Recovering Structural Parameters From Reduced Forms

From Reduced-Form Coefficients Back to Structural Parameters

A structural model describes economic relationships using parameters with a behavioral or technological interpretation. Solving its equations may express observable outcomes as functions of exogenous variables and disturbances. The coefficients in those solved equations are reduced-form coefficients. They summarize observable implications of the structural model, but they do not automatically reveal every structural parameter.

Recovering structural parameters therefore requires more than solving the model forward. One must invert the mapping from structural parameters to reduced-form coefficients, check that the proposed inverse solution lies in the model’s admissible parameter space, and determine whether any other admissible solution produces the same coefficients. The distinction is important: an algebraic formula may produce a candidate parameter vector without establishing that the vector is admissible or uniquely determined.

Definition: Let $\theta$ be a vector of structural parameters in an admissible parameter space $\Theta$, and let $\rho=g(\theta)$ be the vector of population reduced-form coefficients implied by the model. For a given $\rho$, the set of structural parameter values consistent with it is $$ \Theta(\rho)=\{\theta\in\Theta:g(\theta)=\rho\}. $$ The parameters are recoverable from $\rho$ at that value if $\Theta(\rho)$ contains exactly one element. If it is empty, the reduced-form vector is incompatible with the specified model and parameter space; if it contains multiple elements, the reduced form does not uniquely recover the structural parameters.

The mapping $g$ is defined by the structural equations, the restrictions imposed on their parameters, and the solution used to express outcomes in reduced form. Recovery is a property of this mapping over the stated parameter space. It is not a property of a coefficient vector considered without its model.

Existence, Admissibility, and Uniqueness

Three checks should be kept separate. Existence asks whether any structural parameter value maps to the given reduced-form vector. Admissibility asks whether a candidate solution satisfies all parameter restrictions and conditions needed for the model’s reduced form to be defined. Uniqueness asks whether that admissible solution is the only one.

For example, a structural system may have a reduced-form denominator that must be nonzero. An algebraic inverse that divides by a reduced-form coefficient also requires that coefficient to be nonzero. Parameter sign restrictions, parameter bounds, or a normalization may further rule out some mathematical roots. A solution to the equations $g(\theta)=\rho$ is not a valid recovery unless it passes these checks.

A one-to-one mapping on the admissible parameter space guarantees that every reduced-form vector in its image corresponds to at most one structural parameter vector. The image matters: a reduced-form vector outside $g(\Theta)$ has no admissible structural explanation under the specified model, even if a formal inverse expression can be evaluated at it.

Worked Example: A Squared Parameter and a Sign Restriction

Suppose a structural parameter $\theta$ generates the reduced-form coefficient $\rho=\theta^2$. Consider first the unrestricted parameter space $\Theta=\mathbb{R}$, and suppose the population reduced-form coefficient is $\rho=9$. Recover the possible structural values.

The inverse equation is $$ \theta^2=9. $$ Taking both square roots gives $\theta=3$ or $\theta=-3$. Both values are admissible in $\mathbb{R}$, and both generate $\rho=9$. Thus $\Theta(9)=\{-3,3\}$: a reduced-form coefficient exists, but it does not uniquely recover the structural parameter.

Now impose the restriction $\Theta=[0,\infty)$. The negative root is inadmissible, so $\Theta(9)=\{3\}$. Under this restriction, $\theta$ is uniquely recovered as $3$. The restriction has not changed the reduced-form mapping; it has changed the set of structural values allowed to explain the observed coefficient.

The same restriction also affects existence: if $\rho=-1$, there is no real $\theta$ satisfying $\theta^2=-1$, so the reduced-form value is outside the image of the model under either parameter space.

