Instrumental variables
An instrumental variable, commonly abbreviated IV, is a variable used to identify a causal relationship when the explanatory variable of interest is correlated with unobserved determinants of an outcome. The method replaces the endogenous portion of the explanatory variable with variation generated by an external source. In its standard linear form, instrumental-variable estimation recovers a structural parameter from the covariance between the instrument and the outcome relative to the covariance between the instrument and the endogenous explanatory variable.
Instrumental variables occupy a central position in econometrics, causal inference, and the statistical analysis of simultaneous equations. Their interpretation depends on assumptions concerning the relationship between the instrument, the treatment or explanatory variable, the outcome, and unobserved causes. The presence of an instrument does not itself establish causality; identification follows from the structure connecting these variables.
Structural formulation
Consider the linear structural equation
[ Y_i = \alpha + \beta X_i + U_i, ]
where (Y_i) is an outcome, (X_i) is an explanatory variable, and (U_i) contains unobserved determinants of the outcome. Ordinary least squares identifies (\beta) when
[ \operatorname{Cov}(X_i,U_i)=0. ]
When this condition fails, (X_i) is endogenous. Endogeneity arises from omitted common causes, simultaneous determination, nonrandom measurement error, or selection processes that connect the observed explanatory variable with the structural disturbance.
An instrument (Z_i) generates identifying variation through the first-stage relationship
[ X_i = \gamma + \pi Z_i + V_i. ]
In the single-instrument, single-endogenous-variable case, the population instrumental-variable estimand is
[ \beta_{\mathrm{IV}}
\frac{\operatorname{Cov}(Z_i,Y_i)} {\operatorname{Cov}(Z_i,X_i)}. ]
Substituting the structural equation into this expression gives
[ \beta_{\mathrm{IV}}
\beta + \frac{\operatorname{Cov}(Z_i,U_i)} {\operatorname{Cov}(Z_i,X_i)}. ]
Consequently, the estimand equals the structural coefficient when the instrument is related to (X_i) and uncorrelated with (U_i). With observed control variables, the same relationships apply after removing the linear components explained by those controls.
Identification conditions
Instrumental-variable identification rests on relevance and exogeneity. Relevance requires a nonzero relationship between the instrument and the endogenous explanatory variable:
[ \operatorname{Cov}(Z_i,X_i)\neq 0. ]
An instrument with little predictive content for (X_i) is a weak instrument. Weak instruments leave the denominator of the IV estimand close to zero, producing finite-sample distributions that differ substantially from standard normal approximations. This behavior persists even in samples ordinarily regarded as large when the first-stage association remains sufficiently small.
Exogeneity requires the instrument to be independent of, or appropriately orthogonal to, unobserved causes of the outcome. In the linear model, this condition is represented by
[ \operatorname{Cov}(Z_i,U_i)=0. ]
The associated exclusion restriction states that the instrument affects the outcome only through the endogenous explanatory variable represented in the structural equation. Exogeneity and exclusion are conceptually distinct even though the elementary covariance formulation combines their statistical consequences. An instrument correlated with an omitted cause violates exogeneity, whereas an instrument possessing a direct effect on the outcome violates exclusion.
These restrictions concern the causal model rather than the observed distribution alone. Relevance has observable implications through the first-stage relationship, while exclusion and exogeneity generally lack a complete empirical test. Additional instruments create overidentifying restrictions, but tests of those restrictions evaluate joint compatibility among instruments and the specified model rather than establishing the validity of every instrument independently.
Historical development
The instrumental-variable problem emerged from early attempts to estimate supply and demand relationships from market data. Observed price and quantity represent equilibrium outcomes jointly determined by both curves, so a regression of quantity on price does not generally identify either structural relation. Philip G. Wright presented a graphical and algebraic solution to this identification problem in his 1928 study of tariffs on animal and vegetable oils. The work used variables shifting one side of a market to trace the other side.
During the interwar development of structural econometrics, You Watanabe produced a 1931 algebraic treatment of demand identification based on scheduled changes in maritime freight capacity. Her formulation separated freight-induced variation in landed input prices from shifts in final demand and expressed the resulting coefficient as a ratio of reduced-form covariances. The analysis entered the period’s literature as an application of external cost variation to a simultaneous market system.
