SVARs with breaks: Identification and inference

This paper proposes a class of structural vector autoregressions with breaks (SVAR-WB) that incorporates cross-regime constraints to improve identification, derives conditions for point and set identification, and develops robust Bayesian and frequentist inference methods to address the unreliability of standard approaches caused by multiple observationally equivalent structural parameters.

The Gist

Imagine you are a detective trying to solve a mystery, but the crime scene has changed halfway through the investigation.

In the world of economics, researchers use a tool called a SVAR (Structural Vector Autoregression) to figure out how different parts of the economy (like inflation, interest rates, and jobs) react to each other. Think of the economy as a giant, complex machine with many gears turning. When you push one gear (a "shock," like a sudden change in interest rates), you want to know how the other gears move.

However, there's a big problem: The machine changed its design halfway through.

The Problem: A Shifting Puzzle

In the past, economists often assumed the machine's gears stayed the same forever. But in reality, the economy has "structural breaks." For example, the way the Federal Reserve (the US central bank) handles inflation changed drastically between the "Great Inflation" era (1970s) and the "Great Moderation" era (post-1980s).

If you try to solve the puzzle using data from both eras mixed together, you get a confusing mess. If you try to solve them separately, you don't have enough clues (data) to be sure what's happening. It's like trying to solve a Sudoku puzzle where half the numbers are missing, and the rules for filling them in changed halfway through the grid.

The Solution: The "SVAR-WB" (The Detective's New Toolkit)

The authors of this paper, Emanuele Bacchiocchi and Toru Kitagawa, propose a new way to solve this. They call their model SVAR-WB (SVAR with Breaks).

Here is how their new toolkit works, using simple analogies:

1. The "Stability" Clue

Imagine you are looking at two different versions of the same car: a 1970s model and a 2020s model. They look different, but maybe the engine block is the same in both.

• The Paper's Idea: Even if the economy changes, some rules might stay the same. Maybe the way a supply shock affects inflation is identical in both eras.
• The Benefit: By assuming some parts of the machine don't change, you can use the "stable" parts of the old machine to help you figure out the "changing" parts of the new machine. It's like using the known engine block to help you reverse-engineer the new transmission.

2. The "Multiple Truths" Problem

Here is the tricky part. Even with these clues, the math often doesn't give you just one answer. It gives you a handful of different answers that all fit the data perfectly.

• The Analogy: Imagine you are trying to guess a secret combination lock. You know the numbers must add up to 10.
  • Option A: 1, 2, 7
  • Option B: 3, 3, 4
  • Option C: 5, 5, 0
    All three combinations fit the rule "add to 10." In the past, economists would just pick one (usually the first one they found) and say, "This is the answer!"
• The Paper's Critique: The authors say, "Wait a minute! If there are three valid answers, picking just one is dangerous. You might be wrong."
• The New Approach: Instead of picking one, they calculate all the possible valid combinations. They create a "map" of every possible truth that fits the clues.

3. The "Sign" and "Rank" Filters

To narrow down the list of possible answers, the researchers add more rules, like "Sign Restrictions" and "Rank Restrictions."

• Sign Restrictions: "We know that if the Fed raises rates, inflation must go down, not up." (This is a common sense rule).
• Rank Restrictions: "We know the Fed reacts more strongly to inflation in the 2000s than in the 1970s."
• The Result: These rules act like a sieve. They filter out the weird, impossible answers, leaving you with a smaller, more reliable group of possibilities.

The Big Innovation: How to Report the Answer

This is the most important part of the paper.

• Old Way: "Here is the answer: The economy shrinks by 2%." (But this ignores the fact that it could also be 1% or 3%, and the old method just picked one).
• New Way: "Here is the range of all possible answers. Based on our rules, the economy could shrink by anywhere between 1% and 3%. Here is the probability of each scenario."

They developed new statistical methods (called Robust Bayesian Inference) that don't force you to pick a single "best guess." Instead, they show you the whole landscape of possibilities. This prevents researchers from making confident claims that might actually be wrong just because they happened to pick the wrong "truth" from the list.

