Observational Equivalence In Structural Models

Why Different Structural Explanations Can Look the Same

Structural models describe how observed outcomes are generated from economic mechanisms, parameters, and disturbances. The same observed pattern can sometimes be explained by more than one mechanism. If those explanations imply exactly the same probability distribution for everything observed, no amount of data on those observables can distinguish them.

This is not merely a problem of a small sample, noisy measurements, or an imprecise estimate. It is a feature of the model: distinct structural parameter values are compatible with the same observable evidence. The key question is therefore not just whether a parameter can be estimated precisely, but whether the observable data distribution distinguishes its possible values at all.

Definition: Let $\theta$ denote structural parameters and $\eta$ any additional parameters or nuisance features of a model. Write $P_{\theta,\eta}$ for the probability distribution the model implies for the observed data. Two admissible specifications $(\theta,\eta)$ and $(\theta',\eta')$ are observationally equivalent if $$ P_{\theta,\eta}(A)=P_{\theta',\eta'}(A) $$ for every event $A$ defined in terms of the observed data. They are equivalent when their full observable distributions agree, not merely when a few summary statistics agree.

A parameter of interest, written $\psi$, is point identified in a model if any two admissible specifications that imply the same observable distribution must assign the same value to $\psi$. If two equivalent specifications assign different values to $\psi$, then $\psi$ is not point identified. A parameter can fail to be identified even when other features of the observable distribution are fully determined.

The requirement that specifications be admissible matters. Economic restrictions and assumptions about disturbances define which specifications belong to the model. A restriction can rule out one of two otherwise equivalent explanations. Conversely, if the model permits flexible disturbance distributions or dependence, it may include many explanations that fit the same observables. The earlier material on structural parameters and economic restrictions motivates this distinction: restrictions are substantive components of a structural model, not optional adjustments made after seeing the data.

Worked Examples of Observational Equivalence

Worked Example: A Regression Slope With Correlated Disturbances

Suppose a hypothetical population has observed variables $X$ and $Y$ with a bivariate normal distribution satisfying $E[X]=E[Y]=0$, $\operatorname{Var}(X)=\operatorname{Var}(Y)=1$, and $\operatorname{Cov}(X,Y)=0.5$. Consider structural specifications of the form $Y=\beta X+\varepsilon$, allowing the disturbance $\varepsilon$ to have a distribution that varies across specifications. Compare $\beta=0$ with $\beta=0.5$.

For $\beta=0$, define $\varepsilon=Y$. Then $Y=0X+\varepsilon$, and $$ \operatorname{Var}(\varepsilon)=1,\qquad \operatorname{Cov}(X,\varepsilon)=\operatorname{Cov}(X,Y)=0.5. $$ This specification exactly reproduces the stated joint distribution of $X$ and $Y$.

For $\beta=0.5$, define $\varepsilon=Y-0.5X$. The structural equation again holds, and the disturbance moments are $$ \operatorname{Cov}(X,\varepsilon) =\operatorname{Cov}(X,Y)-0.5\operatorname{Var}(X) =0.5-0.5(1)=0, $$ $$ \operatorname{Var}(\varepsilon) =1+(0.5)^2(1)-2(0.5)(0.5) =0.75. $$ Because $X$ and $\varepsilon$ are jointly normal and have zero covariance, they are independent. Thus the same observed bivariate normal distribution is generated by $\beta=0.5$ with an independent disturbance, and by $\beta=0$ with a correlated disturbance.

If the model permits correlated disturbances, these two slopes are observationally equivalent. If instead the model imposes the restriction that $\varepsilon$ is independent of $X$, the $\beta=0$ specification is inadmissible, while the $\beta=0.5$ specification satisfies the restriction. The observed distribution alone does not select between them without regard to the model's assumptions.

Worked Example: Equivalent Supply-and-Demand Systems

Consider a hypothetical market in which observed price $P$ and quantity $Q$ are independent standard normal variables. A linear demand and supply system is $$ Q=-b_dP+u_d,\qquad Q=b_sP+u_s, $$ where $b_d>0$ and $b_s>0$, and the disturbances may have a joint distribution that is not specified in advance. Compare the slope pairs $(b_d,b_s)=(1,1)$ and $(b_d,b_s)=(2,1)$.

For any observed pair $(P,Q)$, the first system is satisfied by defining $u_d=Q+P$ and $u_s=Q-P$. Since $P$ and $Q$ are independent with variance 1, $$ \operatorname{Var}(u_d)=2,\qquad \operatorname{Var}(u_s)=2,\qquad \operatorname{Cov}(u_d,u_s)=1-1=0. $$ These disturbances jointly generate the specified observable law.

For the second system, define $u_d=Q+2P$ and $u_s=Q-P$. The observed equations again hold, while $$ \operatorname{Var}(u_d)=1+4=5,\qquad \operatorname{Var}(u_s)=1+1=2, $$ $$ \operatorname{Cov}(u_d,u_s) =E[(Q+2P)(Q-P)] =1-2=-1. $$ Thus the disturbance distribution changes, but the distribution of observed price and quantity remains exactly the same.

