restrictions on the eigenvectors associated with the multiple eigenvalue satisfies a specific rank condition ($f_{ij} = m_i - j$). Crucially, this requires fewer restrictions than a traditional SVAR because heteroskedasticity has already identified the subspace.
  • Theorems 2 & 3 (Set Identification & Convexity): If zero restrictions are insufficient for point identification, the model remains set-identified. The authors prove that under specific ordering of restrictions and sign constraints, the identified set for impulse responses is convex. This convexity is essential for valid statistical inference.

B. Inference: Robust Bayesian Approach

To conduct inference on set-identified models, the authors adapt the Robust Bayesian approach (Giacomini and Kitagawa, 2021).

• Problem with Standard Bayes: In set-identified models, the posterior distribution depends heavily on the prior chosen for the unidentified orthogonal matrix $Q$.
• Solution: Instead of fixing a single prior for $Q$, they consider a set of priors. The inference is based on the posterior mean bounds (the infimum and supremum of the posterior mean over the set of admissible priors) and a robustified credible region.
• Algorithm 1: The paper provides a computational algorithm to:
  1. Estimate reduced-form parameters via Gibbs sampling.
  2. Perform eigen-decomposition.
  3. Check for eigenvalue multiplicity.
  4. If multiplicity exists, generate admissible $Q$ matrices satisfying zero/sign restrictions via linear projections.
  5. Solve constrained nonlinear optimization problems to find the bounds of the identified set for impulse responses.

3. Key Contributions

1. New Identification Strategy: The paper demonstrates that heteroskedasticity + zero restrictions can achieve point identification even when variance shifts are proportional (statistically indistinguishable). This allows researchers to use HSVARs in cases where standard tests suggest failure.
2. Analytical Results on Multiplicity: It extends the literature on set identification (Giacomini and Kitagawa, 2021) to the HSVAR context, providing necessary and sufficient conditions for point identification and proving the convexity of the identified set under mixed restrictions.
3. Robust Inference Algorithm: It offers a practical, implementable algorithm for computing bounds and robust credible regions for impulse responses in partially identified HSVARs, avoiding the pitfalls of pre-testing for eigenvalue multiplicity on every posterior draw.
4. Economic Interpretation: It shows how to recover structural shocks that are consistent with economic theory even when statistical identification is weak, by requiring fewer economic restrictions than a purely recursive SVAR.

4. Empirical Results

The methodology is applied to the global crude oil market (Kilian, 2009 dataset: production, real activity, real oil price).

• Data Context: The sample covers 1973–2007 with a volatility break in 1987.
• Test Results: Statistical tests (Lütkepohl et al., 2020) confirm that the variance shifts for two of the three shocks are not distinct ($\lambda_2 = \lambda_3$). Standard HSVAR identification fails.
• Model Comparisons:
  • M0 (Recursive): Standard benchmark.
  • M1 (Standard HSVAR): Assumes distinct eigenvalues (incorrectly).
  • M2 (Set-Identified): Uses heteroskedasticity + sign restrictions. Only one shock (oil-specific demand) is point-identified; supply and aggregate demand shocks are set-identified.
  • M3 (Point-Identified): Uses heteroskedasticity + one exclusion restriction ($c_{21}=0$, oil supply shock does not affect activity on impact).
• Findings:
  • Model M3 successfully achieves point identification for all shocks using only one zero restriction, whereas a traditional recursive SVAR requires three.
  • The impulse responses in M3 are economically plausible (e.g., oil supply shocks raise prices and lower activity) and the confidence intervals are tighter than in the set-identified M2 model.
  • The robust credible regions confirm that the results are not driven by arbitrary prior choices.

5. Significance

This paper provides a critical "safety net" for econometricians using heteroskedastic SVARs. It resolves the dilemma where volatility shifts exist but are insufficient for full identification. By showing that minimal additional economic restrictions (zero or sign) can restore point identification, the authors:

• Prevent the need to discard valuable volatility information.
• Reduce the reliance on strong, potentially untestable exclusion restrictions required in traditional SVARs.
• Offer a rigorous framework for inference in "partially identified" settings, ensuring that uncertainty is properly quantified rather than ignored.

The approach is particularly valuable in macro-finance and commodity markets where structural breaks in volatility are common but often affect multiple shocks similarly.