Structural Parameters And Economic Restrictions

From Economic Mechanisms to Observable Distributions

Economic models describe how choices, constraints, and shocks jointly produce outcomes. Econometric analysis, however, uses observed data. The bridge between them is a set of restrictions on the probability distribution of those data. A structural equation is therefore not just a regression written with economic labels: together with its assumptions, it specifies how observed outcomes depend on behavioral parameters and latent disturbances.

A central distinction is between structural parameters, which describe the economic mechanism, and reduced-form coefficients, which describe relationships among observables after the model has been solved. The two may be algebraically related, but they need not be the same. A model can also contain nuisance parameters, such as shock variances, that affect the observable distribution without being the main behavioral quantities of interest.

Definition: A structural model is a specified collection of behavioral equations, equilibrium or choice conditions, and shock assumptions indexed by parameters. Its observable implications are restrictions on the distribution of the variables that are observed. Structural parameters govern the economic mechanism; reduced-form coefficients parameterize relationships among observables implied by that mechanism; nuisance parameters affect the observable distribution but are not the primary behavioral parameters of interest.

Write \(X\) for observed covariates or predetermined variables, \(Y\) for observed outcomes, \(\theta\) for structural parameters, and \(\eta\) for nuisance parameters. A fully specified model implies a conditional distribution \(P_{\theta,\eta}(Y\in A\mid X=x)\) for observable events \(A\). In a model with latent shocks \(\varepsilon\), this distribution is obtained by combining the model’s outcome rule with the assumed distribution of \(\varepsilon\). The restriction might be a conditional mean, a variance, a probability, or the entire conditional distribution.

For example, if the model gives a unique outcome \(Y=g(X,\varepsilon;\theta)\), and the conditional distribution of the shock is \(F_{\varepsilon\mid X}\), then the implied probability of an event \(A\) is

P_{\theta,\eta}(Y\in A\mid X=x) = \int \mathbf{1}\{g(x,e;\theta)\in A\}\,dF_{\varepsilon\mid X=x,\eta}(e).

Here \(\mathbf{1}\{\cdot\}\) equals one when its condition holds and zero otherwise. The equation states precisely how the structural outcome rule and shock distribution combine to restrict observable probabilities. In an equilibrium model, \(g\) is the solution to the equilibrium conditions; the model must specify how that solution is obtained and under what conditions it is unique.

A Behavioral Equation as a Conditional-Mean Restriction

Consider a simple demand model in logarithms. Let \(q\) be quantity demanded, \(p\) price, and \(m\) income. Suppose the behavioral equation is

\log q=\theta_0+\theta_p\log p+\theta_m\log m+u, \qquad E[u\mid p,m]=0.

The parameters \(\theta_p\) and \(\theta_m\) describe how log quantity demanded responds to log price and log income in this model. The conditional mean-zero assumption is essential: it says the demand disturbance has mean zero at each price-income combination. Under this assumption, taking the conditional expectation of both sides gives an observable restriction:

E[\log q\mid p,m] = \theta_0+\theta_p\log p+\theta_m\log m.

The structural price parameter is thus also the slope in this particular conditional-mean relation. That correspondence follows from the specified equation and assumption; it is not true automatically in every economic model. If price responds to demand shocks, for example, the conditional mean-zero assumption may fail, and the displayed relation need not hold.

Worked Example: Translating Log Demand into an Observable Restriction

Suppose a hypothetical demand model is \(\log q=1.2-0.6\log p+0.3\log m+u\), with \(E[u\mid p,m]=0\). Calculate the implied conditional mean log quantity at income \(m=100\) for prices \(p=2\) and \(p=2.2\).

The model implies \(E[\log q\mid p,m]=1.2-0.6\log p+0.3\log m\). At \(p=2\) and \(m=100\),

1.2-0.6\log(2)+0.3\log(100) = 1.2-0.4159+1.3816 = 2.1657.

At \(p=2.2\), holding income fixed,

1.2-0.6\log(2.2)+0.3\log(100) = 1.2-0.4731+1.3816 = 2.1085.

