Robust Generalized Least Squares (Robust GLS)
Robust Generalized Least Squares · Also known as: robust generalized least squares, GLS with robust standard errors, heteroscedasticity-consistent GLS, HC-GLS
Robust GLS extends classical Generalized Least Squares by pairing GLS coefficient estimation with heteroscedasticity- and autocorrelation-consistent (HAC) standard errors, or by using M-estimation within the GLS framework. It corrects for non-spherical errors — heteroscedasticity, autocorrelation, or both — while also guarding inference against misspecification of the error covariance structure.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Robust GLS when your data exhibit clear heteroscedasticity or serial correlation in residuals (detected by Breusch-Pagan, White, or Durbin-Watson tests) AND you have reason to believe the error covariance structure may be partially misspecified. It is particularly valuable in macroeconomic time series, finance, and cross-country panel applications where both non-spherical errors and model uncertainty are common. Do not use Robust GLS when error structure is spherical (homoscedastic, no autocorrelation) — plain OLS is sufficient and simpler. Also avoid it in very small samples (n < 30) where sandwich estimators can be unreliable; cluster-robust or bootstrap inference may be preferable instead.
Strengths & limitations
- Jointly addresses non-spherical errors (heteroscedasticity, autocorrelation) and inference uncertainty under misspecification.
- Coefficient estimates retain the efficiency gains of GLS weighting over plain OLS when errors are non-spherical.
- Robust sandwich standard errors remain consistent under a wide class of error covariance misspecifications.
- Applicable to cross-sectional, time-series, and pooled data where error structure is complex but partially characterised.
- Natural bridge between GLS and HAC-corrected inference; well-supported in standard econometric software.
- Requires a preliminary estimate of the error covariance Omega; misestimating Omega affects coefficient efficiency even if inference is robust.
- Sandwich standard errors are asymptotic; in small samples (n < 50) they can underestimate true uncertainty, leading to over-rejection.
- More complex to implement and interpret than plain OLS or simple robust OLS, adding researcher degrees of freedom in choosing the error structure.
- If the true error structure is spherical, Robust GLS offers no efficiency gain over OLS and may inflate standard errors unnecessarily.
- Computationally heavier than OLS, particularly when Omega must be estimated iteratively.
Frequently asked
What is the difference between Robust GLS and Feasible GLS?
Feasible GLS (FGLS) estimates Omega from the data and plugs it into the GLS formula, but uses classical (non-robust) standard errors that assume Omega was correctly specified. Robust GLS additionally wraps the estimator in a sandwich covariance matrix, so standard errors remain valid even if the estimated Omega was partly wrong.
Does Robust GLS fix endogeneity or omitted-variable bias?
No. Robust GLS corrects for non-spherical errors and guards inference against covariance misspecification, but it does not eliminate bias from endogenous regressors or omitted variables. Instrumental variables or panel fixed effects are needed for those problems.
When should I prefer Robust OLS over Robust GLS?
If residuals show heteroscedasticity but no meaningful autocorrelation, and you have no reliable prior information about the error covariance shape, Robust OLS (OLS with White standard errors) is simpler and avoids the risk of introducing bias from a poorly estimated Omega. Robust GLS is preferred when the error structure is partially known and efficiency gains from GLS weighting are important.
How do I estimate the error covariance matrix Omega in practice?
Common approaches include: (1) fitting OLS first, then using squared residuals to estimate heteroscedasticity weights; (2) assuming a specific parametric form such as AR(1) errors and estimating the autocorrelation coefficient; or (3) using a nonparametric Newey-West HAC kernel. The choice of method should be guided by residual diagnostics from an initial OLS fit.
Is Robust GLS available in standard statistical software?
Yes. In R it is available via the nlme and sandwich packages (gls() combined with vcovHAC()); in Stata via xtgls with robust options; in Python via statsmodels GLS with cov_type='HAC' or 'HC3'. Always verify that the software applies the sandwich correction at the GLS (not OLS) stage.
Sources
- Greene, W. H. (2012). Econometric Analysis (7th ed.). Pearson. Chapter 9: The Generalized Regression Model and Heteroscedasticity. ISBN: 978-0131395381
- White, H. (1980). A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity. Econometrica, 48(4), 817-838. DOI: 10.2307/1912934 ↗
How to cite this page
ScholarGate. (2026, June 3). Robust Generalized Least Squares. ScholarGate. https://scholargate.app/en/econometrics/robust-gls
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Generalized Least SquaresStatistics↔ compare
- OLS RegressionEconometrics↔ compare
- Panel GLSEconometrics↔ compare
- Robust OLSEconometrics↔ compare
- Weighted Least SquaresStatistics↔ compare