Functional-Form Identification And Its Fragility

Why the Source of Identification Matters

A structural parameter is identified when the observable implications of the model determine it uniquely. That uniqueness can come from economic restrictions, such as an excluded variable that shifts one equation but not another, or from the functional form imposed on the equations. The distinction matters: a parameter recovered from a credible economic restriction has a different foundation from one recovered only because the model is assumed to have a particular curvature.

Functional-form identification is not automatically invalid. A genuinely nonlinear economic relationship can generate variation that separates parameters even when the same exogenous variables appear in multiple equations. But the conclusion then depends on the chosen nonlinear specification, the range of observed variation, and the assumed equilibrium relationship. If those ingredients are weakly supported, a unique algebraic answer may still be fragile.

Definition: Let \(\theta\) denote structural parameters and let \(M(\theta)\) denote the model's observable implications, such as its reduced-form coefficients or population moments. The parameters are identified on a parameter space \(\Theta\) if \(M(\theta_1)=M(\theta_2)\), for \(\theta_1,\theta_2\in\Theta\), implies \(\theta_1=\theta_2\). Identification supplied by functional form occurs when uniqueness depends on restrictions on the mathematical shape of the structural equations, rather than on an excluded source of variation.

This is a population property, not a claim that a finite sample will estimate the parameters precisely. Identification asks whether distinct parameter values can generate the same observable implications under the stated model. Estimation asks how to use data to learn those implications and parameters; the identification question comes first.

Exclusion Restrictions as a Contrast

An exclusion restriction supplies a difference between structural equations: a variable affects one equation but is ruled out of another. This can create identifying variation without requiring the equations to have different nonlinear shapes. The following small calculation is only a contrast with the nonlinear examples that follow.

Worked Example: A Structural Slope Identified by Excluded Variation

Suppose a structural outcome equation is \(y=\beta x+u\), where \(x\) is endogenous, and an observed variable \(z\) is excluded from this equation. Assume \(z\) is relevant for \(x\), so \(\operatorname{Cov}(z,x)\ne 0\), and satisfies the exclusion condition \(\operatorname{Cov}(z,u)=0\). In a hypothetical population, suppose \(\operatorname{Cov}(z,x)=2.00\) units and \(\operatorname{Cov}(z,y)=6.00\) outcome-units per unit of \(z\).

Taking covariance with \(z\) in the structural equation gives \(\operatorname{Cov}(z,y)=\beta\operatorname{Cov}(z,x)+\operatorname{Cov}(z,u)\). Under the stated exclusion condition, \[ \beta=\frac{\operatorname{Cov}(z,y)}{\operatorname{Cov}(z,x)} =\frac{6.00}{2.00}=3.00. \]

The slope is identified here because \(z\) moves \(x\) while its direct structural contribution to \(y\) is excluded. The calculation depends on relevance and the exclusion condition; it does not depend on identifying \(\beta\) from a nonlinear curve. A functional-form argument has a different burden: the assumed shape itself must generate enough distinct observable variation to separate the parameters.

How Nonlinear Shape Can Separate Parameters

Consider a simplified structural system in which a common exogenous variable \(z\) enters both equations:

y_1=\alpha z+\beta y_2, \qquad y_2=\gamma z+\delta y_1^2.

There is no excluded \(z\): it appears in both equations. The nonlinear term \(y_1^2\), however, can make the observable equilibrium relationships curve with \(z\). If that curvature gives the structural regressors distinct variation, it may permit separate recovery of the coefficients.

For a transparent algebraic example, treat these as exact population equilibrium relationships over a range of \(z\), with the displayed equilibrium branch selected. This illustrates identification, not a general estimation recipe. In a stochastic application, one would also need appropriate assumptions connecting the structural equations to observables, including assumptions about disturbances and equilibrium selection.

Worked Example: Recovering Coefficients From Curvature

Suppose the population relationships are generated by \(\alpha=1\), \(\beta=1\), \(\gamma=1\), and \(\delta=0.5\), with the equilibrium branch satisfying \(y_1=2z+0.5y_1^2\) and \(y_2=y_1-z\). At \(z=0.10\), this branch gives \(y_1=1-\sqrt{0.60}\approx0.2254\) and \(y_2\approx0.1254\). At \(z=0.20\), it gives \(y_1=1-\sqrt{0.20}\approx0.5528\) and \(y_2\approx0.3528\).

In the first equation, the two observations imply \(y_1/z=\alpha+\beta(y_2/z)\). The ratios are approximately \(2.2540= \alpha+1.2540\beta\) and \(2.7639=\alpha+1.7639\beta\). Subtracting gives \(0.5099=0.5099\beta\), so \(\beta=1.00\), and substitution gives \(\alpha=1.00\).

For the second equation, divide by \(z\): \(y_2/z=\gamma+\delta(y_1^2/z)\). At the two points, \(y_2/z\) is approximately \(1.2540\) and \(1.7639\), while \(y_1^2/z\) is approximately \(0.5081\) and \(1.5279\). Therefore, \[ \delta=\frac{1.7639-1.2540}{1.5279-0.5081} \approx 0.5000, \qquad \gamma=1.2540-(0.5000)(0.5081)\approx1.000. \]

The nonlinear shape makes the relevant ratios vary across values of \(z\), allowing the coefficients in each equation to be separated. Both equations contain \(z\), so this example does not use an exclusion restriction. Its identification comes from the assumed functional form and the resulting curvature in the equilibrium relationships.

The Rank Intuition Behind the Example

The algebraic requirement in the example can be stated directly. Within the first equation, the coefficients \(\alpha\) and \(\beta\) are uniquely determined if the observed values of \(z\) and \(y_2\) are not proportional across the support. Otherwise, different coefficient pairs can give the same values of \(y_1\). Within the second equation, \(\gamma\) and \(\delta\) are uniquely determined if \(z\) and \(y_1^2\) are not proportional. These are full-column-rank conditions for the two structural regressor arrays.

