Locally- but not Globally-identified SVARs

This paper addresses the challenges of Structural Vector Autoregressions (SVARs) with locally but not globally identified parameters by characterizing the resulting observationally equivalent points and proposing novel algorithms for exhaustive parameter computation and robust Bayesian and frequentist inference on multi-modal impulse responses.

The Gist

Here is an explanation of the paper "Locally- but not Globally-identified SVARs" using simple language and creative analogies.

The Big Picture: The "Where's Waldo?" of Economics

Imagine you are an economist trying to understand how the economy works. You have a giant puzzle (the data) and you want to figure out the rules of the game (the structural model). Specifically, you want to know: "If the Central Bank raises interest rates, what happens to inflation and unemployment?"

To answer this, economists use a tool called a SVAR (Structural Vector Autoregression). Think of a SVAR as a machine that takes the messy, jumbled data of the economy and tries to separate it into distinct "shocks" (like a surprise interest rate hike, a supply chain disruption, or a change in consumer confidence).

The Problem:
Usually, this machine works perfectly. You feed it data, and it gives you one clear answer. This is called Global Identification. It's like looking at a fingerprint and knowing exactly whose hand it is.

The Twist:
Sometimes, the machine gets confused. It finds two (or more) different sets of rules that fit the data exactly the same way.

• Scenario A: The interest rate hike causes a small drop in unemployment.
• Scenario B: The exact same data could also mean the interest rate hike causes a massive drop in unemployment.

Both scenarios fit the math perfectly. The data cannot tell you which one is the "real" truth. This is called Local Identification. The machine has found the solution, but it's stuck in a "local" valley and doesn't know there's another valley just as deep nearby.

The Analogy: The Mystery Dinner Party

Imagine you are a detective at a dinner party. You know the ingredients used in a soup (the data), but you don't know the recipe (the structural parameters).

• Global Identification: You taste the soup, and there is only one recipe that could possibly make that flavor. You solve the case immediately.
• Local Identification (The Paper's Focus): You taste the soup, and you realize there are two different recipes that could create that exact same flavor.
  • Recipe 1: Lots of salt, no pepper.
  • Recipe 2: No salt, lots of pepper.

If you just pick one recipe at random (which is what many economists used to do), you might give the wrong advice to the chef. If you pick Recipe 1, you tell them to add more salt. If you pick Recipe 2, you tell them to add more pepper. Both are "correct" based on the taste, but they lead to very different future cooking instructions.

The Old Way vs. The New Way

The Old Way (The "Guess and Check" Method):
In the past, when economists found this "two recipes" problem, they would usually just pick one solution (the one their computer found first) and pretend it was the only truth.

• The Risk: If they picked the wrong recipe, their entire analysis of the economy would be wrong. It's like a GPS telling you to turn left when the only valid route is to turn right, but both routes look the same on the map.

The New Way (Bacchiocchi and Kitagawa's Solution):
This paper says, "Stop guessing! Let's find all the possible recipes."

The authors developed a new set of tools (algorithms) to:

1. Find all the solutions: Instead of stopping at the first answer, their computer method systematically hunts down every possible recipe that fits the data.
2. Show the whole picture: Instead of giving you one answer, they give you a "menu" of possibilities.
   • Bayesian Approach: They show you a probability map. "There is a 50% chance the effect is small, and a 50% chance it's huge." The graph looks like a mountain with two peaks (bimodal).
   • Frequentist Approach: They draw a wide safety net (confidence interval) that covers both possibilities, ensuring you never miss the truth, even if you don't know which one it is.

Why Does This Happen? (The Geometry of Confusion)

The paper explains that this confusion happens when the rules we impose on the model are "non-recursive" or "across-shock."

The Metaphor: The Tangled Necklace
Imagine you have a necklace with beads (the economic variables).

• Recursive (Global): You untangle the beads one by one. Bead A affects Bead B, which affects Bead C. It's a straight line. Easy to solve.
• Non-Recursive (Local): The beads are tangled in a knot. Bead A affects B, but B also affects A.
  • If you pull the knot tight in one direction, it looks like a specific shape.
  • If you pull it the exact opposite way, it looks like a mirror image of that shape.
  • The data (the shape of the knot) looks identical from both sides. You can't tell which way is "up" without an external clue.

