On global identification in structural vector autoregressions

This paper refutes the sufficiency of the rank condition proposed by Rubio-Ramirez, Waggoner, and Zha (2010) for global identification in structural vector autoregressions by demonstrating that it fails to account for redundant restrictions, and subsequently derives a corrected necessary and sufficient condition.

The Gist

Here is an explanation of the paper using simple language, analogies, and metaphors.

The Big Picture: Solving a Puzzle with Missing Pieces

Imagine you are an economist trying to solve a giant, complex jigsaw puzzle. This puzzle represents the economy. You have the picture on the box (the data you can actually see, like unemployment rates and inflation), but you need to figure out how the individual pieces fit together to create that picture.

In the world of economics, this puzzle is called a Structural Vector Autoregression (SVAR). The goal is to separate the "noise" of the economy from the specific "shocks" that cause changes (like a sudden oil price spike or a change in interest rates).

For a long time, economists relied on a famous rulebook written by Rubio-Ramírez, Waggoner, and Zha (let's call them RWZ). Their rulebook, specifically Theorem 7, was like a "magic checklist." It said: "If you count your zero-restrictions (the rules you set about which pieces don't touch), and the number matches a specific formula, then your puzzle is solved uniquely. You have found the one true answer."

This checklist was popular because it was easy. You just counted the rules. No complex math required.

The Problem: The "Redundant Rule" Trap

The authors of this new paper, Bacchiocchi and Kitagawa, discovered a flaw in the magic checklist. They found that sometimes, the checklist gives you a "Green Light" (saying the puzzle is solved) when the puzzle is actually broken.

The Analogy: The Over-Strict Teacher

Imagine you are trying to guess a secret 3-digit code.

1. Rule A: The first digit must be even.
2. Rule B: The second digit must be odd.
3. Rule C: The third digit must be... well, let's say Rule A and Rule B automatically force the third digit to be a specific number, even though you didn't write a rule for it.

Now, imagine you write a fourth rule: Rule D: "The third digit must be 5."

If you just count the rules, you have four rules. The RWZ checklist sees four rules and says, "Great! That's enough to solve the code!"

But here is the catch: Rule D is redundant. It doesn't actually help you solve the puzzle because the first three rules already forced the third digit to be 5. Rule D is just repeating information you already have. Because it's a "ghost rule" (it adds no new information), you haven't actually narrowed down the possibilities enough. You might still have two different codes that fit all the rules, but the checklist thinks you only have one.

In the paper, the authors show a specific economic example where this happens. They impose rules on how variables relate to each other now, and rules on how they react later. It turns out the "now" rules mathematically force the "later" rules to be zero. So, when they try to count the "later" rules as extra help, they are actually counting a rule that was already implied. The checklist gets fooled.

The Solution: A Better Detective

The authors propose a new way to check if the puzzle is truly solved. Instead of just counting the rules, you have to test them.

The New Method: The "Unique Path" Test

Think of the solution process as walking down a hallway with many doors.

• The Old Way (RWZ Theorem 7): You count the doors. If you have enough doors closed off, you assume there is only one path left.
• The New Way (Bacchiocchi & Kitagawa): You actually try to walk through the hallway.
  1. You close the first set of doors based on your rules. Is there only one path left? Good.
  2. You move to the next step. You close the next set of doors. Crucially, you check if the doors you just closed were actually necessary. Did they block a path that was already blocked by the previous doors?

If you find that a new door you closed was actually redundant (it didn't block any new paths), then your hallway still has multiple exits. The puzzle is not solved.

Why This Matters

1. It's Not Just Theory: The authors show that this isn't just a weird math trick. Real-world economists often mix rules about "immediate effects" and "long-term effects." This combination frequently creates these "redundant rules" without the economist realizing it.
2. The Risk: If you use the old checklist, you might publish a paper claiming you have found the "true" effect of a policy, when in reality, your model is ambiguous. You might be drawing conclusions from a puzzle that has multiple solutions.
3. The Fix: The authors provide a simple computer algorithm (a new checklist) that doesn't just count the rules. It checks the rank (the mathematical strength) of the rules at every single step. It asks: "Does this specific rule actually narrow down the possibilities, or is it just echoing what we already know?"

Summary

• The Old Rule: Count your restrictions. If the number is right, you are safe.
• The Flaw: Sometimes, restrictions are "echoes" of each other. Counting them makes you think you have more information than you do.
• The New Rule: Don't just count. Check if every single restriction is doing unique work. If one is a "ghost" (redundant), the whole identification fails.

The paper essentially tells economists: "Stop just counting your rules. Make sure they aren't repeating themselves, or you might be solving the wrong puzzle."

Technical Summary

Here is a detailed technical summary of the paper "On global identification in structural vector autoregressions" by Bacchiocchi and Kitagawa.

1. Problem Statement

The paper addresses a critical gap in the identification theory of Structural Vector Autoregressions (SVARs). Specifically, it challenges the universal applicability of Theorem 7 from Rubio-Ramírez, Waggoner, and Zha (2010) (henceforth RWZ).

• The Context: RWZ (2010) provided necessary and sufficient conditions for the global identification of SVAR structural parameters under general zero restrictions. Their Theorem 7 is widely used by practitioners because it simplifies the identification check to a "counting exercise" (verifying that the number of restrictions in column $j$ equals $n-j$).
• The Flaw: RWZ's Theorem 7 relies on strict technical assumptions, specifically Condition 2 (Regularity), which requires the first derivative of the transformation function $f(\cdot)$ (mapping structural parameters to restricted objects) to have full rank.
• The Issue: The authors demonstrate that when Condition 2 fails, Theorem 7 can yield false positives. A model may satisfy the counting rule ($q_j = n-j$) but still fail to be globally identified due to redundant restrictions. In such cases, one or more imposed zero restrictions are implicitly implied by other restrictions, meaning they do not provide additional identifying information, leading to a rank deficiency in the identification system.

