Homoskedastic 2SLS Standard Errors

Why 2SLS Needs Its Own Standard Error

Two-stage least squares (2SLS) uses instruments to estimate coefficients when one or more regressors are endogenous. Its coefficient estimate is not generally an ordinary least-squares estimate from the observed regressors, so its uncertainty must reflect the way the instruments identify the coefficients. Under homoskedastic errors, the conventional covariance estimator has a compact matrix form. Its simplicity is useful, but depends on assumptions that should be made explicit.

Let \(y\) be an \(n\times 1\) outcome vector, \(X\) an \(n\times k\) matrix of structural regressors, and \(Z\) an \(n\times L\) matrix of instruments, including any exogenous regressors that serve as their own instruments. The structural model is \(y=X\beta+u\). Write \(P_Z=Z(Z'Z)^{-1}Z'\) for the projection matrix onto the columns of \(Z\), assuming \(Z\) has full column rank. The 2SLS estimator is \(\hat{\beta}_{2SLS}=(X'P_ZX)^{-1}X'P_Zy\), provided \(X'P_ZX\) is invertible.

The denominator \(n-k\) is a commonly used degrees-of-freedom convention. It does not make the residual variance estimate exactly unbiased in every finite-sample IV design. For large-sample inference, dividing by \(n\) instead has the same probability limit under standard regularity conditions. The important distinction is that \(\hat{u}\) is formed from the structural regressors \(X\), not from the first-stage fitted regressors.

Assumptions and the Covariance Formula

The formula is used under a correctly specified linear structural model, instrument validity, and identification. In particular, the instruments must satisfy the moment condition \(E[z_i u_i]=0\), where \(z_i\) and \(u_i\) denote the instrument vector and structural error for observation \(i\). The instruments must also be relevant: their population relationship with the structural regressors must be strong enough to identify the coefficients, corresponding in the sample to \(X'P_ZX\) being nonsingular. For the conventional homoskedastic variance, the error variance is constant across instrument values, \(E[u_i^2\mid z_i]=\sigma^2\). Independent observations, suitable finite moments, and a large-sample approximation support the usual standard error interpretation.

The key matrix identity follows from the 2SLS estimation error. Using \(y=X\beta+u\),

For a homoskedastic error vector with covariance matrix \(\sigma^2 I_n\), the corresponding design-based covariance calculation uses \(P_Z'=P_Z\) and \(P_Z^2=P_Z\):

In an IV model, endogenous regressors mean that a literal finite-sample covariance conditional on \(X\) is not generally the sampling argument of interest. The formula is conventionally understood as the homoskedastic model-based estimator, with the standard large-sample justification under valid instruments, identification, and regularity conditions. The variance \(\sigma^2\) is unknown, so it is replaced by \(\hat{\sigma}^2\) from the structural residuals. This yields the stated estimated covariance matrix.

Reading and Computing the Matrix Components

With multiple regressors, \(X'P_ZX\) is a \(k\times k\) matrix. Its inverse contains the variance scaling and covariance between coefficient estimates; multiplying by the scalar \(\hat{\sigma}^2\) gives the estimated covariance matrix. The diagonal entries are coefficient variances, while off-diagonal entries are estimated covariances. Thus, taking a square root of a diagonal entry—not an off-diagonal entry—produces an individual coefficient standard error.

Diagnostics, Alternatives, and Limitations

The conventional formula is useful when its assumptions are credible, but it does not diagnose them automatically. Check that the instrument set and structural regressors provide full rank, and examine first-stage evidence for weak relevance. When \(X'P_ZX\) is nearly singular, its inverse can have large entries: the coefficient estimates and their standard errors may be highly sensitive to small data changes. A standard error formula does not repair weak identification or make an invalid instrument valid.

Homoskedasticity is also substantive. If the conditional error variance changes across observations, \(\hat{\sigma}^2(X'P_ZX)^{-1}\) generally does not estimate the appropriate sampling covariance. The heteroskedasticity-robust 2SLS covariance estimator replaces the common-variance simplification with observation-level residual contributions; cluster-robust or HAC covariance estimators are relevant when dependence follows those structures. These alternatives address different variance patterns and do not fix instrument validity or relevance.

Ordinary least-squares standard errors from a regression of \(y\) on first-stage fitted regressors are not a substitute for the 2SLS covariance estimator. The point estimates can coincide with 2SLS in the appropriate implementation, but treating fitted regressors as directly observed generally gives the wrong uncertainty calculation. Use the structural residuals and the \(X'P_ZX\) matrix in the conventional formula.

The standard error quantifies estimated sampling variability conditional on the maintained statistical model. It cannot account for uncertainty about the choice or validity of instruments, omitted structural variables, or model misspecification. Those concerns require design reasoning and diagnostics in addition to a covariance calculation.