The paper shows that when we have these "knots" (non-zero restrictions or restrictions across different shocks), we end up with these mirror-image solutions.

The Real-World Test: The Great Inflation vs. The Great Moderation

To prove their method works, the authors applied it to a real economic mystery: Monetary Policy Shocks.

They looked at data from the "Great Inflation" (volatile times) and the "Great Moderation" (calm times). They used a technique called Heteroskedastic SVAR (which uses changes in volatility to find the shocks).

• The Result: Their method found two distinct monetary policy shocks that both fit the data perfectly.
  • Shock A: A small, short-lived rate hike.
  • Shock B: A large, long-lasting rate hike.
• The Old Way: An economist would have picked one, ignored the other, and published a paper saying, "This is definitely what happened."
• The New Way: The authors said, "Both are possible. Here is the range of outcomes for both." They showed that regardless of which shock you pick, the conclusion is similar: Interest rates go up, and the economy slows down.

The Takeaway

This paper is a guidebook for economists who are stuck in a "fog of war."

1. Don't panic when your model has multiple answers. It's a common mathematical feature, not a mistake.
2. Don't pick one and ignore the rest. That leads to bad policy advice.
3. Use the new tools to find all the answers.
4. Report the uncertainty. Tell the world, "The data supports two different stories, and here is what happens in both."

In short, the paper teaches us that in economics, sometimes there isn't just one truth hidden in the data. There might be two, and the smartest thing to do is to acknowledge both.

Technical Summary

Here is a detailed technical summary of the paper "Locally- but not Globally-identified SVARs" by Emanuele Bacchiocchi and Toru Kitagawa.

1. Problem Statement

The paper addresses a critical gap in Structural Vector Autoregression (SVAR) analysis: models where structural parameters are locally identified but not globally identified.

• The Issue: In many practical SVAR settings (e.g., non-recursive zero restrictions, calibrated parameters, cross-equation restrictions, or heteroskedastic SVARs), the imposed identifying restrictions do not yield a unique structural solution. Instead, they result in a set of isolated, observationally equivalent parameter points.
• Consequences for Frequentists: Standard practice involves finding a single Maximum Likelihood Estimate (MLE). However, because the likelihood function is multi-modal with peaks of equal height, standard numerical optimization routines arbitrarily select one mode. This leads to impulse response functions (IRFs) that depend entirely on the initial values of the optimizer, potentially yielding contradictory economic conclusions.
• Consequences for Bayesians:
  1. Prior Sensitivity: Without global identification, the posterior distribution does not converge to a single point even asymptotically; it remains sensitive to the prior specification regarding which observationally equivalent point to select.
  2. Computational Failure: Standard Markov Chain Monte Carlo (MCMC) algorithms (e.g., Gibbs or Metropolis-Hastings) often fail to explore all modes of the posterior distribution, leading to inaccurate approximations and a failure to converge to the true posterior.

2. Methodology

The authors propose a comprehensive framework for estimation and inference that explicitly accounts for the discrete nature of the identified set.

A. Theoretical Characterization

• Identification Framework: The paper utilizes the orthogonal matrix parametrization (Uhlig, 2005; Rubio-Ramirez et al., 2010), where the structural matrix $A_0$ is decomposed as $A_0 = Q'\Sigma_{tr}^{-1}$, with $Q$ being an orthogonal matrix.
• Rank Conditions: The authors derive a necessary and sufficient condition for local identification (Proposition 2) based on the rank of a matrix involving the Jacobian of the restrictions and the orthogonal matrix $Q$. This condition is checkable via numerical rank tests.
• Counting Solutions: They establish upper bounds on the number of observationally equivalent solutions:
  • For triangular restrictions: At most $2^n$ solutions.
  • For non-triangular restrictions: At most $2^{n(n+1)/2}$ solutions.
  • Crucially, they show that under local identification, the set of admissible parameters consists of a finite number of isolated points, distinct from the continuous intervals found in standard set-identification literature.