Inverting a Structural System

A common route to reduced forms is to solve a system of simultaneous structural equations. Consider the following two-equation model:

\begin{aligned} y_1 &= \beta y_2+\gamma_1 x+u_1,\\ y_2 &= \delta y_1+\gamma_2 x+u_2. \end{aligned}

Here $y_1$ and $y_2$ are jointly determined outcomes, $x$ is an exogenous variable, $u_1$ and $u_2$ are structural disturbances, and $\beta,\delta,\gamma_1,\gamma_2$ are structural parameters. In matrix form, the coefficient matrix on the outcomes is

\begin{pmatrix} 1 & -\beta\\ -\delta & 1 \end{pmatrix} \begin{pmatrix} y_1\\ y_2 \end{pmatrix} = \begin{pmatrix} \gamma_1x+u_1\\ \gamma_2x+u_2 \end{pmatrix}.

The system has a unique solution for the outcomes when its coefficient matrix is invertible. Its determinant is $1-\beta\delta$, so the required condition is $1-\beta\delta\ne 0$. Under this condition, solving the system yields reduced-form coefficients on $x$:

\pi_1=\frac{\gamma_1+\beta\gamma_2}{1-\beta\delta}, \qquad \pi_2=\frac{\delta\gamma_1+\gamma_2}{1-\beta\delta}.

These formulas define a mapping from the structural parameters to the pair $(\pi_1,\pi_2)$. Whether that pair can recover the structural parameters depends on which parameters are known or restricted. Two reduced-form coefficients alone cannot generally recover four unrestricted structural parameters.

Worked Example: Recovering Two Parameters From a Restricted Mapping

Suppose the structural model above imposes $\gamma_1=1$ and $\gamma_2=0$, while $\beta$ and $\delta$ are unknown. Assume $1-\beta\delta\ne0$. The reduced-form mapping becomes $$ \pi_1=\frac{1}{1-\beta\delta}, \qquad \pi_2=\frac{\delta}{1-\beta\delta}. $$ $$ $$ Suppose the population reduced-form coefficients are $\pi_1=8/7$ and $\pi_2=2/7$. Recover $\beta$ and $\delta$.

First divide the second coefficient by the first. Since $\pi_1\ne0$, $$ \delta=\frac{\pi_2}{\pi_1} =\frac{2/7}{8/7} =\frac14. $$ Next, rearrange the first mapping: $$ \pi_1=\frac{1}{1-\beta\delta} \quad\Longrightarrow\quad \beta=\frac{1-1/\pi_1}{\delta}. $$ Substituting the given values, $$ \beta=\frac{1-7/8}{1/4} =\frac{1/8}{1/4} =\frac12. $$

The recovered values satisfy the required system condition: $$ 1-\beta\delta=1-\frac12\cdot\frac14=\frac78\ne0. $$ They also reproduce both coefficients: $1/(7/8)=8/7$, and $(1/4)/(7/8)=2/7$. To check uniqueness, any solution must have $\delta=\pi_2/\pi_1=1/4$; with this value, the first equation requires $\beta=1/2$. Thus this restricted mapping has exactly one admissible solution for the stated reduced-form vector.

A Direct Justification of the Inverse

The preceding recovery can be verified without guessing the answer. With $\gamma_1=1$ and $\gamma_2=0$, the reduced-form equations imply

\frac{\pi_2}{\pi_1} = \frac{\delta/(1-\beta\delta)}{1/(1-\beta\delta)} =\delta,

where the cancellation is valid because $1-\beta\delta\ne0$, and the ratio is defined because $\pi_1\ne0$. Thus $\delta$ is forced to equal $\pi_2/\pi_1$. Also, the first equation gives $1-\beta\delta=1/\pi_1$, so $\beta\delta=1-1/\pi_1$. If $\delta\ne0$, division yields $\beta=(1-1/\pi_1)/\delta$. Each parameter is therefore forced by the reduced-form coefficients, which establishes uniqueness wherever these inverse operations are defined. The values still need to satisfy the model’s admissibility conditions.