The functional-form contribution is that the squared term can break proportionality that would remain in a linear system. If both structural equations were linear in \(z\) and the equilibrium outcomes, their reduced-form variation might only reveal combinations of structural coefficients. A nonlinear term can generate an additional pattern, but only when that pattern actually varies independently over the observed support.

Identification check: Functional-form identification requires more than writing a nonlinear equation. The observable implications must distinguish parameter values. In particular, the nonlinear term must vary in a way that cannot be replicated by the other terms over the relevant support, and the assumed structural relationship and equilibrium branch must be appropriate.

A short derivation makes the point in the example's second equation. At two nonzero values \(z_1,z_2\), write \(r_i=y_{2i}/z_i\) and \(q_i=y_{1i}^2/z_i\). Then \(r_i=\gamma+\delta q_i\). Subtracting the two equations eliminates \(\gamma\):

r_2-r_1=\delta(q_2-q_1).

If \(q_2\ne q_1\), then \(\delta=(r_2-r_1)/(q_2-q_1)\), and either original equation gives \(\gamma=r_1-\delta q_1\). If \(q_2=q_1\), the two observations provide no separate information about \(\delta\); they identify only the combination \(\gamma+\delta q_1\). Thus curvature identifies the coefficients only insofar as it creates variation in \(q_i\).

Why Functional-Form Identification Can Be Fragile

Uniqueness can depend on small amounts of curvature. If the nonlinear term is nearly proportional to another regressor over the observed range, the equations may technically have a unique solution while small data changes produce large changes in the recovered coefficients. If the nonlinear term is omitted or its shape is wrong, the variation attributed to curvature may instead reflect other features of the data. Identification under a maintained model is not proof that the maintained model is a good economic description.

Worked Example: Small Outcome Changes, Large Curvature Changes

Return to the second equation, \(y_2=\gamma z+\delta y_1^2\), but observe the system only near zero. At \(z_1=0.01\), the specified equilibrium branch gives \(y_1\approx0.02020\) and \(y_2\approx0.01020\). At \(z_2=0.02\), it gives \(y_1\approx0.04083\) and \(y_2\approx0.02083\). The corresponding ratios are approximately \(q_1=y_1^2/z_1=0.04082\), \(q_2=0.08337\), \(r_1=y_2/z_1=1.0204\), and \(r_2=1.0417\).

Using the two-point formula gives \[ \delta=\frac{1.0417-1.0204}{0.08337-0.04082} \approx 0.500, \qquad \gamma=1.0204-(0.500)(0.04082)\approx1.000. \]

Now suppose the recorded \(y_2\) at \(z_2\) is higher by just \(0.001\) outcome-units, while \(y_1\) and \(z\) are unchanged. The second ratio becomes \(r_2=1.0417+0.001/0.02=1.0917\). The recovered coefficients become \[ \delta\approx\frac{1.0917-1.0204}{0.08337-0.04082} \approx1.68, \qquad \gamma\approx1.0204-(1.68)(0.04082)\approx0.952. \]

The point estimates remain algebraically unique, but a small outcome change has more than tripled the inferred curvature coefficient. The denominator \(q_2-q_1\) is small because the observed range offers limited nonlinear variation. This calculation diagnoses fragility: it does not establish that the model is false, but it shows that the structural conclusion is highly dependent on precise outcomes and the assumed quadratic form.

What to Check Before Interpreting a Structural Result

When a structural coefficient is identified mainly by functional form, make the source of uniqueness explicit. Ask which nonlinear terms separate the parameters, whether those terms vary sufficiently over the observed support, and whether the same observable implications could arise from a plausible alternative shape or equilibrium branch. A parameter may be identified under the maintained specification yet remain economically unpersuasive if its identification rests on a shape restriction with little substantive support.

These checks do not replace the model's economic reasoning. Economic theory can motivate curvature, but the functional form still needs to be stated and its identifying role demonstrated. Likewise, an exclusion restriction is not automatically credible merely because it gives a convenient source of variation. The two sources of identification have different assumptions, and each assumption needs substantive justification.

Common Mistake

Wrong reasoning: “The nonlinear system has a unique algebraic solution, so the structural conclusions are reliable.”
Correction: Uniqueness establishes identification only under the specified model and support. Check whether the identifying curvature is substantial, whether small perturbations materially change the parameters, and whether the assumed functional form and equilibrium selection are defensible.
Wrong reasoning: “Because there is no excluded variable, the model cannot be identified.”
Correction: Exclusion is one source of identification, not the only one. Functional form can identify parameters when it creates distinct observable implications, but the conclusion then relies directly on those shape restrictions.

Key takeaway: Exclusion restrictions identify through economic restrictions on which equations variables may affect; functional-form identification identifies through the shape of the structural relationships. Nonlinear shape can supply genuine identifying variation, but uniqueness alone does not show that the resulting parameters are robust or economically credible.

Check Your Understanding

Use the definitions and examples above to test whether you can distinguish the source of identification from the precision or credibility of a structural estimate.

In the covariance example, which two assumptions permit the structural slope to be recovered from the covariance ratio?
In the nonlinear system, why can a common exogenous variable entering both equations still help identify separate structural coefficients?
For the equation \(y_2=\gamma z+\delta y_1^2\), what happens to identification of \(\gamma\) and \(\delta\) if \(y_1^2/z\) is constant across all observed values of \(z\)?
Why does a unique solution to the two-point equations not, by itself, demonstrate that the estimated curvature coefficient is reliable?
Name one substantive question to ask when a structural conclusion depends on a particular nonlinear functional form.