The equilibrium interpretation is consistent in each case: subtracting the supply equation from the demand equation gives $u_d-u_s=(b_d+b_s)P$. For the first specification this is $u_d-u_s=2P$; for the second it is $u_d-u_s=3P$, as follows from the respective disturbance definitions. With unrestricted joint disturbance distributions, the two slope pairs are observationally equivalent. A restriction on disturbance dependence or on the economic mechanism could rule out one specification, but it must be part of the model.

Worked Example: An Unidentified Latent Variance Decomposition

Suppose the only observed variable is $Y=U+V$, where $U$ and $V$ are independent, mean-zero normal latent components. Compare a model with $\operatorname{Var}(U)=1$ and $\operatorname{Var}(V)=3$ to one with $\operatorname{Var}(U)=2$ and $\operatorname{Var}(V)=2$.

A sum of independent normal variables is normal, with variance equal to the sum of their variances. Under the first specification, $$ Y\sim N(0,1+3)=N(0,4). $$ Under the second specification, $$ Y\sim N(0,2+2)=N(0,4). $$ Therefore both specifications imply the same distribution for every observed value of $Y$, even though the variance assigned to $U$ differs. The two component variances cannot be separately point identified from observations of $Y$ alone. Observing additional information about the components, or imposing further restrictions, could change what is identified.

Why Identical Distributions Prevent Identification

If two specifications imply the same distribution, their probabilities for every possible data event are equal. This includes the probability of any sample pattern, any rejection decision based on the observed sample, and any value reported by an estimator. A larger sample does not create distinguishing information if the model continues to assign the same joint law to all the observations.

The implication for consistent estimation can be shown directly. Suppose equivalent specifications assign different values $\psi_0$ and $\psi_1$ to a parameter of interest, with $\psi_0\ne\psi_1$. Choose $\epsilon>0$ small enough that the intervals within distance $\epsilon$ of the two values do not overlap; for example, take $\epsilon<|\psi_0-\psi_1|/2$. An estimator that consistently estimates $\psi$ would have to satisfy $$ P_{\psi_0}(|\hat{\psi}-\psi_0|<\epsilon)\longrightarrow 1 $$ under the first specification, and $$ P_{\psi_1}(|\hat{\psi}-\psi_1|<\epsilon)\longrightarrow 1 $$ under the second. But the observable distributions are identical, so the estimator has the same distribution under both specifications. It cannot put probability tending to 1 in each of two disjoint intervals. This contradiction shows that no estimator based only on those observations can consistently estimate both distinct values across the model.

This argument concerns the structural model, not the skill of a particular estimator. A numerical routine might return one value, perhaps because of an imposed constraint, an initialization, or an unstated convention. That output does not demonstrate that the observed distribution uniquely supports the reported value. To establish point identification, the admissible model must exclude alternative specifications that reproduce the same observable law while changing the parameter of interest.

Key distinction: Observational equivalence concerns equality of the full distribution of observed data. Point identification concerns whether equivalent admissible specifications agree on the parameter of interest. A model may leave some structural parameters unidentified while still identifying other parameters or features of the observable distribution.

Common Mistake: Treating a Good Fit as Identification

A common error is to reason, “This structural specification reproduces the observed distribution, so its parameters are identified.” Reproducing the distribution establishes compatibility, not uniqueness. The regression and market examples each provide more than one compatible structural specification. Identification requires checking whether any distinct admissible specification also reproduces that distribution while changing the target parameter.

A related error is to compare only means, variances, or a handful of moments and then declare two specifications equivalent. Matching those summaries is not enough: the full observable distributions may differ in other ways. The reverse error is also possible. Two specifications can look different in their equations or disturbance descriptions yet imply exactly the same law for the observed variables. Always make the comparison at the level of the observables and their complete probability distribution.

When an equivalence appears, the next question is substantive: which additional observations, restrictions, or economic assumptions would distinguish the specifications? The answer must be justified by the model and the data available. A restriction that simply selects a preferred value without observable or substantive support does not make that value empirically identified.

Specify what is observed.
List the observed variables and the probability distribution the model is intended to explain.

State the admissible model.
Record the structural equations, parameter space, and restrictions on disturbances or other unobserved features.

Compare specifications.
Ask whether distinct admissible parameter values can imply the same full distribution of the observables.

Check the target parameter.
If equivalent specifications disagree on the target, it is not point identified under that model.

Check Your Understanding

Use the definitions and reasoning in this tutorial to answer the following questions.

In the formal definition, why must observational equivalence refer to every event defined in terms of observed data rather than only the observed mean?
In the regression example, which disturbance restriction excludes the specification with $\beta=0$? Show the covariance that violates it.
In the supply-and-demand example, what changes between the two specifications, and what remains the same?
In the latent-component example, what observable distribution is shared by both variance decompositions? Which variance component cannot be point identified from $Y$ alone?
Explain why an estimator cannot be consistent for two distinct parameter values that generate exactly the same observable distribution.

Observational Equivalence In Structural Models

What You'll Learn