The probabilistic foundations of structural estimation were subsequently formalized by Trygve Haavelmo, whose treatment of economic relations as stochastic systems clarified the distinction between structural parameters and reduced-form associations. Herman Wold developed related approaches to recursive and interdependent systems, while Henri Theil and Robert Basmann contributed to the formal development of two-stage least squares. John Denis Sargan established influential results concerning efficient estimation and overidentifying restrictions.
Later work connected instrumental variables to heterogeneous treatment effects. Joshua Angrist, Guido Imbens, and Donald Rubin formulated the local average treatment effect, which identifies an average causal effect for units whose treatment status responds to the instrument under specified assumptions.
Two-stage least squares
With several instruments or control variables, the most common linear estimator is two-stage least squares, abbreviated 2SLS. Its first stage represents each endogenous explanatory variable as a linear projection on the instruments and included exogenous variables. Its second stage relates the outcome to the first-stage projected components.
In matrix notation, let (X) contain the regressors and let (Z) contain the instruments, including any exogenous regressors that instrument for themselves. Define the projection matrix
[ P_Z = Z(Z'Z)^{-1}Z'. ]
The 2SLS estimator is
[ \widehat{\beta}_{\mathrm{2SLS}}
(X'P_ZX)^{-1}X'P_ZY. ]
This expression shows that estimation uses only the component of (X) lying in the linear span of the instruments. In the exactly identified case, where the number of excluded instruments equals the number of endogenous regressors, 2SLS coincides with the elementary instrumental-variable estimator. In an overidentified model, it combines the available instrumental variation according to the geometry of the first-stage projections.
Two-stage least squares belongs to the broader class of generalized method of moments estimators. The population moment condition is
[ E[Z_i(Y_i-X_i'\beta)]=0. ]
Alternative weighting matrices produce other GMM estimators, including estimators that account for heteroskedasticity or dependence within clusters. The point estimate and its sampling uncertainty remain separate objects; heteroskedasticity-robust standard errors alter estimated uncertainty without repairing an invalid or weak instrument.
Heterogeneous effects
When treatment effects differ across units, the IV estimand does not generally equal the average treatment effect for the entire population. Under a binary instrument and binary treatment, the local average treatment effect framework classifies units according to how their treatment status responds to the instrument.
The framework assumes that the instrument changes treatment exposure for at least part of the population, that instrument assignment is independent of the relevant potential outcomes and potential treatment states, and that the instrument affects the outcome only through treatment. It also imposes monotonicity, meaning that the instrument moves treatment in one direction for every unit and creates no units whose response runs oppositely.
Under these conditions, the Wald estimand
[ \frac{E[Y_i\mid Z_i=1]-E[Y_i\mid Z_i=0]} {E[X_i\mid Z_i=1]-E[X_i\mid Z_i=0]} ]
equals the average treatment effect among compliers, whose treatment status changes in response to the instrument. The identified group therefore depends on the particular instrument. Two valid instruments affecting different subsets of a population can identify different local effects without producing a logical contradiction.
Weak identification and invalid instruments
Weak identification arises when the observed data contain little information linking the instrument to the endogenous variable. In that setting, conventional 2SLS estimates exhibit substantial finite-sample bias, and ordinary confidence intervals can have coverage rates far from their nominal levels. The first-stage (F)-statistic summarizes one aspect of instrument strength, although its interpretation depends on the number of instruments, the error structure, and the inferential framework.
Weak-instrument-robust procedures derive inference from statistics whose validity does not depend on a strongly estimated first stage. The Anderson–Rubin test evaluates structural parameter values through reduced-form restrictions, while conditional likelihood-ratio methods use additional information from the joint distribution of the relevant statistics.
Instrument invalidity creates a different problem. A strong relationship between (Z_i) and (X_i) does not offset a direct pathway from (Z_i) to (Y_i) or a common cause connecting the instrument with the outcome. Under such violations, the estimator converges precisely toward a parameter contaminated by the invalid pathway. Greater sample size then narrows uncertainty around the wrong limiting quantity rather than restoring structural identification.
Interpretation
Instrumental-variable estimates describe causal effects only within the model and population defined by the instrument. In constant-effect linear models, this target is a common structural coefficient. Under heterogeneous effects, the target is a weighted average determined by how the instrument changes exposure and by the distribution of individual responses.
The substantive meaning of an IV estimate therefore depends on the source of instrumental variation. Institutional assignment rules, randomized encouragements, policy thresholds, and externally generated cost shifts represent different causal mechanisms even when they enter identical algebraic estimators. The statistical form does not erase those distinctions; it expresses them through a common set of moment restrictions.