2. Methodology

The authors employ a combination of analytical counterexamples, rank condition analysis, and algorithmic modification to address the problem.

A. Counterexample Construction

The authors construct a trivariate SVAR example involving:

1. Contemporaneous restrictions on the structural matrix $A_0$ (specifically, $a_{21}=0$ and $a_{31}=0$).
2. Impulse response restrictions on the contemporaneous impulse response matrix $IR_0 = (A_0^{-1})'$ (specifically, $(IR_0)_{1,2}=0$).

They analytically derive the inverse of $A_0$ to show that the restrictions on $A_0$ mathematically imply the restriction on $IR_0$. Consequently, the restriction on $IR_0$ is redundant.

• Result: While the counting rule of RWZ's Theorem 7 suggests the model is identified ($q_1=2, q_2=1, q_3=0$), the system fails to pin down a unique orthogonal matrix $P$ because the matrix used to solve for the second column vector ($\tilde{Q}_2$) is rank-deficient.

B. Analytical Investigation

The authors investigate why RWZ's Theorem 7 fails in this scenario:

• The transformation function $f(A_0, A_+)$ in the example is not regular (the derivative does not have full rank) because the domain of the transformation (structural parameters) has a lower dimension than the codomain (the space of restricted matrices).
• They show that RWZ's Theorem 6 (which checks the rank of a matrix $M_j$ constructed from the restrictions) would correctly detect the failure if the analyst analytically understood that the restrictions implied a zero in the $(1,3)$ position of the impulse response matrix. However, for large-scale SVARs, manually deriving these implications is often intractable.

C. Proposed Solution: Modified Condition and Algorithm

To resolve this without requiring complex analytical derivations, the authors propose a modified framework:

1. Definition of Non-Redundancy: They define restrictions as "non-redundant" if the previously identified orthogonal vectors ($p_1, \dots, p_{j-1}$) are linearly independent of the row vectors of the current restriction matrix $Q_j f(A_0, A_+)$.
2. Theorem 1 (New Condition): They establish a new necessary and sufficient condition for exact identification:
   • The counting rule must hold ($q_j = n-j$).
   • AND the restrictions must be non-redundant at the specific reduced-form parameter point $(B, \Sigma)$.
3. Algorithm 1: They propose a computational algorithm that:
   • Takes reduced-form parameters $(B, \Sigma)$ as input.
   • Constructs the Cholesky factor to generate a candidate structural parameter set.
   • Sequentially checks the rank of matrices $\tilde{Q}_j$ (which stack the current restrictions and previously identified vectors).
   • Verifies if $\text{rank}(\tilde{Q}_j) = n-1$ for all $j$. If the rank is deficient, global identification fails.

3. Key Contributions

1. Identification of a Critical Failure Mode: The paper proves that the "counting rule" in RWZ's Theorem 7 is insufficient when the transformation function is not regular. It highlights that redundant restrictions (where one restriction is a linear combination of others) can lead to a false conclusion of global identification.
2. Generalization of Identification Theory: The authors extend the identification conditions to a broader class of transformation functions that do not satisfy the strict "regularity" (full rank derivative) assumption required by RWZ.
3. Practical Algorithm: They provide Algorithm 1, a computationally feasible procedure that:
   • Does not require the user to analytically derive implied restrictions (which is difficult for large SVARs).
   • Works directly with estimated reduced-form parameters (e.g., MLE or Bayesian posterior draws).
   • Explicitly checks for the rank deficiency caused by redundancy.
4. Clarification of RWZ's Theorem 6: The paper clarifies that while RWZ's Theorem 6 is theoretically robust, its practical application requires the analyst to know the full set of implied constraints, which is often impossible in complex models. The proposed algorithm automates this check.

4. Results

• Theoretical Result: The condition $q_j = n-j$ is necessary but not sufficient for global identification if the transformation is not regular. The sufficiency holds only if the restrictions are non-redundant.
• Empirical Implication: In the provided trivariate example, standard software implementing RWZ's Algorithm 1 (using QR decomposition) would return a unique solution, misleading the researcher. The authors' modified algorithm correctly identifies the rank deficiency in $\tilde{Q}_2$ and flags the model as unidentified.
• Robustness: The proposed method works for models involving short-run, long-run, and cumulated impulse response restrictions, even when the lag structure of the VAR violates the regularity conditions of RWZ (e.g., when the number of restrictions exceeds the degrees of freedom in the derivative).

5. Significance

This paper is significant for both theoretical econometrics and applied macroeconomics:

• Prevention of Spurious Inference: It prevents researchers from drawing incorrect conclusions about structural shocks and impulse responses in SVAR models where restrictions are redundant.
• Accessibility: It democratizes the check for global identification. Previously, ensuring identification in complex models required deep analytical insight into how restrictions propagate. The new algorithm allows practitioners to verify identification numerically at any point in the parameter space.
• Broad Applicability: The method applies to a wide range of SVAR specifications, including those with narrative restrictions, set-identified SVARs, and models mixing short-run and long-run constraints, making it a vital tool for modern macroeconomic modeling.

In summary, Bacchiocchi and Kitagawa refine the identification theory of SVARs by replacing a purely algebraic counting rule with a rank-based verification that accounts for the geometric redundancy of restrictions, ensuring that global identification claims are robust and reliable.