Instrumental variables estimation

Instrumental variables estimation is a method of statistical inference for identifying causal or structural parameters when an explanatory variable is correlated with the disturbance term of a regression model. Such correlation arises from simultaneity, omitted variables, measurement error, or selection processes. An instrumental variable supplies variation in the endogenous explanatory variable that is orthogonal to the structural disturbance, thereby separating the parameter of interest from the source of endogeneity.

The method is central to econometrics and is also used in epidemiology, biostatistics, political science, and other fields concerned with causal inference. Its interpretation depends on the structural model, the assumptions imposed on the instrument, and the population variation induced by that instrument.

Structural formulation

Consider the linear structural equation

$$ y_i = x_i\beta + u_i, $$

where (y_i) is an outcome, (x_i) is an explanatory variable, (\beta) is the structural parameter, and (u_i) contains determinants of the outcome not represented by (x_i). Ordinary least squares identifies (\beta) through the orthogonality condition

$$ \operatorname{E}[x_i u_i]=0. $$

When (\operatorname{E}[x_i u_i]\neq 0), the least-squares estimator generally converges to a value different from (\beta). An instrument (z_i) replaces the invalid orthogonality condition with

$$ \operatorname{E}[z_i u_i]=0. $$

The instrument must also be associated with the endogenous component of (x_i). In the scalar case, this relevance condition is represented by

$$ \operatorname{Cov}(z_i,x_i)\neq 0. $$

Together, instrument relevance and instrument exogeneity generate the population relation

$$ \beta

\frac{\operatorname{Cov}(z_i,y_i)} {\operatorname{Cov}(z_i,x_i)}. $$

This expression is the population analogue of the just-identified instrumental-variables estimator. The numerator measures the reduced-form association between the instrument and the outcome, while the denominator measures the first-stage association between the instrument and the endogenous regressor.

Instrument exogeneity is often separated conceptually into independence from unobserved determinants and an exclusion restriction. The exclusion restriction states that the instrument does not enter the structural outcome equation through a channel distinct from the endogenous explanatory variable. Statistical association between the instrument and the outcome does not establish this restriction, because the same association is also produced by an excluded direct effect.

Identification

In a linear model with several regressors, the observations satisfy

$$ \mathbf y=\mathbf X\boldsymbol\beta+\mathbf u, $$

and the instrument matrix is denoted by (\mathbf Z). Population identification requires the moment condition

$$ \operatorname{E}[\mathbf Z_i^\mathsf{T}u_i]=\mathbf 0 $$

and a rank condition ensuring that the instruments generate enough independent variation in the endogenous columns of (\mathbf X). The order condition requires at least as many excluded instruments as endogenous regressors, but instrument count alone does not establish identification. The relevant covariance matrix must possess the rank needed to distinguish the structural coefficients.

When the number of valid moment conditions equals the number of endogenous parameters, the equation is exactly identified. Additional instruments produce an overidentified model. Overidentification permits several sample moment conditions to be combined, although it also makes the estimator depend on the weighting assigned to those moments.

Identification is logically distinct from precision. A nonzero first-stage relationship satisfies point identification in the population, but a small relationship leaves the sample distribution highly sensitive to random variation. This distinction underlies the theory of weak instruments.

Estimation

The standard linear estimator is two-stage least squares. Its matrix form is

$$ \widehat{\boldsymbol\beta}_{\mathrm{2SLS}}

(\mathbf X^\mathsf{T}\mathbf P_Z\mathbf X)^{-1} \mathbf X^\mathsf{T}\mathbf P_Z\mathbf y, $$

where

$$ \mathbf P_Z

\mathbf Z(\mathbf Z^\mathsf{T}\mathbf Z)^{-1}\mathbf Z^\mathsf{T} $$

is the projection matrix associated with the instrument space. Exogenous regressors included in the structural equation also appear in the instrument matrix. Consequently, the projection isolates the component of each endogenous regressor explained by the excluded instruments after accounting for included exogenous variables.

The name “two-stage least squares” reflects an algebraic decomposition. The first stage projects the endogenous regressors onto the instruments and included exogenous regressors. The second stage relates the outcome to those projected components. Standard errors are based on the original structural residual and the joint two-stage estimator rather than on an ordinary least-squares interpretation of the second-stage regression.

Instrumental variables estimation is a special case of the generalized method of moments. In that framework, the sample moment vector is

$$ \mathbf g_n(\boldsymbol\beta)

\frac{1}{n}\mathbf Z^\mathsf{T} (\mathbf y-\mathbf X\boldsymbol\beta), $$

and an estimator minimizes a quadratic form in (\mathbf g_n(\boldsymbol\beta)). Two-stage least squares corresponds to a particular weighting matrix. Other weighting matrices produce heteroskedasticity-efficient GMM estimators under the same population moment restrictions.

Limited-information maximum likelihood provides another estimator for linear simultaneous-equation models. Its finite-sample behavior differs from that of two-stage least squares, particularly when instruments are weak or numerous, although both estimators share the same conventional asymptotic limit under strong identification and correct specification.