The difference is \(2.1085-2.1657=-0.0572\), which also equals \(-0.6\log(2.2/2)=-0.6\log(1.1)\). These are conditional means of log quantity, not arithmetic means of quantity. Interpreting the coefficient as an elasticity requires the log-log behavioral specification and its maintained assumptions; it does not turn this calculation into a causal claim if price is not conditionally exogenous.

Solving a Structural System Produces Reduced-Form Coefficients

In a simultaneous market model, price and quantity are jointly determined. Demand and supply coefficients describe behavioral responses, while reduced-form coefficients describe how equilibrium price and quantity vary with observed shifters. To obtain the latter, solve the structural equations rather than treating any one equation as an observed regression.

Suppose demand and supply satisfy

Q^d=\alpha_0+\alpha_pP+\alpha_yY+u_d, \qquad Q^s=\beta_0+\beta_pP+\beta_wW+u_s, \qquad Q^d=Q^s=Q.

Here \(Y\) and \(W\) are observed demand and supply shifters, respectively. Assume \(\alpha_p<0\), \(\beta_p>0\), and \(\beta_p-\alpha_p\ne0\), so the two linear relations determine a unique equilibrium for given shifters and shocks. Equating demand and supply and collecting the price terms gives

(\beta_p-\alpha_p)P = \alpha_0-\beta_0+\alpha_yY-\beta_wW+u_d-u_s.

Dividing by \(\beta_p-\alpha_p\) gives the reduced-form price equation. Substituting that solution into either structural equation gives the reduced-form quantity equation. These coefficients are combinations of structural parameters, not the original demand or supply slopes.

Worked Example: Deriving Equilibrium Price and Quantity Equations

Consider the hypothetical system \(Q^d=10-2P+Y+u_d\) and \(Q^s=2+2P-W+u_s\), with equilibrium \(Q^d=Q^s=Q\). Assume \(Y\) and \(W\) are predetermined and \(E[u_d\mid Y,W]=E[u_s\mid Y,W]=0\). The price determinant is \(2-(-2)=4\), which is nonzero, so the system has a unique linear equilibrium for each realization of the shifters and shocks.

Equating the two quantities gives

10-2P+Y+u_d=2+2P-W+u_s, \qquad 4P=8+Y+W+u_d-u_s.

Therefore the reduced-form price equation is

$$ P=2+0.25Y+0.25W+0.25u_d-0.25u_s. $$

Substitute this expression into demand to obtain equilibrium quantity:

\begin{aligned} Q &=10-2(2+0.25Y+0.25W+0.25u_d-0.25u_s)+Y+u_d\\ &=6+0.5Y-0.5W+0.5u_d+0.5u_s. \end{aligned}

Under the stated conditional mean assumptions, the observable conditional means are \(E[P\mid Y,W]=2+0.25Y+0.25W\) and \(E[Q\mid Y,W]=6+0.5Y-0.5W\). Thus the reduced-form coefficient on \(Y\) in the price equation is \(0.25\), even though neither behavioral price slope is \(0.25\). It is a function of both structural price slopes and the system’s other coefficients.

The shock terms in these reduced-form equations are also not the original structural shocks. For instance, the price disturbance is \(0.25u_d-0.25u_s\), whereas the quantity disturbance is \(0.5u_d+0.5u_s\). If the structural shocks are conditionally uncorrelated, have conditional variances \(4\) and \(9\), and have conditional means zero, then

\operatorname{Var}(P\mid Y,W)=0.25^2(4)+0.25^2(9)=0.8125, \qquad \operatorname{Var}(Q\mid Y,W)=0.5^2(4)+0.5^2(9)=3.25.

The variances are distributional implications involving shock-variance parameters as well as the behavioral coefficients. Those variances are nuisance parameters if the main goal is to describe demand and supply responses.

Means Are Only Part of the Observable Distribution

A structural equation can imply restrictions beyond conditional means. Its shock distribution and variance assumptions determine, for example, conditional variances and tail probabilities. It is therefore important to state which part of the observable distribution a claim concerns. A model that specifies only \(E[u\mid X]=0\) does not, by itself, specify a normal distribution or a particular conditional variance.