B. Computational Algorithms

To overcome the "arbitrary selection" problem, the authors propose algorithms to exhaustively compute all admissible structural parameters given the reduced-form parameters ($\phi$):

1. Algorithm 1 (Standard SVAR): Solves the system of non-linear polynomial equations defined by the equality restrictions and orthogonality constraints ($Q'Q=I$). It uses numerical solvers (e.g., vpasolve in MATLAB) to find all roots, then filters them based on sign and normalization restrictions.
2. Algorithm 2 (Heteroskedastic SVAR - HSVAR): Addresses identification via volatility breaks. It solves an eigen-decomposition problem. Since eigenvalues can be permuted, the algorithm generates all valid permutations of eigenvectors that satisfy economic sign restrictions, identifying all admissible $A_0$ matrices.

C. Inference Procedures

The paper develops three distinct inference strategies:

1. Bayesian Inference (Multi-modal Sampling):

   • Instead of relying on standard MCMC which may get stuck in one mode, the authors propose a constructive sampling approach.
   • Step 1: Draw reduced-form parameters $\phi$ from their posterior.
   • Step 2: Use Algorithms 1 or 2 to compute the entire set of admissible orthogonal matrices $Q$ for that specific $\phi$.
   • Step 3: Draw one $Q$ from this discrete set according to prior weights.
   • Result: This ensures the posterior sampler stochastically moves across all modes, correctly capturing the multi-modal nature of the posterior distribution.

2. Frequentist-Valid Inference (Projection Approach):

   • The authors treat the set of observationally equivalent points as the identified set.
   • They project the confidence set of the reduced-form parameters ($\phi$) through the discrete mapping to the impulse responses.
   • Labeling Challenge: A unique challenge is labeling equivalent points consistently across different draws of $\phi$. The paper proposes two solutions:
     • Switching-Label: Allows labels to change across horizons to capture marginal multi-modality.
     • Fixed-Label: Anchors labels to a specific impulse response to track dependence across horizons and variables.
   • This yields asymptotically valid confidence intervals that cover all admissible solutions.

3. Robust Bayesian Inference:

   • The frequentist projection approach is interpreted as a Robust Bayesian procedure. It performs global sensitivity analysis over a class of priors that assign arbitrary weights to the observationally equivalent points, providing bounds on posterior probabilities and means without committing to a specific prior weight.

3. Key Results

• Geometric Insight: The paper provides a geometric exposition showing how non-homogeneous restrictions (e.g., $c \neq 0$) create multiple intersection points between the restriction plane and the unit sphere (orthogonality constraint), leading to multiple solutions.
• Empirical Application (Monetary Policy):
  • The authors apply their method to a Heteroskedastic SVAR analyzing monetary policy shocks using data from 1954–2007 (covering the Great Inflation and Great Moderation).
  • Finding: The model is locally identified with two observationally equivalent structural shocks that both satisfy the theoretical sign restrictions (positive response of interest rates to a contractionary shock).
  • Implication: Standard methods would arbitrarily pick one, potentially missing the other. The proposed method reveals a bimodal posterior distribution for the impulse responses.
  • Conclusion: Despite the lack of global identification, the robust inference (both Bayesian and Frequentist) confirms that monetary policy shocks have a recessionary effect on output and unemployment, and a positive effect on credit spreads, consistent with economic theory. The uncertainty is correctly quantified by the wide, multi-modal confidence/credible regions.

4. Significance and Contributions

1. Solving the Multi-modal Problem: The paper provides the first systematic computational solution to the problem of local identification in SVARs, moving beyond the "pick one" heuristic to "find all."
2. Bridging Frequentist and Bayesian Paradigms: It unifies inference by showing how projection-based frequentist methods can be interpreted as robust Bayesian procedures, offering a way to perform inference without requiring a unique prior over observationally equivalent points.
3. Applicability to DSGE and Heteroskedastic Models: The framework is highly relevant for DSGE models (which often imply non-recursive restrictions) and modern heteroskedastic SVARs, areas where global identification is frequently unattainable.
4. Computational Feasibility: By leveraging polynomial root-finding and eigen-decomposition, the authors make it computationally feasible to enumerate all solutions in moderate-dimensional systems, a task previously considered intractable for empirical researchers.

In summary, this paper fundamentally changes how econometricians should handle SVARs with local identification: rather than ignoring the multiplicity of solutions, researchers should explicitly compute the full set of admissible parameters and report inference that reflects this inherent ambiguity.