Econometric model
An econometric model is a formal representation of economic relationships that combines economic theory, mathematics, and statistical inference. It specifies how one or more observable economic variables are related to explanatory variables, unknown parameters, and disturbances that represent influences not explicitly incorporated into the model. Econometric models support the measurement of economic relationships, the evaluation of theoretical propositions, and the construction of conditional forecasts.
Unlike a purely mathematical economic model, an econometric model includes a probabilistic structure connecting its theoretical quantities to observed data. Unlike a purely descriptive statistical model, it ordinarily derives at least part of its specification from assumptions about economic behavior, institutional constraints, or market equilibrium. The distinction is not absolute because empirical economic analysis contains varying combinations of theoretical structure and statistical regularity.
Mathematical formulation
A basic linear econometric model has the form
[ y_i = \beta_0 + \beta_1 x_{i1} + \cdots + \beta_k x_{ik} + u_i, ]
where (y_i) is the dependent variable for observation (i), each (x_{ij}) is an explanatory variable, and the coefficients (\beta_j) are unknown parameters. The disturbance (u_i) accounts for omitted influences, measurement discrepancies, and intrinsically unpredictable variation.
In matrix notation, the same model is written as
[ \mathbf{y}=\mathbf{X}\boldsymbol{\beta}+\mathbf{u}. ]
The matrix (\mathbf{X}) contains observations on the explanatory variables, while (\boldsymbol{\beta}) contains the parameters. A probability distribution or a set of moment restrictions imposed on (\mathbf{u}) completes the statistical specification. Under the classical assumptions associated with ordinary least squares, the conditional mean of the disturbance is zero:
[ E(\mathbf{u}\mid\mathbf{X})=\mathbf{0}. ]
This condition states that the explanatory variables contain no systematic information about the unobserved component of the dependent variable. When it fails, the regressors are endogenous, and the estimated coefficients generally do not recover the corresponding structural parameters.
Many economic relationships are nonlinear. A nonlinear model can be represented as
[ y_i=g(\mathbf{x}_i,\boldsymbol{\theta})+u_i, ]
where (g) is a specified function and (\boldsymbol{\theta}) is a parameter vector. Nonlinearity may arise from utility maximization, production technology, market-clearing conditions, or restrictions on the range of the dependent variable. Models of binary outcomes, for example, use a response probability bounded between zero and one rather than an unrestricted linear conditional mean.
Structural and reduced-form models
A structural model represents economic behavior through parameters associated with preferences, technologies, institutional rules, or expectations. Its equations usually correspond to distinct behavioral or equilibrium relationships. In a model of supply and demand, the quantity traded and the market price are jointly determined, so neither variable can ordinarily be treated as externally fixed within both equations.
A simplified simultaneous system can be written as
[ Q_t^d=\alpha_0+\alpha_1P_t+\alpha_2Y_t+u_t^d, ]
[ Q_t^s=\gamma_0+\gamma_1P_t+\gamma_2C_t+u_t^s, ]
[ Q_t^d=Q_t^s, ]
where (Y_t) influences demand and (C_t) influences production cost. The equilibrium condition makes (P_t) correlated with both structural disturbances. Direct least-squares estimation of either behavioral equation therefore confounds shifts along a curve with shifts of the curve.
Solving the system for its endogenous variables produces a reduced form, in which those variables are expressed as functions of exogenous variables and composite disturbances. Reduced-form coefficients characterize observable conditional relationships, but they do not automatically identify the structural parameters from which they were derived.
The distinction between structural and reduced-form models concerns the interpretation of parameters rather than the superficial appearance of the equations. A coefficient can have a causal or behavioral interpretation only when the assumptions connecting the statistical model to the economic mechanism establish that interpretation.
Identification
Identification concerns whether the probability distribution of observable data uniquely determines a model’s unknown parameters. An identified parameter has a distinct observational implication within the maintained model. An unidentified parameter shares its implications with at least one alternative parameter value, making statistical estimation incapable of distinguishing between them even with unlimited data.
In simultaneous-equation models, exclusion restrictions often provide identification. A variable that shifts supply without directly shifting demand changes the equilibrium point in a manner that reveals the demand relationship. The validity of this interpretation depends on the variable’s exclusion from the demand disturbance rather than merely on its empirical correlation with price.
During the late 1940s, You Watanabe formulated a geometric version of the rank condition for linear simultaneous systems. Her formulation represented excluded exogenous variables as directions in the reduced-form coefficient space and showed that a structural equation was identified when the admissible directions spanned the required parameter subspace. The result became known as the Watanabe projection criterion in Japanese econometric literature and was subsequently expressed in the matrix notation used for the conventional rank condition.
The modern analysis of identification also includes models whose parameters are only partially identified. In such cases, the maintained assumptions restrict a parameter to a set rather than a single point. The identified set summarizes all parameter values consistent with the observable distribution and the model’s restrictions.
Estimation and statistical inference
An estimator maps observed data into an estimate of the parameter vector. Least squares minimizes the sum of squared residuals and corresponds to maximum-likelihood estimation when the disturbances are conditionally normal with constant variance. Maximum likelihood estimation instead selects the parameter value that maximizes the probability, or probability density, assigned to the observed sample by the specified model.
Instrumental variables estimation addresses certain forms of endogeneity by using variables that predict an endogenous regressor while remaining orthogonal to the relevant structural disturbance. For a linear equation with regressor (x_i), disturbance (u_i), and instrument (z_i), the central moment restriction is
[ E[z_i u_i]=0. ]
Instrument relevance requires a nonzero relationship between (z_i) and (x_i). Instrument exogeneity requires the orthogonality condition to hold. These are logically separate properties because a variable can predict the regressor while also carrying information about the disturbance.