Worked Example: A Production Model and a Nuisance Variance

Suppose a hypothetical production technology is \( \log Q=\theta\log K+\gamma\log L+u\), where \(K\) and \(L\) are observed capital and labor inputs. Assume \(u\) is independent of the inputs and \(u\sim N(0,\sigma^2)\). The technology parameters are \(\theta\) and \(\gamma\); the shock variance \(\sigma^2\) is a nuisance parameter for an analysis focused on the output elasticities.

The model implies the conditional distribution

\log Q\mid K,L \sim N\!\left(\theta\log K+\gamma\log L,\ \sigma^2\right).

For \(\theta=0.4\), \(\gamma=0.6\), and \(\sigma=0.2\), at \(K=4\) and \(L=9\), the conditional mean of log output is

0.4\log(4)+0.6\log(9) = 0.5545+1.3183 = 1.8728,

and the conditional variance of log output is \(\sigma^2=0.04\). Doubling capital while holding labor fixed changes the conditional mean of log output by \(0.4\log(2)=0.2773\). This is a statement about the specified production function and the conditional distribution under its assumptions.

Because the disturbance is normal, output in levels is lognormal. Its conditional arithmetic mean is not simply the exponential of the conditional mean of log output:

E[Q\mid K,L] = K^{0.4}L^{0.6}\exp(\sigma^2/2).

At \(K=4,L=9,\sigma^2=0.04\), this is \(\exp(1.8728+0.02)\approx6.638\). The variance parameter affects the conditional mean of output in levels through the distribution of the shock, even though it is not a production elasticity. This illustrates why translating a structural model into observable restrictions requires carrying along its distributional assumptions.

A General Moment Implication

A useful derivation applies whenever a model gives an affine outcome rule in a mean-zero shock. Suppose \(Y=a(X;\theta)+B(X;\theta)\varepsilon\), where \(E[\varepsilon\mid X]=0\) and \(\operatorname{Var}(\varepsilon\mid X)=\Omega(X;\eta)\). Taking conditional expectations and covariances yields

E[Y\mid X] = a(X;\theta)+B(X;\theta)E[\varepsilon\mid X] = a(X;\theta),

\operatorname{Var}(Y\mid X) = B(X;\theta)\Omega(X;\eta)B(X;\theta)'.

The first equality is a restriction on the observable conditional mean; the second is a restriction on its conditional covariance matrix. In both, the same structural parameters can enter through \(a\) and \(B\), while nuisance parameters enter through \(\Omega\). These formulas require finite conditional second moments for the covariance statement. If the shock mean is not zero conditional on \(X\), the conditional mean includes \(B(X;\theta)E[\varepsilon\mid X]\) instead, and omitting that term gives an incorrect restriction.

Common Mistake: Calling Every Observable Slope Structural

Wrong reasoning: “The price coefficient in the equilibrium regression is \(0.25\), so the demand curve’s price response is \(0.25\).” This confuses a reduced-form coefficient with a behavioral parameter. In the supply-and-demand example, the equilibrium price coefficient on a shifter combines parameters from both structural equations, while the demand price slope is \(-2\).

Correction: First write the structural equations and their assumptions; then solve the system and label the resulting coefficients as reduced form. Likewise, distinguish a structural shock from the linear combination of shocks that appears in a reduced-form equation, and keep distributional parameters such as shock variances separate from the behavioral parameters of interest.

Finally, deriving observable restrictions is not the same as showing that the structural parameters can be uniquely recovered from observed data. The derivation tells us what distribution the model implies for specified parameter values. Whether different parameter values can produce the same observable distribution is a separate question.

Key takeaway: A structural model becomes empirically meaningful through its implications for observable distributions. Derive those implications from the behavioral equations, equilibrium conditions, and shock assumptions together; do not identify reduced-form coefficients with structural parameters without an explicit derivation.

Check Your Understanding

Use the model definitions and assumptions in each question.

In the log-demand model, which assumption allows the structural equation to imply the displayed conditional-mean restriction?
In the supply-and-demand example, why is the reduced-form price coefficient on \(Y\) not the demand price slope?
What is the reduced-form price disturbance in that example, in terms of \(u_d\) and \(u_s\)?
In the production example, which parameters describe output elasticities, and which parameter controls the conditional variance of log output?
Does deriving a family of observable distributions establish that structural parameters can be uniquely recovered? Explain briefly.