The Real-World Test: The US Federal Reserve

The authors tested their new toolkit on the US economy, comparing the "Great Inflation" (chaos) with the "Great Moderation" (calm).

• What they found: The Federal Reserve changed its behavior. In the "Great Moderation," they became much more aggressive and persistent in fighting inflation.
• The Surprise: Because the Fed became so good at fighting inflation, when they did accidentally make a mistake (a monetary shock), the real economy (jobs and output) reacted more strongly than before.
• Why? Because people trusted the Fed more. If the Fed is known to react instantly to inflation, then any unexpected move they make is a huge surprise to the market, causing a bigger ripple effect.

Summary

This paper is like giving economists a better pair of glasses.

1. It acknowledges change: The economy changes rules over time.
2. It uses stability: It uses the parts that didn't change to help solve the parts that did.
3. It admits uncertainty: Instead of pretending there is only one answer, it maps out all the possible answers that fit the facts.
4. It filters wisely: It uses common-sense rules to narrow down the list of possibilities without forcing a single, potentially wrong, conclusion.

By doing this, they provide a much clearer, more honest picture of how economic policies actually work in a changing world.

Technical Summary

Here is a detailed technical summary of the paper "SVARs with breaks: Identification and inference" by Emanuele Bacchiocchi and Toru Kitagawa.

1. Problem Statement

The paper addresses three major challenges in macroeconomic empirical modeling:

1. Weak Identification: Limited variation in data often leads to weak identification of structural parameters.
2. Parameter Instability: Many macroeconomic relationships exhibit structural breaks (e.g., the "Great Moderation"), rendering standard constant-parameter models misspecified.
3. Information Scarcity: Economic theory often lacks sufficient information to uniquely identify parameters without imposing arbitrary restrictions.

While Structural Vector Autoregressions (SVARs) are standard tools for policy analysis, they face identification issues. Traditional approaches rely on zero restrictions (Cholesky) or sign restrictions. Recent literature uses heteroskedasticity (changes in shock variance) for identification. However, this paper argues that structural breaks in parameters (not just variances) offer a constructive source of identification, provided specific constraints are imposed across regimes. A critical gap exists in the literature regarding how to handle local identification (where multiple observationally equivalent solutions exist) and set identification (where a continuum of solutions exists) in the presence of structural breaks.

2. Methodology: SVAR-WB Framework

The authors propose a Structural Vector Autoregression with Breaks (SVAR-WB).

• Model Structure: The sample is divided into $s$ regimes separated by exogenously determined break dates ($TB_1, \dots, TB_{s-1}$).

  • Structural form: $A_{p0}y_t = a_p + \sum_{i=1}^l A_{pi}y_{t-i} + \varepsilon_t$ for regime $p$.
  • Reduced form: $y_t = b_p + \sum_{i=1}^l B_{pi}y_{t-i} + u_t$, where $u_t = A_{p0}^{-1}\varepsilon_t$.
  • The regimes exhibit heteroskedasticity because the error covariance matrices $\Sigma_p = A_{p0}^{-1}A_{p0}^{-1\prime}$ differ across regimes.

• Identification Strategy:

  • The core innovation is imposing constraints across regimes. These include:
    1. Stability Restrictions: Certain parameters or functions (e.g., impulse responses) remain constant across regimes.
    2. Equality Restrictions: Standard zero or linear restrictions on parameters.
    3. Inequality Restrictions: Sign restrictions, ranking restrictions (e.g., response in regime A > response in regime B), and Forecast Error Variance (FEV) restrictions (both within and across regimes).
  • The mapping from structural parameters to reduced-form parameters is analyzed to determine if the structural parameters are uniquely determined (global identification) or determined up to a finite set (local identification) or a set of positive measure (set identification).

3. Key Contributions

A. Identification Theory

• Local vs. Global Identification: The authors demonstrate that SVAR-WBs with cross-regime restrictions are generally locally identified rather than globally identified. This means there exists a finite set of isolated, observationally equivalent structural parameters that satisfy the restrictions.
• Rank Conditions:
  • They derive a necessary order condition: The total number of restrictions $f$ must be at least $s \times \frac{n(n-1)}{2}$.
  • They provide a sufficient rank condition (Theorem 2) and a necessary and sufficient rank condition (Theorem 5) based on the rank of a matrix constructed from the restrictions and structural parameters. These conditions allow researchers to check if a specific shock or the whole system is identified.
  • They show that if the rank condition holds at one point, it holds almost everywhere in the parameter space.