This argument also makes the limits of the inverse visible. If $\pi_1=0$, the ratio used to recover $\delta$ is undefined; moreover, the restricted mapping has $\pi_1=1/(1-\beta\delta)$, which cannot equal zero for any finite admissible $\beta,\delta$. If $\pi_2=0$, the ratio gives $\delta=0$, but then $\pi_1=1$, independent of $\beta$. In that case $\beta$ is not recovered: different admissible values of $\beta$ produce the same reduced-form coefficients. Invertibility of the outcome system alone therefore does not guarantee recovery of every structural parameter.

Information Lost in the Mapping

A reduced-form mapping can preserve some features of the structural parameters while discarding others. A classic algebraic pattern is to map two labeled parameters to their sum and product. Let

\rho_1=\theta_1+\theta_2, \qquad \rho_2=\theta_1\theta_2.

The reduced-form pair determines the polynomial whose roots are the two structural values, but it need not determine which root belongs to which label. This is a substantive loss if $\theta_1$ and $\theta_2$ have distinct structural interpretations.

Worked Example: Sum and Product Do Not Preserve Labels

Suppose $\Theta=\mathbb{R}^2$ and the reduced-form coefficients are $\rho_1=4$ and $\rho_2=3$. Find the structural parameter pairs that map to these coefficients.

The parameters must have sum 4 and product 3. They are the roots of $$ t^2-\rho_1t+\rho_2=0, \qquad\text{so}\qquad t^2-4t+3=0. $$ Factoring gives $(t-1)(t-3)=0$. Therefore the two parameter values are 1 and 3, and the admissible labeled pairs are $(\theta_1,\theta_2)=(1,3)$ and $(3,1)$. Both yield $\rho_1=4$ and $\rho_2=3$, so the mapping does not uniquely recover the labeled pair.

If the admissible space instead imposes $\theta_1<\theta_2$, only $(1,3)$ remains. That ordering restriction resolves the ambiguity for this coefficient pair. Without it, the reduced form identifies the unordered set of values $\{1,3\}$, not which parameter has which value.

Common Pitfalls and Practical Checks

A frequent mistake is to find one algebraic solution and call the parameters identified. For instance, taking only the positive square root of $\rho=9$ assumes a nonnegative parameter space that may not be part of the model. The correction is to solve for the full inverse image over the stated admissible space, not just to select a convenient root.

A second mistake is to confuse solving the structural equations for outcomes with recovering the structural parameters. A nonzero determinant ensures a unique outcome solution for each admissible parameter vector; it does not ensure that distinct parameter vectors produce distinct reduced-form coefficients. The two-equation example showed that some parameters may remain unrecovered even when the outcome system is solvable.

Finally, a finite-sample estimate of a reduced-form coefficient vector is not necessarily an exact population reduced-form vector in the model’s image. Plugging an estimate into an inverse formula may produce a candidate estimate, but it does not by itself establish uniqueness, admissibility, or the sampling uncertainty of that estimate. The recovery exercise here concerns the structural implications of a specified population reduced-form mapping.

Key takeaway: To recover structural parameters, specify the reduced-form map and admissible parameter space, solve $g(\theta)=\rho$, and inspect the entire set of solutions. Confirm that at least one solution exists, every retained solution is admissible, and exactly one remains if unique recovery is claimed.

Check Your Understanding

For each question, consider both the inverse equations and the admissible parameter space.

If $\rho=\theta^2$ and $\theta\in\mathbb{R}$, how many admissible values of $\theta$ generate $\rho=16$? How does the answer change if $\theta\ge0$?
In the restricted two-equation example, why must $\pi_1\ne0$ to use the ratio $\pi_2/\pi_1$?
For the sum-and-product mapping, what structural information is recovered from $(\rho_1,\rho_2)$ even when the labeled parameter pair is not unique?
Explain why a nonzero determinant of the structural outcome system does not by itself establish unique recovery from reduced-form coefficients.
What does an empty set $\Theta(\rho)$ say about the reduced-form vector and the specified model?