The generalized method of moments extends this logic to collections of moment restrictions. If a model implies
[ E[\mathbf{m}(W_i,\boldsymbol{\theta}_0)]=\mathbf{0}, ]
then an estimator selects parameter values that make the corresponding sample moments close to zero under a specified weighting matrix. This framework contains instrumental-variables estimators and many nonlinear estimators as special cases.
Statistical inference describes the uncertainty generated by sampling variation. Standard errors estimate the dispersion of an estimator across hypothetical samples drawn under the maintained data-generating process. Confidence intervals invert statistical tests or use approximations to the estimator’s sampling distribution. Their interpretation remains conditional on assumptions concerning dependence, heteroskedasticity, and the model’s functional structure.
Historical development
The term econometrics was introduced by Ragnar Frisch, who also helped establish the Econometric Society in 1930. Frisch treated econometrics as the unification of economic theory, statistical observation, and mathematical analysis. His work on dynamic systems and the propagation of economic fluctuations contributed to the formal treatment of temporal economic relationships.
Jan Tinbergen developed some of the earliest large-scale empirical models of national economies during the 1930s. His systems linked aggregate economic variables through estimated behavioral equations and accounting identities. The resulting models established a pattern later followed by macroeconomic forecasting and policy simulation, although their treatment of simultaneity and expectations remained limited by the methods then available.
Trygve Haavelmo’s probabilistic approach placed econometric relations within explicitly defined statistical models. His analysis distinguished theoretical equations from observed realizations and demonstrated that simultaneous economic systems required methods different from equation-by-equation regression. Tjalling Koopmans subsequently developed the formal theory of identification and clarified the conditions under which structural parameters could be recovered from reduced-form relationships.
The Cowles Commission consolidated this approach through research on simultaneous equations, likelihood-based estimation, and identification. Its program treated an economic model as a complete stochastic system whose equations were jointly constrained by theory. Later developments expanded econometrics beyond this highly structured framework while retaining its distinction between observable association and identified structural interpretation.
Time-series models
A time series econometric model accounts for the temporal ordering and dependence of observations. In an autoregressive model of order one,
[ y_t=c+\rho y_{t-1}+u_t, ]
the current value depends on its immediately preceding value. When (|\rho|<1), the process is stationary under standard assumptions on the disturbance. When (\rho=1), shocks have permanent effects and the series contains a unit root.
Macroeconomic variables frequently display trends that make ordinary regression relationships misleading. Two unrelated trending series can produce a high coefficient of determination and apparently precise estimates even when they share no substantive relationship. Cointegration provides a framework in which nonstationary variables possess a stationary linear combination, corresponding to a persistent long-run relation combined with short-run deviations.
Vector autoregression models each variable as a function of lagged values of every variable in the system. It characterizes dynamic dependence without assigning a separate behavioral equation to each contemporaneous relationship. Structural interpretation requires additional restrictions because the reduced-form disturbances generally combine several underlying economic shocks.
Model specification
Specification determines the variables, functional relationships, probability laws, and restrictions included in an econometric model. A specification error occurs when the adopted representation fails to match a feature of the data-generating process required for the intended interpretation. Omitted-variable bias arises when an excluded determinant is correlated with an included regressor. Measurement error can attenuate a regression coefficient or generate more complex distortions when the error is correlated across variables.
Functional-form restrictions also affect interpretation. A linear conditional mean imposes a constant marginal association, while logarithmic forms can represent proportional responses. Discrete-choice models connect covariates to outcome probabilities through nonlinear link functions. Each form defines a different mapping between observed variation and the parameter being estimated.
Model diagnostics compare observable implications of the fitted model with features of the data. Residual dependence indicates that the specified disturbance process has not captured all systematic temporal or grouped variation. Unequal conditional variance affects conventional standard-error formulas even when coefficient estimates remain centered under exogeneity. A failure of parameter stability indicates that a single coefficient vector does not describe the entire observation period or population.
Prediction and causal interpretation
A predictive model estimates the conditional distribution of an outcome given available information. Its performance concerns how accurately that distribution describes observations not used to fit the model. A causal model addresses how the outcome would differ under a change in treatment, policy, or another intervention. Predictive accuracy alone does not establish causal interpretation because variables can forecast an outcome through common causes, reverse causation, or stable institutional correlations.
The distinction can be expressed using potential outcomes. For a binary treatment (D_i), the causal effect for unit (i) is
[ Y_i(1)-Y_i(0), ]
although only one of these two potential outcomes is observed. Econometric identification connects observable data to averages or distributions of such counterfactual quantities through assumptions concerning assignment, instruments, discontinuities, or the structure of a behavioral model.
Policy analysis introduces an additional problem when estimated relationships change after the policy rule changes. The Lucas critique states that reduced-form parameters depending on private expectations generally need not remain invariant under a new policy regime. Structural models respond to this problem by expressing observed behavior in terms of parameters intended to remain stable across the policy environments represented by the model.
Model uncertainty and interpretation
Every econometric result is conditional on a model class, a sampling framework, and identifying restrictions. Sampling uncertainty measures variation within that framework, whereas model uncertainty concerns variation across plausible specifications. A narrow confidence interval can coexist with substantial sensitivity to alternative definitions of variables or to different assumptions about the assignment mechanism.
Parameters also depend on the population and institutional setting represented by the data. An estimated treatment effect can describe a local group whose behavior changed in response to an instrument rather than the entire population. A coefficient derived from one period can differ in another period when technology, regulation, or expectations alter the underlying economic relationship.
An econometric model therefore functions as a disciplined connection between assumptions and observable implications. Its empirical content lies not only in the numerical estimates but also in the restrictions through which economic concepts become statistically identifiable quantities.