White (1980) A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity

Summary

White (1980) solves two related problems simultaneously: how to do valid inference under heteroskedasticity of unknown form, and how to test whether such heteroskedasticity is present. The heteroskedasticity-consistent (HC) covariance matrix estimator — the "sandwich" estimator — is consistent for the true covariance of the ordinary least squares (OLS) estimator regardless of the error variance structure, requiring only that fourth moments are finite. The accompanying $\chi^2$ test for heteroskedasticity regresses squared residuals on the cross-products of regressors; under homoskedasticity the test statistic $nR^2$ is asymptotically $\chi^2$. Both contributions became standard tools in applied econometrics.

Key Claims

• HC sandwich estimator. Let $\hat\varepsilon_i = y_i - x_i'\hat\beta$ be the OLS residuals. The HC covariance estimator is: $$\hat V(\hat\beta) = (X'X)^{-1}\hat\Omega(X'X)^{-1}, \qquad \hat\Omega = \sum_{i=1}^n \hat\varepsilon_i^2\, x_i x_i'$$ This converges in probability to the true asymptotic covariance $\sigma^2(X'X)^{-1}$ under homoskedasticity and to the correct heteroskedasticity-robust covariance under any pattern of heteroskedasticity satisfying mild regularity conditions (finite fourth moments, no perfect multicollinearity). Standard errors computed from $\hat V(\hat\beta)$ yield asymptotically valid $t$- and $F$-statistics regardless of the error variance structure.

• White test for heteroskedasticity. Regress $\hat\varepsilon_i^2$ on $x_i$, $x_i \otimes x_i$ (cross-products and squares), and a constant. Under the null of homoskedasticity (or more precisely, under the null that $E[\varepsilon_i^2 \mid x_i] = \sigma^2$), the test statistic $nR^2 \sim \chi^2_q$ where $q$ is the number of regressors excluding the constant. The test is consistent against any form of heteroskedasticity that causes the variance to depend on $x_i$.

• Justification. Under heteroskedasticity, the standard OLS covariance estimator $\hat\sigma^2(X'X)^{-1}$ is inconsistent: the estimated standard errors are biased and inference is invalid. The sandwich estimator replaces the scalar $\hat\sigma^2$ with the matrix $\hat\Omega$ that directly estimates the cross-product $E[\varepsilon_i^2 x_i x_i']$ without assuming a common variance.

• Finite-sample refinements. The original $\hat\varepsilon_i^2$ estimator (HC0) is biased downward in small samples; later literature introduced HC1 ($n/(n-k)$ correction), HC2 ($\hat\varepsilon_i^2/(1-h_{ii})$), HC3 ($\hat\varepsilon_i^2/(1-h_{ii})^2$) leverage corrections. HC3 has the best small-sample coverage in simulations.

• Information matrix test. White also establishes that under correct specification the two natural estimators of the information matrix — the expected Hessian and the outer product of scores — must agree. A test of this equality is a general misspecification test (extended in White 1982 to the quasi-maximum likelihood estimation (QMLE) setting).

Concepts Introduced or Extended

• Misspecification

Entities Mentioned

• Halbert White

Quotes

"A consistent estimator for the covariance matrix of the least squares estimator is proposed which is consistent in the presence of heteroskedasticity of unknown form." (p. 817)

"A test for heteroskedasticity is also proposed which does not require the specification of the alternative hypothesis." (p. 817)

My Take

One of the most-cited papers in econometrics. The HC estimator is now the default in virtually every applied paper — the "robust standard errors" reported in nearly all empirical work derive directly from this formula. The White test is less universally used (practitioners often just use robust SEs rather than testing for heteroskedasticity first) but remains a clean diagnostic. The deeper contribution is conceptual: White shows that correct inference does not require a correctly specified error distribution, only a correctly specified conditional mean — a separation that underlies the whole robust-inference literature and connects directly to the QMLE theory in White (1982).