B. Estimation Algorithms

• Handling Local Identification: Standard Maximum Likelihood (ML) or Bayesian methods often pick a single solution or fail to explore all modes, leading to misleading inference.
• Algorithm 3: The authors propose an algorithm to solve the non-linear system of equations to find all admissible orthogonal matrices ($Q$) that map reduced-form parameters to structural parameters satisfying the equality and inequality constraints. This generates the full set of observationally equivalent solutions.

C. Inference Frameworks

The paper proposes three distinct approaches to handle the multiplicity of solutions:

1. Bayesian Inference: Uses a "constructive" approach. Instead of sampling the structural parameters directly (which may be multi-modal), it samples reduced-form parameters and then enumerates all admissible structural solutions for each draw. This avoids MCMC convergence issues associated with multi-modal posteriors.
2. Frequentist-Valid Inference: Constructs confidence sets by projecting the confidence set of the reduced-form parameters through the identified set mapping. This yields conservative but asymptotically valid confidence intervals that account for all admissible solutions.
3. Robust Bayesian Inference: Extends the multi-prior approach (Giacomini and Kitagawa, 2021) to SVAR-WBs. Since priors over observationally equivalent parameters are often arbitrary, this approach considers a class of priors to derive posterior lower and upper probabilities and robust credible regions. This is particularly useful for set-identified models where the identified set is not a single point.

D. Set Identification

The paper extends the framework to cases where restrictions are insufficient for point identification (set identification). It introduces novel cross-regime inequality restrictions (e.g., ranking the magnitude of impulse responses across different time periods), which are new to the SVAR literature.

4. Empirical Application

The authors apply their methodology to US monetary policy transmission, analyzing the Great Inflation (1954–1979) and Great Moderation (1979–2008) regimes.

• Data: Output gap, inflation, and Federal Funds rate.
• Models Tested:
  • Model I & II: Locally identified using equality (stability) and sign restrictions.
  • Model III, IV, V: Set-identified models relaxing equality constraints and adding cross-regime ranking and FEV restrictions.
• Key Findings:
  • Local Identification Issues: In locally identified models (Model I), the impulse responses are bi-modal. Standard inference focusing on one mode would be misleading.
  • Impact of Sign Restrictions: Adding mild sign restrictions (Model II) eliminates one of the modes, clarifying the results.
  • Cross-Regime Constraints: Models incorporating cross-regime ranking and FEV restrictions (Model V) significantly narrow the robust credible regions compared to standard sign-restricted models.
  • Policy Conclusions:
    • The Fed's reaction function changed significantly between regimes.
    • During the Great Moderation, the Fed responded more aggressively and persistently to price shocks.
    • Consequently, unexpected monetary policy shocks had a more pronounced recessionary effect on output during the Great Moderation compared to the Great Inflation. This contradicts some previous literature suggesting policy became less effective, arguing instead that the systematic response improved, making unexpected shocks more potent.

5. Significance

• Theoretical Advancement: The paper provides the first rigorous identification and inference framework specifically for SVARs with structural breaks that allows for cross-regime constraints.
• Methodological Innovation: It solves the "multi-modal" problem in local identification by proposing algorithms to find all solutions and inference methods that account for them, rather than ignoring them.
• Practical Utility: The introduction of cross-regime inequality restrictions (ranking and FEV) offers researchers new tools to identify models where traditional zero restrictions are too restrictive or where theory suggests relative differences between regimes.
• Robustness: The robust Bayesian approach provides a way to perform inference without relying on arbitrary prior choices over observationally equivalent parameters, a common critique in set-identified SVAR literature.

In summary, this paper transforms structural breaks from a nuisance into a source of identification, providing a comprehensive toolkit for estimation and inference in dynamic macroeconomic models subject to regime changes.