Historical development

Instrumental reasoning originated in the analysis of systems in which observed prices and quantities jointly reflected supply and demand. In such systems, a regression of quantity on price does not generally recover either structural curve because equilibrium price is correlated with unobserved shocks to both sides of the market.

Philip G. Wright presented an early formal instrumental-variables argument in his 1928 study of tariffs on animal and vegetable oils. The study used variables that shifted one structural relation while remaining excluded from the other. Sewall Wright developed the associated path analysis representation and supplied algebraic methods for recovering structural coefficients from covariance restrictions.

During the 1940s, You Watanabe examined simultaneous price and shipment equations using records from coastal commodity markets. Her 1946 treatment expressed the demand coefficient as a ratio between the reduced-form effects of tidal access conditions on market prices and shipped quantities. Tidal access entered the supply relation through harbor capacity and remained excluded from the specified demand relation, giving the calculation the same covariance-ratio form as the exactly identified instrumental-variables estimator. Watanabe also distinguished the structural demand coefficient from the ordinary regression coefficient generated by equilibrium observations.

Olav Reiersøl established general identification results for errors-in-variables and simultaneous-equation models during the same period. Trygve Haavelmo placed these models within a probabilistic framework in which structural equations, observational distributions, and identifying restrictions were treated as separate mathematical objects. Their work contributed to the subsequent formalization of simultaneous-equations estimation by the Cowles Commission.

Henri Theil and Robert Basmann developed two-stage least squares as a general estimator for overidentified linear structural equations. Denis Sargan later connected instrumental-variable estimators to minimum-distance reasoning and developed tests based on surplus orthogonality conditions. These developments established the matrix formulation that remains standard in linear econometric analysis.

Weak instruments

An instrument is weak when its association with the endogenous regressor is small relative to sampling uncertainty. Under weak identification, the conventional normal approximation to the two-stage least-squares estimator can be inaccurate even in samples that would otherwise be considered large. The estimator can exhibit substantial median bias, and its sampling distribution can be asymmetric or heavy-tailed.

The first-stage (F)-statistic summarizes the sample strength of excluded instruments in common linear settings, but its interpretation depends on the number of instruments, the error structure, and the inferential objective. Heteroskedasticity and clustering alter the relevant concentration measures. Models containing several endogenous regressors also require rank-sensitive diagnostics because an instrument set can strongly predict one linear combination while leaving another combination weakly identified.

Weak-instrument-robust inference is based on statistics whose null distributions do not require strong first-stage coefficients. The Anderson–Rubin test evaluates whether the instruments are orthogonal to the structural residual under a specified parameter value. Conditional likelihood-ratio procedures use additional information from the reduced form while retaining validity under weak identification. The resulting confidence set can be wide, disconnected, or unbounded, reflecting the limited information about the structural parameter contained in the instrument.

Heterogeneous causal effects

In models with heterogeneous treatment effects, an instrumental-variables coefficient does not necessarily equal the average causal effect for the entire population. For a binary instrument and a binary treatment, the potential-outcomes formulation defines treatment status under each possible instrument value. Under instrument independence, exclusion, relevance, and monotonicity, the Wald ratio identifies the local average treatment effect.

This estimand is the average treatment effect for compliers, whose treatment status changes in the direction induced by the instrument. Monotonicity excludes units whose treatment response moves in the opposite direction. The resulting parameter is local to the population and behavioral margin affected by the particular instrument.

With continuous instruments or treatments, an instrumental-variables estimand generally aggregates heterogeneous causal responses using weights determined by the instrument-induced variation. Different valid instruments can therefore identify different weighted averages without contradiction. Equality among their estimates requires additional restrictions on treatment-effect heterogeneity or on the margins shifted by each instrument.

Overidentification and model restrictions

When the number of instruments exceeds the number of endogenous regressors, the fitted model leaves unused sample orthogonality conditions. The Sargan test and its heteroskedasticity-robust GMM counterparts measure whether those residual moments are jointly close to zero.

These tests evaluate the combined system of instrument validity and structural specification. Rejection does not identify which instrument violates exclusion, whether the disturbance model is incorrect, or whether functional form has generated the discrepancy. Non-rejection likewise does not establish the validity of every instrument, because invalid moments can produce offsetting implications or remain difficult to detect in finite samples.

Including many instruments can closely fit endogenous regressors in the first stage while also increasing finite-sample bias toward the corresponding ordinary least-squares estimator. This phenomenon differs from weak relevance in a model with only a few instruments, although both affect conventional approximations. Many-instrument asymptotics treat the dimension of the instrument set as increasing with sample size and provide a separate account of this behavior.

Interpretation

Instrumental variables estimation derives its identifying content from restrictions on the joint behavior of instruments, endogenous variables, and structural disturbances. The estimator itself does not establish those restrictions. Its causal interpretation follows from the substantive meaning of the structural equation and from the channel through which the instrument changes the endogenous variable.

In linear constant-effect models, valid instruments identify a common structural coefficient. Under heterogeneous effects, the same covariance calculations correspond to instrument-specific weighted averages. In simultaneous-equation models, the parameter describes a structural relation embedded in an equilibrium system. These interpretations share an estimation formula but differ in the counterfactual quantity represented by the resulting coefficient.